S1576
Physics - Autosegmentation
ESTRO 2026
Digital Poster 3003 External validation of a deep learning model for autosegmentation of shoulder muscles on radiotherapy planning CT-images of breast cancer patients Mariam Hammad, Lotte N Veldt, Anna M Dinkla, Desiree H.J.G. van den Bongard, Omar Bohoudi Department of Radiation Oncology, Amsterdam UMC, location Vrije Universiteit Amsterdam, Amsterdam, Netherlands Purpose/Objective: Accurate delineation of individual shoulder muscles is essential to evaluate radiation exposure and potential treatment side effects in breast cancer patients, as muscle dose may be related to treatment-induced side effects. This study aims to externally validate TotalSegmentator [1], for automatic muscle segmentation. CT-images were acquired from patients positioned in radiotherapy-position, i.e. in supine position with both arms abducted in an arm support. Additionally, the corrected segmentations were used to train a refined model. Material/Methods: CT-images of 40 breast cancer patients referred for postoperative radiotherapy, included in the BREAST- ART trial (ClinicalTrials.gov ID:NCT05727553), were segmented using TotalSegmentator, focusing on upper-body muscles. Auto-segmentations were reviewed and manually corrected to establish reference contours. Segmentation accuracy was evaluated using the Dice coefficient, 95th percentile Hausdorff distance (HD95), and absolute volume bias against these reference contours. A refined model was trained on the corrected data to adapt the AI-model to the anatomy during radiotherapy, i.e. with arms abducted. Both the original and refined models were applied to an independent cohort of 40 additional patients, and their performance metrics were compared. Results: Segmentation accuracy of TotalSegmentator varied across muscles. Large, well-defined structures such as the infraspinatus and supraspinatus achieved the highest accuracy (median Dice ≈0.98; HD95 ≈2 mm). Similarly, the trapezius, subscapularisandpectoralis minor showed consistently high overlap (≈0.96-0.98; HD95< 3 mm). Intermediate performance was observed for the deltoid (0.94; ≈7 mm) and teres major (0.76; ≈9 mm) and triceps brachii(0.72; ≈14 mm), while the serratus anterior (0.51; ≈24 mm) and coracobrachialis (0.34; ≈21 mm) showed the lowest performance and widest variability. An nnU-Net model was used to retrain on the corrected contours; evaluated on an independent 40-patient cohort, segmentation performance remained stable for most
Conclusion: Uncertainty-guided thresholds allow selective correction of the most uncertain regions of an auto- segmentation output. Applying T > 0.8 provided an actionable uncertainty map that captured the majority of dosimetric improvements while halving manual editing effort. Although APL serves as an indicator of time savings, future studies should validate the actual
clinical workflow efficiency gained through uncertainty-guided editing with thresholded uncertainty maps. References:
[1] Aoi Shimomura, et al. Monitoring intrafraction motion of the prostate during radiation therapy: suggested practice points from a focused review. Practical Radiation Oncology, 14(2):146–153, 2024.[2] M Huet-Dastarac, et al. Quantifying and visualising uncertainty in deep learning-based segmentation for radiation therapy treatment planning: What do radiation oncologists and therapists want? Radiotherapy and Oncology, 201:110545, 2024.[3] González, Patrick J., et al. "Explaining the dosimetric impact of contouring errors in head and neck radiotherapy." Biomedical Physics & Engineering Express 8.5 (2022): 055001. Keywords: Uncertainty, auto-segmentation, prostate MRgRT
Made with FlippingBook - Share PDF online