Exploring Segment Anything Model Applications for Medical Imaging

Medical image segmentation is a machine-learning task seeking to identify regions of interest, such as tumors and organs, within a medical image. It remains a complex task because of the need for precise knowledge of anatomical structures and abnormalities. Being trained on a vast dataset comprising 11 million diverse images and over 1 billion high-quality segmentation masks, the Segment Anything Model (SAM) seems to have the potential to address issues facing current segmentation models. However, this hypothesis needs to be verified especially since SAM was trained on natural image data, which differs from medical image data. This work explores the studies and applications of SAM on medical image data. The results show that SAM generally performs lower than the state-of-the-art segmentation models because of its weakness in detecting objects with a low background contrast, indefinite borders or shapes, or ambiguous prompts. To overcome its inconsistency of performance and generalizability limitation, in addition to the scarcity of medical image data problems, several adaptations and customized versions of SAM were proposed.