S336
Clinical - Breast
ESTRO 2026
accuracy (Excellent/Good vs. Fair/Poor) and macro F1- score, and compared against BCCT.core, the current standard for objective CO assessment.External validation was performed using the FAST-Forward trial (UK), comprising 4,300 standardized photographs from 1,314 patients at 2- and 5-year follow-up. Clinicians rated change in appearance from pre- radiotherapy photographs (None, Mild, Marked) and surgical deficit (Small, Medium, Large). These scores were mapped to the DBCG binary system: Mild/Marked changes or large deficits were classified as Fair/Poor, and None with Small/Medium deficits as Excellent/Good. Results: Models trained with both image-derived and additional features consistently outperformed image- only models. Accuracy and F1-scores improved across classifiers, suggesting CO depends on more than visual asymmetry and colour differences. Clinical markers like induration captured CO aspects not visible in photos. The eXtreme Gradient Boosting (XGB) model with additional features yielded the highest F1- score at 0.812 & 87.3% accuracy, the XGB model without additional features reached 0.728 & 83.1% while external validation reached 0.710 & 80.5%, all substantially outperforming BCCT.core (F1-score 0.647, accuracy 75.8%).
Oncology, Aalborg University Hospital, Aalborg, Denmark. 7 Department of Radiation Oncology and OncoRay Centre, University Hospital and Faculty of Medicine, Carl Gustav Carus; Technische Universität Dresden; Ger-man Cancer Consortium Dresden; Helmholtz-Zentrum Dresden-Rossendorf; National Centre for Tu-mour Diseases, Dresden; and German Cancer Research Centre, Heidelberg, Germany. 8 Department of Plastic Surgery, Odense and University Hospital of Southern Denmark SLB-Vejle, Odense, Denmark. 9 Department of Regional Health Research, University of Southern Denmark, Odense, Denmark. 10 Deparments of Breast and Plastic Surgery, Aalborg University Hospital, Aalborg, Denmark. 11 Clinical Trials and Statistics Unit, Institute of Cancer Research, Sutton, London, United Kingdom. 12 Department of Oncology, University of Cambridge, Cambridge, United Kingdom. 13 School of Medicine, University of Keele, Keele, Staffordshire, United Kingdom. 14 School of Medical Sciences, University of Manchester, Manchester, United Kingdom. 15 Division of Radiotherapy and Imaging, The Institute of Cancer Research, Sutton, London, United Kingdom. 16 Department of Oncology, Aarhus University Hospital, Aarhus, Denmark Purpose/Objective: Cosmetic outcome (CO) after breast-conserving treatment (BCT) is a key determinant of long-term patient satisfaction, yet systematic clinical follow-up has largely been discontinued. We aim to develop an automated, selfie-based CO assessment integrated with patient-reported outcome measures (PROMs). As PROMs are increasingly implemented in breast cancer follow-up, an objective image-based tool could complement patient-reports, support clinical decision- making, and help identify women needing additional care. This study evaluates whether machine learning (ML) models can classify CO from photographs and how inclusion of clinical and patient-reported features influences performance. Material/Methods: Data originated from two Danish Breast Cancer Group (DBCG) trials (HYPO and PBI), comprising 22,228 standardized photographs from 2,359 patients across five centres, collected pre-radiotherapy and at follow- up (1–10 years). Clinicians rated CO on a four-level scale (Excellent, Good, Fair, Poor) based on physical examination, along with scores for dyspigmentation, telangiectasia, induration, scar visibility, and oedema. Patient-reported outcomes included satisfaction, pain, and analgesic use.ML classifiers were trained using 122 image-derived features quantifying asymmetry and colour dissimilarity between the treated/untreated breast. Models were trained with and without additional clinical and patient-reported variables. Performance was evaluated using binary
Made with FlippingBook - Share PDF online