Abstract
BACKGROUND: Gastric cancer is a biologically aggressive disease, accounting for a substantial proportion of cancer-related deaths globally. Accurate localization of the lesion by artificial intelligence techniques helps timely and efficiently diagnose and treat. Segment Anything Model (SAM) has demonstrated considerable potential in medical image segmentation by displaying high performance in numerous image benchmark tests. However, its resource-intensive nature limits feasibility in embedded medical contexts. METHODS: This study proposed GC-SAM, a lightweight model for tumor segmentation. The architecture of GC-SAM is innovatively proposed, including a knowledge distillation image encoder, prompt encoder, and mask decoder, which effectively replaces the conventional fixed and computationally intensive network components. RESULTS: Extensive experiments demonstrate that GC-SAM significantly outperforms both classical segmentation models and recent state-of-the-art networks. On the internal test set, GC-SAM achieves 0.8186 Dice and 0.6504 mIoU, while reducing inference time and parameter count by over 80% compared to the original SAM. On the external dataset, GC-SAM maintains superior performance (Dice 0.8350), demonstrating excellent generalization. CONCLUSIONS: The proposed GC-SAM model shows strong capability in segmenting gastric cancer tissue, while also demonstrating practical potential for deployment in embedded medical imaging devices.