Multi-Scale SAM for Cell Segmentation

Addressing the insufficient generalizability of most existing algorithms for cell segmentation, which often fail to maintain the same high-level segmentation performance when facing cells with different staining methods, imaging modes, and types, this study proposes a multi-scale segmentation network based on an attention mechanism and adapter combination. This network, derived from the Segment Anything Model (SAM), incorporates a multi-scale convolutional module with a fused attention mechanism, optimizing the feature maps in both channel and spatial dimensions. An aggregation connection module is also introduced to effectively fuse cross-branch information, enhancing the perception of multi-scale information. Furthermore, an adapter combination is introduced to better adapt the model to the feature distribution in the medical image domain, thereby improving its performance in pathological image segmentation tasks. Validation on the public dataset DSB reveals that the proposed network achieves an average Dice coefficient of 90.48 %, representing a nearly 20 % improvement compared to the original SAM. Meanwhile, it outperforms the benchmark networks UNet, TransUnet, and MedSAM by 8.88 %, 5.93 %, and 9.86 %, respectively, demonstrating the effectiveness and advancement of the proposed method.