EARLY DETECTION OF KNEE OSTEOARTHRITIS USING DEEP LEARNING ON KNEE MRI

Abstract

Majority of the previous studies using deep learning approaches to predict knee OA incidence have used a radiography-based outcome variable. However, an MRI-based outcome variable could provide a more comprehensive view of OA. The aim of this study was to investigate the ability of three different deep learning algorithms to predict MRI-based knee OA incidence within 24 months from MR images. The intermediate-weighted turbo spin echo (IW-TSE) sequence of 593 OAI participants was used to predict knee OA incidence. Knee OA was defined using the MRI OA knee score (MOAKS) features except for osteophytes which were evaluated on radiographs. The IW-TSE MRIs were cropped to a region with a size of 250 × 320 voxels containing the knee joint. The position of the cropping box was defined as follows. First, a 2D U-Net model was used to segment tibial and femoral bone on a DESS sequence. The U-Net model was trained using the manual segmentations of femoral and tibial bone of 507 knees from the OAI-ZIB dataset. The performance of the segmentation algorithm was evaluated using the Dice similarity coefficient. Subsequently, IW-TSE and DESS sequences were registered using the Elastix software and the DESS segmentations were transformed to the corresponding IW-TSE scans. Finally, the coordinates of the tibial and femoral bones were used to define the position of the cropping box. Three different deep learning architectures using 3D MRI data as input were used to extract features from the cropped IW-TSE scans: a residual network (ResNet-50), a densely connected convolutional network (DenseNet-121), and a convolutional variational autoencoder (CVAE). Several combinations of batch size, learning rate, regularization, and data augmentation were tested. Furthermore, different combinations of discriminative penalties and latent space dimensions were tested for CVAE. 70% of the knees were included in the training set, 15% in the validation set, and 15% in the holdout test set. For comparison, a logistic regression model was trained using only age, gender, and BMI of the participants. Furthermore, a logistic regression model that combined the aforementioned basic variables and the image features was trained. Performance of the basic model (age, gender, BMI), image features model, and basic+image features model to predict knee OA incidence was assessed using the area under the receiver operating characteristic curve (ROC AUC) and the precision-recall curve (PR AUC) metrics using Python. The Dice similarity coefficients for the automatic segmentation of tibial and femoral bones were 0.985 and 0.987, respectively. The basic model had an ROC AUC of 0.639 and an PR AUC of 0.364 to predict knee OA incidence in the test set. Of the models based on the images only, CVAE was the best performing model according to the ROC AUC values (ROC AUC: 0.669, PR AUC: 0.346). Of the models combining basic variables and images, the basic + CVAE model was the best performing model according to the ROC AUC values (ROC AUC: 0.670, PR AUC: 0.359). When combined with the basic model, the best performing ResNet-50 model had an ROC AUC of 0.651 and an PR AUC of 0.379 and the best performing DenseNet-121 model had an ROC AUC of 0.656 and PR AUC of 0.371. According to the PR AUC values, the CVAE model that included kernel regularization and had lower number of latent space dimensions than the CVAE model with the highest ROC AUC, was the best performing model (PR AUC: 0.416, ROC AUC: 0.601). CVAE deep learning models had the highest performance to predict OA incidence using IW-TSE MRIs. Inclusion of age, sex, and BMI in the models improved the performance of the prediction algorithms. In general, the performance of all models was limited showing that the prediction of knee OA incidence is a complex problem. The impact of other MRI sequences, modalities, and data (e.g., genetics) will be investigated in the future.

Topics & Concepts