Adaptive Segment Anything Model for Spatial Transcriptomic Cell Segmentation

With the development of spatial transcriptomics, now we could have a much better understanding of the spatial genomic profile of complex tissue. However, the cell segmentation remains a critical step for data analysis processing. Here we developed an adaptive SAM model for cell segmentation in the spatial transcriptomics. SAM model has been developed as a powerful basic model for segmentation. In the application of biomedical image segmentation, SAM did not perform equally well for different dimension of images. In our study, we first exhibited that SAM model segmentation is sensitive to image dimension. We provided minimal human annotation to initiate the optimization for finding the best dimension images for SAM model. We found that adaptiveSAM could perform better or equally well in the biomedical image segmentation without finetuning, compared with Faster RCNN and Mask RCNN based detection head.