Distributed deep learning across multisite datasets for generalized CT hemorrhage segmentation.

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Deep neural networks (DNNs) excel in computer vision but require large datasets for generalization.
• Medical imaging datasets are often limited in size and difficult to share due to privacy concerns (protected health information - PHI).
• Single-site learning (SSL) can lead to overfitting and poor generalization on external data.

Purpose of the Study:

• To implement and evaluate a multi-site learning (MSL) approach using cyclic weight transfer for medical imaging.
• To assess if MSL can improve model generalization without data consolidation or compromising PHI.
• To compare the performance of MSL against single-site learning (SSL) models.

Main Methods:

• Implemented cyclic weight transfer for training DNNs on independent datasets from geographically disparate sites.
• Trained both single-site learning (SSL) and multi-site learning (MSL) models.
• Evaluated models on training sites and two external institutions.

Main Results:

• The MSL model achieved a higher average Dice Similarity Coefficient (DSC) of 0.690 compared to SSL models (0.646).
• MSL demonstrated a statistically significant improvement in DSC (7%) and volume correlation (5%) on holdout datasets.
• MSL models showed improved generalization capabilities on unseen data from external institutions.

Conclusions:

• Cyclic weight transfer enables efficient training of neural networks across multiple sites without data sharing.
• MSL significantly improves model generalization and performance on external medical imaging datasets.
• This federated learning approach is a viable solution for training robust medical imaging AI models.