A Hybrid Approach Incorporating Superpixels for Diabetic Foot Lesion Segmentation Using YOLOv5 and SAM

The diagnosis and monitoring of diabetic foot lesions present a substantial challenge, with traditional techniques susceptible to inaccuracies and potential medical complications due to manual handling. To address this, we introduce a novel methodology that combines computer vision methodologies and advanced transformer-based technologies, specifically Meta’s Segment Anything Model (SAM). Our approach starts with lesion identification via YOLOv5, followed by segmentation utilizing superpixels within the area of interest and the creation of feature vectors. This process enables classification through logistic regression to discern between healthy and damaged skin superpixels. Following this, we integrate the bounding boxes, the pinpointed locations, and the original image as prompts for the SAM model to conduct a final comprehensive segmentation. Our methodology, evaluated using a sample of 83 lesion images, outperformed other state-of-the-art methods, with an Intersection over Union (IoU) of 94.10% and a Dice Similarity Coefficient (DSC) of 96.90%. Hence, the combination of conventional machine learning techniques with cutting-edge models like SAM can yield to precise results while ussing less computational resources when compared to the fine-tuning of advanced models like MedSAM.