An Adaptive Weighted Interpretable Multi-Scale CNN Model for Multi-Class Skin Lesion Classification

Abstract

Health issues are among the major concerns for people nowadays. Cancer remains one of the leading causes of mortality worldwide, and many are suffering from it. Out of these, skin cancer is also one of its kind and the deadliest. According to estimates, the number of new melanoma cases has surged in recent years, and the death rate from these instances has also sharply increased. By 2040, the number of melanoma cases will have sharply increased if this trend continues. So, finding and diagnosing it in the earliest stage is one of the most powerful and best ways to improve a skin cancer patient’s chance of survival. It is also essential to determine the type of cancer to ensure effective treatment. Artificial Intelligence (AI) has shown remarkable progress in medical image analysis, especially in Machine Learning and Deep Learning, for detecting and classifying skin cancer types. There are many revolutionary changes with the CNN models. Even though CNNs have become popular, these networks are called black boxes because we are unsure where the models focus in input images.These results are insufficient for medical practitioners to confirm it.Therefore, we require additional information on the model's explainability. Visual Interpretability is one of the techniques we used to address this issue. Grad-CAM is one of the XAI tools for visualizing where a CNN Network focuses. In this work, we focus on seven types of skin cancer. We trained the proposed IMSCNN+Grad-CAM model, along with state-of-the-art image classification models.The proposed model, IMSCNN+Grad-CAM, outperforms well, achieving an accuracy of 96.87%. Wealso compared and cross-verified the results of CNN models with Grad-CAM-generated images to ensure them. From this, we analysed our results and found that our proposed IMSCNN+Grad-CAM Model is a more suitable approach for skin cancer image classification and explainability problems.