Feature Fusion Model for Skin Lesion Classification Using Grad-CAM

Abstract

Many skin lesion detection methods have been proposed in recent years, which rely on rules such as asymmetry and irregular borders of skin lesions. However, further improvement is necessary for robust automated diagnosis due to the significant variations in skin lesion images, including noise and low contrast. Ensemble-based methods have been proposed to enhance robustness, but these methods utilize label outputs from multiple convolutional neural networks (CNNs) and exhibit relatively low performance. In this study, we propose a Feature Fusion model that directly utilizes image features extracted from each model. Furthermore, we incorporate feature maps from Grad-CAM to enable the utilization of important features. In our performance evaluation using the ISIC open-source dataset, we achieved a high accuracy of 93% in classifying four classes of skin lesions.