Efficient Fine-Tuning of SAM for Interactive Medical Image Multi-Organ Segmentation

Recently, Segment Anything Model (SAM) has significantly advanced interactive segmentation techniques, particularly its fine-tuned variants in medical imaging. Although previous approaches can adapt SAM to diverse downstream tasks with additional data, two challenges persist. On the one hand, organ ambiguities render SAM susceptible to human factor prompts, leading to unsatisfactory segmentation results. On the other hand, fine-tuning with massive data and the segment-single-region paradigm incur considerable computational and time costs. To this end, we propose to efficiently fine-tune SAM for interactive multi-organ medical segmentation by parallel heuristic prompt generation (PHPG) and class-balanced data pruning (CBDP). Specifically, PHPG generates prompts for diverse human behaviors guided by error prediction, effectively enhancing the consistency of prompts between training and testing. At the same time, it also offers a segment-multi-region paradigm to accelerate fine-tuning of SAM. Furthermore, considering that the contributions of training samples are dynamically variable and organ-related, CBDP is designed to reduce fine-tuning iterations. Experimental results on the FLARE and Synapse datasets indicate that the proposed method outperforms existing strategies. With only approximately 56.25% of the iterations, the segmentation performance is comparable to that of full fine-tuning, possessing better generalization ability and insensitivity to human factors.