Ensemble of Deep Convolutional Neural Networks for Multi-Class Skin Lesion Recognition Using Soft Attention

Precise diagnosis of skin lesions at an early stage is crucial for thwarting skin cancer. Physicians often require substantial time to inspect skin lesions due to their similarity across diverse types. Automated systems utilizing deep Convolutional Neural Networks (CNN) for skin lesion recognition have recently gained considerable attention. However, it is essential to ensure that these models effectively emphasize the significant areas of an input image to extract distinguishing features. This paper explores the potency of incorporating a Soft Attention (SA) layer in deep CNN models to focus on the most critical regions during feature extraction. The experimental results reveal that the SA layer enables models to concentrate on critical areas, enhancing their discriminative capabilities. Additionally, the efficacy of five deep CNN models is compared, and their collective intelligence is combined through the Softmax Averaging (Savg) ensemble method. An empirical evaluation operating a publicly available dataset demonstrates that our approach achieved high accuracy, precision, and recall in recognizing skin lesions. Therefore, the findings from this work help devise more efficient techniques for identifying skin lesions and preventing skin cancer.