Federated learning is a training framework built around a simple but powerful idea: instead of pooling sensitive medical records on a central server, each participating institution keeps its data locally and trains a shared model on its own machines. In each communication round, a central server distributes the current global model to every client, each client performs local training using its own data, and the server then aggregates the updated parameters to produce a new global model. This approach, known as FedAvg, has attracted intense interest in medical imaging because it promises the statistical benefits of large, diverse datasets while respecting privacy regulations and institutional boundaries that often make data sharing impossible.

The research team, led by Yukihiro Nomura of Chiba University and The University of Tokyo Hospital, tested this idea using two distinct computer-aided detection systems. The first was a three-dimensional U-Net model designed to detect cerebral aneurysms in magnetic resonance angiography images, a task where missing a bulging vessel wall can be catastrophic. The second was a single-shot multibox detector for spotting brain metastases in contrast-enhanced T1-weighted MR images, a critical aid for planning stereotactic radiosurgery. For each task, the team assembled four heterogeneous datasets differing in scanner vendor or institution, totaling 423 MRA images for aneurysm detection and 398 contrast-enhanced T1-weighted images for metastasis detection, with voxel-level annotations available throughout.

The central question was deceptively simple: if a hospital network starts with two institutions and then adds a third and a fourth one at a time, does the detection software actually get better? The researchers hypothesized that it would, but the answer turned out to depend heavily on how the model is retrained when a newcomer arrives. They compared three strategies. From-scratch training reinitialized all model parameters and trained anew on the combined data of all participating clients. Full fine-tuning started from the previously trained global model and updated every parameter. Partial fine-tuning also started from the existing model but froze most layers, updating only a carefully chosen subset, such as the decoder layers of the U-Net or the upper and head portions of the single-shot detector.

The results revealed a clear pattern. For both detection tasks, sequential institution addition combined with fine-tuning generally produced higher median performance gains than training from scratch. Performance improvements tended to accumulate as institutions were added one by one, whereas adding two institutions simultaneously, a condition the team called At-once addition, yielded less consistent improvements. This distinction matters because the simultaneous addition of multiple clients introduces a larger distributional shift in one step, forcing the model to absorb more change at once. When the researchers extended the At-once stage to 200 communication rounds, gains improved in some settings but remained inconsistent across tasks and strategies, suggesting the problem is not simply a matter of insufficient training time.

Perhaps the most striking finding concerns parameter efficiency. In the aneurysm task, full fine-tuning and from-scratch training each updated roughly 0.80 million parameters, while the partial fine-tuning strategies updated only about 0.31 million and 0.08 million parameters respectively. In the metastasis task, the corresponding numbers were approximately 7.95 million for full updates versus 2.66 million and 0.53 million for the two partial strategies. Despite touching a fraction of the network, partial fine-tuning achieved performance changes comparable to full fine-tuning across both tasks. For real-world deployments, this efficiency translates directly into lower computational load and reduced communication overhead, both of which are significant practical constraints when hospitals with varying IT resources participate in a federated network.

The study also confronted an uncomfortable trade-off known in machine learning circles as catastrophic forgetting. When a model adapts to newly added institutions, its performance on the original institutions’ data can decline, which would be unacceptable if the software were already deployed and trusted at those sites. The researchers evaluated this by measuring performance changes on test sets from the initial clients separately from test sets spanning all four institutions. For cerebral aneurysm detection, performance decreases on the initial clients’ test sets were less frequent with full and partial fine-tuning than with from-scratch or centralized training. For brain metastasis detection, however, no strategy showed a clearly distinct pattern, underscoring that the forgetting problem remains task-dependent and unresolved.

Statistical caution tempers the enthusiasm. The team measured performance using the competition performance metric, which averages sensitivity across seven false-positive operating points, and quantified changes as differences from the corresponding initial model. Confidence intervals estimated by paired bootstrap resampling with 1,000 iterations included zero in every evaluated setting on the all-clients test sets, and direct pairwise comparisons among the four training strategies found no statistically significant differences after adjustment for multiple testing. The researchers also ran additional analyses with the FedProx aggregation method, designed for heterogeneous non-IID data, and found broadly similar patterns, suggesting the trends were not an artifact of the FedAvg baseline. Still, the honest conclusion is that the observed benefits are promising tendencies rather than proven guarantees.

The authors frame their setting as a domain-incremental learning problem, a concept from continual learning research in which models must accommodate new data distributions while retaining performance on previously learned ones. This lens highlights both the promise and the limits of the work. The study did not address scenarios such as institutions withdrawing from a network or participating intermittently, nor did it incorporate privacy-enhancing technologies like differential privacy or homomorphic encryption, which would likely be required in production deployments. The clients were also defined by acquisition site or scanner vendor rather than random patient-level splits, capturing realistic heterogeneity in devices and protocols but not the full breadth of differences in patient populations, clinical practice, and annotation policies that large international networks would encounter.

Even with these limitations, the study offers a roadmap for a future in which hospital AI systems improve continuously throughout their operational life. Regulatory frameworks have already approved commercially available detection software bearing CE marks and FDA clearances, and guidelines increasingly recommend multi-institutional data for building robust medical AI. What this research adds is evidence that the network can grow organically, hospital by hospital, with fine-tuned updates that are computationally light and privacy-preserving. As federated learning infrastructure matures, the modest but consistent gains documented here suggest that the collective intelligence of a medical network may indeed be greater than the sum of its parts, provided each new member is welcomed with the right training strategy.

Subject of Research: Federated learning strategies for improving computer-aided detection of cerebral aneurysms and brain metastases as institutions sequentially join a collaborative training network

Article Title: Performance changes in automated lesion detection under federated learning with sequential institution addition

Article References: Performance changes in automated lesion detection under federated learning with sequential institution addition. (n.d.). https://doi.org/10.1007/s11548-026-03795-w

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03795-w

Keywords: federated learning, computer-aided detection, cerebral aneurysm, brain metastasis, magnetic resonance imaging, fine-tuning, deep learning, FedAvg, catastrophic forgetting, medical imaging, multi-institutional data, privacy-preserving AI