Related Experiment Video
Updated: Jun 30, 2026
11:29
Real-time Video Projection in an MRI for Characterization of Neural Correlates Associated with Mirror Therapy for Phantom Limb Pain
Published on: April 20, 2019
MedNet-FS: a few-shot learning framework for 3D MRI-based knee injury classification
Xu Lu1, Hongming Lin1, Shanhua Sun2
1Department of Orthopedics, Xiangyang No.1 People's Hospital, Hubei University of Medicine, Xiangyang, Hubei, 441000, China.
Scientific Reports
|June 1, 2026
Summary
Deep learning for knee MRI analysis is limited by data scarcity. This study introduces MedNet-FS, a few-shot learning framework achieving competitive ACL tear detection with minimal data, outperforming generic methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning for 3D knee MRI analysis requires large annotated datasets, which are scarce.
- Few-shot learning (FSL) shows potential for data-limited scenarios but is underexplored in volumetric medical imaging.
Purpose of the Study:
- To introduce MedNet-FS, a 3D FSL framework for knee MRI analysis in data-scarce environments.
- To evaluate the effectiveness of domain-specific pre-training and a Generalized End-to-End (GE2E) loss for enhancing FSL performance.
Main Methods:
- Developed MedNet-FS, a 3D FSL framework integrating domain-specific pre-training on knee MRI data.
- Utilized a Generalized End-to-End (GE2E) loss function.
- Evaluated performance on internal (MRNet) and external (KneeMRI) datasets for ACL tear detection.
Main Results:
- MedNet-FS significantly outperformed models with generic pre-training or standard cross-entropy loss.
- Achieved an AUC of 0.76 for ACL tear detection on MRNet using only 40 samples per class.
- Demonstrated generalizability on KneeMRI (AUC 0.62 for clear cases) but showed reduced performance on ambiguous partial tears (AUC 0.58).
Conclusions:
- The combination of domain-specific pre-training and GE2E loss is crucial for effective FSL in knee MRI.
- MedNet-FS provides a practical, scalable framework reducing annotation dependency for medical image analysis.
- While not yet suitable for autonomous clinical use, it serves as a strong baseline for data-efficient deep learning in radiology.
