A Comparative Study of SAM’s Performance Against Three U-Net Architectures on Retinal Vessels Segmentation Task

Semantic Segmentation is a pivotal undertaking within the domain of medical imaging analysis. It involves partitioning of a medical image into distinct regions of interest. Many deep learning architectures has led to the automation of the segmentation task, yielding performance levels close to those of domain experts. The U-Net architecture and many of its variants, has showcased their effectiveness for this task especially for biomedical images. Recently, the Segment Anything Model (SAM) has been released showing impressive performance in segmenting natural images. This paper is about challenging SAM on the task of retinal vessels image segmentation in comparison of three U-Net architectures, namely vanilla U-Net, Attention U-Net, and Spatial Attention U-Net. The evaluation is conducted on five diverse retinal vessel datasets. Our findings reveal that the inference mode of the SAM model, based on zero-shot learning, demonstrates inferior segmentation performance. However, when fine-tuned, it surpassed the performance of the U-Net models across all the studied datasets.