Distributed deep learning networks among institutions for medical imaging.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Deep learning models require substantial data for high performance in clinical diagnosis.
• Sharing patient data across institutions is often hindered by privacy, legal, and ethical constraints.
• Collaborative approaches are essential when individual datasets are limited.

Purpose of the Study:

• To investigate the efficacy of distributing deep learning models as an alternative to sharing sensitive patient data.
• To compare the performance of distributed model training heuristics against a centrally trained model.
• To address the challenges of data scarcity in collaborative medical AI development.

Main Methods:

• Simulated distribution of deep learning models across four institutions.
• Investigated training heuristics: ensembling single models, single weight transfer, and cyclical weight transfer.
• Evaluated performance on image classification tasks using retinal fundus photos, mammography, and ImageNet datasets.

Main Results:

• Cyclical weight transfer demonstrated performance comparable to models trained on centrally pooled patient data.
• Increased frequency of weight transfer in cyclical weight transfer led to improved model performance.
• Distributed deep learning models proved effective across diverse medical imaging datasets.

Conclusions:

• Distributing deep learning models is a viable and effective strategy for collaborative medical AI research.
• This approach overcomes barriers associated with direct patient data sharing.
• Findings support the broader adoption of federated learning principles in clinical settings.