ESTRO 2025 - Abstract Book

S3848

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

3831

Digital Poster AI-based metrics detecting the impact on outcome of individual variability in contouring CTV Gabriele Palazzo 1 , Maria Giulia Ubeira-Gabellini 1 , Andrei Fodor 2 , Cecilia Riani 1,3 , Monica Vincenzi 1 , Robert Jeraj 4,5 , Antonella del Vecchio 1 , Nadia Gisella Di Muzio 2,6 , Claudio Fiorino 1 1 Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy. 2 Radiation Oncology, IRCCS San Raffaele Scientific Institute, Milan, Italy. 3 Physics, Università degli Studi di Pavia, Pavia, Italy. 4 Faculty of Mathematics and Physics, University of Ljubljana, Ljubljana, Slovenia. 5 Department of Medical Physics, University of Wisconsin, Madison, USA. 6 Faculty of Medicine and Surgery, Vita-Salute San Raffaele University, Milan, Italy Purpose/Objective: To explore if the inconsistency between clinical contours of CTV and AI-predicted contours may identify class of patients at higher risk of toxicity and/or of local relapse. This first proof of principle study was focused on oedema after whole breast Radiotherapy. Material/Methods: Planning CT scans and clinical information of 797 breast cancer patients previously treated in the period 2009-2017 with whole breast post-operative Radiotherapy at 40Gy/15fr were available. Patients with silicone implants were previously excluded. CTV contours were predicted using an in-house U-Net model, trained on a separate, more recent, cohort from the same institute. These auto-segmented CTVs were compared against the clinical CTVs using three metrics: Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and the difference in predicted versus treatment CTV volumes (vol diff) The cohort was stratified into two subgroups: 32 patients with oedema within 6 months from the end of radiotherapy, prospectively registered, and 765 patients without oedema. Statistical significance of the differences between these subgroups was evaluated using the Mann-Whitney U-test, aiming to detect if individual variability from the predicted CTV reflects in higher/lower toxicity risk. Results: As can be seen from Figure 1 (the dashed line represents the mean), for patients with edema, all predicted CTVs well match planning CTVs. Patients with oedema show a DSC distribution different (mean: 0.896) compared to that of patients without oedema (mean: 0.878), p < 0.05. Importantly, only 3% patients with oedema showed a DSC value< 0.85 against 20% of patients without oedema, including all the outliers with low DSC values. Similarly, volume differences for patients with oedema were lower (mean: -55 cm3) than non-oedema (mean: -111 cm3), p < 0.05. Negative vol diff values mean that the clinical CTV is larger than the predicted one.