An Explainable Deep Learning Framework for Multi-Class Skin Lesion Classification While Resolving Class Imbalance
Israt Jahan, Anwar Hossain Efat, S. M. Mahedy Hasan, Md. Farukuzzaman Faruk, Azmain Yakin Srizon, Md. Rakib Hossain, Md. Al Mamun
Abstract
Skin lesions are indicative of various skin diseases, including serious conditions like melanoma, which can be life-threatening if not detected early. Automated systems for analyzing dermatoscopic images have shown promise in improving diagnostic accuracy. However, many existing studies rely on pre-trained convolutional neural network (CNN) architectures without customization, leading to potential biases and suboptimal performance. In this study, we propose a customized DenseNet121 architecture tailored for skin lesion classification. We address class imbalance issues through data augmentation and fine-tune the model to extract fine-grained features. Our method surpasses other cutting-edge models in terms of accuracy, precision, recall, specificity, sensitivity, and root mean square (ROC) values. Furthermore, we utilise Gradient-weighted Class Activation Mapping (Grad-CAM) to visually represent the decision-making process of the model. The next study will prioritise the validation of the model using a wide range of datasets and the integration of lesion-related information to enhance its performance.