Implementation of explainable AI for detection of skin cancer using image processing leveraging Grad-CAM

Abstract

In the ever-evolving landscape of medical diagnostics and treatment, the early detection and precise classification of skin cancer is crucial for effective and successful treatment outcomes. Using deep learning models in cancer diagnostics have been proven to be remarkably successful and transformative, and in some cases, computer- aided diagnosis (CAD) has even been observed to outperform conventional diagnosis techniques and human expertise. However, the complexity of artificial intelligence (AI) and machine learning models used for such diagnosis often leads to a lack of transparency in decision-making, and hence, adversely affects the confidence of medical professionals as well as patients in computer-aided diagnosis. Acknowledging the vital need for transparency in the process of CAD, our study explores the application of Explainable AI techniques in the detection of skin cancer. Using transformer models like Vision Transformer and Swin Transformer and deep Convolutional Neural Networks (CNN) based learning architectures such as VGG-16 and ResNet-50, and by integrating Grad-CAM with image processing, we aim to enhance the techniques of skin lesion classification. Grad-CAM provides a visual explanation of the model prediction allowing for medical professionals to judge the decision-making and prediction process behind the AI’s assessments and predictions. Consequently, our research aims to improve the trustworthiness of the black-box AI models in the field of medical diagnostics. Using the knowledge of the aforementioned models, we have developed our own hybrid model, AResNN (Attention-based Residual Convolutional Neural Network), that incorporates attention mechanisms and skip connections to improve accuracy and computational efficiency. We leveraged the power of transfer learning to extract maximum performance from our model. Furthermore, Grad-CAM has been applied to provide visual explainability of our model. Ultimately, our study intends to close the rift between complex AI models and their medical application in skin cancer diagnosis by making the models more transparent.