Explainable Attention-Based Deep Learning Model for Stage-Wise Classification of Chronic Venous Disease

Abstract

Chronic Venous Disease (CVD) is a progressive vascular condition that develops in the lower limbs that may cause critical complications unless diagnosed and treated accordingly. Stage-wise classification is essential to successful treatment, but a lot depends on timely and accurate classification, and the traditional diagnostic tools normally depend on subjective assessment which is likely to be subject to inter-observer errors. A Deep Learning Model of automated stage-wise CVD classification using explainable attention. This model is built on a backbone of convolutional neural network (CNN) with an attention mechanism to concentrate the attention on the clinically significant areas, including varicose veins, edema, and skin lesions. The framework offers visual description in the form of attention maps, Grad-CAM, with help of which the clinician can interpret and legitimize the predictions of the model. This method makes the diagnostic less subjective and increases the confidence in AI-aided clinical decision-making by emphasizing medically important characteristics and promoting the diagnosis at the stage level. The current project provides an interesting framework of the real-world CVD testing and may be extended in the future by including multimodal data in the evaluation of a patient.