Classification of Skin Disease Dermoscopy Images Utilizing Multi-headed CNN

Skin cancer has a significant mortality rate, particularly in Western nations. It frequently affects people’s lives all around the world. Early screening of it increases human survival and helps ward off disease. Manual screening of skin disease needs more time for inspection. Hence, it is essential to develop automatic diagnostic methods that can classify skin lesions more precisely. Most of the existing automated models have larger sizes, have more configurable parameters, and take longer time to train networks. To enhance diagnosis performance of skin lesions this work proposed a multi-headed convolutional neural network (MHCNN) using dermoscopy images. At first, Tikhonov-based unsharp masking is utilized on dermoscopy images to capture high-frequency details, which may contain more appropriate information. The obtained unsharpened images are fed as input to the proposed MHCNN to classify multiclass skin lesions. The efficacy of the proposed MHCNN is assessed against other notable existing approaches. The proposed network obtained an accuracy of 90.25%, precision of 89.70%, sensitivity of 89.58%, specificity of 92.74%, and F1 score of 0.8964 from the experimental analysis.