A Hybrid CNN and Morphological Feature-based Framework for Varicose Vein Detection and CEAP Severity Estimation

Abstract

Varicose veins are one of the most prevalent venous diseases that occur due to the minor morphological changes in superficial veins where the early identification is clinically difficult. This study suggests a hybrid image-based classification framework that combines a contour-based method for CEAP clinical staging (C0-C4) with deep learning for varicose vein diagnosis. The method initially leverages a Convolutional Neural Network (CNN) to discriminate normal images from varicose-affected leg images. To address limited dataset size, image augmentation techniques were employed and class imbalance got minimized through controlled sampling and pre-processing. An enhanced contour-density analysis was introduced for CEAP classification. A quantitative vein-density measure was calculated following the extraction of venous outlines using morphological filtering and Canny edge detection. Statistical quantiles were derived from training-set venous density measurements to create adaptive threshold boundaries for CEAP classes C0-C4. This strategy avoids manual threshold selection and allows the classifier to adapt to dataset-specific intensity distributions. Experimental results reveal that the CNN obtained 70% overall accuracy with 89% precision for normal cases and 86% recall for varicose veins. In order to find meaningful patterns, the contour-based CEAP approach effectively assigned severity levels which categorizes the severity from C0 to C4. This work can support future clinical validation and emphasizes the potential of lightweight computer vision techniques for early varicose vein screening, especially in resource-limited settings.