Enhancing Knee Disease Diagnosis via Multi-View Graph Representation With Multi-Task Pre-Training.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Graph Neural Networks

Background:

• Magnetic resonance imaging (MRI) is crucial for knee examination, typically involving 2D slices from multiple views.
• Current computer-aided diagnosis struggles with fusing multi-view information, impacting performance and data requirements.
• Radiologists manually integrate information across views for diagnosis, a process that can be time-consuming and prone to error.

Purpose of the Study:

• To develop a novel framework for improved multi-view MRI analysis in knee diagnosis.
• To introduce the Knee Graph Network (KGNet) for enhanced identification and fusion of local regions in knee scans.
• To leverage multi-task pre-training to boost the diagnostic capabilities of the proposed network.

Main Methods:

• Representing multi-view MRI scans as a unified knee graph structure.
• Developing and implementing the Knee Graph Network (KGNet) for graph-based diagnosis.
• Employing multi-task pre-training involving masked local patch reconstruction and segmentation tasks.

Main Results:

• The KGNet framework demonstrated superior performance in diagnosing cartilage defects, anterior cruciate ligament tears, and general knee abnormalities.
• The proposed method outperformed existing approaches on both public and in-house clinical datasets.
• The graph representation effectively fuses information from multiple MRI views.

Conclusions:

• The novel framework effectively enhances knee disease diagnosis by utilizing graph representation and multi-task pre-training.
• KGNet shows significant potential for improving the accuracy and efficiency of radiological assessments of the knee.
• The integration of graph networks and advanced pre-training offers a promising direction for medical image analysis.