Feature Fusion Model for Skin Lesion Classification using Grad-CAM

Abstract

Many methods have been proposed for computer analysis of skin lesions in recent years. Some widely used methods employ rules based on asymmetry, irregularity of boundaries, etc., in skin lesions. However, it is difficult to determine disease from skin lesion images due to their considerable variability, such as noise and low contrast. Therefore, more robust and superior automated skin lesion diagnosis methods are required. Ensemble-based methods have been proposed to improve their robustness, but these methods roughly use the final decision of each model. They can’t effectively use precise image features, which are extracted using each model for the final decisions. In this paper, based on the ISIC open-source dataset, we utilize feature maps from multiple Convolutional Neural Networks（CNN） and construct a Feature Fusion Model by applying Attention to classify skin lesion images into four classes. By applying Grad-CAM to various features extracted from each CNN and concatenating them, we achieved an accuracy of 93%. The proposed model is expected to learn valid features from skin lesion images.