Artificial Intelligence for Contouring and Treatment Planning in Locoregional Radiotherapy Including Internal Mammary Nodal Irradiation for Breast Cancer

Hong Kong J Radiol Radiol 2025 Mar;28(1):e23-34 | Epub 24 March 2025

 

 

Department of Clinical Oncology, Queen Mary Hospital, Hong Kong SAR, China

 

  Full Article in PDF

Breast cancer is the most prevalent cancer affecting women worldwide. According to the World Health Organization, in 2022, there were approximately 2.3 million new breast cancer diagnoses made globally, and 670,000 deaths.[1] In Hong Kong, approximately 5500 new breast cancer cases were reported in 2021, with a crude incidence rate of 138.1 per 100,000 female population.[2]

 

The treatment of breast cancer typically involves a multimodal approach, including surgery, radiotherapy, and systemic therapies. Adjuvant radiotherapy plays a crucial role in reducing the risk of cancer recurrence and improving survival, particularly for patients who have undergone breast conserving surgery or have been diagnosed with locally advanced breast cancer.

 

The Danish Breast Cancer Group Internal Mammary Node study in 2022[3] showed that elective irradiation of the internal mammary nodes (IMNs) reduced the risk of distant recurrence and improved long-term survival of patients with node-positive early breast cancer, regardless of breast cancer laterality. However, the irradiation of IMNs has remained a controversial topic, and guidelines for IMN irradiation vary significantly across different institutes. One controversial issue is the potential for long-term cardiotoxicity,[4] as IMN irradiation tends to result in a higher cardiac dose.[5] The technical complexities involved in radiotherapy treatment can be challenging, leading to an increase in the time and resources required to treat each patient.

 

Over the past decade, there has been rapid advancement in artificial intelligence (AI) and deep learning technology, which have been increasingly utilised in radiation oncology.[6] The AI-powered tools can perform segmentation, treatment planning, and delivery,[7] with the aim of improving the quality and consistency of the treatment delivered, and enhancing the efficiency of the work of oncologists and radiation therapists. Ultimately, the incorporation of AI in radiation oncology seeks to streamline hospital operations, thus benefiting a greater number of patients within the demanding clinical environment. In the past few years, the use of deep learning and convolutional neural networks on segmentation as well as treatment planning in breast cancer radiotherapy has been explored. These studies have shown promising results, with the automated techniques demonstrating comparable quality to manual approaches.[8] [9] [10] [11] [12]

 

In this article, we sought to compare the performance of AI to that of manual approaches, in the context of segmentation and treatment planning in radiotherapy of IMNs in breast cancer.

 

The study included the data from 27 breast cancer patients, comprising 14 patients with left-sided breast cancer and 13 with right-sided breast cancer. All patients had undergone either mastectomy or breast conserving surgery (including lumpectomy and cryotherapy), as well as axillary surgery (including sentinel lymph node biopsy and axillary dissection). Five of the mastectomy patients also had undergone breast reconstruction. All 27 patients had received postoperative locoregional radiotherapy including irradiation of clinically evident IMN (i.e., lymph nodes that are suspected to be involved by malignancy on imaging) at the dose of 40 Gy in 15 fractions over 3 weeks between 2017 and 2023. A tumour bed electron boost was administered at a dose of 10 Gy in five fractions over 1 week to patients who had undergone breast conserving surgery. Patient characteristics (including cancer staging[13]) are presented in Table 1.

 

Table 1. Patient characteristics (n = 27).

 

All patients were positioned in the supine orientation on a breast board (Breastboard MT-350N; CIVCO, Orange City [FL], US), with the arms abducted and holding onto a central rod. Simulation was performed using computed tomography (CT), with a slice separation of 2.5 mm, covering the region from the jaw to the mid-abdomen. For patients who had undergone mastectomy without reconstruction, a 0.5-cm skin bolus was administered to the chest wall in the planning process.

 

The CT data were retrieved and exported to treatment planning software Eclipse[14] (version 15.6; Varian Medical Systems, Palo Alto [CA], US) and RayStation 12A deep learning segmentation software[15] (RaySearch Laboratories, Stockholm, Sweden) to perform contouring and treatment planning.

 

This study comprises two parts. In the first part of the study, two sets of contours were generated from the patients’ CT data. The performance of auto-segmentation was assessed by comparing its results to the manual contouring done by the oncologists.

 

Contouring of target volume and organs at risk (OARs) were done on Eclipse with reference to the European Society for Radiotherapy and Oncology consensus guidelines.[16] They were reviewed and approved by two senior clinical oncologists, each having >10 years of experience in breast cancer.

 

Automatic contouring of target volume and OARs were created using the deep learning segmentation (RSL Breast CT_1) planning model within RayStation 12A. A 5-mm expansion was applied to the breast or chest wall, axillary levels 3 to 4, and the IMNs from manual contours to become the planning target volume (PTV). The PTV was then trimmed by 5 mm underneath the skin, except in slim patients where the thoracic chest wall was too thin, causing the target volume to disappear when cropped. The time taken for generation of auto-segmentation was recorded.

 

The list and number of target volumes that were created are presented in Table 2.

 

Table 2. Number of target volumes created.

 

In the second part of the study, four virtual treatment plans were created. These included two plans created with machine learning (ML): a volumetric modulated arc therapy ML [VMAT (ML)] plan and an intensity modulated radiation therapy ML [IMRT (ML)] plan using a dynamic multileaf collimator with the ML model RSL-Breast-L-nodes-4005(1.0). Manual VMAT [VMAT (manual)] and manual IMRT [IMRT (manual)] plans were generated by the usual manual treatment planning methods.

 

The dosimetric objectives, constraints for target volumes and OARs used for treatment planning dose optimisation are presented in Table 3, in accordance with departmental consensus, The Royal College of Radiologists’ UK Consensus Statements,[17] and the Danish Breast Cancer Cooperative Group HYPO II protocol.[18] We recorded the time needed for automated and manual treatment planning.

 

Table 3. Dosimetric objectives and constraints of target volumes and organs at risk for radiotherapy treatment planning.

 

The manually contoured and auto-segmented volumes were compared using the following four geometrical metrics:

 

1. Volume ratio (VR)[19]: This metric evaluates the ratio of the volumes calculated by the test contour (auto-segmentation) and reference contour (manual contouring). A VR close to 1 indicates good agreement between the two contours.

 

2. The 95th percentile Hausdorff distance (95% HD)[19] [20]: HD is a surface-based metric of the maximum perpendicular distance between two sets of points (in mm). It is used to quantify the differences between the field edges by manual and auto-contouring. The purpose of using the 95% HD instead of maximum HD is to eliminate the impact of small outliers.

 

3. Surface Dice similarity coefficient (sDSC)[19] [21] [22]: sDSC is a metric of spatial overlap between two volumes as a measure of estimated editing. It specifically looks at the overlap between the surfaces of the reference and test contours, providing an assessment of the contouring agreement. It has been shown to have a good correlation with the time required to edit contours.[23] [24]

 

4. Volume Dice similarity coefficient (vDSC)[19] [22]: vDSC evaluates the overall volumetric overlap between the reference and test contours. This gives an indication of how well the total volumes agree. However, it is not sensitive to complex boundaries and tends to give higher scores for larger volumes. It is generally accepted that volume DSC >0.7 indicates excellent agreement.[25] [26]

 

Dose-volume histograms (DVHs) were generated in each treatment plan for review of dosimetric parameters, including the PTV of the whole breast or chest wall (V_95% [i.e., volume (%) of the PTV receiving at least 95% of the prescription dose] of 40 Gy, V_105% of 40 Gy, V_107% of 40 Gy, and maximum dose [D_0.5cc]), the PTV of the nodal volumes (V_90% of 40 Gy and maximum dose [D_0.5cc]), the mean dose to the ipsilateral lung, the volume receiving 17 Gy in the ipsilateral lung, the mean dose to the contralateral lung, the mean dose to the heart, the volume receiving 17 Gy in the heart, and the mean dose to the contralateral breast.

 

Descriptive analyses were stated as mean ± standard deviation; descriptive tables were created for the comparison of the contouring performance by AI. The dosimetric parameters for the target volumes and OARs in different treatment plans were compared using the Wilcoxon signed rank test for statistical significance. A p value < 0.05 was considered statistically significant.

 

AI successfully performed auto-segmentation of parts of the target volumes including axillary level 2, axillary level 3 (infraclavicular volume), axillary level 4 (supraclavicular volume), interpectoral nodes, breast (in breast conserving surgery patients), as well as OARs including the heart and the lungs for all patients. Nevertheless, it failed in identifying axillary level 1 in two patients (one on the left side and one on the right side) and IMNs in four patients (one on the left side and three on the right side). AI misclassified the region of the chest wall in mastectomy patients as breast due to the absence of a specific chest wall segmentation function on RayStation 12A (Table 4). Thus, the result of comparison between manual and auto-contouring of the chest wall was disregarded. Table 2 shows the number of structures created by auto- and manual contouring, respectively.

 

Table 4. Results of comparison of manual and auto-contouring.

 

Table 4 shows the results of the VR, 95% HD, sDSC, and vDSC for all structures.

 

Overall, AI demonstrated excellent performance in delineating the lungs and heart across all patients. In contrast, the performance for the target structures, i.e., nodal volumes and the breast, was less robust compared to the OARs, based on the four evaluation metrics used.

 

Figure 1 shows the manual contouring and auto-segmentation of the axillary nodal volume and breast on selected CT slices. When comparing the contouring of the nodal volumes, looking at VRs, the AI-generated structures tended to have smaller volumes compared to the manual contours, with VRs ranging from 0.56 ± 0.18 to 0.92 ± 0.03. This discrepancy was more pronounced for contouring of axillary level 4, axillary level 1, and IMNs (Table 4).

 

Figure 1. Selected cuts of the radiotherapy plans comparing auto-segmentation (red contours) and manual contouring (green contours) of target volumes on a patient following left lumpectomy and sentinel lymph node biopsy.

 

The largest 95% HD measurements were observed at axillary level 1, reaching up to 19.29 ± 12.11 mm for the left axillary level 1 and 19.89 ± 10.04 mm for the right axillary level 1, and left breast with the result of 21.00 ± 0.99 mm. This was followed by the IMNs, with measurements of 15.62 ± 16.94 mm for the left side and 13.37 ± 12.26 mm for the right side. The 95% HD measurements for the remaining structures ranged from 6.74 ± 2.73 mm to 10.52 ± 6.34 mm (Table 4).

 

High sDSC values were observed for interpectoral nodes, axillary level 2 and the breast. The lowest agreement was found for axillary level 1. The best results for vDSC were again observed in the interpectoral nodes, axillary level 2 and the breast. Conversely, both axillary level 1 and axillary level 4 had the lowest vDSC values (Table 4).

 

Results of nodal volumes in patients who underwent axillary dissection and sentinel lymph node biopsy (SLNB) are shown in Tables 5 and 6, respectively. In patients who underwent SLNB instead of axillary dissection, there was a tendency towards better performance of auto-segmentation, as indicated by higher sDSC and vDSC, in comparison to patients who underwent axillary dissection. However, it is important to note that a limitation of this analysis is the small sample size of patients who underwent SLNB.

 

Table 5. Results of comparison of manual and auto-contouring of nodal volumes in patients who underwent axillary dissection.

 

Table 6. Results of comparison of manual and auto-contouring of nodal volumes in patients who underwent sentinel lymph node biopsy.

 

The time required for auto-segmentation was recorded, showing an average of 27.39 ± 2.53 seconds for generation of one contouring.

 

DVH parameters of VMAT (ML) and VMAT (manual) are presented in Table 7, while DVH parameters of IMRT (ML) and IMRT (manual) are shown in Table 8. Figure 2 illustrates the dose distribution comparisons between VMAT (ML), VMAT (manual), IMRT (ML) and IMRT (manual) on selected slices.

 

Figure 2. Images of the same computed tomography slice at two different levels in the same patient showing the dose distribution achieved by (a) volumetric modulated arc therapy (VMAT) created by machine learning, (b) VMAT created manually, (c) intensity modulated radiation therapy (IMRT) created by machine learning and (d) IMRT created manually.

 

Table 7. Parameters of dose-volume histograms of volumetric modulated arc therapy created with machine learning and created manually.

 

Table 8. Parameters of dose-volume histograms of intensity modulated radiation therapy created with machine learning and created manually.

 

All of the four virtual plans achieved satisfactory performance in covering the target volumes, including the breast or chest wall and nodal volumes, by fulfilling the dose constraints. For OARs, IMRT (ML) failed to achieve the dose limit for mean dose to the ipsilateral lung, while the other parameters were fulfilled in all of the plans.

 

Both VMAT (ML) plans and VMAT (manual) plans demonstrated clinically acceptable treatment plans. VMAT (manual) showed a significantly better performance for target coverage of the breast or chest wall and the nodal volumes [p < 0.001 for both PTV (IMN) and PTV (axillary levels 3 and 4); p = 0.034 for PTV (whole breast/chest wall)]. VMAT (ML) plans had significantly lower mean doses to the contralateral lung and the heart compared to VMAT (manual) [p = 0.027 and p = 0.009, respectively] (Table 7). The time required for treatment planning of a VMAT (ML) plan was 11.65 ± 2.00 minutes, while that of a VMAT (manual) plan was 15.42 ± 5.24 minutes (p = 0.023).

 

IMRT (ML) plans and VMAT (manual) plans had comparable performance in terms of target coverage for the breast or chest wall and PTV (axillary levels 3 and 4); however, IMRT (ML) has a higher coverage of the PTV (IMN) than IMRT (manual) [99.06% vs. 96.96%; p = 0.013]. For OARs, IMRT (ML) plans had a higher dose to the ipsilateral lung in terms of both mean dose and V_17Gy (both p < 0.001). The mean dose to ipsilateral lung (14.81 ± 1.26 Gy) failed to fulfil the dose constraint. The differences in mean dose and V_17Gy to the heart were not statistically significant (Table 8). The time required for treatment planning of an IMRT (ML) plan was 51.47 ± 12.46 minutes, while that of a IMRT (manual) plan was 30.84 ± 5.20 minutes (p < 0.001).

 

AI has been revolutionising various sectors, including the medical field, impacting the system from daily patient care to policy formulation.[27] [28] The advancements in AI have significantly enhanced radiological imaging through improved image recognition using deep learning algorithms. The clinical utilisation of AI in radiotherapy delivery is on the rise, encompassing tasks such as segmentation and treatment planning, in malignancies arising from different regions including head and neck, thorax and pelvis.[29] The primary goals are to enhance quality, decrease inter-observer variability, standardise contouring, and increase efficiency.

 

Based on our study findings, AI demonstrated encouraging results in segmentation of OARs. The results of 95% HD and vDSC for heart and lungs were comparable to those of other AI models.[10] [30] This achievement can be attributed to the clear visualisation and differentiation of the heart and lungs, which also exhibit minimal variability across different patients. For clinical target volume, favourable performance was observed for axillary level 2 and interpectoral nodes. Delineation of these structures is closely associated with identification of the pectoralis minor muscle, which is a readily identified structure on CT, thus contributing to improved consistency in contouring.

 

The deep learning segmentation model employed in our study exhibited limitations in segmentation in some of the patients, failing to delineate axillary level 1 in two patients (7.4%) and the IMN in four patients (14.8%). In addition, the model’s inability to identify the chest wall underscores its constraints in clinical application. The agreement between auto-segmentation and manual contouring was found to be less consistent for axillary level 1 compared to axillary levels 2 to 4, with larger 95% HD and smaller sDSC. From the analysis of SLNB and AD patients, better outcomes were more likely in those who underwent SLNB instead of AD, despite the small number of SLNB cases. Contouring of axillary level 1 is typically influenced by axillary surgery to encompass the postoperative OARs or areas with surgical clips based on clinical judgement rather than solely relying on anatomical structures, leading to possible inter-observer variability. The results suggest that disrupted anatomy may also pose a challenge for auto-segmentation.

 

Apart from individual factors, variations between manual contouring and auto-segmentation can also be attributed to systemic factors. For example, patients from our institute may have had different physiques compared to the patients recruited in the AI training dataset. Additionally, the quality of the CT imaging can be a contributing factor, as it plays a significant role in the visibility of structures such as the axillary and subclavian vessels. These structures are crucial for defining the boundaries of the axillary levels and consequently influence the precision of target volume delineation by both manual and auto-contouring.

 

Regarding the time spent for auto-segmentation and the potential of AI to improve clinical efficiency, in our study, the mean duration for auto-segmentation was found to be <30 seconds. Although the time spent on manual contouring was not documented, clinical experience from our institute and existing literature[31] [32] suggest that it typically requires approximately 35 minutes for manual contouring of a breast radiotherapy plan, which is significantly longer than the time required by AI.

 

At present, although vDSC with 0.7 indicates excellent agreement in general,[25] there are no established acceptable thresholds for the four metrics for clinical use. In this study, the metrics provided only objective numerical analysis, without incorporating qualitative review by oncologists to gauge user satisfaction. This is a limitation of this study, as the ultimate goal of reviewing performance of AI in this domain is to augment manual processes, where user acceptance is crucial for clinical implementation of the technology.[33] Additionally, the small number of oncologists reviewing the manual contours as ground truth may not adequately represent the intra- and inter-observer variability in reality. This could be better addressed by creating a consensus manual contour developed by a larger group of oncologists.

 

Another limitation of this study is that, while the findings indicate that auto-segmentation requires significantly less time than manual contouring, further assessment of the time needed for manual review and correction of auto-segmentation is necessary to fully evaluate its time-saving benefits. Surface DSC is a relatively novel metric for assessing the performance of auto-segmentation and has been demonstrated to be a more effective indicator of correction time compared to traditional metrics, such as vDSC.[23] [24] Generally, it is anticipated that no time savings will be realised if an observer expects to alter an automatically generated contour by more than approximately 40%.[23] Future research is needed to investigate the relationship between surface DSC and other relevant metrics with time savings, and establish acceptable thresholds for clinical practice.

 

Adjuvant radiotherapy for breast cancer has been associated with long-term complications. One notable concern is the increased risk of cardiac toxicity,[34] and the risk is particularly pronounced in women with existing cardiac risk factors, or women receiving radiotherapy for left-sided breast cancer. Additionally, there is a risk of pneumonitis[35] and the development of second primary cancers.[36] Although a recent meta-analysis found no significant differences in the occurrence of cardiac events and pneumonitis between patients who received IMN irradiation and those who did not,[37] it is still recommended to prioritise minimising the radiation dose to OARs as much as possible, following the ALARA (As Low As Reasonably Achievable) principle.[38]

 

Facing the expected growing numbers of patients undergoing radiotherapy including the IMN, a shift from three-dimensional conformal radiation therapy to inverse planning is needed to improve the quality of radiotherapy plan. As it is likely to impose a substantial increase in the workload associated with treatment planning, there is a need to see if AI can facilitate the radiotherapy planning process. The goal is to carry out treatment planning efficiently while maintaining established standards.

 

The virtual plans in this study, including VMAT (ML), VMAT (manual), IMRT (ML), and IMRT (manual), were created based on contouring performed by oncologists, and all of them demonstrated satisfactory target coverage. VMAT (ML), VMAT (manual) and IMRT (manual) successfully met all of the dose constraints, while IMRT (ML) exceeded the objective for the ipsilateral lung mean dose, which was measured at 14.81 ± 1.26 Gy, surpassing the desired limit of ≤13 Gy.

 

In the evaluation of VMAT plans, VMAT (manual) demonstrated superior target coverage compared to VMAT (ML) plans, which exhibited a trend towards reduced doses to OARs. Both planning approaches, however, are deemed acceptable from a dosimetric perspective.

 

In the context of IMRT plans, IMRT (manual) is regarded as more acceptable, as it successfully meets the specified dose objectives and constraints. In contrast, IMRT (ML) will require manual adjustments and optimisation before clinical application.

 

In our study, the time required to generate a VMAT (ML) plan was faster by 3.77 minutes compared to VMAT (manual) plan. The time taken to create an IMRT (manual) plan was faster than the IMRT (ML) plan by 20.63 minutes. Notably, during the AI planning process for both IMRT and VMAT, planners can attend to other responsibilities while the system generates the treatment plans; thus, the actual clinical time dedicated to AI planning may be less than the recorded duration. On the other hand, it is also important to acknowledge that AI-generated plans are not inherently perfect and may necessitate further adjustments to tailor the treatment for individual patients. A limitation of this study is that it does not quantify the time required for optimising an AI-generated plan after its initial creation. These factors may influence the actual clinical time saving benefits derived from the use of AI in treatment planning.

 

In the first part of the study, AI has demonstrated reliable and excellent performance in auto-segmentation of OARs including heart and lungs for breast cancer radiotherapy. While the technology may also be applied to the delineation of the breast and regional nodal targets, it is crucial to bear in mind the need for case-by-case adjustment, particularly for postoperative changes.

 

In the second part, the study showed that satisfactory VMAT plans can be efficiently generated using an automated approach, achieving performance comparable to that of manual planning in a relatively short timeframe. This method holds potential for clinical application, warranting further investigation into its impact on improving workflow in radiotherapy practices.