Federated Learning-Based Skin Cancer Detection Using CNN

Abstract

The primary categories of skin cancer are melanoma, squamous cell carcinoma, and basal cell carcinoma rank as some of the most commonly occurring tumors. An emphasis on early detection is essential in improving survival rates in patients. Recently, computeraided diagnostic systems using deep learning techniques for automation of skin lesion classification from dermoscopic images have emerged with considerable success. However, conventional approaches adopt the centralized paradigm wherein patient data are collected from multiple hospitals in one database. The issue of patient privacy remains significant in the health sector, given strict privacy policies. The limitation associated with the centralization method pertains to poor generalization of models owing to variation in imaging equipment, patient populations, and data distribution. Furthermore, the inability to ensure interpretability of the model makes it less suitable for computer-aided diagnostics since results produced by deep learning models are not explainable to physicians participating in clinical decision-making. Attempts at designing explainable artificial intelligence (AI) systems have proven ineffective in practice, with solutions that are too scattered for application. To overcome these limitations, a federated learning approach has been introduced for skin cancer detection. This method allows multiple healthcare institutions to jointly train the model while keeping patient data secure and stored locally, thereby preserving privacy and reducing the risk of sensitive information exposure. EfficientNet-B0 Convolutional Neural Network is used as an architecture for achieving accurate multi-class detection of skin cancer. Optimization of the model is done using Federated Averaging technique while interpretation of the results is attained using Grad-CAM visualization method. Evaluation experiments were conducted using various dermoscopy image datasets such as HAM10000 dataset. Results obtained from the analysis indicated better classification results compared to those obtained from existing models The suggested system has made some key contributions in this context. The use of a federated deep learning technique has been used to detect skin cancer. An explainable artificial intelligence method has been applied to help detect skin cancer. This method has been tested for skin cancer detection using dermoscopy images