AI and Arthritis Imaging: The Path to Automation

Background:
Magnetic Resonance Imaging (MRI) plays a critical role in evaluating joint health and tracking disease progression in arthritis, especially osteoarthritis (OA). Traditional MRI scoring systems such as WORMS, MOAKS, and BLOKS help quantify joint abnormalities but require time-intensive interpretation by experts, often introducing variability. With nearly 600 million people affected by OA globally, there is growing interest in using artificial intelligence (AI) to automate these MRI assessments, aiming to enhance efficiency, objectivity, and scalability.
This systematic literature review investigated the feasibility of automating quantitative and semi-quantitative (Q/SQ) MRI scoring systems in arthritis. A total of 129 randomized controlled trials published between 2014 and 2024 were reviewed, with a strong focus on knee OA. Researchers evaluated which MRI biomarkers—such as cartilage loss, effusion/synovitis, bone marrow edema (BME), erosions, osteophytes, and meniscal or labral injuries—are most amenable to automation. The analysis was guided by expert radiologists and AI specialists who assessed technical complexity, reproducibility, and suitability for machine learning models.

What Does Previous Literature Tell Us?
Convolutional Neural Network-Based Automated Scoring Systems for Inflammatory Arthritis2
A particularly relevant 2025 study published in RMD Open developed a convolutional neural network (CNN) for automated scoring of hand MRI in rheumatoid arthritis (RA) and psoriatic arthritis (PsA). The researchers trained and validated their AI system using the OMERACT-validated RA MRI Scoring System and PsA MRI Scoring System. Their model achieved impressive diagnostic accuracy:
• High mean macro-AUC of 92%±1% for erosions
• 91%±2% for osteitis (bone marrow edema)
• 85%±2% for synovitis
The system demonstrated strong correlation with human annotation, achieving Spearman correlations of 90±2% for erosions, 78±8% for osteitis, and 69±7% for synovitis. Importantly, this CNN-based approach achieved these results using fewer MRI sequences than conventional scoring methods, highlighting its potential for more efficient clinical implementation

AI Models for Ultrasound Image Analysis in Osteoarthritis3
A 2024 study expanded AI applications to ultrasound imaging in hand osteoarthritis. This research developed an AI model for both segmentation of joint ultrasound images and severity scoring of osteophytes following the EULAR-OMERACT grading system. The system demonstrated strong agreement with physician assessments:
• 76% Percent Exact Agreement (PEA) and 97% Percent Close Agreement (PCA) for MCP joints
• 70% PEA and 97% PCA for PIP joints
• 59% PEA and 94% PCA for DIP joints
• 50% PEA and 82% PCA for CMC joints
Overall, the model achieved 68% PEA and 95% PCA across all joints, demonstrating AI’s potential beyond MRI analysis.

The Role of AI in Peripheral Joint MRI Assessment4
A systematic review and meta-analysis from 2022 provides additional context on the reliability and validity of computer-aided quantification in inflammatory arthritis. This analysis found excellent intra- and inter-reader reliability for:
• Bone erosion volume (BEV): r=0.97 (95%CI:0.92–0.99) and 0.93 (0.87–0.97)
• Synovial membrane volume (SV): r=0.98 (0.90–0.99) and 0.86 (0.78–0.91)
• DCE-MRI perfusion parameters: r=0.99 (0.82–1) for maximum enhancement
Importantly, meta-regression analysis showed that computer-aided and manual methods provide comparable reliability for quantifying bone erosion, synovitis, and perfusion parameters

Results:
• Scoring systems: WORMS and MOAKS were the most used for knee OA. Several in-house and emerging AI-integrated systems were also identified.
• Effusion/synovitis: High T2 contrast on MRI makes this biomarker readily automatable.
• Cartilage loss: A key OA biomarker but remains difficult to automate due to anatomical complexity and variable MRI quality.
• Bone marrow edema (BME): Potential for automation exists, though challenges remain in accurate lesion localization and segmentation.
• Erosions and osteophytes: Can be detected by AI with high-resolution imaging and training on standardized definitions.
• Meniscus/labrum tears: AI has shown good performance in detecting meniscal abnormalities; integration with cartilage and BME data may improve clinical insights.
• Fat metaplasia and fat pad changes: Readily visualized on T1-weighted MRI and considered feasible targets for AI-based analysis.

Conclusion:
This review highlights strong potential for AI to automate interpretation of MRI-based scoring in arthritis, especially for effusion/synovitis and meniscal tears. However, accurate cartilage and BME quantification remains a challenge. Existing scoring systems can serve as a foundation for training AI models, with future tools likely to enhance clinical workflow, standardization, and longitudinal disease tracking in arthritis management.

Editorial Comment:
An important step toward integrating AI into musculoskeletal MRI reporting, this review outlines the feasibility of automating complex semi-quantitative arthritis scores. With refined imaging and training datasets, AI holds promise in supporting clinical trials and patient care.

Teaching Points for Residents:
• AI can automate select MRI biomarkers in arthritis, including effusion and meniscal tears.
• Existing scoring systems (WORMS, MOAKS) provide a valuable framework for AI development.
• Cartilage loss and BME require further refinement for reliable automation.
• High-resolution MRI sequences enhance AI performance in osteophyte and erosion detection.
• AI can improve diagnostic consistency, enable large-scale analysis, and support personalized arthritis care.