To examine the impact of progressively adding participating institutions under federated learning on lesion detection software performance.
Approach:
Datasets: The study utilized 423 MRA images for cerebral aneurysm detection and 398 contrast-enhanced T1-weighted MR images for brain metastasis detection, partitioned into four heterogeneous datasets.
CAD Algorithm: A 3D U-Net model was used for cerebral aneurysm detection, while a single-shot multibox detector (SSD) was employed for brain metastasis detection.
Federated Learning Protocol: The FL framework involved a server-client architecture, where clients trained models locally and shared parameters with a central server for aggregation.
Training Approaches: Three integration approaches were evaluated: from-scratch training, fine-tuning (FT), and full fine-tuning (FFT).
Client Addition Steps: The study involved initial training with two clients, followed by sequentially adding a third and then a fourth client, assessing various training configurations.
At-once Condition: An alternative 'At-once' condition was assessed where the remaining clients were added concurrently.
Centralized Reference Training: Centralized models were trained using pooled data from participating clients at each stage.
Key Findings:
Sequential addition of institutions improved CAD software performance.
Different training approaches (from-scratch, FT, FFT) yielded varying results.
The 'At-once' condition did not show significant performance differences compared to sequential addition.
Limitations:
The study was retrospective and limited to specific datasets.
Performance may vary with different types of CAD algorithms and imaging modalities.
Conclusion:
The study highlights the potential benefits of sequential institution integration in federated learning for improving CAD software performance, while also acknowledging the need for further research to validate findings across diverse datasets and clinical settings.
