Assessment of knee pain from MR imaging using a convolutional Siamese network

Area of Science:

• Orthopedics
• Radiology
• Artificial Intelligence

Background:

• Characterizing knee joint pain sources via radiographs or MRI is challenging.
• Advanced machine learning offers potential for improved diagnostic capabilities.

Purpose of the Study:

• To determine if deep neural networks can distinguish knees with pain from those without.
• To identify structural features associated with knee pain using MRI.

Main Methods:

• A convolutional Siamese network was developed to analyze MRI scans from the Osteoarthritis Initiative (OAI).
• The network learned from paired MRI slices of painful and non-painful knees.
• Class activation mapping (CAM) generated saliency maps to highlight pain-associated regions.

Main Results:

• The model achieved an Area Under Curve (AUC) of 0.808, improving to 0.853 when excluding non-discordant pain scores.
• Radiologist review confirmed effusion-synovitis in 86% of correctly predicted cases within highlighted regions.
• Deep learning successfully associated MRI findings with unilateral knee pain.

Conclusions:

• Deep learning provides a novel approach for assessing knee pain from MRI scans.
• This study demonstrates the feasibility of using AI to identify pain-related structural abnormalities in the knee.
• The findings pave the way for more accurate knee pain diagnosis and characterization.