Adaptation of Segment Anything Model enhanced by contrastive-manifold-regularised SSL and YOLO prompting for liver tumour segmentation.

Abstract

Liver tumours are a major global cause of cancer deaths, yet their segmentation remains challenging due to annotation complexity and high computational costs. While the Segment Anything Model (SAM) has shown promise in certain medical image segmentation applications, its performance remains suboptimal for liver tumour segmentation. Current three-dimensional models face scalability issues from computational demands and limited annotated data, restricting real-time clinical use. Here, we present MED-SAM, a 2D semi-supervised framework employing the medical SAM adapter (Med-SA), trained with contrastive manifold regularisation (CMR) and YOLO-guided click prompts. The method achieves a tumour Dice score of 93.26% on the MSD Challenge dataset-a 30-point improvement over the nnU-Net baseline (62.77%)-while requiring only 30% of the available labelled data (81 of 271 labelled volumes; the remaining 190 and all 272 unannotated volumes are used as unlabelled SSL data). YOLO-generated prompts are fully automatic, reducing annotation time by 29.0% in a within-rater controlled pilot (97.8s → 69.4s, same annotator, N=25slices). With 10-hour training and 0.33s per image inference time, the framework achieves an effective balance between computational cost and accuracy. These results indicate that MED-SAM can support treatment-planning workflows in hepatobiliary surgery while addressing the challenges of scarce annotations and real-time processing.