ESTRO 2025 - Abstract Book

S2502

Physics - Autosegmentation

ESTRO 2025

segmentation accuracy across varied anatomical structures and conditions. Some exceptions should be known, leaving room for future fine-tuning.

Keywords: head and neck, flap, surgery

References: Carsuzaa F. Recommendations for postoperative radiotherapy in head & neck squamous cell carcinoma in the presence of flaps: A GORTEC internationally-reviewed HNCIG-endorsed consensus.2021. Thariat J. Reconstructive flap surgery in head and neck cancer patients: aninterdisciplinary view of the challenges encountered by radiation oncologists inpostoperative radiotherapy. 2024. Beddok A. International assessment of interobserver reproducibility of flap delineation in head and neck carcinoma.2022. Ronneberger, O. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI 2015. Isensee F. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. 2021. JT/AF co1st

3013

Digital Poster Comparing DL-based systems for automated CTV contouring in Breast Cancer Radiotherapy Gabriele Palazzo 1 , Francesca Saveria Maddaloni 1,2 , Maria Giulia Ubeira-Gabellini 1 , Cecilia Riani 1,3 , Sara Broggi 1 , Andrei Fodor 4 , Marcella Pasetti 4 , Roberta Tummineri 4 , Nadia Gisella Di Muzio 4,5 , Antonella del Vecchio 1 , Claudio Fiorino 1 1 Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy. 2 Physics, Università degli Studi di Milano Statale, Milan, Italy. 3 Physics, Università degli Studi di Pavia, Milan, Italy. 4 Radiation Oncology, IRCCS San Raffaele Scientific Institute, Milan, Italy. 5 Faculty of Medicine and Surgery, Vita-Salute San Raffaele University, Milan, Italy Purpose/Objective: To test the performances of two commercial AI-based auto-contouring systems in delineating the Clinical Target Volume (CTV) for whole breast irradiation and comparing their performance against an in-house deep learning (DL) model. Additional aims were: (a) quantifying potential improvements in contouring consistency among different Radiation Oncologists, and (b) assessing reductions in segmentation time for clinical target volumes (CTVs). Material/Methods: Three expert Radiation Oncologists manually contoured CTVs on CT scans of 40 breast cancer patients previously treated at our Institute, using the Eclipse TPS system (v13). Then, planning CTs were exported to two commercial AI based auto-contouring systems (named MODEL_1 and MODEL_2): in both cases, the structure “breast” was considered as potentially usable to auto-contour CTV. The automatically generated contours were reviewed and edited by radiation oncologists to their needs. The same CTs were also automatically contoured by an in-house U Net, recently developed using the Python libraries MONAI and PyTorch, and trained on an internal cohort of 861 patients. Variability was assessed using Dice Similarity Coefficient (DSC), Average Surface Distance (ASD), Hausdorff Distance (HD) and its 95th percentile variant (HD_95) between manual and the three automatic contours without considering editing. Statistical significance was evaluated using the Wilcoxon test. Editing times for MODEL_1 generated contours were also measured. Results: The commercial systems produced high-quality contours, though major editing was necessary. The in-house U-Net demonstrated good performance (see Figure 1), achieving DSC scores ( ∼ 0.9) comparable to inter-observer variability. MODEL_1 achieved significantly better DSC than MODEL_2 while MODEL_2 obtained better performances for HD metrics and ASD (p<0.05 for all metrics). Performances were significantly less accurate than the U-Net model (p<0.05), as shown in Figure 1, with the exception of HD that showed similar values between U-Net and MODEL_2.