SAM2-Aug: Prior knowledge-based Augmentation for Target Volume Auto-Segmentation in Adaptive Radiation Therapy Using...

Area of Science:

• Medical imaging
• Artificial intelligence in oncology
• Radiation therapy

Background:

• Accurate tumor segmentation is crucial for effective adaptive radiation therapy (ART).
• Current methods are time-consuming and user-dependent, impacting treatment efficiency.
• The Segment Anything Model 2 (SAM2) shows potential but requires improved tumor segmentation accuracy.

Purpose of the Study:

• To enhance the Segment Anything Model 2 (SAM2) for accurate tumor segmentation in ART.
• To develop prior knowledge-based augmentation strategies to improve SAM2's performance.
• To create a robust and efficient tool for ART tumor segmentation.

Main Methods:

• Introduced two augmentation strategies: contextual input of prior MR images/annotations and enhanced prompt robustness.
• Implemented random bounding box expansion and mask erosion/dilation for prompt refinement.
• Fine-tuned the enhanced model (SAM2-Aug) on the One-Seq-Liver dataset and evaluated on Mix-Seq-Abdomen and Mix-Seq-Brain datasets.

Main Results:

• SAM2-Aug achieved superior Dice scores: 0.86 (liver), 0.89 (abdomen), and 0.90 (brain).
• Outperformed existing convolutional, transformer-based, and prompt-driven segmentation models.
• Demonstrated strong generalization across diverse tumor types and imaging sequences, with improved boundary-sensitive metrics.

Conclusions:

• Prior image integration and prompt diversity significantly enhance segmentation accuracy and generalizability.
• SAM2-Aug provides a robust and efficient solution for tumor segmentation in ART.
• The developed code and models will be publicly released to facilitate further research and application.