Automated Morphology Detection of Nail-Fold Capillaries Through Enhanced Object Detection Network

The analysis of nail-fold anatomy can effectively evaluate microcirculation and diagnose vascular-related diseases. Early detection of these conditions is crucial due to the risk of severe complications if intervention is delayed. Extensive research supports the notion that nail-fold capillary morphology serves as a critical biomarker for various disease processes, with the degree of capillary structural damage potentially reflecting the involvement of internal organs. This study proposes a non-invasive methodology for detecting nail-fold capillary morphology by integrating an object detection model for improvement within a deep learning framework. We conducted an ablation study to enhance YOLOv8’s performance in detecting nail-fold capillaries and classifying their morphology. Our enhancements included adding a detection layer to improve the detection of various-sized objects, implementing Efficient Channel Attention (ECA) mechanisms, and incorporating data augmentation techniques and hyper-parameter tuning. These modifications yielded a notable improvement in mean Average Precision at IoU 0.50 (mAP@50), with increases of 3.7% in mAP, 3.6% in precision, and 2.5% in recall compared to the baseline YOLOv8 model. This culminated in a mAP@50 score of 79.9%. We also utilized Slicing-Aided Hyperinference (SAHI) to enhance inference performance on untrained multi-scale images and smaller capillaries, demonstrating significant effectiveness in real-time testing scenarios. The results from this research are promising for advancing early-stage diabetes detection using nail-fold image analysis and could potentially enable real-time applications in clinical environments.