Performance Improvement of Convolutional Neural Network Architectures for Skin Disease Detection

Abstract

Convolutional neural networks are one of the most important techniques used in classification processes, specifically digital image classification. In this paper, work has been done to apply a set of different deep CNN architectures in classifying skin disease images. The performance of those deep networks or training methods was also improved which improved the image classification result. The main objective of this research paper is to improve the performance of convolutional neural network architectures for the detection of skin diseases, as the data set of images of skin diseases was adopted from the International Collaboration for Skin Imaging (ISIC) 2020, where the number of images that were used 5224 digital images of five skin diseases included 1327 Nevus pictures, 1098 pictures of basal cell carcinomas, 1099 pictures of pigmented benign keratoses, 1046 pictures of seborrheic keratoses, and 654 pictures of squamous cell carcinomas. The performance of AlexNet, ZfNet, VGG16, and VGG19 deep networks has been improved by generating new seed weights for each network based on Artificial Bee Algorithm, Bat Algorithm, Gray Wolf Optimization, Bacterial Foraging Optimization, and Particle Swarm Optimization. After obtaining the results from the improved architectures, it was found that the performance accuracy increased significantly, and the architectures gave clear stability in training the deep network data set.