Federated Learning in Radiology

Background and aims: Artificial intelligence, particularly deep learning, has shown considerablepromise in medical imaging. The models may be used to interpret radiological images toassist clinicians with clinical activities including disease diagnosis, medical intervention, treatmentplanning, and prognosis, to mention a few. However, the development of adaptable, robust,and accurate deep learning-based models frequently relies on the collection and time-consumingcuration of large amounts of high-quality annotated training data, which should ideally comefrom diverse sources and patient populations to account for the heterogeneity that exists in suchdatasets. Federated learning enables different institutions to work together to create a machinelearning algorithm without transferring any of their data. We did a review to assess the state offederated learning in radiology and to discuss its drawbacks and potential.Method: A comprehensive search of valid English scientific databases such as PubMed, Web ofScience, and Scopus was conducted from ۲۰۱۶ to ۲۰۲۲ using a combination of keywords such as“machine learning”, “federated learning”, “distributed learning”, “medical imaging”, and “radiology”.The inclusion criteria for this study were articles that focused on FL as the main topic intheir research and used medical image datasets. However, articles that used private datasets forthe model, review papers, abstracts, short articles, pre-prints, books or book parts were excludedfrom this review.Results: A total of ۱۹ papers were included in this review based on inclusion and exclusion criteria.The quantity of papers on federated learning has been constantly growing since ۲۰۱۶. Thelargest collaborative effort contained data from ۵۰ different hospitals/institutions, and all includedresearch that featured at least two collaborating institutions. Interdisciplinary teams made up ofclinicians and technical professionals carried out all of the studies. The majority of the studieswere carried out in an international collaborative context, while the others were undertaken just indeveloped nations, all of which were developed. The dataset sizes of study individuals or deriveddata ranged from hundreds to tens of thousands. The most widely used models are convolutionalneural networks (CNNs), but neural networks (NNs) and recurrent neural networks (RNNs) arealso quite common. The majority of studies used offline learning to perform a binary classificationprediction task.Conclusion: Researchers face a number of open difficulties, including privacy-preserving hyperparameteroptimization, entity resolution for vertically divided data, and efficient encryptionmethods. We anticipate that federated learning for medical applications will gain popularity in thenear future, resulting in more advanced security and privacy assurances that will enable real-worldimplementation of federated learning systems. In comparison to other areas, the healthcare sectoris in desperate need of the potential breakthroughs made available by machine learning and, inparticular, federated learning.

کلیدواژه ها:

artificial intelligence ، distributed learning ، federated learning ، radiology ، machine learning