Comparative Analysis of CNN Architectures in a Federated Learning Environment for Skin Lesion Diagnosis

Machine learning has emerged as a transformative force in the healthcare industry, enabling the development of intelligent systems that support early diagnosis. In particular, deep learning models have shown significant promise in the automated analysis of medical images. Some skin diseases can be malignant, which can lead to serious cancer risks. Therefore, accurate and timely diagnoses play a vital role in preventing the progression of the disease by offering the chance of early intervention. In this study, we focus on skin lesion diagnosis and evaluate the performance of various Convolutional Neural Network (CNN) architectures for accurate and reliable lesion classification. To address privacy concerns associated with centralized data collection, we adopt a federated learning (FL) approach that allows decentralized training across multiple institutions while keeping patient data local. We compare the diagnostic performance of multiple CNN architectures, within the FL framework using benchmark skin lesion datasets. According to the findings, the federated learning-based VGG16 model (FL-VGG16) achieved an accuracy of 83.48%, which is very close to the centralized learning approach. Furthermore, the analysis with the Vanilla CNN architecture shows that the federated learning approach has significantly higher classification performance compared to client-based learning.