ESTRO 2025 - Abstract Book

S2501

Physics - Autosegmentation

ESTRO 2025

multidisciplinary teams. Future work will focus on model optimization with a larger cohort for improved generalizability in cases with missing teeth.

Keywords: orodental, head and neck, osteoradionecrosis

References: 1. Watson EE et al. Development and Standardization of a Classification System for Osteoradionecrosis: Implementation of a Risk-Based Model. medRxiv . Published online September 13, 2023:2023.09.12.23295454. doi:10.1101/2023.09.12.23295454

2973

Digital Poster Comparison of training strategies for fully automated detection and segmentation of flaps volumes in computed tomography scans Juliette Thariat 1 , abir fathallah 2 , alice blache 3 , arnaud beddok 4 , romain mony 5 , mathieu hatt 2 1 radiation oncology, centre baclesse, caen, France. 2 laTIM, INSERM UMR 1101, brest, France. 3 radiation oncology, university hospital, amiens, France. 4 radiation oncology, institut godinot, reims, France. 5 radiation oncology, centre henri becquerel, rouen, France Purpose/Objective: Postoperative radiotherapy (poRT) after reconstructive surgery is challenging due to a anatomy modifications. Flaps introduce additional complexities. Traditional methodologies have demonstrated limitations in segmenting flaps due to tissue composition and intricate morphology, hence we rely on Deep Learning frameworks such as the state of-the-art nnU-Net. Our goal was to optimize its training strategy. Material/Methods: Four datasets were used: GORTEC (trial 33 centers, n=148) and Xflap real-world cohorts from 3 centers, A (n=70), C (n=46) and R (n=20), including anatomy-changing outliers (tongue displacement devices). CT scans (1-3mm slices, contrast enhancement, artefact minimization algorithms, left to physician's appreciation; some including plaques, screws, dental materials or bone left unresected) and RT structure sets were collected. The contouring process was conducted by two expert radiation oncologists blind to patient characteristics and outcomes. Head and neck surgeons and a radiologist helped for difficult cases. The nnU-Net framework was employed. Interpolation to a common voxel size and data augmentation was performed. NnU-Net was first trained using random split (80/20) of the GORTEC set, and evaluated on A, C and R datasets. Then, a new training set was created by selecting 50% of patients with characteristics (anatomical location, flap volume, artifact severity, pedicled/free flap, bone resection/not) associated with lower Dice score (Dsc), in order to increase the nnU-Net generalizability. The two models were finally compared on the same testing set (the remaining 50% of the dataset). Results: Mean±std) Dsc for dataset GORTEC, A, C and R were 0.80±0.02, 0.69±0.05, 0.52±0.07, and 0.74±0.04 (overall 0.69±0.04, median 0.73) with the first training. These improved to 0.80±0.02, 0.74±0.03, 0.69±0.06, 0.75±0.03 (overall 0.75±0.03, median 0.88). Although factors such as volume, anatomical location, and the presence of artifacts exhibited some influence, their correlations with Dsc were generally weaker in the second training, indicating limited impact thanks to better training data representativeness. Notably, regions such as the oropharynx and higher volume segments showed mild positive correlations, suggesting that larger volumes and specific anatomical sites may facilitate better model performance. The presence of bite blocks in 5 patients resulted in poor Dsc.

Conclusion: Our results illustrate the importance of exposing the model to a more diverse training set in achieving consistent