A Combination of Soft Attention-aided CNN Models using Dempster-Shafer Theory for Skin Cancer Classification

Skin cancer is one of the most deadly forms of cancer in the world, causing hundreds of deaths every year. Researchers have been developing various computer-aided diagnosis (CAD) systems to help medical professionals classify the types of skin cancer using different image modalities. In this paper, we propose a model, which combines the outcomes of three separately trained attention-based convolutional neural networks (CNNs) using the Dempster-Shafer theory. The three base learners used in this paper are ResNet50, InceptionResNetV2 and Dense-Net201, each of which is aided by a soft attention module. For evaluating the proposed model, we have considered a publicly accessible skin cancer dataset, called HAM10000. This is a challenging dataset to work with as it is a class imbalanced dataset. Our proposed approach yields an accuracy of 0.932, which is better than many recently published methods found in the literature. The source code is available at: https://github.com/Cmatermedicalimageanalysis/DST_Combination