3179 - Multi Model Uncertainty Guided and Prompt Enhanced Segment Anything Adaptation for Head and Neck Radiotherapy
Presenter(s)
Y. Xia1, S. Jabor2, P. New1, R. Salazar1, A. Roy3, N. Kirby1, and N. Papanikolaou1; 1University of Texas Health Science Center at San Antonio, San Antonio, TX, 2St. Mary's University, San Antonio, TX, 3University of Texas at San Antonio, San Antonio, TX
Purpose/Objective(s):
Deep learning based auto contouring is increasingly integrated into head and neck radiotherapy workflows. However, accurate delineation of organs at risk remains essential, particularly for small and low contrast structures where minor geometric differences can influence dose distribution. As these tools transition into routine clinical practice, evaluation should extend beyond segmentation accuracy to include assessment of prediction reliability. Conventional UNet based models provide strong performance but require prolonged training and inference times, significant computational resources, and offer limited interactive flexibility. Foundation models such as the Segment Anything Model(SAM) offer prompt-based flexibility, yet their stability in small and anatomically complex regions can vary. For safe implementation, segmentation systems must provide both accurate contours and interpretable confidence estimates. We developed a prompt enhanced fine tuning framework SAM-FT-HN that incorporates multi model uncertainty estimation to improve both segmentation accuracy and clinical reliability.
Materials/Methods:
An institutional cohort of anonymized head and neck CT datasets with expert delineations of 24 OARs and target volumes was used for adaptation and evaluation. The framework incorporates region focused enhancement for small structures, including localized prompt initialization and anatomy informed calibration, to improve consistency in challenging regions. Uncertainty was characterized using three approaches. Monte Carlo dropout captured stochastic variability. Snapshot ensemble aggregation across training checkpoints represented parameter uncertainty. Prompt uncertainty analysis evaluated sensitivity to prompt variation. These estimates were combined to generate spatial confidence maps identifying boundary ambiguity and variability. Performance was compared against SAMMed3D and nnUNet using Dice, Jaccard, HD95, recall, and precision.Results:
For larger OARs, SAM-FT-HN achieved Dice comparable to nnUNet (0.802 versus 0.788, p = 0.0020). However, the Dice was significantly higher when compared to SAMMed3D(0.605, p < 10?4). For smaller structures, the model improved over SAMMed3D (?Dice = +0.355, p = 0.0156) and remained similar to nnUNet ?Dice =-0.066. Across all OARs, SAM-FT-HN achieved higher recall than both SAMMed3D(p = 0.0011) and nnUNet(p = 0.0315). The integrated uncertainty framework significantly correlated with segmentation errors and supported clinical quality assurance.
Conclusion:
This uncertainty integrated refined implementation, improves segmentation of small and complex head and neck OARs while maintaining performance comparable to nnUNet. Spatial confidence information supports quality assurance and risk informed review, facilitating safer clinical implementation in radiotherapy planning.