Journal

Hong Kong J Radiol 2025 Sep;28(3):e160-2

 

 

Department of Clinical Oncology, School of Clinical Medicine, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong SAR, China

 

  Full Article in PDF

With their insidious onset and indolent clinical behaviours,[1] unresectable or metastatic neuroendocrine tumours (NETs) have long been regarded as one of the most intractable malignancies. The digestive system is the most common site of involvement. In particular, the gastroenteropancreatic (GEP) location of NETs (GEP-NETs) as the most common well-differentiated NET has been of great interest to clinicians. Well-differentiated NETs are classified into grade 1 (Ki-67 index <3%), grade 2 (Ki-67 index 3-20%), and grade 3 (Ki-67 index >20%).[2] The abundant expression of somatostatin receptors (SSRTs), especially subtype 2 (SSRT2) on the NET cell surface makes SSRT-directed therapy a promising treatment option.[3] [4]

 

Traditionally, somatostatin analogues, including octreotide, lanreotide, and pasireotide, alone or in combination with targeted therapy, have been the most commonly used treatments for unresectable or metastatic NETs, showing durable and effective tumour control with favourable safety profiles.[5] [6] [7] Most recently, peptide receptor radionuclide therapy (PRRT) with either lutetium-177 (¹⁷⁷Lu) or yttrium-90 has established itself as a safe and effective treatment for unresectable or metastatic GEP-NETs.[5] The NETTER-1 study, which compared ¹⁷⁷Lu-dotatate plus standard-dose long-acting octreotide to high-dose long-acting octreotide alone as second-line therapy in patients with advanced midgut NETs demonstrated a significantly higher objective response to ¹⁷⁷Lu-dotatate and a better progression-free survival (PFS), although the secondary endpoint of overall survival (OS) was not met.[8] [9] In view of such encouraging result using ¹⁷⁷Lu-dotatate PRRT as second-line treatment, the NETTER-2 study was subsequently conducted to evaluate the efficacy and safety of first-line ¹⁷⁷Lu-dotatate PRRT plus long-acting repeatable (LAR) octreotide versus high-dose LAR octreotide alone in patients with higher grade 2 (Ki-67 indices ≥10% and ≤20%) and grade 3 (Ki-67 indices >20% and ≤55%) NETs.[10] The results again demonstrated a better objective response (43% vs. 9%) and longer PFS (22.8 months vs. 8.5 months; hazard ratio 0.28, p < 0.0001) but not a lengthened OS, when compared to LAR indium-111 octreotide (¹¹¹In-octreotide) alone.[10] The absence of a significant OS benefit is often observed in cross-over studies, where patients in the control group may later receive investigational treatment. It should be noted that almost all published studies have mainly recruited Caucasians. The efficacy and safety of PRRT has been less well assessed in the Chinese population.

 

Wong et al[11] reported the outcomes of PRRT in their retrospective study of 21 Chinese patients with metastatic NETs treated in a single institution in Hong Kong, including one patient who had grade 3 NET. Most of them (85.7%) had received at least one prior line of systemic therapy, including somatostatin analogues, targeted therapy, and/or chemotherapy, while the remaining three patients (14.3%) received PRRT as their first-line systemic treatment. All patients had undergone ¹¹¹In-octreotide scintigraphy or gallium-68 dotatate (⁶⁸Ga-dotatate) positron emission tomography with computed tomography (PET/CT) scanning prior to treatment, which confirmed that the amount of SSRT uptake by tumour cells was equal to or greater than that of normal tissues. Both ¹⁷⁷Lu and yttrium-90 were used in two patients. The recruited patients underwent an average of four PRRT sessions, ranging from one to six sessions.

 

The study by Wong et al[11] reported an objective response rate of 47.6% was observed, in addition to 23.8% stable disease. The median PFS and OS were 22.3 months and 45.2 months, respectively, after a median follow-up duration of 19.0 months. Multivariable analysis revealed that bone metastasis and a high liver tumour burden of more than 50% were significant negative prognostic factors in OS. Lymphopenia, followed by anaemia, neutropenia, thrombocytopenia, and hepatotoxicity were the most common adverse events after PRRT. Grade 3/4 toxicities with lymphopoenia and hepatotoxicity were reported in 42.9% and 4.8% of patients, respectively. Of interest, two patients received retreatment with PRRT after their initial courses of PRRT, with tolerable and manageable grade 3 lymphopenia noted in one patient. No patient developed myelodysplastic syndrome, which had been seen in one patient who had this possibly PRRT-related toxicity approximately 14 months after the first dose of PRRT in the NETTER-2 study.[10]

 

Despite the relatively small number of patients and retrospective nature of this study, Wong et al[11] demonstrated for the first time that PRRT is a safe and effective treatment for metastatic GEP-NETs in a Chinese population, echoing the results of the NETTER-1[8] [9] and NETTER-2 trials.[10] Patient selection and eligibility screening based on the uptake of ¹¹¹In-octreotide in octreotide scintigraphy or ⁶⁸Ga-dotatate in ⁶⁸Ga-dotatate PET/CT scans may merit discussion. The Krenning score, a semi-quantitative tool, is commonly used to assess SSRT uptake based on octreotide scintigraphy and is defined as follows[12]: Grade 1 as uptake less than normal liver background activity; grade 2 as uptake equal to normal liver background activity; grade 3 as uptake greater than normal liver background activity; and grade 4 as uptake greater than spleen or kidney background activity. In the NETTER-1 study,[9] pretreatment screening was performed using octreotide scintigraphy and the Krenning score, with patients eligible if their score was grade 2, 3, or 4. The NETTER-2 study,[10] used either ¹¹¹In-octreotide scintigraphy with the Krenning score or ⁶⁸Ga-dotatate PET/CT with a modified Krenning score (an adaptation of the original Krenning score applied to ⁶⁸Ga-dotatate PET/CT) for eligibility. However, it is still unclear whether the modified Krenning scores are equivalent between these imaging modalities. A post-hoc head-to-head comparison study of ⁶⁸Ga-dotatate PET/CT and ¹¹¹In-octreotide scan-based Krenning scores in 150 patients from a phase 2 prospective study (NCT01967537) revealed that the Krenning score was significantly higher with PET/CT than with two-dimensional scintigraphy or ¹¹¹In-octreotide scintigraphy.[13] In patients with a Krenning score of 3 or above on PET/CT, the detection rates of two-dimensional scintigraphy and ¹¹¹In-octreotide scintigraphy were significantly lower for lesions smaller than 2 cm compared to lesions of 2 cm or larger: 15% and 24% versus 78% and 89%, respectively (p < 0.001). On the other hand, for lesions greater than 5 cm, the Krenning scores between PET/CT scan and octreotide scintigraphy were comparable. Lesion size did not affect PET/CT-based Krenning scores. In other words, octreotide scintigraphy may miss smaller lesions (<2 cm) that would otherwise be detected on ⁶⁸Ga-dotatate PET/CT. Prospective studies are warranted to standardise the use of a single imaging modality for eligibility screening.

 

Of concern, a recent notification was announced by the drug sponsor of the NETTER-2 study that the application for using ¹⁷⁷Lu-PRRT as first-line systemic therapy to the European Medicines Agency was withdrawn.[14] The lack of OS prolongation because of immature OS data and potentially unfavourable risks including myelodysplastic syndrome, radiation-associated second malignancies, and haematological and renal toxicities in a treatment-naïve population were the key concerns.[10] For now, the current European Medicines Agency approval for ¹⁷⁷Lu-dotatate is confined to advanced, progressive, grade 1 and grade 2 GEP-NETs.[15] Patients with grade 3 NETs are still denied access to PRRT.

 

In summary, PRRT is a novel and promising treatment modality for grade 1 and 2 unresectable or metastatic GEP-NETs after failure of prior systemic therapy. More prospective and mature data for grade 3 NETs and OS are awaited to confirm whether PRRT can also work favourably in this histological subgroup and as first-line therapy for treatment-naïve patients.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Hong Kong J Radiol 2025 Sep;28(3):e163-71 | Epub 12 September 2025

 

 

¹ Department of Clinical Oncology, Queen Elizabeth Hospital, Hong Kong SAR, China

² Department of Nuclear Medicine, Queen Elizabeth Hospital, Hong Kong SAR, China

 

  Full Article in PDF

Neuroendocrine tumours (NETs) are a heterogeneous group of neoplasms originating from neuroendocrine cells located in various anatomical sites, predominantly the gastrointestinal (GI) tract, pancreas, and lungs.[1] Approximately 20% of cases present with metastatic disease at the time of diagnosis.[2] [3] For cases not amenable to local treatment, systemic treatments commonly include somatostatin analogues, targeted agents such as sunitinib or everolimus, chemotherapy, and peptide receptor radionuclide therapy (PRRT).[4] [5] [6]

 

PRRT exploits the high expression of somatostatin receptors on NET cells,[7] enabling the targeted delivery of radionuclides conjugated to somatostatin analogues.[8] The two primary radionuclides used are Yttrium-90 (⁹⁰Y) and Lutetium-177 (¹⁷⁷Lu). While ⁹⁰Y emits high-energy beta particles to induce cytotoxic effects, ¹⁷⁷Lu emits lower-energy beta particles with a shorter path length of 1 to 2 mm, allowing more precise radiation delivery to smaller metastases and reducing the overall toxicity.

 

The NETTER-1 phase 3 randomised trial demonstrated that patients with well-differentiated, metastatic midgut NETs treated with ¹⁷⁷Lu in combination with a somatostatin analogue had a progression-free survival (PFS) rate of 65.2% at 20 months, compared to 10.8% in those receiving octreotide long-acting repeatable alone.[9] The treatment was well tolerated, with significant myelosuppression occurring in fewer than 10% of patients and no observed renal toxicity during the study period.[9] Subsequently, the US Food and Drug Administration approved ¹⁷⁷Lu-Dotatate in 2018 for the treatment of somatostatin receptor–positive NETs, and it has since become a standard of care in clinical practice.[10]

 

Despite the extensive data on PRRT from Europe and the US,[11] [12] [13] [14] [15] there is a lack of local data for the Chinese population. This retrospective study aimed to address this gap by reporting treatment responses, survival outcomes, and toxicity associated with PRRT in Chinese patients treated at a hospital in Hong Kong.

 

A retrospective cohort study was conducted on patients with NETs treated with either ¹⁷⁷Lu alone or in combination with ⁹⁰Y at the Department of Nuclear Medicine, Queen Elizabeth Hospital, Hong Kong, between August 2018 and August 2022. The intended PRRT regimen consisted of four to six cycles administered, with 8 to 12 weeks between each cycle. An amino acid infusion was given for renal protection, reducing kidney radiation by limiting reabsorption and enhancing clearance of the radiotracer. A post-therapy scan was performed on day 4 following each treatment cycle.

 

Eligible patients were Chinese individuals aged 18 years or above with histologically confirmed metastatic NETs. All patients underwent either a baseline octreotide scan or a ⁶⁸Gallium-dotatate positron emission tomography scan to confirm somatostatin receptor expression, defined as tumour uptake equal to or greater than that of normal liver tissue. Patients receiving PRRT for paraganglioma were excluded, as these tumours exhibit different biological behaviour compared to epithelial NETs.

 

Electronic medical records were reviewed to extract patient demographics, including sex, date of birth, date of death (if applicable), date of NET diagnosis, primary tumour location, World Health Organization grade, Ki-67 index, sites of metastases, hepatic metastatic burden, and date of last follow-up. Data on previous treatments was also collected, including surgical resection, locoregional therapies such as transarterial chemoembolisation and radiofrequency ablation, somatostatin analogues (octreotide or lanreotide), targeted therapies (everolimus or sunitinib), and chemotherapy (e.g., temozolomide or capecitabine plus oxaliplatin). Laboratory parameters were recorded before and after each course of PRRT, including haemoglobin, neutrophil and lymphocyte counts, creatinine levels, bilirubin, and alanine transaminase levels. Symptoms after each course of PRRT were extracted from the medical records.

 

Treatment response was assessed by comparing pre- and post-PRRT imaging using RECIST (Response Evaluation Criteria in Solid Tumors) 1.1. Survival outcomes were analysed using Kaplan–Meier curves. PFS was defined as the date from the first PRRT to disease progression or death from any cause, and overall survival (OS) was defined as the time from the first PRRT to death. Toxicities were evaluated according to the CTCAE (Common Terminology Criteria for Adverse Events) version 4.03, with follow-up blood tests recorded at each visit. Patients alive at the time of final analysis or lost to follow-up were censored at the date they were last known to be alive.

 

Two patients who demonstrated an initial response following their course of PRRT subsequently developed disease progression and underwent retreatment with PRRT. Retreatment PFS for these two patients was presented separately and was not pooled with the cohort PFS and OS analyses.

 

Normally distributed data are shown as means (standard deviations), whereas non-normally distributed data are shown as medians (ranges). Univariable Cox proportional hazards regression was performed to assess the association of baseline factors with PFS and OS. Variables showing at least a trend towards significance (p < 0.1) in the univariate analysis were subsequently included in the multivariate analysis using the Cox proportional hazards model. All tests were two-sided, with a significance threshold of p < 0.05. Analyses were conducted using SPSS (Windows version 24.0; IBM Corp, Armonk [NY], US).

 

PRRT was performed on 23 Chinese patients between August 2018 and August 2022 at Queen Elizabeth Hospital. Two patients were excluded from this study because they had paragangliomas, leaving a total of 21 patients included in the study. Patient characteristics are summarised in Table 1. The median age at diagnosis was 55 years (range, 31-72), with 11 male and 10 female patients. All patients had an Eastern Cooperative Oncology Group performance status score of 0 to 1. The most common primary tumour site was the pancreas (71.4%), followed by the rectum (23.8%) and stomach (4.8%). According to World Health Organization tumour grading, 23.8% were Grade 1, 71.4% were Grade 2, and 4.8% were Grade 3. Liver metastases were present in 90.5% of patients, followed by bone metastases (19.0%) and peritoneal metastases (4.8%). About 43% of patients had a hepatic metastatic burden of 50% or more of liver volume on baseline imaging.

 

Table 1. Patient demographics (n = 21).

 

The median interval between diagnosis and initiation of PRRT was 12.5 months (range, 2.5-86.2). A total of 42.9% of patients had received prior locoregional treatment, including surgery or transarterial chemoembolisation, before undergoing PRRT. Overall, 85.7% of patients received PRRT after the failure of at least one systemic treatment, which included somatostatin analogues (n = 12), everolimus (n = 9), and chemotherapy agents such as temozolomide plus capecitabine or oxaliplatin plus capecitabine (n = 8). The median number of systemic treatments before PRRT was 1 (range, 0-3). PRRT was used as first-line treatment in 14.3% of patients, as second-line in 61.9%, and as third-line or beyond in 23.8%. No patient received additional local or systemic treatment between PRRT cycles.

 

¹⁷⁷Lu was used as monotherapy in 90.5% of patients, while a combination of ¹⁷⁷Lu and ⁹⁰Y was used in 9.5%. In the ¹⁷⁷Lu alone group, patients received a median dose of 7622 MBq per injection (range, 4662-8140). In the combination group, the median dose was 7844 MBq of ¹⁷⁷Lu (range, 5920-8140) and 3633 MBq of ⁹⁰Y (range, 3330-3848). Patients underwent a mean of 4.2 PRRT cycles. Five patients were unable to complete at least four cycles: two received one cycle, two received two cycles, and one received three cycles, due to disease progression or death.

 

All patients were assessed for radiological response using the RECIST 1.1 criteria. Among the 21 patients, partial response was observed in 47.6%, stable disease in 23.8%, and disease progression in 28.6%. No patient achieved a complete response. The disease control rate (defined as the proportion of patients with either partial response or stable disease) was 71.4%.

 

The median follow-up duration after the last PRRT treatment was 19 months (range, 2-44). Median PFS was estimated at 22.3 months (95% confidence interval [95% CI] = 16.0-28.7) [Figure 1a]. Univariate analysis (Table 2) identified significantly shorter PFS in patients with bone metastasis (p = 0.006) and GI primary NETs (p = 0.039). Bone metastases remained an independent prognostic factor in multivariate analysis, with a hazard ratio (HR) of 4.96 (95% CI = 1.28-19.20; p = 0.02) [Figure 1b]. Age, sex, prior treatment with somatostatin analogues, targeted therapy, chemotherapy, and maintenance treatment were not significantly associated with PFS.

 

Figure 1. Kaplan–Meier plots of progression-free survival (PFS) following peptide receptor radionuclide therapy. (a) PFS for the entire cohort. (b) PFS comparison based on bone metastasis status.

 

Table 2. Univariate and multivariate analyses of factors associated with progression-free survival and overall survival.

 

The median OS was 45.2 months (95% CI = 33.4-57.0) [Figure 2a]. Univariate analysis of baseline factors potentially associated with OS is shown in Table 2. In the unadjusted analysis, significantly shorter OS was associated with bone metastasis (p = 0.016), liver metastatic burden of 50% or more (p = 0.008), and GI primary tumours (p = 0.042). Following multivariate analysis, both bone metastasis (HR = 5.27, 95% CI = 1.09-25.4; p = 0.038) and a high liver metastatic burden of 50% or more (HR = 5.80, 95% CI = 1.07-31.4; p = 0.042) remained independently associated with poorer OS (Figure 2b and c).

 

Figure 2. Kaplan–Meier plots of overall survival (OS) following peptide receptor radionuclide therapy. (a) OS for the entire cohort. (b) OS comparison based on bone metastasis status. (c) OS comparison based on liver metastatic burden.

 

Patients were evaluated for haematologic, hepatic, or renal toxicity using CTCAE version 4.03 (Table 3). Among haematological toxicities, lymphopenia was the only Grade 3/4 event in 42.9% of patients. No cases of myelodysplastic syndrome were reported. Hepatotoxicity (defined as elevation of alanine transaminase or bilirubin level) was observed in 33.3% of patients at any grade, with 4.8% experiencing Grade 3/4 toxicity. No renal toxicity (defined as elevated plasma creatinine level) was observed. Additionally, nausea was reported in 19.0% and fatigue in 23.8% of patients.

 

Table 3. Adverse events.

 

Two patients in our cohort underwent retreatment with PRRT after initial tumour progression. The first patient, who achieved a PFS of 19.9 months after the initial five cycles of ¹⁷⁷Lu PRRT, received a further five cycles of PRRT upon progression and achieved a partial response, with a subsequent PFS of 20 months. The second patient was undergoing a fourth cycle of PRRT retreatment, and the initial post-therapy scan showed partial response. No Grade 3/4 toxicities were observed except for Grade 3 lymphopenia in one patient.

 

To the best of our knowledge, this is the first local study to examine how effective PRRT is for treating NETs in Chinese patients from Hong Kong. An objective response rate (ORR) of 47.6% was observed, with a median PFS of 22.3 months, and a median OS of 45.2 months. The OS outcome closely mirrors that reported in the NETTER-1 trial, which included only patients with well-differentiated midgut NETs and reported a median OS of 48 months in the ¹⁷⁷Lu group, compared to 36.3 months in the control group.[16] In contrast, our study included a more heterogeneous group of primary tumours, particularly characterised by a significant proportion of pancreatic NETs. A recent large retrospective study by Mitjavila et al,[17] which evaluated 522 patients with a heterogeneous group of NETs including pancreatic, midgut, and bronchopulmonary subtypes, reported an ORR of 33.9%, a median PFS of 24.3 months and a median OS of 42.3 months. Our results align with these findings, further supporting the efficacy of PRRT in treating metastatic NETs across various primary sites in a Chinese patient population.

 

Our study demonstrated that bone metastases were significantly associated with poorer PFS and OS in multivariate analysis. This finding aligns with the results from Sitani et al,[18] who analysed a cohort of 468 patients and reported that bone metastasis significantly impacted OS. Similarly, Abou Jokh Casas et al[19] observed an inverse relationship between OS and the presence of bone metastases in their cohort of 36 patients treated with PRRT for gastroenteropancreatic NETs. The concordance of our findings with these studies supports the role of bone metastases as a potentially negative prognostic factor in metastatic NETs. The challenges posed by bone metastases may be attributed to the more aggressive nature of the disease, as bone metastases are considered a late event in NETs[20] and predisposes patients to serious skeletal events such as spinal cord compression.[21]

 

Our study also showed that a high liver metastatic burden of 50% or more, as assessed by radiological evaluation, was associated with poorer OS. A visual semi-quantitative assessment method for liver tumour burden was used, as recommended by the European Neuroendocrine Tumor Society,[22] and has been shown to be reliably reproducible.[20] The presence of neuroendocrine liver metastases is one of the most significant negative prognostic factors for long-term survival in patients with NETs.[23] A retrospective study by Ezziddin et al[11] analysed 68 patients with pancreatic NETs treated with ¹⁷⁷Lu PRRT and showed that a liver metastatic burden greater than 25% or above was associated with reduced OS. Our findings also align with a recent retrospective study by Swiha et al,[24] which demonstrated that liver metastases involving more than 50% of liver volume was associated with poorer OS. Although there was a trend towards poorer PFS with high liver metastatic burden, this was not statistically significant in multivariate analysis, likely due to the limited sample size and confounding factors.

 

Previous studies have suggested that tumours of GI origin generally have a better prognosis compared to those of pancreatic origin.[25] [26] However, our univariate analysis showed that pancreatic origin might have a better prognosis, though this was not significant in the multivariate analysis, indicating the influence of potential heterogeneity and confounding factors. Notably, prior research has shown that GI-origin NETs are typically associated with lower-grade tumours.[27] In contrast, within our cohort, 83.3% of GI-origin cases were classified as Grade 2 tumours or higher (including one Grade 3 tumour), whereas 73% of pancreatic-origin cases were all Grade 2 tumours and no Grade 3 tumours. Because only patients with progressive disease were referred for PRRT, our GI subgroup likely represents a selection of more aggressive cases. The inclusion of more aggressive histological subtypes within the GI-origin group in our cohort may explain the unexpected findings in our analysis.

 

Grade 3 NETs are associated with a worse prognosis compared to their lower-grade counterparts,[25] making treatment more challenging due to their aggressive nature. In our study, we had only one case of Grade 3 NET treated with PRRT, limiting further statistical analysis. The use of PRRT in Grade 3 NETs has mainly been supported by retrospective studies of heterogeneous groups.[12] [17] [18] [19] However, the recent Phase 3 NETTER-2 trial demonstrated significant improvements in PFS and ORR with ¹⁷⁷Lu-dotatate as a first-line treatment for higher-grade Grade 2 and Grade 3 NETs.[28] The primary analysis showed a median PFS of 22.8 months in the ¹⁷⁷Lu-dotatate group compared to 8.5 months in the control group, and an ORR of 43.0% versus 9.3%, respectively.[28] These findings underscore the potential efficacy of PRRT in treating high-grade NETs and suggest a promising therapeutic option for this challenging subgroup.

 

The toxicity profile observed in our study was generally favourable. The NETTER-1 study reported no Grade 3 anaemia, 1% neutropenia, and 2% thrombocytopenia.[16] Our study similarly found no Grade 3/4 anaemia or neutropenia. However, while NETTER-1 reported 9% Grade 3/4 lymphopenia,[16] our study observed a higher rate (42.9%), which may be attributed to the greater number of PRRT cycles administered. The toxicities were reversible, and none of the patients developed myelodysplastic syndrome. Only one patient developed Grade 3/4 liver derangement, and none developed any grade of renal toxicity, aligning with published retrospective studies that showed minimal Grade 3 toxicities of liver and renal function.[13] [17]

 

Finally, two patients in our cohort underwent retreatment with PRRT after tumour progression, receiving additional distinct PRRT treatment courses. Both patients showed a partial response. One patient is still undergoing treatment at the time of writing, and the other demonstrated a similar PFS between the two series of PRRT. No significant toxicities other than Grade 3 lymphopenia were observed.

 

According to a meta-analysis by Strosberg et al,[29] the median PFS following PRRT retreatment was 12.5 months, and the median OS was 26.8 months. The study also reported a low rate of Grade 3/4 toxicities,[29] confirming that PRRT retreatment generally maintains a manageable safety profile. It was also observed that PFS decreased after the second treatment course compared to the first.[30] Despite a limited number of retreatment cases in our cohort, the results are encouraging.

 

Our study has several limitations. First, as a retrospective study, it is subject to inherent selection bias and recall bias. Selection bias may have occurred because PRRT is an expensive treatment, making it more accessible to patients with better overall health and higher socio-economic status, potentially skewing the results towards more favourable outcomes. Recall bias may have influenced our findings since the data relied on medical records and follow-up reports, which might not have consistently captured all relevant information. Second, the sample size is relatively small and the study population was heterogeneous, which may limit the generalisability of the findings and introduce variability that could confound the results. Third, the estimation of PFS may have been affected by variability in the timing of CT scans across different patients, potentially leading to inconsistencies and bias in assessing disease progression. Finally, the small number of patients undergoing PRRT retreatment limited our ability to draw robust conclusions about the efficacy and safety of repeat treatments.

 

This study provides important insights into the use of PRRT in Chinese patients with metastatic NETs. Our findings suggest that PRRT is an effective and generally well-tolerated treatment option for this population, supporting its use in local clinical practice. We identified bone metastasis as a significant factor for worse PFS and OS, while a high liver metastatic burden was significantly associated with worse OS. These results can help clinicians in Hong Kong optimise patient selection and management strategies for the treatment of NETs. However, larger prospective studies are needed to further validate these findings.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Hong Kong J Radiol 2025 Sep;28(3):e179-86 | Epub 12 September 2025

 

 

¹ Department of Radiology, University of Ottawa, Ottawa, Canada

² Systems Integration and Architecture, The Ottawa Hospital, Ottawa, Canada

 

  Full Article in PDF

Intracranial tumours can pose a diagnostic challenge in clinical practice. Recent advancements in artificial intelligence in radiology may provide additional diagnostic information. Radiomics techniques such as texture analysis can reveal grey-level patterns beyond what is possible through expert human visual perception alone.[1] Numerous quantitative textural features can be derived, including simple statistics based on grey-level histograms, and higher-order features based on spatial relationships among pixels.[2] These radiomic features, derived from conventional magnetic resonance imaging (MRI) sequences, can be leveraged to train various machine learning (ML) models to detect and classify brain tumours, as well as predict prognosis and treatment response.[3] [4] [5]

 

Previous applications of radiomics in brain tumour research have aimed to improve diagnosis and post-treatment imaging of gliomas, including grading, distinguishing tumour progression from pseudoprogression, patient survival, and genetic expression.[3] [4] Other neuro-oncologic ML advances include differentiating gliomas from mimics such as meningiomas, pituitary tumours, and solitary metastases,[5] [6] diagnosis of paediatric tumours,[7] [8] and detecting metastases.[9] Radiomic approaches can offer more consistent results with good external validity, compared to the interrater variability of human readers.[6] More recently, deep learning models trained to detect brain metastases have shown advantages over classical ML in terms of lower false-positive rates albeit with greater training data requirements.[9]

 

Accurately determining whether a lesion arises from within the brain parenchyma (intra-axial, IA) or from the surrounding structures (extra-axial, EA) is crucial for diagnosis and treatment planning. This study aimed to determine whether ML of MRI radiomic features can differentiate between IA and EA locations.

 

This retrospective study performed at a single tertiary-care academic centre. Medical record review was conducted per the guidelines of the Institutional Review Board. We identified 92 consecutive adult patients (age >18 years) who underwent brain MRI for a newly diagnosed solitary brain lesion from January 2017 to December 2018. Patients with multiple lesions or prior surgery or chemo/radiotherapy were excluded. Data on age, sex, and final histopathology were recorded.

 

MRI of the brain was performed using 1.5 T (25 IA and 9 EA tumours; Siemens Magnetom Symphony, Siemens Medical Systems, Erlangen, Germany) or 3 T (45 IA and 13 EA tumours; Siemens Tim Trio or GE DISCOVERY MR 750w, GE Medical Systems, Milwaukee [WI], US) systems. Using a dedicated head coil, three-dimensional (3D) axial stacks of post-contrast T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI) were acquired with the following parameters: 1.5 T: T2WI (fast spin echo with fat saturation): TR/TE 3510/97, echo train length 9, flip angle 180°, section thickness 5.5 mm; post-contrast T1WI (fast spoiled gradient echo with fat saturation): TR/TE 6.73/2.71, flip angle 15°, section thickness 1 mm. 3 T: T2WI (fast spin echo with fat saturation): TR/TE 6700/97, echo train length 18, flip angle 120°, section thickness 3 mm; post-contrast T1WI (fast spoiled gradient echo): TR/TE 8.48/3.21, flip angle 12°, section thickness 1 mm. Post-contrast T1WI were acquired after hand injection of 0.1 mmol/kg of gadobutrol (Gadovist; Bayer Healthcare, Hong Kong, China), followed by a 10-20 mL saline flush, using a 4–5-minute delay before acquisition at 3 T and a 6-to-7-minute delay at 1.5 T. All images were reviewed using a Picture Archiving Computed System (PACS; Horizon Medical Imaging, McKesson Corporation, San Francisco [CA], US).

 

The IA or EA designation was confirmed by final histopathology following biopsy, extracted from the electronic medical records. The specimens were obtained by the neurosurgeons and analysed by the neuropathologists at our centre.

 

Tumour volumes of interest (VOI) were manually segmented on both T2WI and post-contrast T1WI using ImageJ version 1.52r (National Institutes of Health, US, https://imagej.net/) by a neuroradiology fellow, under supervision of a staff neuroradiologist with over 30 years of experience. VOI contours were subsequently submitted to a blinded medical imaging scientist (redacted/blinded for review) for texture analysis. Examples are shown in Figure 1.

 

Figure 1. Manual segmentation in two sample patients. (a) T1-weighted post-contrast and (b) T2-weighted magnetic resonance imaging (MRI) sequences showing an extra-axial mass arising from the left frontal falx. (c) T1-weighted post-contrast and (d) T2-weighted MRI sequences from a different patient with a right anterior temporal mass.

 

First- and second-order statistical textural features were computed for each VOI and MRI sequence using MaZda software (version 4.6.0; Institute of Electronics, Technical University of Lodz, Lodz, Poland).[10] First-order features included grey-level histogram mean, variance, skewness, kurtosis, and percentile values (1st to 99th). Second-order features included grey-level co-occurrence matrix (GLCM[11]) and run-length matrix (RLM[12]) features (11 GLCM and 5 RLM features per sequence) Before computing GLCM and RLM features, signal intensities were normalised between μ ± 3σ (where μ was the mean value of grey levels inside the VOI [or VOI subzone] and σ was the standard deviation) and decimated to 32 grey levels to minimise inter-scanner variability.[13] [14]

 

We used XGBoost,[15] an open-source ML algorithm, to train models on the textural features extracted from T2WI, post-contrast T1WI, and their combination. Hyperparameters were tuned using an 100-trial Bayesian optimisation experiment via GPyOpt[16] (http://github.com/SheffieldML/GPyOpt), guided by Gaussian process modelling and an exploration-exploitation heuristic. Each model was evaluated using stratified ten-fold cross-validation, repeated 10 times.[17] The SHAP (Shapley Additive exPlanations) framework[18] was used to estimate each feature’s relative importance, normalised to sum to 1.0.

 

Statistical analysis was performed using RStudio, an open-source software (version 1.3.1093; PBC, Boston [MA], US). Mann-Whitney U tests were used to compare IA and EA groups for each feature. Stepwise Holm-Bonferroni correction was applied for multiple comparisons.[19] Model performance was assessed using accuracy, the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1 score. ROC confidence intervals (95% CI) were calculated using 5000 bootstrap iterations, and AUC differences were assessed using DeLong’s method.[20]

 

A total of 92 lesions were analysed (70 IA and 22 EA). Table 1 summarises demographics and histopathological data. Glioblastomas (44%) and metastases (27%) were the most common diagnoses in the IA group, while meningiomas (68%) were the most common in the EA group. The patients aged from 23 to 84 years, with a mean age of 58.8 years. No significant difference was noted between groups (IA median, 62 years [interquartile range, 18] vs. EA median, 55 years [interquartile range, 10]; p = 0.09). Males were more common in the IA group (66% vs. 23%; p < 0.01).

 

Table 1. Demographics and histopathologies (n = 92).

 

Table 2 provides median and interquartile range values for individual 3D textural features. EA tumours showed higher histogram mean, kurtosis, and 10th and 50th percentiles on post-contrast T1WI, and lower skewness (p < 0.001 for all). On T2WI, only kurtosis was significantly higher in the EA group (p < 0.001), while the 90th and 99th percentiles were significantly lower (p = 0.002 and 0.001, respectively).

 

Table 2. Median values of 25 individual textural features used in machine learning classification.

 

Among the second-order features computed from post-contrast T1WI, GLCM correlation, sum of squares, sum variance, and sum entropy associated with EA tumours were significantly lower than those computed for the IA group (p = 0.002 for correlation and p < 0.001 for the rest; Table 2). The T1 GLCM sum average and difference variance were both significantly greater in EA tumours compared to the IA group (p = 0.001 and p < 0.001, respectively). Similar to post-contrast T1WI, the GLCM correlation, sum of squares, and sum variance features evaluated from T2 images were significantly lower than those in the IA group (p < 0.001 for each). The T2 GLCM difference variance was significantly greater in EA tumours compared to the IA group (p < 0.001). Among the group differences assessed for RLM features, none was found to be significant after Holm–Bonferroni correction for multiple comparisons (Table 2).

 

The classification performance metrics for each of the three ML models are summarised in Table 3. Receiver operating characteristic curves and feature attribution scores are depicted in Figures 2, 3, and 4. The AUC for the identification of IA tumours was 0.91 (95% CI = 0.89-0.93; Figure 2a) for the model based on T1-weighted post-contrast MRI features, 0.81 (95% CI = 0.78-0.84; Figure 3a) based on T2-weighted MRI features, and 0.92 (95% CI = 0.90-0.94; Figure 4a) based on all features.

 

Table 3. Classification performance metrics for each magnetic resonance imaging textural feature of the machine learning model.

 

Figure 2. (a) Receiver operating characteristic curve showing the area under the curve (AUC) for the XGBoost model, and (b) SHAP (Shapley Additive exPlanations) feature importance plot for the same model trained on T1-weighted post-contrast magnetic resonance imaging textural features to identify intra-axial tumours. SHAP values are aggregated across samples and normalised to sum to 1.0.

 

Figure 3. (a) Receiver operating characteristic curve showing the area under the curve (AUC) for the XGBoost model, and (b) SHAP (Shapley Additive exPlanations) feature importance plot for the same model trained on T2-weighted magnetic resonance imaging textural features to identify intra-axial tumours. SHAP values are aggregated across samples and normalised to sum to 1.0.

 

Figure 4. (a) Receiver operating characteristic curve showing the area under the curve (AUC) for the XGBoost model, and (b) SHAP (Shapley Additive exPlanations) feature importance plot for the same model trained on combined T1-weighted post-contrast and T2-weighted magnetic resonance imaging textural features to identify intra-axial tumours. SHAP values are aggregated across samples and normalised to sum to 1.0.

 

Table 3 shows that the three models (T1WI, T2WI, and combined) yielded AUCs that were significantly greater than 0.5 (p < 0.0001 for each comparison). The model produced by the T2WI features alone resulted in an AUC that was significantly lower than the model produced by either the T1-weighted post-contrast or combined sequences (p < 0.0001 for each comparison). The combined model AUC was not significantly greater than the T1 model alone (p = 0.135). All models had high sensitivity in identifying intra-axial tumours (>90% for each), but none achieved high specificity (39%-64%). Nevertheless, the models attained acceptable levels of accuracy (>80%) and precision (>88%).

 

The feature importance attribution profile associated with the model trained using T1-weighted post-contrast MRI features reveals that grey-level kurtosis, skewness, sum entropy, 10th (histogram) percentile, and sum variance contributed two thirds of the total proportional importance score (0.67/1.00) for this model (Figure 2b). Sum variance and kurtosis also contributed strongly towards the total proportional importance of the T2-based model (0.40/1.00), with sum of squares, difference variance, and 1st percentile contributing an additional 0.36/1.00 towards the total proportional importance for the T2-based model (Figure 3b). In the combined model, the T1WI skewness, kurtosis, sum entropy, and 10th percentile were found to represent four of the five ‘most important’ features, contributing 0.50/1.00 of the total proportional importance (Figure 4b).

 

Classically described features of EA lesions include broad-based dural attachment, adjacent bony changes, formation of a cerebrospinal fluid cleft, deviation of pial vessels, and buckling of the grey-white junction.[21] The most common primary EA lesions include meningiomas, schwannomas, pituitary adenomas, and Rathke’s cleft cysts, all of which have characteristic locations and signal properties that can aid the radiologist’s diagnosis.[21] For example, pituitary adenomas always arise in the sella or suprasellar location and may be completely T1 hypointense, or may also contain areas of cystic change and haemorrhage, and demonstrate delayed enhancement. Conversely, primary IA lesions are dominated by gliomas and lymphoma, with gliomas showing an infiltrative T2/fluid attenuated inversion recovery hyperintense pattern of spread along white matter tracts, and lymphoma characterised by greater diffusion restriction of its solid component.

 

Nevertheless, lesions can be challenging to localise visually, especially if large and associated with mass effect or oedema. These cases may require advanced imaging such as MRI perfusion and spectroscopy. In certain morphologically complex cases, the tumour origin may not be known until after biopsy. Utilising all available data, including quantitative radiomic features, could potentially improve treatment planning and avoid unnecessary procedures. Another practical benefit of radiomic models, should they achieve parity with human readers, would be the ability to reduce the number of imaging sequences required (e.g., axial T2WI and post-contrast T1WI, as used in our study).

 

We analysed 3D radiomic features of brain tumours using a ML framework (XGBoost), which determined IA versus EA location with high accuracy and AUC. The textural features used in the present study have been well described.[2] The T1-weighted post-contrast model achieved a classification performance comparable to the combined T2WI and post-contrast T1WI model, and both performed significantly better than the model trained on T2WI-based features alone.

 

Various studies have utilised ML for similar tasks with excellent results. In a meta-analysis of 29 ML studies in neuro-oncology focused on patient outcomes, tumour characterisation, and gene expression, Sarkiss and Germano[4] reported a pooled sensitivity ranging from 78% to 98%, specificity from 76% to 95%, and greater accuracy compared to conventional imaging analysis in predicting clinical outcomes such as survival, high-versus low-grade tumours, and future progression. Tetik et al[5] developed an automated deep learning model to distinguish among gliomas, meningiomas, and pituitary tumours, achieving over 88% on all performance metrics including sensitivity, specificity, and Matthews correlation coefficient. A direct comparison of two human readers, a traditional ML model and a deep ML in differentiating glioblastoma from solitary metastases yielded similar performance, with AUCs of 0.77 and 0.90 in human readers and 0.89 and 0.96 for the traditional and deep ML models, respectively.[6] This study also highlighted the potential for robust generalisability with validation performed at a different institution, and greater inter-rater agreement between the ML models than the two human readers (albeit with differing levels of experience).[6]

 

Our study has several limitations. While all three models demonstrated high sensitivity for identifying IA tumours (>90% for each), none achieved high specificity (T1: 61%, T2: 39%, and combined: 64%). This is most likely due to the case imbalance between the IA (n = 70) and EA (n = 22) groups, which can artificially raise the sensitivity and accuracy for predicting IA tumours. The imbalance reflects the incidence of these tumours[21] and the nature of consecutive data acquisition. Second, the small sample size clearly limits the generalisability of our models. The current study was designed to assess whether MRI textural features contain sufficient predictive information for our XGBoost models to generate effective classifiers. Accordingly, the use of an established ten-fold cross validation method[17] was appropriate for estimating the generalisation errors of the models trained on post-contrast T1WI, T2WI, and combined MRI features. This preliminary stage is distinct from the development and validation of a single model intended for clinical deployment, which would require a much larger dataset and evaluation on external or ‘out-of-distribution’ data.[22] With improved optimisation, increased sample size and external validation, specificity could be enhanced and such a model could theoretically augment or complement the radiologist’s assessment in ambivalent cases. Finally, it should be noted that our dataset may be skewed towards aggressive or large lesions requiring resection, as we used a pathological reference standard. As a result, the extracted features could be skewed towards tumours that required resection, rather than asymptomatic lesions such as non-aggressive meningiomas.

 

A location-based ML classification model for differentiating IA from EA tumours is feasible based on this preliminary study, demonstrating good sensitivity. However, specificity was low to moderate, likely due to the imbalanced dataset. Further study with a more balanced cohort and external validation is required to optimise performance.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Hong Kong J Radiol 2025 Sep;28(3):e187-97 | Epub 12 September 2025

 

 

¹ Department of Clinical Oncology, Tuen Mun Hospital, Hong Kong SAR, China

² Department of Radiology, Tuen Mun Hospital, Hong Kong SAR, China

³ Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong SAR, China

 

  Full Article in PDF

Statistical analysis shows adverse events of magnetic resonance imaging (MRI) growing at nearly three times the rate of MRI procedure volume growth.[1] The potential risks in magnetic resonance (MR) are related to the three types of magnetic fields used in MRI: the static magnetic field (B₀), the radiofrequency (RF) field (B₁), and the time-varying magnetic field gradients. Each of the three creates both their own and combined safety risks including projectile forces, torque force, biological effects, biomedical implant and device risk, cryogen-related bodily harm and asphyxiation, heat deposition and acoustic noise—all of which have the potential to cause significant harm or even death.

 

The potential risks in MRI are associated with the three major electromagnetic fields: the B₀, the varying magnetic field gradients, and the time-varying B₁.[2] The ultra-low temperature helium found in superconductive magnets presents a risk. With the new generation of sealed, low-volume helium scanners, handling cryogenics may not be required.

 

Most clinical MR scanners in use today are superconducting electromagnets with a superconducting solenoid coil (niobium-titanium) immersed in liquid helium at -269℃ (4°K). Even without an external power supply, the magnet’s magnetic field remains unchanged because the electrical resistance of superconductors is negligible; therefore, the risk associated with the superconductive magnetic field is always present. Clinically available scanners have magnetic fields of typically 1.5 or 3.0 T. It is estimated that approximately 100 ultra-high field 7-T MR scanners have been released for clinical use in Europe and the US. The potential risks associated with a B₀ and its spatial field gradients with sharp slope near the scanner include biological effects on humans (such as vertigo, nausea or magnetophosphenes), as well as the translational and rotational forces acting on objects, with the associated device displacement and medical device disruption. For current-carrying objects, Lenz’s force applied to the objects can result in movement in the magnetic field; however, patients with ferromagnetic heart valves are typically excluded from MRI.

 

The interaction between the magnetic field and objects greatly depends on the magnetic properties of the materials and their shape. Based on the behaviour of materials in the magnetic field, materials are generally classified into three categories: (1) diamagnetic substances such as calcium produce negative magnetisation when placed in an external magnetic field; (2) paramagnetic substances acquire magnetisation in the direction of the applied external magnetic field; and (3) ferromagnetic materials are strongly attracted by the applied magnetic field. Magnetic susceptibility is defined as the magnitude of the extent to which an object becomes magnetised when placed in a magnetic field.[3] Table 1 lists the magnetic properties of a number of materials.[2] [3] Most biological tissues contain a high proportion of water (H₂O) and have weakly diamagnetic susceptibility (χ), typically around -11 × 10^-6 to -7 × 10^-6. Among the paramagnetic and diamagnetic materials, the χ of most substances encountered in routine clinical imaging lies in the range of approximately -10^-5 < χ < 10^-5. Most modern implants that claim to be MRI-safe are either diamagnetic such as copper, or paramagnetic such as titanium.

 

Table 1. Magnetic properties of materials.[2] [3]

 

The two types of forces exerted on metal objects are translational and rotational. The forces on diamagnetic or paramagnetic materials are generally weak to negligible, regardless of whether gravitational force is considered. Forces on ferromagnetic objects are of paramount concern, as they experience the greatest forces in the MR environment. The translational force (F_t) in Equation 1 increases when there are rapid changes in the magnetic field with high spatial field gradients. It is strongest at the edge of the magnet bore with a very sharp slope, inversely related to the third power of the distance (1/r³). The rotational force (F_r) is generally greatest at the centre of the magnet bore, as it is proportional to the square of the B₀, as shown in Equation 2. V in the equation means the volume of the metal device. Elongated objects experience stronger torques compared with isotropic objects. Ferromagnetic medical implants may move rapidly in the B₀, and the temporary or permanent B₀ field–induced current may be substantial enough to hinder the normal function of electronically powered or magnetically programmed active implanted medical devices, such as disabling the reed switch of an MR Conditional pacemaker.

 

 

For small asymmetrically shaped ferromagnetic objects implanted in the body, the rotational force may become the dominant safety issue. In a 10-year review of 1548 adverse MRI-related events reported to the US Food and Drug Administration (FDA), 133 (9%) involved projectiles.[4]

 

Patients or medical staff might experience vertigo, dizziness or nausea when approaching or moving towards the scanner. Different theories[5] [6] [7] [8] have been proposed to explain this phenomenon in which the Lorentz force concept is the favoured explanation.[9] According to the Lorentz force law, Lenz’s force is applied to the current-carrying objects when placed inside and moving within the magnetic field. The normal potassium-based ionic current within the middle inner ear endolymph will experience Lenz’s force with head movement in the B₀. This force is transmitted to the ampulla, displacing the crista and hair cells of the canal, stimulating them to generate impulses within the vestibular nerve and resulting in vertigo. This is the predominant source of the physiological response associated with transient sensations of vertigo, dizziness or nausea in MRI. The Lorentz forces are also responsible for the magnetohydrodynamic effect. Human blood is conductive. A Lorentz force is created when ionic currents in the thoracic aorta flow through the magnetic field B₀. The Lorentz force deflects positive and negative ions towards opposite sides of the vessel when blood flows through a magnetic field. A voltage is induced. This voltage is superimposed on the T-wave of the electrocardiogram (ECG) used to monitor the patient and elevates it. The distortion of the recorded ECG by the magnetohydrodynamic effect results in faulty cardiac triggering for cardiac MR scanning and makes the cardiac MR examination quite challenging. This interferes with the interpretation of the ECG that renders it unreliable, especially when patients experience chest pain inside the scanner.

 

Some MRI patients might observe the flickering lights known as magnetophosphenes. They are generally considered to be the result of motion-induced currents when the eyes or head move through a B₀. According to the Faraday–Lenz law, an electric field (or current) is induced in a conductor whenever it moves through a B₀. The induced currents directly stimulate the retina when there is physical movement of a person’s head within the B₀. The generation of electric currents in the tongue due to magnetically induced electric fields is viewed as the true cause of the metallic taste experienced by patients who undergo routine MRI examinations.[10]

 

There are three orthogonal linear gradient magnetic fields (expressed in mT/m) generated by three sets of coils (one set for each of the x-, y-, and z-directions) for image spatial encoding and contrast manipulation during image acquisition. There is no concern about the static effects generated by the gradient magnetic field as the strength of the magnetic fields (the maximum amplitude per axis 40-80 mT) generated by gradients is much weaker than that of the main magnetic field (B₀). However, there are three potential MR safety concerns associated with the time-varying gradient magnetic field. Modern MR scanners are equipped with powerful gradients to facilitate rapid, high-resolution imaging or shorter echo times and echo spacing. Gradient coils are powered by high voltage up to 1500 V and high current of several hundred amperes on one side. On the other side, the gradients are switched on and off quickly with slew rates as high as 200 T/m/s in practice. Two major physical effects and the associated three potential MR safety concerns are produced by the rapidly changing currents flowing through the gradient coils.[11] The two major physical effects are mechanical vibration of the MR system and the induced currents in nearby conductive materials (the induced currents are proportional to dB/dt, i.e., the rate of change of the gradient field), respectively. The MR safety concerns include noise, nerve/cardiac stimulation, and tissue heating.

 

The movement and vibration of the coils due to mechanical forces are the primary sources of acoustic noise in MR scanners.[12] The sound pressures generated during routine MRI can reach as high as 100 to 130 dB depending on which pulse sequences are used. It is required and mandatory to provide hearing protection when acoustic threshold exposure limits exceed 99 dB by the International Electrotechnical Commission (IEC).[13] [14] [15] Some patients have headaches and hearing loss following MRI examinations when not wearing appropriate hearing protection.[16] The hearing protection should reduce noise to at least 99 dB for patients and 85 dB for personnel in the examination room. Although concerns have been raised about MRI scans in pregnant women due to potential risks to fetal hearing or other effects,[17] [18] no harmful effects have been reported over the past 30 years for those scanned during the first trimester. Despite limited data on fetal hearing risks, it is still recommended to establish institutional policies for MRI exposure in pregnant patients. Pregnant healthcare practitioners are permitted to work in and around the MR environment throughout all stages of pregnancy.[19] Although permitted to work in and around the MR environment, pregnant healthcare practitioners should be advised not to remain within the MR scanner bore during actual data acquisition or scanning.[20]

 

According to Faraday’s law of induction as mentioned above, time-varying magnetic fields result in the generation of electric fields in conducting materials and an electromotive force. The gradient switching-induced electric fields in a human subject stimulate the nerves and muscle fibres and may cause what is referred to as peripheral nerve stimulation (PNS).[11] [21] It is generally reported as a tingling or tapping sensation, although the severity of discomfort ranges from barely noticeable to physically dangerous at the other extreme, depending on the subject’s physiological conditions. The patient’s overall health, nerve sensitivity, and even stress or anxiety can affect their perception of the stimulation.[22] Meanwhile, the intensity of nerve and muscle fibre excitation is proportional to the dB/dt and the duration of its application. The IEC has established limits for gradient exposure to protect patients and subjects against PNS and cardiac stimulation,[23] which have been adopted by the US FDA and many other organisations. However, PNS stimulation limits for both whole-body and regional scans can be determined by averaging the individual stimulation thresholds of test subjects (at least 11 volunteers), based on studies conducted with appropriate ethics committee approval, rather than using derived values. The first-level controlled operating mode is defined such that 50% of all patients experience at least mild stimulation after reaching the stimulation threshold, while the normal operating mode limits the scanner to 80% of this threshold.[24] [25] Cardiac muscle contraction requires levels of stimulation at least 10 to 100 times higher than those required for PNS, and a subject accidentally exposed to very high levels of dB/dt would almost certainly experience warning signs of PNS before reaching levels that pose a risk to the heart.[22]

 

Changing magnetic fields from both RF pulses and switched gradient fields generate electric (eddy) currents. In the presence of a conducting medical device or implant, thermal energy is produced both within the implant itself and in the adjacent tissues by these eddy currents. Heating of conducting devices and the adjacent tissues will be discussed later for RF pulses, while the heating due to the instantaneous power deposited by the eddy currents from the switched gradient field (dB/dt) will be covered here.

 

The degree of energy deposition can be quantified by the specific absorption rate (SAR), which is often expressed in units of power per mass of tissue (watts/kg). Each manufacturer provides a conservative estimate of SAR for all commercially available MR scanners. The SAR values are estimated automatically using a specific imaging protocol and patient-specific information as input, and a warning message will appear if regulatory limits are likely to be exceeded. Considering the factors contributing to SAR, it can be approximated by a simple model for the switched gradient field (Equation 3)[26] [27]:

 

 

where σ is the tissue conductivity, A is the volume of the body size, D is the duty cycle (representing the percentage of time the gradient operates at maximum amplitude during a sequence), and ρ is the tissue density.

 

Because gradient frequencies are quite low, lying in the range of kHz, gradients do not generate appreciable eddy currents in tissues. The thermal effects due to heat diffusion from the implant itself may be considered, and these effects are likely to come into play only near the regions of maximum dB/dt for large-volume implants.[28] [29]

 

B₁ is applied perpendicular to the main magnetic field (B₀) on the order of milliseconds. It tips the net magnetisation out of alignment with B₀ and MR signals are produced. B₁ is weak (μT) and oscillates at a frequency in the MHz range matching that of a proton, with resonance frequencies of approximately 64 MHz and 128 MHz for 1.5 T and 3.0 T, respectively. The primary safety concerns at these frequencies are whole-body and localised heating from the deposition of the RF energy.[30] In a 10-year review of 1548 adverse MRI-related events reported to the US FDA, 906 (59%) involved thermal injury, making it the most prevalent reported injury.[4] According to Maxwell’s Laws, the time-varying B₁ is the source of an induced changing electric field. Such field deposits energy into tissues, and the power applied to tissue is generally a function of field strength, pulse sequence, and patient size. The primary safety concerns are the whole-body temperature increases due to heating absorbed in the patient and the potential for tissue damage from localised high-temperature exposures.[31] [32] [33] [34] As internal temperature measurement is not easily performed during routine clinical MRI, SAR or specific energy dose (SED), which reflects the total energy delivered into the patient during the active scan period, is used to control system power output in modern MRI. This is approximated by Equations 4 and 5.

 

 

Some MR manufacturers now compute and report both SAR and SED to limit scanning during a full exam if the accumulated SED is too high. In addition to the dosimetric unit used for diffusion heating over a large volume, B⁺_+1rms (the root mean square value of B⁺₁) is used as an supplemental metric to SAR, which may be a better exposure measure for focal heating because it is more closely related to the induced electrical field and is less dependent on the patient. The major use of B⁺_+1rms is for MR Conditional implants. Implant manufacturers are responsible for providing the value for the safe use of their devices in an MR scanner.

 

It is generally believed that three physical mechanisms underlie RF-induced thermal injury.[31] [35]

 

Both the human body and metallic foreign bodies are conductors. The currents induced by RF excitation in modern MRI reside almost entirely along the surface of the conductive materials, or along the conductive loop if there are no areas of high resistance. The eddy currents induced by the changing RF magnetic fields are channelled into areas of high resistance (such as a metal-skin interface or breaks in the loop); however, the primary concern is that these current distributions can lead to resistive heating of tissue and RF burns.[36] This is analogous to resistive energy loss in a conventional electrical circuit governed by Ohm’s Law. Resistive heating in tissue is a function of material conductivity, geometry, and location within the excitation coil. Implants located closer to the edge of the coil tend to experience higher electric fields. The skin itself is conductive, and skin-to-skin contact can lead to a high current concentration. The associated energy deposited may be substantial enough to cause tissue damage.[37] The point of contact is a potential region of high resistance where significant heating can occur.[38] For example, crossing of legs, ankle to ankle, thigh to thigh, and so on. This phenomenon is particularly relevant when patients are under general anaesthesia in an intraoperative MRI.

 

For smaller-sized conducting materials (<2 cm), there is no great concern for significant heating issues if there are no adjacent conductors within approximately 3 cm.[39] Otherwise, there may be enhanced heating due to coupling effects. Larger and smoother conductors can generate a significant amount of current. These currents flow through only a tiny fraction of the total implant mass at its surface and do not cause significant heating of the implant itself. In soft tissues immediately adjacent to the implant or at sharp corners or disconnects, or when in close proximity to another conductor, however, RF-induced currents can become concentrated and resistive heating may occur.[28] [29] [40] Several cases of thermal tissue damage caused by implants have been reported, such as from a deep brain stimulator and MR Conditional intracranial pressure monitoring devices.[41] [42] These examples highlight the importance of strictly adhering to the manufacturer’s guidelines.

 

In some cases, certain electrical circuits might exhibit resonance absorption and release energy at a specific resonance frequency. It is a relatively uncommon situation. However, if the electrical circuits contain both capacitance and inductance elements to form resonant loops, they may generate a very large amount of current and a high level of inductive heating through resonant absorption and energy release at a resonance frequency.[43] This situation applies to the use of ECGs with cables or similar loops.

 

Another mechanism for RF heating is the so-called antenna effect. Straight wires and elongated conductive objects can act like antennas, capturing electromagnetic waves to extract power from them. According to antenna theory, the length of the wire or object must be sufficient to support the formation of standing waves and produce standing-wave patterns of voltage and current that are concentrated near their tips. Typically, when the length is close to one-half of the RF wavelength, maximum heating may be produced at the tips of the device. For MRI, the relevant length is approximately 26 cm at 1.5 T and 13 cm at 3.0 T. There have been incidents resulting in fire and patient burn injuries ascribed to the antenna effect of ECG leads and cardiac pacing leads as well.[44]

 

For MR safety, evaluation of impact on patient, fetus, family and staff, as well as interaction with auxiliary equipment and medical devices, is of constant concern. The electromagnetic field–related hazards are summarised in Table 2.

 

Table 2. Summary of primary safety concerns related to magnetic fields.

 

Given the risks associated with the MR environment, it is essential to take effective measures and procedures to keep all personnel including patients, accompanying family members, and staff safe, and to ensure all auxiliary medical equipment and devices remain functional.

 

The American College of Radiology (ACR) has divided the MRI suite into four zones corresponding to the potential safety concerns.[45] [46] The purpose of this definition is to prevent unqualified staff and unscreened patients from accessing hazardous areas and to restrict MR-unsafe medical equipment or devices from being wrongly brought into the MR scanner room. There are other alternative schemes such as the three-area definition from the United Kingdom or Netherlands[47]; however, the ACR zone definition is widely adopted throughout the world.

 

As shown in Table 3,[45] [46] Zone I is a public area that the general public can access freely without supervision, where the fringe field is less than 5 gauss. Zone II is a buffer area between Zones I and III for patient preparation and safety screening. Zone III is the area near the magnet room with potential hazards to unscreened patients and personnel; physical barriers are used to help control access. Zone IV is the MR scanning room with the highest risk, where all ferromagnetic objects are forbidden. Only properly screened personnel and patients are permitted to enter this area.

 

Table 3. Safety zones in the magnetic resonance environment.[45] [46]

 

During the early implementation of MRI technology, the 5 gauss (0.5 mT) line or area was established as the threshold to define the limit beyond which ferromagnetic objects and the general unscreened public are strictly prohibited. It also served as a reminder that one is within a region where active medical device might pose a hazard due to exposure to the electromagnetic fields produced by MR equipment and accessories. Recently, the 9-gauss line has been updated to indicate the standard for identifying the ‘magnet mode’ area for certain active implantable medical devices, particularly cardiac devices, to prevent accidental activation or functional changes. The International Standard IEC 60601-2-33 for MRI safety requirements was amended to reflect this change.[23] Magnetic fields extend in all directions, and the 9-gauss line may extend into non-MR areas above, below, or adjacent to the MR magnet room, which should be carefully evaluated and clearly marked to restrict access by unauthorised personnel.[48]

 

The ACR Manual on MR Safety[49] suggests that all patients and non-MR personnel must undergo MR safety screening when entering Zone III.[50] Trained MR personnel have the responsibility and authority to decide whether a patient may be cleared for scanning. For non-emergent patients, the ACR recommends performing at least two separate screenings before granting access to the MR scanner room. Screening, in the form of a questionnaire, should be available. One of the screenings should be conducted when the examination is requested and should include questions such as: (1) any implanted cardiac devices (pacemakers, defibrillators, valves, stents, wires, etc.); (2) intracranial vascular coils or aneurysm clips; (3) neurostimulators; (4) bone growth or bone fusion stimulators; (5) cochlear implants; (6) the possibility of intraorbital metallic foreign bodies; (7) implanted infusion devices such as those for insulin; and (8) orthopaedic implants.

 

The second level of screening should be performed when patients themselves present to the MR suite for MR examination. Conscious patients should be screened at least twice, using metal detectors and verbal questioning, before being allowed to enter the MR scanning room (Zone IV). The screening form similar to the MR safety screening form provided by the ACR should be reviewed by staff with MR training, such as an MR nurse. MR radiographers or technologists should then review and evaluate the form in detail; ideally, both parties should sign it. MR radiographers should ask whether any device, foreign body, or implant is present that could pose a danger in the MR room, including both passive and magnetically or electrically active items. For patients who are unable to answer screening questions, such as children, it is acceptable to question family members, guardians, health carers or any decision makers. All patients should be asked to change into MR-safe hospital gowns and to remove all watches, hearing aids, hairpins, jewellery, drug delivery patches, eye makeup, artificial lenses (especially high-technology ones, such as implanted contact lens monitors intraocular pressure with a micro-sensor), and so on. For emergent patients, screening with a metal detector is acceptable but it must be performed by MR-trained personnel or MR radiographers. Before entering the MR scanner room, all patients should undergo final screening. The use of ferromagnetic detection systems is recommended as an adjunct to enhance detection of ferromagnetic materials.[49] If a patient with a metalworking history reports the presence of metal in the orbital area, an X-ray should be taken or their previous X-ray history reviewed. The handheld metal detector should have a strength of at least 1000 gauss with the ability to detect ferromagnetic or magnetic objects. Metal detection equipment is helpful in screening non-MR personnel, especially those from other medical departments who may or may not have MR training, since they can easily forget to remove their personal items before entering the MR magnet room (Zone IV).[51]

 

For all non-MR personnel entering the MR scanner room, for example a family member or carer who wishes to accompany the patient, they should be screened using the same criteria as those applied to patients. For cleaning of scanning room, personnel who have received basic MR safety training may perform their duties under MR personnel supervision and only after undergoing the same screening procedure as patients. All MR personnel should undergo the same screening as patients to ensure their own safety in the MR scanner room and to protect the non-MR personnel under their supervision. Pregnant healthcare practitioners are permitted to work in and around the MR environment throughout all stages of their pregnancy. Although they are allowed to work in and around the MR environment, pregnant healthcare practitioners are advised not to remain inside the scanning room while data acquisition is in progress. These recommendations are based on the preponderance of data relating to 3T magnetic fields.

 

The US FDA introduced guidelines on testing and labelling medical devices and implants for safety in the MR environment, which apply to all medical devices that might be used in the MR environment.[52]

 

The square green MR Safe label indicates that the object or device is safe in all MR environments. It is non-magnetic, non-conductive and non-metallic, posing no known hazards in any MR environment. Caution should be taken with products marked as ‘MR SAFE’ as some of these products have been found with metallic components according to our experience, and should therefore be treated with care. The round red MR Unsafe label applies to objects or devices that pose potential harm to MRI patients or staff under all MR circumstances, for example, ferromagnetic objects. The triangular yellow MR Conditional label is for devices or objects that may be used safely in an MR environment, provided that the conditions for safe use are fulfilled.

 

The MR safety profiles for all accessory devices used in the MR suite must be well-established before being brought into Zone IV to avoid potential safety risks to patients undergoing MRI. All devices being used in the MR suite should be clearly marked with their MR safety status (i.e., MR Safe, MR Conditional, or MR Unsafe). Any new device or replacement must be tested for MRI safety before use in the MR scanner room. Non-clinical incidental objects, such as ladders or home-made phantoms (e.g., custom-built imaging test objects) with no manufacturer or third-party MR safety test results, as per American Society for Testing and Materials standard,[48] should be site-tested prior to use in the MR magnet room. Unmarked or unknown items must not be allowed into the MR scanner room. Never assume a device’s MR Conditional or MR Safe status unless it is clearly documented in writing. All accessories used within the MR magnet room must be labelled either with MR Safe or MR Conditional. The operating conditions of the MRI system must be fully complied with, including limits on magnetic field strength, coils, spatial field gradient, gradient slew rate, SAR, or/and B⁺_1rms as shown in the example below.

 

Different types of MR Conditional equipment have varying requirements for safe use in the MR suite. It is quite complex and impractical to label all gauss lines and spatial gradient magnet fields on the floor. In our practice and according to our data analysis, most MR Conditional devices can be used outside the 150/200 gauss (yellow) line. A simple rule we recommend is to place medical devices as far away from the magnetic bore as possible, provided this does not affect the physical connection with the patient. For example, infusion pumps can be located outside the 150/200 gauss line even if the manufacturer’s safety label permits use within a higher gauss field, provided the tubing is long enough to reach the patient. This approach also accounts for the steep slope of spatial field gradients, allowing safe control space. It is vital that medical equipment requiring placement close to the patient and magnet core, that is, near the 1000 gauss line (the red line), such as ventilators, be checked by MR safety personnel to ensure that the manufacturer’s label confirms compliance with the maximum magnetic field or spatial gradient requirement, as specified in the instruction for use. In addition to the FDA MR Conditioned label, for daily operational convenience, it is good practice for MR safety officers to apply clear secondary labels to frequently used MR Conditional equipment. For example: “Don’t exceed the 200 gauss line (the YELLOW line).” This helps avoid misplacement or confusion when equipment must be moved during patient transfer, as shown in Figure 1. These labels are for user convenience only; there is no need for concern over the gauss line issues. In practice, assessment and management must be carried out on a case-by-case basis.

 

Figure 1. Gauss line labels for floor and magnetic resonance conditional devices.

 

Other suggestions for using MR Conditional devices or objects in the MR magnet room (Zone IV) include: (1) Consider the limits on connector tubing length and patient positioning when placing MR Conditional or MR Safe devices. (2) Ensure that the auto-lock brake of the device is engaged. (3) Secure the MR Conditional device to a non-movable object if necessary, e.g., using a plastic safety belt as shown in Figure 2.

 

Figure 2. Securing the magnetic resonance Conditional device to a non-movable object.

 

It is recommended that MR-trained personnel manage all emergency events, such as fire, quench, resuscitation, etc.[53]

 

Resuscitation should not take place within the magnet room (Zone IV). On-site MR-trained personnel should remove the patient from the magnet room immediately. Consider quenching only if there is a threatening situation in which someone is trapped inside the magnet by a heavy object. Close the MR magnet room door (Zone IV) to prevent accidental access. Call for help according to local guidelines.

 

If smoke or fire is coming from the scanner, equipment room or console, on-site MR radiographers or technologists should stop any examination procedure immediately and evacuate the patient from the MRI suite. Perform an emergency electrical shutdown. Do not activate the magnet quench button unless absolutely necessary. Close and lock the scan room door to prevent inadvertent entry of any ferromagnetic materials into the scan room. The incident should be announced immediately via the intercom system. Activate the nearest fire alarm pull station, if available. Escort all patients in the MRI suite to a safe location. Only MR Conditional fire extinguishers may be brought into and kept within Zones III and IV. Firefighters must be informed of the existence of the MRI facility and the compatibility requirements of their fire-fighting equipment.

 

A quench is the procedure by which the magnetic field is removed through the release of liquid helium. The large volume of gaseous helium displaces oxygen in the MRI examination room and poses a risk of asphyxiation. In the event of a quench, stay calm and evacuate the patient. Call for help and report the event to supervisors. Service engineers should be contacted to assess the origin of the fire and the condition of the scanner.

 

In the event of all incidents or near-incidents, on-site trained MR radiographers and technologists should notify supervisors and the relevant parties.

 

Introducing MRI technology into the operating theatre and radiotherapy settings presents new challenges, which may be explored further in a future article.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Hong Kong J Radiol 2025 Sep;28(3):e198-203 | Epub 12 September 2025

 

 

¹ Department of Diagnostic and Interventional Radiology, Queen Elizabeth Hospital, Hong Kong SAR, China

² Department of Pathology, Queen Elizabeth Hospital, Hong Kong SAR, China

 

  Full Article in PDF

A 39-year-old male presented in December 2022 with a fever of unknown origin. Positron emission tomography–computed tomography (PET-CT) was performed to search for a potential source of sepsis. The patient did not report any symptoms of pain or swelling throughout the evaluation. Nonetheless, an incidental finding of a bone lesion in the left ischiopubic ramus prompted further investigation and an orthopaedic referral.

 

The lesion appeared subtle and poorly defined on the pelvic radiograph. PET-CT revealed a lytic intramedullary lesion in the left ischiopubic ramus, with no periosteal reaction or soft tissue mass (Figure 1). No pathological fracture was observed. The lesion was hypermetabolic with a maximum standard uptake value of 13.5. Magnetic resonance imaging (MRI) revealed predominantly T1-weighted hypointense signals and T2-weighted hyperintense signals with contrast enhancement (Figure 2). Perilesional bone marrow oedema was noted, along with cortical disruption and adjacent soft tissue oedema at the medial aspect of the left ischial bone.

 

Figure 1. Positron emission tomography–computed tomography (PET-CT) demonstrating a lytic intramedullary lesion at the left ischiopubic ramus (arrows), with no periosteal reaction or soft tissue mass. The lesion exhibits hypermetabolic activity. (a) CT and (b) PET images.

 

Figure 2. Magnetic resonance imaging of the lesion showing T1-weighted hypointense signal and T2-weighted hyperintense signal, with contrast enhancement and adjacent bone marrow oedema (arrows). (a) Axial T1-weighted image. (b) Axial T2-weighted and short-tau inversion recovery image. (c) Axial post-gadolinium T1-weighted fat-saturated image. (d) Coronal post-gadolinium T1-weighted fat-saturated image.

 

Results of the CT-guided biopsy of the lesion suggested a giant cell-rich spindle cell lesion, most compatible with an aneurysmal bone cyst. The biopsy showed random fascicles of spindle cells admixed with scattered osteoclast-type giant cells, accompanied by occasional lymphocytes, plasma cells and haemosiderin-laden macrophages. No atypical mitoses, marked nuclear pleomorphism, or necrosis were noted. Immunohistochemical study showed that the lesional cells were negative for H3G34W and H3K36M. Ubiquitin-specific protease 6 (USP6) gene translocation was identified by fluorescence in situ hybridisation.

 

Follow-up CT and MRI were performed prior to surgery, 2 months after the biopsy and 8 months after the initial imaging. The follow-up CT revealed interval enlargement of the lesion in the left ischiopubic ramus, with marked thinning of the medial cortex of the left ischium (Figure 3). Follow-up MRI demonstrated changes to the signal characteristics of the lesion, now showing multiloculated cystic spaces with fluid-fluid levels and peripheral contrast enhancement along the wall and internal septa (Figure 4). There was no perilesional bone marrow oedema or extraosseous soft tissue extension.

 

Figure 3. Follow-up computed tomography (CT) images showing interval enlargement of the lytic lesion (arrows) in the left ischiopubic ramus, with marked thinning of the medial cortex of the left ischium. (a) Axial CT image in bone window. (b) Coronal CT image in bone. (c) Coronal CT image in soft tissue window.

 

Figure 4. Follow-up magnetic resonance imaging of the lesion demonstrating changes in signal characteristics. The lesion now consists of multiloculated cystic spaces with fluid-fluid levels and peripheral contrast enhancement along the wall and internal septa (arrows). (a) Axial T1-weighted image. (b) Axial T2-weighted and short-tau inversion recovery image. (c) Axial post-gadolinium T1-weighted fat-saturated image. (d) Coronal post-gadolinium T1-weighted fat-saturated image.

 

Subsequently, the patient underwent curettage of the lesion, with allograft chips packed into the bone cavities. Pathological examination revealed tumour tissue featuring haphazardly arranged spindle cell fascicles and osteoclast-like multinucleated giant cells, set against a background of anastomosing woven bone trabeculae rimmed by osteoblasts and fibrous septa formed by bland fibroblasts (Figure 5). No cytological atypia, increased mitosis, or necrosis were observed. Immunohistochemical studies were once again performed, with the lesional cells being negative for H3G34W and H3K36M. The features were consistent with an aneurysmal bone cyst. To date, the patient has remained asymptomatic with no clinical evidence of tumour recurrence.

 

Figure 5. Computed tomography (CT)–guided biopsy and subsequent curettage reveal a giant cell-rich spindle cell lesion featuring fascicles of spindle cells admixed with scattered osteoclast-type giant cells (arrows in [a] and [b]), set against a background of anastomosing woven bone trabeculae. No atypical mitoses, marked nuclear pleomorphism, or necrosis are noted. The features are consistent with aneurysmal bone cyst. (a) CT-guided biopsy on haematoxylin and eosin stain (H&E) [× 40]. (b) Curettage specimen on H&E stain [× 100]. (c) Immunohistochemical study [× 100] showing negative staining for H3G34W in tumour cells. (d) Ubiquitin-specific protease 6 gene translocation detected by fluorescence in situ hybridisation break-apart probes.

 

Conventional aneurysmal bone cyst (ABC) refers to an expansile cystic lesion of bone composed primarily of blood-filled spaces. Although some solid areas may be present, these are not the predominant feature. Histologically, conventional ABC is characterised by blood-filled cystic spaces with septa containing fibroblasts, osteoclast-like giant cells, and reactive woven bone.[1] The solid variant of ABC is a rare entity that exhibits distinct radiological and pathological features. The cystic components may be completely absent, with the lesion demonstrating a predominantly solid architecture. ABC was initially thought to represent a reactive and inflammatory response to intraosseous haemorrhage.[2] Nonetheless, with the identification of the USP6 gene rearrangement, a translocation on chromosome 17p13, both conventional ABC and its solid variant are now considered true bone neoplasms.[3] In the latest World Health Organization Classification of Tumors of Bone, ABC is classified as a benign osteoclastic giant cell-rich tumour.[4] The term “aneurysmal bone cyst” is preferred over “primary aneurysmal bone cyst”, while “ABC-like change” is used instead of “secondary aneurysmal bone cyst”, the latter commonly observed in giant cell tumours of bone and chondroblastomas.

 

Reported cases of solid ABC (S-ABC) are primarily found in the metaphyseal and diaphyseal regions of long bones in children and young adults, typically during their second or third decade of life.[5] [6] [7] [8] These cases usually present with an insidious onset of pain, swelling, or a palpable mass. Although conventional ABC can involve the flat bones of the pelvis, S-ABC arising from pelvic bones in adults has not been reported, to the best of our knowledge. ABC in the pelvis is often asymptomatic due to its deep location, and by the time a patient develops symptoms, the lesion has typically reached a significant size. In our case, the lesion was an incidental finding and relatively small without any complications; thus, the patient remains asymptomatic.

 

S-ABC can be confidently differentiated from conventional ABC on imaging. Although both can appear as lytic lesions with sclerotic margins on radiography and CT, conventional ABC typically exhibits an expansile soap bubble appearance with internal septa, whereas S-ABC often lacks septa and is non-expansile in up to one third of cases.[5] Matrix mineralisation and periosteal reaction are generally absent in both forms.[9] MRI is the modality of choice for aiding diagnosis. S-ABC is predominantly solid with uniform contrast enhancement, in contrast to the multiloculated cystic appearance and peripheral and septal enhancement seen in conventional ABC. Fluid-fluid levels can be observed in up to 70% of conventional ABC cases but are not a consistent feature in S-ABC.[7] Characteristically, S-ABC may be associated with mild bone marrow and soft tissue oedema in up to 50% of cases.[8] The expansile nature of the lesion can lead to marked thinning and focal disruption of the bony cortex, as seen in our case during the initial presentation, mimicking an aggressive bony lesion. Interestingly, follow-up MRI after biopsy revealed multiple fluid-fluid levels, while the previously noted bone marrow and soft tissue oedema had resolved. Possible explanations for these changes include interval haemorrhage within the lesion due to the biopsy that could give rise to fluid-fluid levels. The bone marrow and soft tissue oedema may have resulted from an inflammatory response to the rapidly growing lesion in its initial phase, leading to cortical disruption.[10]

 

The diagnostic challenge of S-ABC lies in distinguishing it from ABC-like changes and other solid bone tumours. The main differential diagnoses include giant cell tumour of bone (GCTB), chondroblastoma, and primary bone malignancies such as telangiectatic osteosarcoma. GCTB typically occurs in a slightly older patient population, demonstrates non-sclerotic margins, and often extends to the subarticular surface of long bones. Immunohistochemical staining for H3G34W and H3K36M are specific markers that aid differentiation; they are present in GCTB and chondroblastoma with or without ABC-like changes, but are absent in ABC.[11] In contrast, telangiectatic osteosarcoma is characterised by geographic bone destruction, wide zones of transition, and dense osteoid mineralisation. Extensive soft tissue involvement and cortical destruction, along with periosteal reaction on MRI, raise suspicion for a malignant tumour rather than S-ABC.[12]

 

Biopsy and pathological examination are often necessary when evaluating an aggressive-appearing lesion, and radiological-pathological correlation plays a pivotal role in reaching a diagnosis. S-ABC is characterised by fibroblastic proliferation, osteoclast-like giant cells, and focal osteoid production. In contrast to conventional ABC, the blood-filled cystic spaces are present only in small amounts, if at all. Histologically, S-ABC resembles giant cell reparative granuloma and brown tumours of hyperparathyroidism, as all three entities exhibit giant cells, haemorrhagic areas, and reactive osteoid formation.[8] Nonetheless, they lack the USP6 gene rearrangement.[13] Historically, the term S-ABC was used interchangeably with giant cell reparative granuloma; however, they are now regarded as distinct entities, with the latter term reserved for lesions in the gnathic location.[14] It is important to note that USP6 gene rearrangement is not exclusive to ABCs; it has also been identified in several other lesions, including nodular fasciitis, myositis ossificans, fibro-osseous pseudotumour of the digits, and cellular fibroma of the tendon sheath. The diagnosis of S-ABC can be established through a combination of compatible radiological features and pathological examination.

 

S-ABC presents a unique diagnostic challenge due to its similarities to other aggressive bone lesions. Accurate diagnosis requires a comprehensive approach that includes imaging, histological evaluation, and identification of the USP6 gene rearrangement. Radiologists should be aware of the distinct features of S-ABC, particularly in contrast to conventional ABC and other lesions such as giant cell tumour and telangiectatic osteosarcoma.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Hong Kong J Radiol 2025 Sep;28(3):e204-16 | Epub 11 September 2025

 

 

Department of Radiology, Pamela Youde Nethersole Eastern Hospital, Hong Kong SAR, China

 

  Full Article in PDF

Contrast-enhanced mammography (CEM) is a modality used as an adjunct to non-contrast mammographic imaging for multiple clinical indications, including evaluation of indeterminate abnormalities on mammography, workup of symptomatic patients, staging of breast cancer, and monitoring response to neoadjuvant chemotherapy.

 

Our centre provides diagnostic imaging for patients with breast symptoms through combined full-field digital mammography with tomosynthesis and ultrasound. We introduced CEM in 2019. This pictorial essay highlights the uses of CEM in daily clinical practice through selected cases.

 

We use a Selenia Dimensions 3D Digital Mammography system (Hologic, Glasgow [DE], US) to acquire CEM images. Two minutes prior to image acquisition, iohexol (Omnipaque 300; GE Healthcare, Milwaukee [WI], US) is administered intravenously at 1.5 mL/kg and a rate of 3 mL/s, followed by a saline flush. The breasts are not compressed during injection to facilitate blood flow.

 

A pair of images (one low-energy image and one high-energy image) is acquired in the standard craniocaudal and mediolateral oblique views for each breast. Additional views, including magnification or spot compression, can be acquired if needed using conventional mammographic technique. The ideal imaging window is within 10 minutes after contrast administration before contrast washout commences.[1] [2]

 

The low-energy images used in subtraction have been shown to have equivalent diagnostic value compared to full-field digital mammography, as their K-edge is lower than that of iodine, eliminating the need for an additional set of conventional images.[3] These images are reported using the latest BI-RADS (Breast Imaging Reporting and Data System) mammography lexicon.[4] The high-energy images are used to subtract the low-energy images, emphasising the areas of iodine uptake. The subtracted images are interpreted using the BI-RADS 2022 CEM lexicon.[5]

 

There is currently no universal consensus on the imaging sequence.[1] Our centre prefers to start with the symptomatic breast as it presumably has the highest concentration of contrast agent and, therefore, better lesion conspicuity at the beginning of image acquisition.[2] No consensus has been reached regarding the optimal timing of CEM within the menstrual cycle.[6] [7]

 

Full-field digital mammography and/or tomosynthesis and ultrasound remain the mainstays of breast assessment. However, the sensitivity of mammography decreases in dense breast parenchyma. Magnetic resonance imaging (MRI) is an advanced imaging modality known for its high sensitivity and negative predictive value (NPV), though its use is compromised by high cost and limited availability.

 

CEM, a relatively recent and more affordable adjunct for assessing both lesion morphology and vascularity, has gained popularity. Various studies have evaluated its diagnostic performance. CEM has been shown to have superior clinical performance compared to full-field digital mammography and/or ultrasound.[8] [9] [10] Cheung et al[8] showed that CEM increased sensitivity from 71.5% to 92.7% and specificity from 51.8% to 67.9% in dense breasts, compared to mammography alone. In a meta-analysis by Cozzi et al,[11] CEM had a pooled sensitivity of 95% and specificity of 81% in breast cancer detection. CEM shares MRI’s sensitivity for diagnosing breast cancer.[12] [13] It may be a reliable alternative when MRI is contraindicated or not tolerated, or when simultaneous assessment of suspicious calcifications and contrast enhancement is needed. However, lesion location may affect the feasibility of CEM as an alternative. Lesions in obscured areas on conventional mammography and nodal status are not well assessed by CEM since they share same field of view and might be overlooked.[14]

 

The BI-RADS lexicon defines architectural distortion (AD) as “a distortion of breast tissue with no definite visible mass but with spiculations that radiate from a point with focal retraction or distortion at the edge of the parenchyma”.[4] Differential diagnoses include both malignant and benign entities, such as radial scars, complex sclerosing lesions, and postoperative changes. Image-guided biopsy or surgical excision is typically recommended for suspicious AD.

 

With the increasing use of digital breast tomosynthesis, the detection rate of AD has risen.[15] One study reported that nearly 60% of AD foci are benign.[16] Identifying features that support or discourage biopsy may be helpful. Contrast-enhanced modalities such as MRI enable further evaluation of AD, especially if the finding is equivocal, by assessing areas of contrast enhancement, presumably related to angiogenesis and vascular leakage in malignancies.[17] MRI has shown a high NPV (98%) when there is no enhancement in AD,[18] but it is not easily accessible due to high costs and long queuing times in a busy clinical institute setting.

 

CEM is a promising alternative for AD assessment, with lower cost and shorter acquisition time. Patel et al[19] reported a high NPV (92%) for malignancy when primary AD showed no enhancement on CEM. While this supports using CEM as an additional tool for assessing tomosynthesis-detected AD, our centre still recommends image-guided biopsy as the likelihood of malignancy remains in BI-RADS category 4A lesions, and some low-grade malignant lesions may not enhance.

 

Conversely, the high NPV in non-enhancing primary AD could be applied in cases where a histological diagnosis had already been obtained to help confirm imaging-histopathologic concordance in cases with benign biopsy results.[20] The increasing number of AD biopsies yielding non-malignant pathology, due to the better detection rate of AD on digital breast tomosynthesis, adds complexity. The absence of contrast enhancement on CEM can help confirm concordance in these cases (Figure 1).

 

Figure 1. Case of a 53-year-old woman. (a) Low-energy mediolateral oblique view showing architectural distortion (arrow) in the upper left breast. (b) Corresponding recombined images showing no associated enhancement or focal mass. (c) Stereotactic-guided vacuum-assisted biopsy was performed and the postprocedural image demonstrates a marker within the region of the biopsied architectural distortion. Histopathological analysis revealed proliferative fibrocystic changes and a possible sclerosing lesion.

 

The approach to calcifications varies with the degree of malignancy risk based on BI-RADS descriptors.[4] Biopsy is offered for suspicious cases, while short-interval follow-up imaging is recommended for those deemed probably benign.

 

Sometimes suspicious calcifications can be challenging to manage, especially when unaccompanied by soft-tissue abnormalities and with no corresponding sonographic lesion to allow further actions such as localisation or biopsy in a readily feasible way. As CEM consists of both low-energy and subtracted images, this allows for the simultaneous delineation of calcifications and associated contrast enhancement. The presence of associated contrast enhancement correlates well with the likelihood of malignancy.[21] [22] Therefore, it may be useful for selecting lesions for CEM-guided localisation, particularly in cases of nonpalpable lesions that are invisible on sonography (Figure 2). Again, however, the absence of contrast enhancement does not exclude malignancy.[21] [22] The morphology and distribution of calcifications remain the key determinants. Further exploration of this application is warranted.

 

Figure 2. A 48-year-old woman recalled from screening for suspicious calcifications; clinically, no palpable lesion was noted. (a) Magnification view of low-energy image of right breast upper portion. (b-e) Low-energy and subtracted contrast-enhanced mammography (CEM) images of the right breast in craniocaudal (CC) and mediolateral oblique (MLO) views. CEM images demonstrated segmental fine pleomorphic and coarse heterogeneous calcifications (arrow), with associated clumped non-mass enhancement (open arrows) in the upper outer quadrant of the right breast. These findings were highly suspicious for malignancy and biopsy-confirmed ductal carcinoma in situ. (f, g) Mammography of the right breast in CC and MLO views. Breast-conserving treatment was planned. Stereotactic-guided marker placement (arrowheads) was performed for lesion bracketing.

 

It is often difficult to determine the optimal surgical approach, namely, breast-conserving treatment or mastectomy, in patients with suspected additional tumour foci in the ipsilateral or contralateral breast. In addition, the Asian population often has dense breast tissue,[23] which lowers the sensitivity of mammography and complicates the surgical decision making.

 

The application of CEM in assessing tumour extent has been compared to conventional mammography, ultrasound, and MRI. Both CEM and MRI have higher sensitivity for cancer detection compared to conventional mammography and ultrasound alone (Figures 3, 4, 5, and 6).[24] [25] Compared to MRI, CEM exhibits similar sensitivity in detecting the index cancer and secondary cancer.[25] [26] Preliminary results from the studies[24] [25] [26] demonstrate that CEM may be a feasible and cost-effective modality for preoperative staging.

 

Figure 3. A 53-year-old woman with preoperative staging for newly diagnosed left breast malignancy. (a) Contrast-enhanced mammography (CEM) magnification view of low-energy image of the left breast inner portion shows an oval circumscribed lesion with internal calcification (open arrow) and heterogeneous enhancement (arrow) at the lower outer quadrant, which was subsequently biopsied and confirmed as malignant. (b-e) CEM low-energy and recombined images of bilateral breasts in craniocaudal and mediolateral oblique views demonstrate coarse heterogeneous and fine pleomorphic calcifications in segmental distribution at the lower central portion of the left breast, with associated clumped non-mass enhancement extending towards the left nipple (arrows), suggestive of malignancy. An irregular lesion with homogeneous enhancement is noted in the upper inner quadrant in the right breast (arrowheads), with low suspicion for malignancy; this lesion was not seen on prior breast imaging.

 

Figure 4. Same patient as Figure 3. (a) Magnetic resonance images showing similar findings to those on contrast-enhanced mammography. (b) Enhancing lesion in the upper inner quadrant of the right breast (arrows) with restricted diffusion and a type III curve, suspicious for malignancy. (c) Lower central enhancing lesion in the left breast corresponding to the biopsy-proven malignancy (not shown in [c], arrow in [a]), associated with segmental clumped non-mass enhancement (dashed arrows).

 

Figure 5. Second-look sonographic images of the right breast in the same patient as Figure 3. (a) An irregular hypoechoic lesion with angulated margin is seen at 1 o'clock position of the right breast, 4 cm from the nipple. This lesion was previously identified on contrast-enhanced mammography and magnetic resonance imaging, with low suspicion for malignancy. (b) Ultrasound-guided biopsy was performed, and histopathology showed invasive carcinoma. (c) Insertion of the deployment needle preloaded with Magseed into the targeted right breast lesion (left). Postprocedural image indicating successful marking of the targeted right breast lesion with Magseed, indicated by the arrow (right).

 

Figure 6. In same patient as Figure 3, a subsequent mammogram was performed to confirm the Magseed marker (arrows) located deep in the upper inner quadrant of the right breast, corresponding to the site of the previously noted irregular enhancing lesion seen in Figure 3.

 

Neoadjuvant chemotherapy (NAC) has been an important strategy for patients with locally advanced breast cancer. NAC helps improve surgical and cosmetic outcomes by shrinking tumour size, downgrading the nodal status, and increasing the likelihood of successful breast conservation.

 

Accurate imaging assessment of treatment response is therefore essential, and MRI is a reliable tool, superior to the combination of clinical examination, conventional mammography, and ultrasound.[27] Nonetheless, MRI is not always readily available due to limited access and patient-related factors such as long wait times, allergy to contrast agents, kidney problems, claustrophobia, or the presence of metallic devices (Figures 7 and 8). CEM has a lower cost in terms of examination time and resources; it has been an increasingly popular tool to assess treatment response. The results are promising, showing comparable performance between MRI and CEM in evaluating the pathological response of breast cancer to NAC (Figures 9 and 10).[28]

 

Figure 7. Magnetic resonance images of a 40-year-old woman with biopsy-proven left breast malignancy and prior Magseed insertion, for treatment response after neoadjuvant chemotherapy. Most of the left breast was obscured due to significant metallic artifact (arrows) produced by the Magseed.

 

Figure 8. Contrast-enhanced mammographic images of the same patient as Figure 7. (a, b) Pretreatment recombined images of the left breast showed a spiculated enhancing mass (arrows) in the upper outer quadrant, corresponding to the biopsy-proven malignancy. (c, d) Post-treatment recombined images of the left breast revealed interval reduction in the size of the index cancer (arrows).

 

Figure 9. Contrast-enhanced mammographic images of the right breast in a 57-year-old woman with biopsy-proven malignancy in the upper outer quadrant for neoadjuvant chemotherapy. (a) Baseline images. (b) Post-treatment images. Serial images showing interval reduction in the size of the index cancer. Associated non-mass enhancement beyond the confines of index cancer had largely resolved (arrows in [a]).

 

Figure 10. Magnetic resonance images of the same patient as Figure 9 at baseline (a) and post-treatment (b). Interval resolution of right breast index cancer and associated non-mass enhancement (arrow in [a]) was noted.

 

MRI is recommended as a supplemental screening tool in breast cancer screening for high-risk populations, defined as women with a lifetime risk of more than 20% according to the American Cancer Society and the American College of Radiology.[29] For women at intermediate risk, defined as those with a lifetime risk between 15% and 20%, breast MRI is suggested for those with dense breasts and a history of breast cancer diagnosed before the age of 50 years, according to the American College of Radiology.[29] The introduction of CEM has aroused radiologists’ interest in its role in screening and surveillance, particularly given the large number of intermediate-risk women who could benefit. In a pilot study by Jochelson et al,[30] 307 patients at increased risk for breast cancer underwent both screening CEM and MRI. Both modalities detected additional invasive cancers that were occult on conventional mammography, with comparable specificity and positive predictive value.[30] Two other studies showed that CEM outperformed two-dimensional full-field digital mammography when screening women with higher-than-average risk for breast cancer, with greater sensitivity (e.g., 87.5% vs. 50%,[31] 90.5% vs. 52.4%[32]). Preliminary study results are encouraging but the role of CEM warrants further research.

 

Adverse reaction to iodinated contrast agent is a concern, including the risks of extravasation, allergy, and contrast-induced acute kidney injury. Volume expansion by 0.9% normal saline prior to the contrast administration is a feasible preventive measure for those at risk.[33] The image quality of CEM can be degraded by patient motion. CEM is more prone to motion artifacts (Figure 11) due to its longer exposure and compression time, resulting in blurred images. There are many technical artifacts that are specific to CEM. For instance, the use of an undersized compression paddle may cause horizontal lines across the axilla (Figure 12). Suboptimal breast compression may lead to air trapping within skin folds or scars, resulting in poor contact between the skin and the detector or compression paddle (Figure 13). Macrocalcifications, cysts or, post-biopsy haematomas may not enhance and appear as low-density areas compared to background enhancement on subtracted images, which is known as negative contrast enhancement[14] [34] (Figure 11).

 

Figure 11. Contrast-enhanced mammographic images of the left breast in a middle-aged woman. (a) Recombined image of mediolateral oblique view showing ripple artifact with alternating black and white lines in the lower portion (arrow), attributed to patient motion during low- and high-energy image acquisition. (b) Low-energy image showing coarse calcification in the retro-areolar region (arrow) with corresponding negative contrast enhancement in the recombined image (arrowhead in [a]).

 

Figure 12. Contrast-enhanced mammographic image of the right breast in mediolateral oblique view showing axillary line artifact (arrow).

 

Figure 13. Contrast-enhanced mammographic image of the left breast in mediolateral oblique view showing air trapping in the axilla (arrow), which appears as black lines on the recombined image.

 

Subtracted CEM allows delineation of contrast enhancement based on the degree of angiogenesis in the lesions. However, contrast enhancement may also be seen in benign lesions such as fibroadenomas, intraductal papillomas, and fat necrosis,[14] making it difficult to determine the nature of the lesion and potentially resulting in false-positive findings. Varying degrees of angiogenesis and contrast enhancement are also seen among different subtypes of malignancy. Lesions with less pronounced enhancement, such as ductal carcinoma in situ or lobular carcinoma, may be overlooked,[14] leading to false-negative findings.

 

Additionally, the intensity of background parenchymal enhancement can affect interpretation, as it may obscure underlying lesions[14] (Figure 14). Since the field of view in CEM is identical to that of conventional mammography, lesions located in blind spots, such as those near the chest wall, might not be visualised[14] and should be further evaluated with MRI.

 

Figure 14. Contrast-enhanced mammographic images of the right breast in a 65-year-old woman with newly diagnosed right breast malignancy. (a) Low-energy images showing fine pleomorphic, coarse heterogeneous, and amorphous calcifications in segmental distribution in the upper outer quadrant of the right breast (arrows). (b) Subtracted contrast-enhanced images revealed clumped segmental non-mass enhancement and irregular enhancing mass of high conspicuity (open arrows), extending beyond the span of calcifications, in the setting of marked background parenchymal enhancement. Determining exact extent of non-mass enhancement is limited by marked background parenchymal enhancement, rendering precise bracketing localisation difficult. If breast conservation therapy was elected, there would be an increased risk of incomplete excision. Eventually, the surgical team planned for mastectomy.

 

Currently, there are not good biopsy tools that work directly with CEM. Lesions identified on CEM should be correlated with other imaging modalities (e.g., standard digital mammography, ultrasound, or breast MRI) if biopsy is planned. Ultrasound is often preferred due to its accessibility, lower cost, and suitability for ultrasound-guided biopsy.

 

Combined full-field digital mammography with tomosynthesis and ultrasound remains the mainstay of breast assessment in our centre. Incorporation of CEM into daily clinical practice provides further additional information in certain circumstances and is commonly used for evaluating disease extent and monitoring treatment response. CEM is increasingly regarded as a more affordable and accessible modality in settings with limited resources, or as an alternative when MRI is contraindicated or not tolerated. Further exploration of its role in breast imaging is anticipated.