ESTRO 2025 - Abstract Book

S2445

Physics - Autosegmentation

ESTRO 2025

1813

Mini-Oral Incorporation of CTV delineation uncertainty through deep learning-based probabilistic segmentation maps Maria Giulia Ubeira Gabellini 1 , Cecilia Riani 2,1 , Gabriele Palazzo 1 , Martina Mori 1 , Alessia Tudda 1 , Roberta Castriconi 1 , Andrei Fodor 3 , Antonella del Vecchio 1 , Nadia Gisella Di Muzio 1,4 , Claudio Fiorino 1 1 Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy. 2 Physics Department, University of Pavia, Pavia, Italy. 3 Radiation Oncology, IRCCS San Raffaele Scientific Institute, Milan, Italy. 4 Medicine department, Vita-Salute San Raffaele University, Milan, Italy Purpose/Objective: Manual delineation is time-consuming and prone to significant inter-observer variability (IOV). Deep Learning (DL) techniques, particularly supervised methods, have shown great potential for auto-contouring. This work aims to enhance the reliability and effectiveness of DL model prediction by incorporating the impact of IOV through probabilistic, automatically generated, segmentation maps in the relevant case of CTV for whole breast irradiation. Material/Methods: The dataset-used for training and validation of DL models- comprised 3D planning CT images and CTV contours delineated by clinicians of 861 patients treated at our Institute. All patients underwent breast-conserving surgery followed by radiotherapy between 2017 and 2021. A separate set of 100 post-2021 cases was used for temporal validation. Each clinician (A–N) contoured from a minimum of 9 images up to 316 (average 144 and median 75). A CNN-based UNet model, selected from the MONAI library (v1.3), was trained over 200 epochs to predict left/right breast CTV. Validation performance was measured using Dice Similarity Coefficient (DSC), Average Surface Distance (ASD), and Hausdorff distances (HD and HD 95). Then, the best model was fine-tuned with transfer learning (TL, 200 additional epochs) using data labeled by individual clinicians, including the temporal validation set, yielding seven clinician-specific models (A, C, D, G, I, M, N), corresponding to observers with more than 70 available contours. Each model generated 3D binary predictions (1 meaning CTV voxels), averaged afterwards to create a probabilistic segmentation map representing the probabilities of CTV delineation, consistent with the systematic component of IOV. Results: The UNet model attains DSC ∼ 0.9, ASD ∼ 2.2 mm at epoch 200 with a computational training time of ∼ 16 hr. As reported in Fig.1 all TL models improve performance on all metrics compared to UNet model for the corresponding observer, making their predictions more representative of each single observer. The resulting segmentation probability maps (see Fig.2) highlight the variabilities in the cranial caudal axis (sagittal slice) and lateral side (axial slice). Conclusion: To address IOV, clinician-specific segmentation maps were created by retraining the optimal model on datasets from various clinicians. This yielded customized models with consequent diverse predictions leading to a probabilistic segmentation map for every patient, enhancing reliability and interpretability in segmentation applications. The visualization of probabilistic maps in combination with the “mean” auto-contour may translate into direct, individual, estimate of the impact of IOV. The method has also clear potential applications in clinical trials, worse/bad contours identification, tutoring and education.