S1600
Physics - Autosegmentation
ESTRO 2026
nutritional indices, which do not directly assess body composition. The temporalis muscle, located in the temporal fossa, is a known indicator of muscle mass and is visible on brain MRI [2]. This study evaluates the feasibility of using an open-source auto-segmentation tool, TissUNet [3], to measure temporalis muscle thickness (TMT) from routine MRI follow-up in children and TYA patients treated with radiotherapy for brain tumours. Material/Methods: We retrospectively identified child and TYA patients treated with photon radiotherapy for brain tumours at a single institution and selected a sample with T1- weighted volumetric MRI available in our anonymised database. Baseline MRI was defined as the scan closest to first radiotherapy treatment date. TissUNet was used to segment the temporalis muscle, and all segmentations were visually checked by a single observer. TMT was calculated by automatically identifying the slice at the superior orbital rim and taking average measurements from several slices inferior to this landmark. Trends in TMT were assessed over time, and by age and sex. Results: Sixty-one MR scans from thirteen children (6 females, 7 males) were analysed. Median (range) age at baseline was 15.1 (4.17 – 33.1) years. Patients had a median (range) of 4 (2 – 9) scans over a median (range) follow-up of 0.69 (0.07 – 1.40) years. One TMT measure, from a single patient, was excluded due to segmentation failure caused by large brain mass disrupting anatomical identification of the superior orbital rim (Figure 1A). All other segmentations were deemed acceptable (example in Figure 1B).
Results: Without uncertainty awareness, SE-ResUNet achieved median DSC, sDSC, and HD95 of 0.95, 0.94 and 3.77 mm for liver segmentation, and 0.86, 0.84, and 4.89 mm for GTV delineation. AIC showed eight cases with GTV-DSC < 0.6, mostly in multi-metastatic scenarios with undetected lesions. The uncertainty-aware pipeline detected these cases with 85% sensitivity. Of 52 designated metastases, 44 were initially identified by SE-ResUNet, and additional 7 after incorporating our uncertainty-aware module.
Conclusion: Our uncertainty-aware segmentation pipeline achieved high contouring accuracy and systematically identified potential segmentation errors and missed GTVs, representing a promising approach to enhance the trustworthiness of AI-driven auto-segmentation in clinical settings, reducing risks, and supporting safer treatment planning. Keywords: uncertainty quantification, trustworthy AI Digital Poster 4308 Auto-segmentation of routine MRI to assess sarcopenia post-radiotherapy in childhood and TYA brain cancer survivors. Angela Davey, Martin G McCabe, Marianne Aznar Division of Cancer Sciences, The University of Manchester, Manchester, United Kingdom Purpose/Objective: Sarcopenia, the progressive loss of muscle mass, is linked to reduced quality of life in child, teenage and young adult (TYA) cancer survivors [1]. Monitoring sarcopenia often relies on body mass index or
Made with FlippingBook - Share PDF online