A novel nonlinear automated multi-class skin lesion detection system using soft-attention based convolutional neural networks

Abstract

Skin cancer is one of the most common cancer types that negatively affect human life worldwide and can result in death. So early diagnosis is essential for patient treatment. However, even by expert dermatologists, accurate skin lesion detection is complex. Therefore, developing computer-aided skin lesion detection and classification systems will help dermatologists provide early diagnosis and more effective treatment of skin lesion disease. Recent studies show deep learning networks have great potential for recognizing and classifying different skin lesions. This study aims to benefit from data processing, analysis, and a deep learning model base on soft attention to improve the performance for skin lesion classification. This research proposes an attention mechanism-based deep learning approach supported by data balancing methods. The HAM10000 dataset containing 10,015 labeled images belonging to 7 different skin lesion types was used. The dataset is balanced with SMOTE, ADASYN, RandomOverSampler, and Data Augmentation methods. A soft attention module was preferred as an attention mechanism to focus on the input data features and obtain a feature map. The proposed model consists of two convolution layers, a soft attention module, and four convolution layers. By combining convolutional neural networks and the attention mechanism, we extract the image features of the convolutional neural networks. In contrast, the soft attention module focuses on the relevant areas of the image. The accuracy rate of the proposed model is increased by the soft attention module and balance techniques in skin lesion classification problems. Various studies have been conducted using convolutional neural networks and attention mechanisms on open-source datasets for skin lesion classification. The originality of our study can be stated as using the attention mechanism in the middle of the neural network and performing experiments on the balanced dataset with various balancing methods. The training was carried out on the original, and the balanced versions of the HAM10000 dataset and the test results of the proposed model were presented. As a result of the training, 85.73 % training, 70.90 % validation, and 69.75 % test accuracy rates were obtained on the unbalanced HAM10000 dataset. The balanced dataset with SMOTE methods obtained 99.86 % train, 96.41 % validation, and 95.94 % test accuracy rates. SMOTE method performs better results than other balancing methods. It is seen that the proposed model, with data balancing techniques, has achieved high accuracy rates. It was concluded that using attention mechanisms and data balancing methods in deep neural networks increases the network's success.