Diagnosing osteoarthritis from T2 maps using deep learning: an analysis of the entire Osteoarthritis Initiative baseline cohort

Abstract

Objective: We aim to study to what extent conventional and deep-learning-based T 2 relaxometry patterns are able to distinguish between knees with and without radiographic osteoarthritis (OA). Methods: T 2 relaxation time maps were analyzed for 4,384 subjects from the baseline Osteoarthritis Initiative (OAI) Dataset. Voxel Based Relaxometry (VBR) was used for automatic quantification and voxelbased analysis of the differences in T 2 between subjects with and without radiographic OA. A Densely Connected Convolutional Neural Network (DenseNet) was trained to diagnose OA from T 2 data. For comparison, more classical feature extraction techniques and shallow classifiers were used to benchmark the performance of our algorithm's results. Deep and shallow models were evaluated with and without the inclusion of risk factors. Sensitivity and Specificity values and McNemar test were used to compare the performance of the different classifiers. Results: The best shallow model was obtained when the first ten Principal Components, demographics and pain score were included as features (AUC ¼ 77.77%, Sensitivity ¼ 67.01%, Specificity ¼ 71.79%). In comparison, DenseNet trained on raw T 2 data obtained AUC ¼ 83.44%, Sensitivity ¼ 76.99%, Specificity ¼ 77.94%. McNemar test on two misclassified proportions form the shallow and deep model showed that the boost in performance was statistically significant (McNemar's chi-squared ¼ 10.33, degree of freedom (DF) ¼ 1, P-value ¼ 0.0013). Conclusion: In this study, we presented a Magnetic Resonance Imaging (MRI)-based data-driven platform using T 2 measurements to characterize radiographic OA. Our results showed that feature learning from T 2 maps has potential in uncovering information that can potentially better diagnose OA than simple averages or linear patterns decomposition.

T2 analysis of the entire osteoarthritis initiative dataset

“…Our predictive models showed that T tuning the hyperparameter for the tree depths can be performed. 33 In reality, building highly accurate predictive and diagnostic models for OA requires rigorous feature engineering methods 34,35 work which is beyond the scope of this paper. We limited this initial study to the T 2 analysis of average values across compartments, however, previous studies reported that spatial assessment of the knee cartilage relaxation times using laminar and sub-compartmental analyses could lead to better and probably earlier identification of cartilage matrix abnormalities 36,37 Extraction of second-order statistical information or texture analysis 38,39 has been widely used to overcome the limitation of the average-based approaches.…”

“…Voxelbased relaxometry techniques have also been previously proposed 40 ; this technique allows for the investigation of local cartilage composition differences, through voxel-based statistics as statistical parametric mapping, 41 principal component analysis, 42 or deep learning feature extractions. 34 TKR is the most effective intervention for end-stage OA 12 ;…”

Improving Quantitative Magnetic Resonance Imaging Using Deep Learning

“…One pioneer study by Pedoia et al investigated the diagnostic difference of distinguishing knees with and without radiographic OA using conventional and deep learning-based T2 relaxometry measures. 58 In their study, the T2 maps from 4,384 subjects from the baseline OAI data sets were extracted and analyzed. A deep learning method using DenseNet 59 was trained to diagnose OA directly from the T2 maps with a training, validation, and holdout testing set of a 65%-20%-15% split of the entire data sets.…”

Deep Learning for Lesion Detection, Progression, and Prediction of Musculoskeletal Disease

“…Using a hold-out testing group consisting of 83 pelvic radiographs, the machine had an AUC of 0.94 for determining the presence (KL grades 0 and 1) or absence (KL grades 2, 3, and 4) of radiographic OA with a sensitivity and specificity of 95% and 91%, respectively. 50 A recent study by Pedoia et al 51 used a deep-learning method combined with voxel-based relaxometry for the analysis of T 2 relaxation time maps to determine the presence or absence of radiographic knee OA using the interpretation of experienced radiologists as the reference standard. A shallow random forest classifier model trained on handcrafted features consisting of average cartilage T 2 values on different articular surfaces of the knee joint was compared to a DenseNet classification CNN model trained on the raw T 2 data.…”