Real-Time Nail-Fold Capillary Detection Using YOLO-RELAN With Saliency-Based Explainability Model

Resumo

The accurate identification of nail-fold capillaries is essential for the early detection of systemic diseases such microvascular abnormalities and systemic sclerosis. The complexity of capillary design and variations in image quality present significant challenges for precise identification in automated detection systems. An enhanced Deep Learning (DL) architecture that incorporates morphological preprocessing, shape-based feature enhancement, and a robust object recognition foundation. To emphasize essential vascular features, shape focus and extraction modules are performed after input images have previously undergone preprocessing with morphological filters to decrease noise and maintain capillary structures. YOLOv12 is utilized to accomplish the detection task using an enhanced backbone architecture based on Residual-Expanded Layer Aggregation Network (ReLAN), which improves feature reuse, gradient flow, and spatial representation. While significantly improving detection accuracy, this backbone maintains real-time functionality. To provide transparency and clinical reliability, the system makes use of Explainable AI (XAI) tools, especially saliency heatmaps, which highlight the critical areas influencing model predictions and provide interpretability for medical experts. The proposed approach achieves the highest accuracy of 88%, 90% of precision, 89% of recall, 86% of map-50 and 88% of map50-95. These results of the experimental evaluation demonstrate that a reliable and understandable solution for nail-fold capillary detection in clinical applications were achieved using the proposed approach.