Recent advances in medical imaging have transformed patient diagnosis and treatment. Radiologists and other clinicians rely on imaging techniques such as MRI, CT scans, X-rays, OCT scans, and a host of other imaging modalities to diagnose many conditions, and pharmaceutical companies and research organisations use imaging data to drive new discoveries. The sheer volume of imaging procedures that are performed around the world underlines their impact and utility. By the World Economic Forum’s estimation, nearly 4 billion imaging procedures are performed worldwide every year with a market size of USD 40 billion in 2023.
Federated Learning in Radiology
This rapidly increasing amount of data presents an opportunity. The opportunity to harness its power to train machine learning (ML) models that will support the delivery of better, more efficient health care. For example, well-trained machine learning models can improve the efficiency of time-consuming cardiovascular imaging tasks such as image classification, anomaly detection, and patient selection. Or they can be used to extract robust information from cardiac images, allowing the redistribution of physicians’ time toward patient interaction and complex decision-making tasks. An ML model can also be trained to identify patterns in lung CT scans indicative of early-stage cancer, potentially leading to earlier detection and improved patient outcomes.
However, machine learning models achieve their best performance only after they are trained with large data sets. For example, training an AI-based tumour detector requires a large database including the full spectrum of indicators such as possible anatomies and pathologies. But accessing this data is hard, in part due to its sensitivity and regulations governing its use. The challenge is further compounded when these large datasets reside within different institutions, hosted at different data centres or on-premise, and in different countries. Accessing these datasets requires compliance with institutional and country-specific norms in addition to compliance with data protection and privacy laws. This makes the collection, maintenance, and sharing of large datasets expensive and time-consuming. Putting in place the pipelines required to transfer and keep copies of large datasets up-to-date is also extremely labour- and resource intensive. So how do we balance the training of ML algorithms that can enable innovation in healthcare against the limitations of accessing large datasets that are essential for creating effective ML models?
Federated Learning: Enabling the training of ML models on decentralised datasets
An exciting solution to this challenge is Federated Learning, an approach that enables the training of machine learning models, including deep neural networks and large language models, across decentralised datasets via successive rounds of local training and global parameter updates. An example of this is hospitals participating in the training of a machine learning model on their local data and sharing the updated model with the central server, as explained in this JACR publication.
Such an approach allows researchers and data scientists to harness the power of decentralised data – data hosted in different locations by different clients – without the need for data sharing. Algorithms can be trained on a diverse data set while protecting patients’ privacy and anonymity, and ensuring compliance with privacy and data protection regulations.
Other applications of Federated Learning
Beyond medical imaging, Federated Learning can allow the training of ML models that can be applied to a range of use cases. For example, well-trained machine learning models can be used to predict which patients are at risk of long-term hospital stays. Or deep neural networks can aid in the early diagnosis of sepsis, where early diagnosis can reduce mortality.
At Bitfount, we have built a powerful federated learning platform and are using it to boost the speed and accuracy of clinical trials, among other things. We do this by making site selection decisions faster and more informed. We also help trial sites identify patients meeting inclusion/exclusion criteria, thereby reducing screening failures. We enable efficient patient pre-screening by simultaneously analysing imaging and electronic health record (EHR) data using ML models and bespoke algorithms. Our federated system therefore results in better patient matching and significantly decreases a new treatment’s time to market. By using Bitfount to send algorithms and ML models to the data within routine-care settings such as clinics and hospitals, we also eliminate privacy risks.
Bitfount’s solution can also be used for trial feasibility studies. Our federated data science and data partnership solutions can be used to collaborate with trial sites, and to generate real-world data insights, without the need for any data sharing.
Our Federated Learning platform offers a step towards solving the challenge of training deep neural networks on large, diverse, decentralised datasets. By safely unlocking the value of sensitive data, we can transform healthcare – from diagnostic imaging to clinical research.
Interested in learning more about this transformative technology and seeing it in action? Book a demo now.