S1577
Physics - Autosegmentation
ESTRO 2026
muscles. The serratus anterior showed the most improvement, with the median Dice increasing from 0.39 to 0.86 (p<0.001), the median HD95 decreasing from 24.7 mm to 16.5 mm, and the absolute volume bias reduced from 75% to 25%, reflecting better delineation of this structure. The coracobrachialis remained the lowest-performing muscle.
relationships in muscles and reducing treatment- related side effects. References: 1. Wasserthal J, et al. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images. Radiol Artif Intell. 2023;5(5):e230024. Keywords: Breast cancer, Muscle segmentation, Deep learning Digital Poster 3011 Is an average CT enough for accurate deep learning ITV segmentation in advanced lung cancer? A comparison of full vs average breathing phase 4D- CT Luis R. DelaO-Arevalo, Lisanne V. van Dijk, Robert van der Wal, Johannes A. Langendijk, Robin Wijsman, Peter M.A. van Ooijen, Nanna M. Sijtsema Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands Purpose/Objective: Internal Target Volumes (ITV) for radiotherapy of lung tumours are generally defined by combining the GTV segmentations from all 4DCT Breathing Phases (BPs), capturing the tumour’s full breathing motion. Several studies showed that the ITV can also be segmented directly on the average scan [1]. This study aims to compare the performance of deep learning (DL) ITV auto-segmentation from full 4DCT BPs with direct segmentation from the Average CT, in a large cohort of locally advanced stage lung cancer patients including both primary tumours and involved pathological lymph nodes. Material/Methods: A total of 717 patients with locally advanced-stage Non-Small Cell Lung Cancer (n=577) and Small Cell Lung Cancer (n=140), treated with curative-intent radiotherapy were included. Data were split into 80% for development (5-fold cross-validation) and 20% for independent testing. Each patient had 4DCT scans without contrast enhancement that were reconstructed into ten breathing phases and an average CT scan, together with clinical GTVMaxExp (including both primary tumour and pathological lymph nodes) )segmented in the max expiration phase and ITV contours used as ground truth A 2.5D DL UNET LSTM segmentation network, inspired by de Biase et al. [2], using five consecutive CT slices was implemented. This network was trained separately for both segmentation of GTVMaxExp on the max expiration phase scan, as well as ITVAvgCT-DL on the average scan. The model trained for GTVMaxExp segmentation was applied to all breathing phase CT scans, and the ITVBPS-DL was constructed from the resulting DL-segmented GTV across phases.
Conclusion: TotalSegmentator performs well for AI-based
definition of muscles in arms-abducted radiotherapy planning CT-images. Segmentation inaccuracies were pronounced in muscles with shifted arm position due to arm abduction, especially anterior thoracic muscles. Re-training with treatment-position data improved segmentation performance. This adaptation enhanced auto-contouring performance, though expert review remained necessary before clinical use. This study represents a step toward determining dose-effect
Made with FlippingBook - Share PDF online