A Robust Multi-class Skin Cancer Classification using an Attention-based CNN Model

Abstract

Among lesions having high mortality rate, skin lesion is at top of the list. The patient‟s life could be rescued through early diagnosis. The manual identification period could be prolonged, as different categories of skin cancer have high similarities between their structure, color, and size which could lead to misclassification. An automatic and robust system is crucial for the timely and accurate categorization of skin cancer. In the last few decades, deep learning emerges as revolutionizing field, especially in the area of medical imaging. Most state-of-the-art work adopted transfer learning and ensemble learning based techniques for this problem, in which models used for TL and EL are designed for different types of problems and mostly trained on the huge amount of data which sometimes don‟t perform well for distinctive skin cancer problems due to very challenging datasets. This work proposed novel architecture based on a deep learning technique, which is designed especially for skin cancer classification problem. For learning, this method used dermoscopic images, which are excellent quality images captured through a high-level magnifying device, to obtain clear insights of skin containing cancer cells. The proposed architecture‟s main contribution includes, enhancing performance by embedding inception residual (IR) blocks in which, inception block is warily designed with several parallel layers that are merged together. Moreover, residual connections are established in each block to learn from low and high-level features concomitantly. These connections accelerate the learning process along with coping with the vanishing gradient problem. Deployment of the attention unit in the network boosted performance by suppressing the value of noise-containing features and assigning maximum value to important and relevant features during the learning process. Reducing the number of parameters by keenly optimizing the size and no of filters is another milestone attained through suggested model. The presented architecture achieved promising results by attaining 91.63% accuracy, 91.60% x precision, 91.60% sensitivity, 91.52% specificity, and 91.60% f1-score on the ISIC-19 dataset.