An efficient deep learning approach to detect skin cancer using image data

Abstract

We propose and demonstrate an efficient deep learning approach to classify skin can cer using image data. The proposed approach is composed of several stages which are data acquisition, preprocessing and classification. For classifying skin cancer using image data and deep learning, four different convolutional neural network ar chitectures, EfficientNetV2B3, EfficientNetV2s, InceptionNetV3 and DenseNet121 were used on this work. The CNN models achieved accuracies of 83%, 86%, 84% and 88% respectively on a testing split of the HAM10000 dataset. Moreover, each of the CNN models were ensembled in two different ways, one is where all the predictions from the four models were averaged and the other one is based on K-Nearest Neigh bors approach where features from each of the CNN models were combined to fit a KNN model. The ensemble through averaging predictions achieved an accuracy of 90% and the ensemble based on K-Nearest Neighbors achieved an accuracy of 92%. Moreover, we demonstrated each of the CNN models using Explainable AI.