Can AI predict pain progression in knee osteoarthritis subjects from structural MRI

Abstract

Purpose: Despite the widely perceived association between structural change and pain in knee OA, a direct relationship has not been well established. Recent studies suggest that there are distinct subgroups of knee osteoarthritis pain trajectories. While some patients suffer progressively worsening pain, others can stabilize the pain for a long-term period. Establishing a direct relationship between image-based biomarkers and pain progression will be valuable as a prognostic tool. The goal of this study consists of two parts: 1) to identify the distinct pain trajectories among knee OA patients and 2) to investigate the association between MRI image biomarkers learned using the 3D convolutional neural network and the identified pain trajectories. Methods: Datasets: A total of 4,796 subjects’ repeated measures of KOOS pain score for both knees over 10-year study were obtained from the Osteoarthritis Initiative. 3D Double Echo Steady State (DESS) images of the knee for the subjects at baseline were used for the image biomarker discovery. Parameters of sagittal 3D sequence are: TR/TE 16.2/4.7 ms, FOV=14cm, matrix size=307x348, bandwidth=248 Hz/pixel, image resolutions =0.346x0.346x0.7 mm. Data processing overall pipeline is described in Figure 1. Modeling pain trajectory: We temporally smoothed individual’s pain curve by fitting a regression model using the orthogonal polynomials of degree one and two as the basis in order to reduce the inherent noise in the dataset and to handle missing data. The estimated parameters were standardized to be used as input into the Bayesian Gaussian mixture model for clustering analysis. Silhouette approach was used to choose the optimal model that captures the distinct pain patterns best. The parameters from regression were re-fitted against the selected Gaussian mixture model’s means to obtain the posterior probabilities of each pain curve that describe the cluster membership. Deep learning architecture and training details: Our network is a 3D extension of the DenseNet 121 architecture. The training dataset images were augmented by randomly applying image operations including horizon flipping, translation (up to 18 pixels vertically and horizontally), and rotation ([-15, +15] degrees). In doing so, images those are most likely be the rare classes were boosted by applying the transformation nine times per image to limit the effect of class imbalance. The weights were initialized using the He initialization. The network was trained to learn the posterior probabilities of pain trajectory membership. Mean squared error function was used as the loss function for regression. Results: Pain trajectory clustering analysis: A total of 24 candidate models of varying number of components (2, 3, 4, 5, 6, 7) and different types of covariance (full, tied, diagonal, spherical) were optimized using the expectation-maximization algorithm. The Gaussian Mixture of three components with tied covariance, which had the highest mean silhouette score, was selected as the optimal clustering model. The clustering analysis based on the GMM model identified three distinct pain trajectories: stable, worsening, progressively worsening (Figure 2). The distribution of age (Cluster 1: 61.3 ±9.10, Cluster 2: 60.7 ±9.04, Cluster 3: 61.2 ±9.47) and gender (M/F Cluster 1: , Cluster 2: 442/712, Cluster 3:460/652) were very similar across the clusters, while the Kellgren-Lawrence grades (0/1/2/3/4, Cluster 1: 2491/1278/1775/588/146, Cluster 2:195/153/437/298/76, Cluster 3: 266/178/397/211/60) and OA status (Y/N: Cluster 1: 2,509/3,769, Cluster 2: 806/348, Cluster 3: 668/444) were more severe in the progressively worsening cluster. Training Results: The mean squared error scores were 0.0148, 0.1556, 0.1549 for training(n=5,470), validation(n=1,368), and test(n=1,710) set, respectively. The mean absolute error scores were 0.0174, 0.1661, 0.1654 for training, validation, and test set, respectively. The accuracy, which measures whether the network correctly predicts the most probable clusters, were 0.9853, 0.8041, 0.7830 for training, validation, and test set, respectively. Pain curve prediction: We evaluated the model’s performance by comparing the simulated pain trajectories with the posterior probabilities of fitted GMM to the ones with the posterior probabilities the deep learning learned from MRI images. We simulated 1,000 sets of a random sample to obtain the confidence intervals and measured the overlap over the union of the intervals. The results were 0.8489, 0.5935, 0.5991 for the train, validation, test set, respectively. Figure 3 shows some examples of DL performed well in predicting future pain. Conclusions: We built a deep learning model that relates automatic learned MRI imaging biomarkers to temporal information of the knee pain. With our design we can provide, not only the point estimate but also the uncertainty incorporated into the problem. As the analysis was focused on the relationship between image biomarker and pain, other clinical and demographic patient information was not examined. We plan to expand our model to consider other factors that can impact the OA pain management.View Large Image Figure ViewerDownload Hi-res image Download (PPT)View Large Image Figure ViewerDownload Hi-res image Download (PPT)