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 T₂ relaxometry patterns are able to distinguish between knees with and without radiographic osteoarthritis (OA).

Methods: T₂ 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 voxel-based analysis of the differences in T₂ between subjects with and without radiographic OA. A Densely Connected Convolutional Neural Network (DenseNet) was trained to diagnose OA from T₂ 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 = 77.77%, Sensitivity = 67.01%, Specificity = 71.79%). In comparison, DenseNet trained on raw T₂ data obtained AUC = 83.44%, Sensitivity = 76.99%, Specificity = 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 = 10.33, degree of freedom (DF) = 1, P-value = 0.0013).

Conclusion: In this study, we presented a Magnetic Resonance Imaging (MRI)-based data-driven platform using T₂ measurements to characterize radiographic OA. Our results showed that feature learning from T₂ maps has potential in uncovering information that can potentially better diagnose OA than simple averages or linear patterns decomposition.

Keywords: Convolutional neural network; Deep learning; Quantitative MRI; T(2) relaxation times; Voxel based relaxometry.

Figure 1:

Experimental design overview.

Figure 2:

Example of the 2D flatten T₂ map used as input to the convolutional neural network

Figure 3:

Design of the Densely connected neural network used for the OA classification from T₂ relaxation times maps.

Figure 4:

Bland-Altman and correlation plots showing a comparison between manual and automated average T₂ relaxation time computed for 1799 cases in the OAI dataset.

Figure 5

Voxel-based statistical parametric map analysis of the baseline OAI dataset in distinguish subjects with and without sign of OA average and standard deviation maps are shown for OA and controls. T₂ average prolongation observed in OA subjects and p-value map are also shown (N=1937). The maps show just voxel that reaches significance after adjustment for multiple comparison. (A) VBR analysis showed in a representative lateral slice. (B) VBR analysis showed in a representative medial slice.

Figure 6:

(A) ROC curves comparing the Random Forest results between different feature combination. (B) Comparison of the best performant shallow classifier with the deep learning model.

Figure 7

(A) Modeling of the most significant T₂ relaxometry patterns associated with radiographic OA. Subjects with KL>1 exhibit a decreased difference between superficial and deep layer of the cartilage. (B) Modeling of the first Principal component which describe the most variation in the dataset and it is related with global T₂ averages but was not the first contributor for the OA vs Control distinction.