Using FastSAM for Creating Custom Automatic Segmentation Models for Medicine and Biology

Abstract

FastSAM, a publicly available image segmentation model designed for general image segmentation, is turned into a highly adaptable and advanced segmentation tool that requires minimal training in two distinct scenarios. In the first case, we examine macroscopic X-ray images of the knee, in the second case, we focus on microscopic images of the zebra fish embryo retina, which have a significantly smaller spatial scale. We determine the minimum number of images needed to maintain state-of-the-art segmentation quality in each case. Finally, we evaluate the impact of image filtering and the unique considerations of segmenting 3D retinal volumes.