SAM-Glomeruli: Enhanced Segment Anything Model for Precise Glomeruli Segmentation

SAM-Glomeruli: Enhanced Segment Anything Model for Precise Glomeruli Segmentation

연구 분야: Software Development

학회: International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis

Chronic kidney disease (CKD) affects a significant portion of the population, necessitating early detection and intervention. In this work, we propose a novel Segment Anything Model-based Glomeruli Segmentation (SAM-Glomeruli) network tailored for Kidney Pathology Image Segmentation (KPIs). First, we adopt the pretrained ViT encoder of the large scale pre-trained Segment Anything Model (SAM) as our backbone to enhance the feature extraction capability of SAM-Glomeruli, providing robust representations for subsequent segmentatlon process. Second, in order to effectively transfer the natural images pre-trained SAM to the medical image domain, we employe Low-Rank Adaptation (LoRA) for efficient fine-tuning of the backbone to enhance its suitability for our specific task. SAM-Glomeruli demonstrates superlor performance, achieving 1st place in the instance detection task of the KPIs challenge. This work contributes to advancing precise pixel-level glomeruli segmentation across diverse CKD models and tissue conditions, potentially improving CKD diagnosis and research. The code is available in https://github.com/jj-ccc/KPIs2024.git.