Artificial intelligence in medical imaging has achieved remarkable results, but these successes rely on large, diverse, high-quality datasets. The problem is that medical images are among the most sensitive and tightly regulated types of personal data, governed by laws like HIPAA in the United States and GDPR in Europe. These regulations severely limit how data can be moved across institutional or national boundaries.
A model trained on data from a single hospital often reflects that hospital's specific patient demographics, imaging equipment, and clinical protocols. When deployed elsewhere, it can fail in unexpected ways. This problem -- called data bias or domain shift -- is one of the central obstacles to generalizable medical AI.
The traditional solution is to centralize data by aggregating scans from many institutions into one large repository. But this approach is slow, expensive, legally complex, and carries real risks of data breaches. A fundamentally different approach is needed -- one that lets models learn from distributed data without ever moving the data itself.
Federated Learning (FL), introduced in 2016 by McMahan et al. at Google, solves the data-sharing problem by inverting the traditional approach. Instead of bringing data to the model, FL brings the model to the data. Each participating institution trains the model locally on its own patients, then sends only the model's learned parameters (not the patient data) to a central server.
The central server then aggregates these local updates -- most commonly by averaging the model weights from all participating sites -- to create an improved global model. This updated global model is distributed back to all participants, and the cycle repeats. The raw patient data never leaves any institution, which satisfies privacy regulations while still enabling collaborative learning.
FL can be deployed in several architectural configurations. In a centralized setup, a single server coordinates all communication. In a decentralized setup, institutions communicate directly with each other without a central server, eliminating single points of failure but adding complexity. The choice of architecture depends on the trust relationships between institutions and the technical resources available.
FL comes in three main variants based on how data is structured across institutions. Horizontal FL is used when all sites collect the same types of measurements but on different patients -- for example, multiple hospitals each collecting prostate MRI scans. This is the most common setup in medical imaging because different institutions tend to see different patients but use similar imaging protocols.
Vertical FL is used when the same patients are studied at different institutions using different tests. For example, one clinic might have MRI scans and another might have blood tests from the same patients. Vertical FL can combine these complementary data sources to train richer models. Transfer FL applies when neither the patients nor the tests overlap between institutions, relying on knowledge learned from related tasks or domains.
A persistent challenge across all FL variants is that real-world medical data is rarely independently and identically distributed (IID). Variations in scanner hardware, imaging protocols, patient demographics, and annotation practices create systematic differences between institutions. These non-IID conditions can cause the global model to perform poorly at specific sites and have motivated considerable research into specialized aggregation and normalization techniques.
Although FL avoids sharing raw data, it is not automatically secure. Researchers have shown that it is theoretically possible to reconstruct training images from the model weights alone through what are called model inversion attacks. This means that even the parameters transmitted during FL can leak sensitive information if intercepted or misused.
Differential privacy (DP) addresses this by adding carefully calibrated noise to the model updates before they are transmitted, making it mathematically impossible to determine whether any specific patient's data was included in training. The tradeoff is that adding noise can reduce model accuracy, particularly when datasets are small.
Homomorphic encryption (HE) offers a stronger security guarantee by allowing the central server to aggregate model parameters while they remain fully encrypted, without ever decrypting them. This is considered the gold standard for FL security but comes with significant computational overhead. Head CT scans pose particular identity risks because facial contours can be reconstructed from them, making strong encryption especially important for neuroimaging.
The first application of FL to medical imaging was for brain tumor segmentation on MRI, published by Sheller et al. in 2018. Using the BraTS public dataset, the study showed that FL could match or approach the performance of centralized training while keeping data at each site. This foundational work established the template that many subsequent studies followed.
The field's largest FL study to date was published by Pati et al., involving 71 institutions collaborating to detect glioblastoma sub-compartment boundaries. This at-scale demonstration confirmed that FL genuinely improves over locally trained models when enough diverse institutions participate, validating the core premise of the approach.
FL has also been applied to Alzheimer's and Parkinson's disease detection from structural MRI, using features like brain region volumes and cortical thickness. A key challenge in this domain is that scanner variability between institutions creates domain shift, causing models trained at one site to generalize poorly elsewhere. Specialized FL frameworks have been developed to explicitly model and correct for these inter-site differences.
The COVID-19 pandemic generated the largest surge of FL medical imaging research, driven by the urgent need to share diagnostic knowledge across nations without sharing patient files. FL studies used chest X-rays and CT scans to detect COVID-19 patterns, with multiple multinational collaborations demonstrating that federated models can outperform locally trained ones when combining data from geographically diverse institutions.
In breast cancer imaging, a major challenge is that screening datasets are overwhelmingly negative -- only about 11% of screened women have any finding. This severe class imbalance makes standard machine learning difficult. FL allows institutions to pool their rare positive cases across sites, significantly improving the training signal for detection algorithms without centralizing sensitive data.
Prostate cancer has been explored in a smaller number of FL studies. One notable contribution introduced a variation-aware FL framework that normalizes images into a common space before federated training, reducing the performance gap caused by different MRI acquisition protocols at different institutions. Another demonstrated FL-based prostate segmentation across three academic centers while preserving patient privacy.
Despite rapid growth -- from 1 FL paper in 2016 to over 3,500 in 2022 -- most medical imaging FL studies remain proof-of-concept demonstrations rather than production clinical systems. The field has consistently shown that FL can match centralized performance in controlled experiments, but real-world deployment faces additional hurdles around data standardization, institutional trust, and workflow integration.
A recurring obstacle is that finding institutions willing to participate in a federated network is difficult even when the technology works. Concerns about competitive disadvantage, liability, and the effort required to integrate FL into existing clinical systems all limit adoption. Better frameworks for establishing trusted federations and sharing the costs and benefits equitably will be needed.
Researchers have identified several priority areas for future work: improving security against increasingly sophisticated attacks, developing better methods to handle extremely non-IID data, reducing communication costs for resource-constrained institutions, and expanding FL to new modalities such as surgical video. As understanding of how models learn and transfer knowledge matures, federated models are expected to regularly surpass centralized approaches in the medical domain.
For cancer detection and diagnosis, FL represents a potential solution to one of the field's most persistent problems: no single institution sees enough rare cancer subtypes to train a highly generalizable AI model. By enabling hospitals in different countries to train together without sharing patient records, FL could dramatically expand the effective training dataset for rare malignancies.
For prostate cancer specifically, FL offers a path to models that generalize across the diverse MRI equipment, imaging protocols, and patient populations found globally. A model trained federally across dozens of institutions could potentially match or exceed radiologist performance across a much broader range of clinical settings than any single-site model.
The broader lesson from this review is that the privacy-preserving properties of FL are not just a regulatory nicety -- they are an enabling technology. By making it safe and practical for institutions to collaborate, FL has the potential to unlock decades of retrospective clinical data that currently sits unused in hospital servers worldwide, inaccessible to AI researchers because of legitimate privacy concerns.