SkinNet: A Hybrid Convolutional Learning Approach and Transformer Module Through Bi-directional Feature Fusion

In light of melanoma's aggressive nature and complex structure, deep learning algorithms are the most effective way to diagnose the disease. Using bi-branch networks and the Transformer Module (TM), this study proposes to diagnose many lesions from dermoscopy images (DIs). Transformer's global input field features allow us to extract global features by selecting network branches rather than modifying the convolution kernel size. The bi-directional feature fusion architecture, Transformer and CNN branches are combined in this paper to extract more detailed features from the network. Further, our architecture proposes a bi-directional, two-branch configuration of features. In order to test the validity of the proposed technique, we examined large number of DIs. Dermoscopy images correctly classified 96.85% of the eight tissue classes in ISIC-2019. As opposed to current methodologies that categorize skin lesions into two categories, we classified all skin lesions in our study, including melanoma, as multi-class lesions. Additionally, it is possible to simplify and expedite melanoma detection by automating and speeding up the identification process.