Comparison of Evaluation Metrics of Deep Learning for Imbalanced Imaging Data in Osteoarthritis Studies

Objective To compare the evaluation metrics for deep learning methods in the imbalanced imaging data in osteoarthritis (OA) studies. Method We first divided MOAKS (MRI Osteoarthritis Knee Score) grades into the presence (MOAKS > 0) and absence (MOAKS = 0) categories. Second, a deep-learning model was trained to the sagittal intermediate-weighted (IW) fat-suppressed (FS) knee MRI images with MOAKS readings from the Osteoarthritis Initiative (OAI) study to predict the presence of bone marrow lesions (BMLs). After the deep learning models were trained, we obtained probabilities of the presence of BMLs from MRI images at the sub-region (15 sub-regions), compartment, and whole-knee levels. We compared different evaluation metrics (e.g., receiver operating characteristic (ROC) and precision-recall (PR) curves) of the deep learning model in the testing data with various class ratios (presence of BMLs vs. absence of BMLs) at these three data levels to assess the model’s performance. Results We have demonstrated that the commonly used ROC curve is not sufficiently informative when evaluating the performance of deep learning models in the imbalanced data in OA studies. Conclusion The class ratios coupled with results of ROC, PR, and Matthews correlation coefficient (MCC) should be reported in OA studies.