Abstract
Current methods for assessing knee osteoarthritis (OA) do not provide comprehensive information to make robust and accurate outcome predictions. Deep learning (DL) risk assessment models were developed to predict the progression of knee OA to total knee replacement (TKR) over a 108-month follow-up period using baseline knee MRI. Participants of our retrospective study consisted of 353 case-control pairs of subjects from the Osteoarthritis Initiative with and without TKR over a 108-month follow-up period matched according to age, sex, ethnicity, and body mass index. A traditional risk assessment model was created to predict TKR using baseline clinical risk factors. DL models were created to predict TKR using baseline knee radiographs and MRI. All DL models had significantly higher (p < 0.001) AUCs than the traditional model. The MRI and radiograph ensemble model and MRI ensemble model had the highest AUCs of 0.90 (80% sensitivity and 85% specificity) and 0.89 (79% sensitivity and 86% specificity), respectively, which were significantly higher (p < 0.05) than the AUCs of the radiograph and multiple MRI models. DL models using baseline MRI had a higher diagnostic performance for predicting TKR than a traditional model using baseline clinical risk factors and a DL model using baseline knee radiographs.
Introduction
Knee osteoarthritis (OA) is a leading cause of disability worldwide, yet predicting its progression to end-stage disease remains a significant clinical challenge. Currently, clinicians rely on clinical risk factors such as BMI and pain scores, or standard radiographs, which often fail to capture the early structural changes within the joint. While Magnetic Resonance Imaging (MRI) provides a comprehensive view of cartilage, bone, and ligaments, manual analysis is time-consuming and prone to human error. There is a critical need for automated, objective methods to identify high-risk patients likely to progress to total knee replacement (TKR) to facilitate early intervention and more efficient clinical trials.
Methods
The researchers developed several Deep Learning (DL) models based on 3D Convolutional Neural Networks (CNNs) to predict TKR up to nine years in advance. The approach utilized three specific MRI contrasts: Coronal intermediate-weighted (IW-TSE), Sagittal fat-suppressed (FS-IW-TSE), and 3D sagittal dual-echo steady-state (DESS).
Three main modeling strategies were compared:
- Individual Models: Training a CNN on a single MRI sequence.
- Multi-input Models: A single architecture processing multiple sequences simultaneously.
- Ensemble Models: Averaging risk predictions from several independent models (e.g., MRI + Radiographs).
The models were trained and validated using the Osteoarthritis Initiative (OAI) dataset (706 subjects) and externally tested on the Multicenter Osteoarthritis Study (MOST) dataset to evaluate real-world performance across different imaging protocols.
Key Results
The DL models demonstrated significantly superior performance compared to traditional clinical risk assessments:
- Superior Accuracy: The MRI and radiograph ensemble achieved the highest diagnostic performance with an AUC of 0.90.
- Beating the Baseline: All DL models significantly outperformed the traditional clinical risk model, which managed an AUC of only 0.77.
- Individual Winner: Among standalone MRI sequences, the DESS model performed best (AUC 0.88), likely due to its high-resolution 3D volumetric data.
- High Odds Ratio: The MRI ensemble had an adjusted odds ratio of 10.81, meaning patients with higher risk scores are nearly 11 times more likely to undergo a knee replacement.
Key Insights and Ideas
A key finding was that the "Ensemble" approach outperformed the "Multi-input" approach, suggesting that different MRI sequences capture independent prognostic information that can be obscured if fed into a single network too early. Furthermore, the AI's analysis of joint structure was a much stronger predictor of surgery than traditional clinical factors.
To understand the model's decision-making, the team generated saliency maps. These maps revealed that the AI focused on peripheral bone-cartilage interfaces (where osteophytes form) and the medial meniscus in patients who eventually required surgery. Conversely, in healthy controls, the AI focused on the central bone-cartilage interface, likely identifying healthy, intact cartilage.
Limitations and Future Work
Performance dropped when tested on the external MOST dataset (AUCs ~0.71), highlighting the challenge of "domain shift" across different hospital systems and scanner strengths. Additionally, the decision to undergo TKR is subjective and influenced by non-structural factors like pain tolerance and healthcare access. Future work will focus on attention-based fusion strategies and training on more diverse datasets to improve generalizability.
Summary updated