SKIN CANCER RECOGNITION USING DEEP LEARNING ON ORIGINAL AND AUGMENTED DERMOSCOPIC DATASETS

Abstract

Skin cancer has been listed among the most common and deadly diseases in the world and the diagnosis of the cancer in the earliest and accurate manner is quite crucial in increasing the probability of patient survival. Recent advances in deep learning have demonstrated that there is much that can be done to extract skin cancer diagnosis by automatically analyzing dermoscopic images. The article is an elaborate implementation and comparative study of deep learning-based skin cancer classification on original and augmented dermoscopic images datasets. The three types of models, which were investigated, were a classical Convolutional Neural Network (CNN), a deep residual network (ResNet-101), and a proposed attention-based CNN. To deal with the problem of class imbalance and contribute to the overall model generalization, data augmentation tools, including rotation, flipping, zooming, and brightness adjustment were used. Based on the preliminary experimental findings on the original data, it came out that the baseline CNN and the ResNet -101 had an accuracy of 77 and 75 respectively. In comparison, the CNN based on attention had a significantly higher accuracy of 98 that indicates greater feature discrimination. After their training on the augmented datasets, the accuracy increased significantly in all models with CNN and ResNet-101 achieving 99% and 98 accuracy, respectively. Attention-based CNN repeatedly provided the best performance of the highest accuracy of up to 99 percent with the high precision, recall and F1-score on all types of lesions, including the minority clinically vital subsets. These results confirm that data augmentation and attention will be very effective in enhancing robustness, reliability thus the proposed methodology would be most relevant in automated and clinically useful management of skin cancer diagnosis.