A Robust Skin Cancer Classification using Deep Learning

Abstract

In the medical field, detecting skin cancer from pictures can be difficult. Skin cancer identification is a laborious process in modern medicine that could eventually result in patient death. Early detection of skin cancer is essential to the effectiveness of full treatment. It is difficult to diagnose skin cancer effectively. Consequently, there are not enough qualified dermatologists around the globe to meet the demands of modern healthcare. Data imbalance problems are caused by the wide diversity in data from several healthcare industry categories. Deep learning models are frequently trained in one category more than others due to issues with data imbalance. Using an unbalanced dataset, this study suggests a revolutionary deep learning-based skin cancer detector. Data augmentation was used to balance different skin cancer categories and overcome data imbalance. The MNIST: HAM 10000 skin cancer dataset, consisting of seven skin lesion categories, was used. Deep learning models are widely used in image-based disease diagnosis. Deep learning-based models (Custom CNN, DenseNet201, ResNet-101, and Xception) have been used to classify skin cancer. The proposed framework was fine-tuned using different sets of hyperparameters. The results show that Xception outperformed ResNet-101, DenseNet201, and Custom CNN in terms of accuracy, Fl score, and receiver operating characteristic (ROC) curve. It achieved an accuracy of 0.9148. Our proposed framework could help in disease identification, which could save lives, reduce unnecessary biopsies, and reduce costs for patients, dermatologists, and healthcare workers.