Background Subtraction-based CNN with Dual-energy Dependent Active Contour for the Extraction and Classification of Skin Lesion Descriptors

Abstract

This work proposes a background subtraction-based convolutional neural network (BS-CNN) for extracting and classifying the color and texture descriptors of the skin lesion region. The proposed skin lesion classification approach has three phases namely preprocessing, lesion segmentation, and descriptor extraction with classification process. In the preprocessing phase, the skin lesion images are enhanced using contrast-limited adaptive histogram equalization, while the hair artifacts present in the image are removed using the morphological and thresholding process. In the second phase, the lesion region is segmented using the hue, saturation, and value (HSV) transformation followed by a dual energy-dependent active contour model (DED-ACM) based segmentation process. Finally, the proposed BS-CNN extracts and classifies the skin lesion features. The BS-CNN approach compensates for the background skin intensity on the lesion region by estimating a background-mapped skin template region. This skin template region is calculated using the average intensity of the mask region around the boundary points of the lesion region. The evaluation of the proposed skin lesion classification approach was evaluated with the dermoscopic images taken from the datasets namely the PAD-UFES-20 and HAM-10000 with evaluation measures such as Sensitivity, DCI, F1-score, accuracy, and specificity. The proposed approach results in an accuracy of 98.46% and 98.91% when evaluated in the PAD-UFES-20 and HAM-10000 datasets in classifying five different types of skin lesions.