Development of a Deep Learning Model Inspired by Transformer Networks For Multi-class Skin Lesion Classification

Abstract

Skin cancer is one of the most common and life-threatening diseases in humans, where early and accurate diagnosis is a critical challenge in the medical field and can significantly affect the course of treatment for patients. In this study, after preprocessing images using Gabor filtering and color channel weighting, deep features are extracted with a deep learning model based on the advanced DenseNet-121 architecture. The key innovation of the proposed method is the design of a Feature Reinforcement Block (FRB) to enhance the extracted features and to improve the accuracy of the detection of various types of skin lesions. Wavelet transform, Multi-Head attention, and LSTM-Similarity module (LSM) comprise the feature reinforcement block. The wavelet transform helps extract local features such as edges and textures more effectively, and the Multi-Head attention mechanism, inspired by transformer networks, enables the model to focus on more prominent and important features, increasing classification accuracy. Additionally, the LSTM-Similarity module analyzes feature similarities and variations, further enhancing the model's ability to identify key characteristics along with an attention mechanism. HAM10000 and ISIC benchmark datasets were used to test the proposed model, and it was found to be highly accurate in classifying skin lesions into different categories. According to the results, the method is comparable to state-of-the-art approaches in terms of skin lesion classification.