Federated knee injury diagnosis using few shot learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Knee injuries, including Anterior Cruciate Ligament (ACL) and meniscus tears, are prevalent and impact quality of life.
• Early diagnosis is crucial to prevent complications like osteoarthritis.
• Deep learning for MRI analysis faces challenges with scarce, privacy-sensitive labeled data.

Purpose of the Study:

• To analyze a hybrid few-shot and federated learning methodology for knee injury diagnosis from MRI scans.
• To address the limitations of data scarcity and patient privacy in medical image analysis.
• To enhance the diagnostic accuracy and generalization of knee injury detection models.

Main Methods:

• A hybrid framework combining centralized and federated few-shot learning was developed.
• A 3DResNet50 architecture was employed for feature extraction and embedding.
• Episodic-intermittent training with Prototypical Networks, SGD, Cross-Entropy Loss, and a MultiStep Learning Rate scheduler was utilized.

Main Results:

• The model demonstrated strong performance on the MRNet dataset across different MRI views.
• Centralized setting accuracies: 85.3% (axial), 82.1% (sagittal), 71% (coronal).
• Federated setting accuracies: 83% (axial), 83.9% (sagittal), 65% (coronal).

Conclusions:

• The hybrid approach effectively diagnoses knee injuries, improving accuracy and generalization while preserving patient privacy.
• Federated learning mitigates privacy concerns associated with distributed medical data.
• Limitations include lower performance on coronal views and significant computational requirements.