S209
Clinical - Breast
ESTRO 2026
reproducibility, ensure patient safety, and maintain equitable access. References: 1. Du J, Zhang L, Li H, et al. Comparison of machine- learning models for predicting breast-cancer-related lymphoedema. Front Oncol. 2024;14:1398475.2. Sun X, Yu C, Wu Y, et al. Predictive modelling of breast- cancer-related lymphoedema using machine-learning algorithms. Gland Surg. 2024;13(2):112–120.3. Yu J, Lee H, Park S, et al. Development of an early-warning model for breast-cancer-related lymphoedema using postoperative symptom monitoring. Breast J. 2024;30(4):512–519.4. Li C, Zhang Y, Wang J, et al. Development of an interpretable machine-learning model for predicting late radiotherapy-induced toxicities in breast cancer. Sci Rep. 2024;14(1):8752. Keywords: AI-assisted radiotherapy, Toxicity risk prediction Partial Breast Irradiation in Low-Risk Breast Cancer Using Stereotactic Body Radiotherapy Patricia Murina, Milla Galetto, Daniela Angel Schutte, Valentina Gregorat, Deborah Middendorff, Monica Martinez, Daniel Venencia, Silvia Zunino Radiation Oncology, Instituto Zunino, Cordoba, Argentina Purpose/Objective: Partial breast irradiation (PBI) has emerged as a suitable alternative to conventional whole-breast irradiation (WBI) in selected patients with early-stage, low-risk breast cancer. Moderately hypofractionated radiotherapy has shown comparable outcomes in local control to conventional fractionation, with increasing interest in stereotactic body radiotherapy (SBRT) as a precise and potentially less toxic approach. The primary objective of this study is to evaluate early and late toxicity associated with PBI using SBRT; the secondary objective is to assess cosmetic outcomes. Material/Methods: A retrospective analysis was conducted of 28 patients with 29 early-stage breast tumors (one bilateral case) treated between June 2022 and April 2024. Eligible patients had undergone breast-conserving surgery with negative margins, placement of 4–5 titanium clips in the tumor bed, and sentinel lymph node biopsy. The treatment protocol included virtual simulation with CT imaging, fiducial and reflective marker placement, and planning via Eclipse (Varian). Volumetric modulated arc therapy (VMAT) was used to deliver a total dose of 30 Gy in 5 consecutive daily fractions to the PTV_Eval, defined as the PTV with a 5 mm retraction from the skin surface. Results: Digital Poster 520
Digital Poster 508 Machine learning-based toxicity prediction for breast-cancer-related lymphedema in radiotherapy Rupa Das, Thuraya Al Hajri, Jyothy Kalesh Department of Radiation Oncology, Royal Hospital, Muscat, Oman Purpose/Objective: Breast-cancer-related lymphedema (BCRL) remains a significant survivorship issue after surgery and radiotherapy (RT), affecting up to 30% of patients. Traditional risk models lack predictive accuracy. Machine learning (ML) offers the ability to analyse complex, nonlinear factors and generate personalised toxicity-risk profiles. This systematic mini-review summarises recent evidence (2023–2025) on ML- based prediction of BCRL in radiotherapy and its clinical potential. Material/Methods: Published English-language studies up to October 2025 were reviewed using PubMed, Scopus, and ClinicalTrials.gov. The review focused on clinical and dosimetric studies that applied machine-learning techniques to predict lymphedema after breast radiotherapy. Only peer-reviewed or prospectively validated models were considered for inclusion to ensure clinical relevance and scientific reliability. Results: Ten studies were identified. ML models achieved area- under-curve (AUC) values between 0.73 and 0.85, outperforming conventional logistic regression (AUC 0.65–0.70). Key predictors included body mass index, extent of axillary dissection, nodal irradiation, and mean dose to axillary levels I–III. Incorporating dosimetric features and early postoperative symptoms, such as arm heaviness and swelling, improved predictive accuracy and enabled earlier identification of patients at higher risk. Explainable AI tools improved clinical interpretability by highlighting how individual variables contributed to patient-specific risk. Two prospective studies, the CINDERELLA and LOCATOR trials, were identified as current efforts investigating the clinical application of AI in breast radiotherapy, focusing on patient counselling, aesthetic outcome evaluation, and contouring standardisation. These insights can help clinicians tailor dose constraints and individualise follow-up. Conclusion: Machine-learning-based toxicity prediction for BCRL is rapidly emerging as a practical tool in precision radiotherapy. Integration of clinically validated AI models into treatment planning may enable early risk detection, adaptive dose optimisation, and proactive management. Collaborative multi-centre studies and adaptive-RT trials will be essential to confirm
Made with FlippingBook - Share PDF online