Deep learning-based identification of skin lesions associated with chronic venous insufficiency in biopsy-proven dermoscopic images

Chronic venous insufficiency (CVI) often presents with skin lesions that are difficult to distinguish from other dermatological conditions, typically requiring invasive biopsies for accurate diagnosis. This study proposes a non-invasive approach using a deep learning model based on dermoscopic images. A dataset of 839 images from 242 patients with histopathologically confirmed diagnoses was used. The images were categorized into CVI-related and non-CVI conditions. The Swin Transformer architecture, which incorporates hierarchical patch merging and shifted window attention, was trained to improve performance on difficult samples. The model achieved high diagnostic performance, with an AUC of 0.994, accuracy of 95.8%, and F1-scores of 0.955 and 0.960 for CVI and non-CVI classes, respectively. It outperformed both CNN-based models such as ResNet-50 and EfficientNet-B4, and earlier transformer models like ViT-B/16 and ViT-L/16. Misclassifications were primarily due to lesions with atypical secondary changes or overlapping features. These findings suggest that the Swin Transformer is a promising tool for improving diagnostic precision in CVI-related skin conditions, offering a less invasive alternative to current methods. However, further validation with larger datasets is necessary to ensure broader applicability.