AUTOMATIC PREDICTION OF KNEE OSTEOARTHRITIS PROGRESSION THROUGH CARTILAGE T2 MAPPING

Longitudinal cartilage assessment using quantitative T2 maps from MRI can significantly contribute to a personalized Knee Osteoarthritis (OA) assessment and progression. Advancements in deep learning (DL) algorithms with improved computational power and efficient predictive capabilities have led to their use in automated medical image analysis. Here we propose a DL methodology focusing on automating the T2 mapping assessment for predicting knee OA progression using DL-based architectures. A multi-centre longitudinal MRI dataset of females with incidence knee OA over a period of four years was curated from a publicly available Osteoarthritis Initiative (OAI) ( https://nda.nih.gov/oai/ ) dataset for this study. Manual segmentation was performed to extract cartilage region for T2 Map computation using the standard mono-exponential curve fitting approach. DL models were trained and evaluated for this dataset consisting of three longitudinal data points: Baseline (00 Months), 24 th Month, and 48 th Month. The models were trained on 70% dataset, validated on 10% dataset, and tested on 20% dataset. DL models were used to predict the progression of cartilage degeneration using quantitative T2 maps. The predictive capabilities of two DL architectures (U-Net and Transformers) were compared to understand their clinical utility. Based on the cartilage region of interest, the pixel-wise predicted T2 maps were assessed using standard classification evaluation metrics of specificity, sensitivity, and accuracy. Transformer-based architecture showed promising results in comparison with U-Net architecture. To improve the outcome, better tuning of hyperparameters and a novel optimization approach will be analyzed. The current accuracy of the predictions does not provide clinical reliability due to a smaller number of datasets and only three longitudinal data points (each 24 months apart). We recommend testing the system with larger dataset including more time points to ensure reliability. Since cartilage properties are sensitive to subtle changes during the onset of knee OA, this proposed methodology may contribute to the early detection of cases, with slow progression. This may fall under the scope of assessing various treatment aspects to investigate reversing knee OA in the early stages.