ESTRO 2025 - Abstract Book

S2462

Physics - Autosegmentation

ESTRO 2025

2055

Digital Poster Clinical validation of an Artificial Intelligence (AI) based auto-segmentation tool for breast radiotherapy planning Francesca S. Maddaloni 1,2 , Gabriele Palazzo 2 , Maria Giulia Ubeira-Gabellini 2 , Claudio Fiorino 2 , Paola Mangili 2 , Andrei Fodor 3 , Marcella Pasetti 3 , Sara Saufi 3 , Roberta Tummineri 3 , Antonella del Vecchio 2 , Nadia G. Di Muzio 3,4 , Sara Broggi 2 1 Specialization School in Medical Physics, Università degli Studi di Milano Statale, Milan, Italy. 2 Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy. 3 Radiotherapy, IRCCS San Raffaele Scientific Institute, Milan, Italy. 4 Vita-Salute, San Raffaele University, Milan, Italy Purpose/Objective: Modern radiotherapy planning requires precise anatomical contouring. However, manual delineation is time intensive and subject to inter-observer variability (IOV) among oncologists [1]. This study assessed the feasibility of a commercial AI-based segmentation system for breast Clinical Target Volume (CTV) and Organs-at-Risk (OARs). Manual, automatic, and edited automatic contours were compared, with IOV and dose delivery impact also evaluated. Material/Methods: A dataset of 20 left-sided and 20 right-sided breast cancer patients was analyzed. Four physicians, three expert radiotherapists and one resident, manually contoured the breast CTV and OARs, including the contralateral breast, heart and lungs (these last semi-automatically segmented using threshold-based tools). Similarly, automatic contours were obtained from MIM Protegè v.1.2.1, considering the structure “breast” as the CTV; afterwards, if necessary, physicians edited the automatic contours, assigning a score for quality [2]. IOV, manual and automatic structures comparison were assessed using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), 95th percentile Hausdorff Distance (HD95), and Average Surface Distance (ASD). Contouring time (manual vs edited contours) was also compared. To assess the clinical impact of OARs editing, automatic treatment plans were generated and planning data referred to automatic and edited OARs were compared [3]. Results: Contours drawn by the resident differ significantly from the ones of other clinicians (p < 0 . 0001) and were excluded from this analysis. Table 1 shows IOV results for the CTV among observer combinations, comparing them against automatic structures. For all metrics, manual vs. automatic variations were significantly greater than IOV (p < 0 . 05). Figure 2 shows the differences among manual, automatic, and edited contours. When looking to plan data, mean dose differences between automatic and edited contours were -0.02Gy (±0.06Gy, right; ±0.12Gy, left) and -0.08Gy (±0.14Gy) for the contralateral breast and for the heart (left) respectively, showing that editing impacts on contralateral breast/heart are negligible. Editing times for CTV ( ≃ 7min) were longer than manual contouring ( ≃ 6min), suggesting automatic contouring for breast CTV may not save time in clinical practice.