Parameter-Efficient Adaptation of SAM 3 for Automated ITV Generation from 4DCT Images
Researchers have developed a new method to adapt the Segment Anything Model 3 (SAM 3) for generating Internal Target Volumes (ITVs) from 4DCT images. This parameter-efficient fine-tuning approach, utilizing Low-Rank Adaptation (LoRA) and a hard negative mining strategy, significantly improves segmentation accuracy and reduces artifacts in medical imaging. The framework demonstrates high performance with minimal annotated data and can be trained on a single consumer-grade GPU, offering a scalable and data-efficient solution for adaptive radiotherapy. AI