Deep learning for large scale MRI-based morphological phenotyping of osteoarthritis
Nikan K Namiri1, Jinhee Lee1, Bruno Astuto1
1Department of Radiology and Biomedical Imaging and Center for Intelligent Imaging, University of California, San Francisco, 1700 Fourth St, Suite 201, QB3 Building, San Francisco, CA, 94107, USA.
Scientific Reports
|May 26, 2021
Summary
Artificial intelligence identifies knee osteoarthritis phenotypes from MRIs. These phenotypes predict disease progression and total knee replacement, aiding targeted osteoarthritis treatment development.
Area of Science:
- Orthopedics
- Radiology
- Artificial Intelligence
Background:
- Osteoarthritis (OA) pathogenesis is heterogeneous, lacking disease-modifying drugs.
- Current OA treatments lack specificity due to varied disease pathways.
- MRI-based phenotyping could enable targeted OA therapeutics.
Purpose of the Study:
- To train AI models to classify knee MRI phenotypes.
- To assess the association of these phenotypes with OA incidence and progression.
- To evaluate the utility of phenotyping for clinical trial stratification.
Main Methods:
- Convolutional neural networks trained on 4791 knee MRIs from the Osteoarthritis Initiative.
- Classification of bone, meniscus/cartilage, inflammatory, and hypertrophy phenotypes.
- Analysis of associations between phenotypes, structural/symptomatic OA, and total knee replacement.
Main Results:
- High classifier performance (AUCs 0.89-0.96) for all phenotypes.
- Bone and hypertrophy phenotypes predicted incident structural and symptomatic OA, respectively.
- All phenotypes except meniscus/cartilage predicted total knee replacement risk.
Conclusions:
- AI can effectively stratify knees into distinct structural phenotypes.
- These phenotypes correlate with OA incidence and progression.
- AI-driven phenotyping may optimize patient selection for OA clinical trials.
