Image Visibility Filter-Based Interpretable Deep Learning Framework for Skin Lesion Diagnosis

Abstract

Computer-aided diagnosis have made a significant breakthrough in skin lesion diagnosis employing deep learning (DL) frameworks over the years, but it hardly reveals the transparency of the DL architecture. To mitigate this issue, in this article, we propose an image visibility filter (IVF) based DL framework for skin lesion diagnosis. The proposed IVF-DL network employs a ResNet architecture where visibility patches, extracted from the image visibility graph (IVG), are used as the convolutional kernels to extract salient features from dermoscopic images. The primary aim of this article is not only to classify skin lesions but also to depict the interpretable results after each residual block in a supervised manner. An optimal performance has been obtained by tuning three hyperparameters of the proposed method. Furthermore, the final interpretable result has been analyzed via IVG to resemble its spatial characteristics. Experimental results reveal that the proposed system outperforms the state-of-the-art classification methods quantitatively in terms of four performance metrics (accuracy, sensitivity, specificity, and area under the receiver operating curve) and qualitatively in terms of class activation map and relevance map considering two benchmark datasets.