S1564
Physics - Autosegmentation
ESTRO 2026
Proffered Paper 2217 Clinical validation of AI-based probabilistic CTV breast maps incorporating Inter-Observer Variability Maria Giulia Ubeira Gabellini 1 , Cecilia Riani 1,2 , Gabriele Palazzo 1 , Davide Monticelli 1 , Andrei Fodor 3 , Nadia Gisella Di Muzio 3,4 , Antonella del Vecchio 1 , Alessandra Palma 5 , Anna Balsamo 6 , Angela Coniglio 6 , Claudio Fiorino 1 1 Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy. 2 Physics Department, Radiation Biophysics and Radiobiology Laboratory, Pavia, Italy. 3 Radiation Oncology, IRCCS San Raffaele Scientific Institute, Milan, Italy. 4 Faculty of Medicine and Surgery, Vita-Salute San Raffaele University, Milan, Italy. 5 Centro Nazionale Intelligenza Artificiale, HTA e Tecno- assistenza, Istituto Superiore di Sanità,, Rome, Italy. 6 Department of Human Health, Animal Health and Ecosystem (One Health) and International Relations (DOHRI), Ministry of Health, Rome, Italy Purpose/Objective: Manual delineation of Clinical Target Volume (CTV) in breast Radiotherapy is time-consuming and prone to substantial inter-observer variability (IOV). To address this challenge, a deep learning (DL) framework leveraging transfer learning (TL) generated probabilistic clinician-based segmentation maps, to capture and quantify the systematic component of IOV in whole breast irradiation. This study validated it on a separate cohort with multiple observers. Material/Methods: The dataset included 961 patients treated at San Raffaele Institute (2017–2022) with breast-conserving surgery followed by radiotherapy, including 3D planning CTs and clinician-delineated CTV contours. An in-house UNet model (MONAI v1.3), previously trained and externally validated [1], was fine-tuned using TL (200 epochs) on data from seven clinicians with ≥70 cases each, yielding seven clinician-specific models averaged to produce a probabilistic segmentation map capturing inter-observer variability. A validation set of 40 patients contoured by three expert clinicians and one resident was analyzed. CTV iso-probabilities (from 14%, 1 out of seven models, to 100%, all models) were compared against clinical contours. ECDF (Empirical Cumulative Distribution Function) under/over residuals (Fig. 1A), and standard metrics such as Dice Similarity Coefficient (DSC) or Average Surface Distance (ASD), evaluated training map quality by comparing iso-probabilities against the four clinicians’ labels. Dosimetric differences were also investigated by comparing automatically optimized plans referring to isoprobabilities/clinical CTVs on 5 left breast patients.
Conclusion: Multiple 3D DL architectures were optimized, trained and validated on a large cohort of patients, successfully predicting the CTV for right\left breast. Model-based probability maps were generated from the four best performing models (in line with clinical IOV for the current dataset), allowing the computation of model uncertainty. Supported by CCM 2024 (Ministry of Health). References: [1] Ubeira-Gabellini M. G., Palazzo G., et al. “Development and external multicentric validation of a deep learning-based clinical target volume segmentation model for whole-breast radiotherapy”. Physics and Imaging in Radiation Oncology 34 (Apr. 2025), p. 100749. doi: 10.1016/j.phro.2025.100749.[2] Maddaloni F. S., Broggi S., Fodor A., Palazzo G., Ubeira- Gabellini M., Pasetti M., et al. “Clinical validation of an Artificial Intelligence (AI) based auto-segmentation tool for breast radiotherapy planning”. ESTRO 2025 - Abstract Book (2025). Keywords: deep ensemble, breast CTV, radiotherapy
Made with FlippingBook - Share PDF online