A Comprehensive Study on Pre-trained Models for Skin Lesion Diagnosis in a Federated Setting

C. Siddarth, Ajay Kumar Reddy Poreddy, Priyanka Kokil

Abstract: Privacy has been one of the main concerns when it comes to the application of deep learning in the medical domain. Medical institutes prioritizing the privacy of their patients do not make their data public, making it difficult to build better models to diagnose rare diseases. But, after the advent of federated learning, there have been immense improvements toward building better models that employ patient’s private data without compromising their privacy. In this paper, we comprehensively study multiple models to diagnose skin lesions in a federated setting. Replicating real-life scenarios, we experiment in different settings where the number of clients or hospitals that participate varies. Further, we explore if the pre-trained weights obtained from natural image datasets could assist in building a better model for diagnosing skin lesions.