SAM-OCTA: Prompting segment-anything for OCTA image segmentation

Abstract

Detailed analysis of a local specific biomarker in optical coherence tomography angiography (OCTA) images is essential for medical diagnosis, yet current methods primarily focus on global segmentation, such as of retinal vessel (RV) network. We propose SAM-OCTA, which fine-tunes the Segment Anything Model (SAM) with low-rank adaptation (LoRA) for segmentation tasks in OCTA. Our method enhances the semantic comprehension and prompt mechanism of SAM for OCTA en-face images and achieves a more flexible segmentation approach. The experiments explore the impact of prompt points with both global and local segmentation modes with the OCTA-500 and ROSE-O datasets, using random selection and special annotation prompt generation strategies. Considering practical usage, we evaluate model feasibility at smaller scales and demonstrate the necessity of fine-tuning. Comprehensive experiments demonstrate that SAM-OCTA achieves state-of-the-art performance in RV and FAZ segmentation and excels in artery–vein and localized single-vessel segmentation. The code is available at https://github.com/ShellRedia/SAM-OCTA-extend.