Addressing Overfitting in Dermatological Image Analysis with Bayesian Convolutional Neural Network

Abstract

VGG, ResNet, and DenseNet are popular convolutional neural network (CNN) designs for transfer learning (TL), aiding dermatological image processing, particularly in skin cancer categorization. These TL-CNN models build extensive neural network layers for effective image classification. However, their numerous layers can cause overfitting and demand substantial computational resources. The Bayesian CNN (BCNN) technique addresses TL-CNN overfitting by introducing uncertainty in model weights and predictions. Research contributions are (i) comparing BCNN with three TL-CNN architectures in dermatological image processing and (ii) examining BCNN ability to mitigate overfitting through weight perturbation and uncertainty during training. BCNN uses flipout layers to perturb weights during training, guided by the KL divergence and Binary Cross Entropy (BCE) loss function. The dataset used is the ISIC Challenge 2017, categorized as malignant and benign skin tumors. The simulation results show that three TL-CNN architectures, namely VGG-19, ResNet-101, and DenseNet-201, obtained training accuracies of 96.65%, 100%, and 97.70%, respectively. However, all three were only able to achieve a maximum validation accuracy of around 78%. In contrast, BCNN can produce training and validation accuracy of 81.30% and 80%, respectively. The difference in training and validation accuracy values produced by BCNN is only 1.3%. Meanwhile, the three TL-CNN architectures are trapped in an overfitting condition with a difference in training and validation values of around 20%. Therefore, BCNN is more reliable for dermatological image processing, especially for skin cancer images.