ESTRO 2024 - Abstract Book

S3077

Physics - Autosegmentation

ESTRO 2024

[2] van Rooij W, Dahele M, Ribeiro Brandao H, Delaney AR, Slotman BJ, Verbakel WF. Deep Learning-Based Delineation of Head and Neck Organs at Risk: Geometric and Dosimetric Evaluation. Int J Radiat Oncol Biol Phys 2019;104:677– 84. https://doi.org/10.1016/j.ijrobp.2019.02.040. [3] Costea M, Zlate A, Durand M, Baudier T, Grégoire V, Sarrut D, et al. Comparison of atlas-based and deep learning methods for organs at risk delineation on head-and-neck CT images using an automated treatment planning system. Radiother Oncol 2022;177:61–70. https://doi.org/10.1016/j.radonc.2022.10.029.

1886

Mini-Oral

Validation of Deep Learning based CTV breast segmentation model on a large multicentric dataset

Gabriele Palazzo 1 , Maria Giulia Ubeira-Gabellini 1 , Alessia Tudda 1 , Martina Mori 1 , Roberta Castriconi 1 , Elisabetta Cagni 2 , Giulia Rambaldi Guidasci 3 , Eugenia Moretti 4 , Aldo Mazzilli 5 , Caterina Oliviero 6 , Lorenzo Placidi 7 , Andrei Fodor 8 , Antonella Del Vecchio 1 , Nadia Gisella Di Muzio 8,9 , Claudio Fiorino 1 1 IRCCS San Raffaele Scientific Institute, Medical Physics, Milan, Italy. 2 Azienda USL-IRCCS di Reggio Emilia, Medical Physics Unit, Department of Advanced Technology, Reggio Emilia, Italy. 3 Fatebenefratelli Isola Tiberina - Gemelli Isola, UOC di Radioterapia Oncologica, Roma, Italy. 4 University Hospital, Department of Medical Physics, Udine, Italy. 5 University Hospital of Parma AOUP, Medical Physics Dept, Parma, Italy. 6 University Hospital, “Federico II”, ., Naples, Italy. 7 Fondazione Policlinico Universitario A. Gemelli IRCCS, uosd medical physics and radioprotection, Rome, Italy. 8 IRCCS San Raffaele Scientific Institute, Radiotherapy, Milan, Italy. 9 Vita-Salute San Raffaele University, Medicine and Surgery, Milan, Italy Segmentation of clinical target volume (CTV) and organs at risk (OARs) are key steps for a successful radiotherapy treatment. Nowadays, the OARs/CTV segmentation is still done manually, which represents a task time-consuming for clinicians and prone to inter and intra-observer/institute variability [1]. ATLAS based methods typically fail to account for inter-observer variation [2]. With deep learning (DL) techniques, instead, the consistency between manual and automatic segmentation is higher and variability is considered. If OARs DL automatic segmentation is becoming more and more present in clinical practice, CTV auto-segmentation has been considered just by few works [5], especially because the inter-observer/institute variability in breast CTV segmentation is large [3-4]. Additionally, the multi-institute external validation of a DL model on the CTV auto-segmentation has not been described yet. This work aims to internally and externally validate an in-house model, developed and trained at a single institution using 611 patients through a multicentric cohort of seven institutes. Purpose/Objective:

Material/Methods:

As external validation dataset consists of 789, planning CTs and OARs/CTV segmentation from seven centers. The external validation datasets have been enrolled for the generation of multi-institutional knowledge-based plan prediction model, in the context of right and left whole breast irradiation [6]. Patients considered are respectively 428 for right-sided breast cancer (BC) and 361 for left-sided BC, with a minimum patients’ number per center of 54. Of