Prior Knowledge-Based Augmentation for Target Auto-Segmentation In Adaptive Radiation Therapy Using Segment Anything Model 2
Abstract
Purpose
Accurate tumor segmentation is essential for adaptive radiation therapy (ART) but remains time-consuming, labor-intensive, and subject to considerable inter-user variations. Recent advances in foundation models, such as Segment Anything Model 2 (SAM2), show strong potential for prompt-based automatic segmentation but still face notable accuracy limitations in tumor delineation. This study proposes novel prior-knowledge-based augmentation strategies to enhance SAM2’s performance for tumor segmentation within the ART workflow.
Methods
We propose two prior knowledge-driven augmentation strategies to enhance the tumor segmentation performance of SAM2 in MR-LINAC-based ART. The first strategy integrates prior MR images and their corresponding annotations with the current MR scan to provide contextual guidance. The second strategy enhances prompt robustness by applying random bounding box expansion and mask erosion/dilation during model fine-tuning. The resulting SAM2-Aug model was fine-tuned and evaluated on an in-house One-Seq-Liver dataset, which comprises 115 MRI scans acquired using a single MR sequence from 31 liver cancer patients. Furthermore, the fine-tuned model was directly evaluated, without additional adaptation, on two independent datasets: Mix-Seq-Abdomen, consisting of 88 mixed-sequence MRI scans from 28 patients with abdominal tumors, and Mix-Seq-Brain, comprising 86 mixed-sequence MRI scans from 37 patients with brain tumors.
Results
SAM2-Aug outperformed state-of-the-art convolutional, transformer-based, and prompt-driven segmentation models across all datasets. It achieved mean(±s.d.) Dice scores of 0.86±0.08, 0.89±0.06, and 0.90±0.07 for ITV(One-Seq-Liver), ITV(Mix-Seq-Abdomen), and CTV(Mix-Seq-Brain), respectively, compared with 0.78±0.13, 0.84±0.11 and 0.88±0.08 obtained by the original SAM2. Moreover, SAM2-Aug demonstrated robustness to variable tumor boundaries and strong cross-modality generalization without the need for large-scale retraining. The integration of prior contextual information and prompt augmentation significantly improved segmentation quality in challenging ART scenarios.
Conclusion
By incorporating prior imaging information and enhancing prompt diversity, SAM2-Aug delivers accurate, robust, and generalizable tumor segmentation. These advancements offer a potential path toward more efficient and precise ART workflows.