PREDICTION OF OSTEOPHYTE PROGESSION FROM CLINICAL DATA AND MRI-BASED 3D KNEE MORPHOLOGY

Abstract

Osteophytes (OSTs) are one of the most prominent radiographical features of knee OA progression. They are strongly associated with presence of pain and often with other symptoms. However, clinical picture of OSTs in OA and their pathophysiology are not well understood. Prior research on predicting OA progression, particularly from the imaging data, has considered TKR, change in KLG, or JSN as the surrogate outcomes. Thus, the feasibility of using individual radiographic features, such as OSTs, as a potential surrogate endpoint has not been investigated. 1) To investigate the predictive power of the commonly available demographic and radiographic variables in prediction of osteophyte progression over 4 years. 2) To develop an MRI-based (3D tissue morphology) end-to-end model for the same task and evaluate its performance and added value. We used the data from the baseline of the OAI. Progression criterion was based on radiographic OARSI grade for OSTs [0-3]. Knees with an increase in OARSI grade in any compartment at the 4-year follow-up were set as progressors – the rest – as non-progressors. We excluded the knees with OARSI grade 3 in all regions, the non-progressors that dropped out before the last visit, and the samples with missing data. Final sample had 4600 knees (2476 subjects), with 2218 progressed and 2382 non-progressed knees. The baseline models were based on Logistic Regression and the following variables: age, sex, BMI, KLG, and OARSI OST grades. The latter were available for FL, FM, TL, and TM. For MRI-based model, we used the validated automatic segmentations by Tack et al. 2021 (Zuse Institute, Berlin, Germany) derived from 3D DESS MRI. The data included the masks for femur, tibia, corresponding cartilage tissues, and menisci. The ROIs were cropped from the images and resampled to resolution of 95x192x140 (pixel spacing 1.1x0.73x0.73mm^3). The model was based on 3D ResNet-10 CNN and trained in two experiments – from bone masks only, and from all the listed tissues. Eventually, the MRI models were also fused with the baseline ones. Here, the predicted probabilities of the MRI model were used as an additional variable for the baseline model. To develop the predictive models, we used 5-fold cross-validation. For model evaluation, we used a hold-out set of 920 knees (443 progressors and 477 non-progressors). Performance of the models were assessed using area under the ROC curve (AUC), and average precision (AP). Over 4 years, OARSI OST grade changes of (1, 2, and 3) were observed, for (20.22, 2.85, 0.37) percent of knees in TM, (12.54, 2.13, 0.24) in TL, (12.82, 1.48, 0.30) in FL, and (17.28, 3.56, 0.76) in FM, respectively. The age, sex, BMI model yielded AUC=0.60[0.56-0.63], AP=0.56[0.51-0.61]. Adding KLG resulted in similar performance as with added OARSI OST grade. The baseline model with both grades was the best -AUC=0.68[0.65-0.71] and AP=0.67[0.62-0.71]. The MRI (CNN) models showed AUC=0.55[0.51-0.58], AP=0.53[0.48-0.57] (bones-only) and AUC=0.58[0.54-0.61], AP=0.55[0.50-0.60] (all tissues). Fusion of the best baseline and MRI models did not further improve the performance. The model based on the clinical data including KLG and OARSI OST grade showed the highest performance in predicting OST progression in 4 years. The imaging biomarkers, automatically extracted from 3D bone, cartilage, and menisci morphology, showed only limited predictive power and did not improve the final performance. Interestingly, the imaging model including cartilage- and menisci-based biomarkers showed higher metrics than bone-only, which may suggest contribution of soft tissue morphology to the progression of OSTs. Supported by the Academy of Finland (334771 & 334773), under the frame of ERA PerMed. Alexander Tack and ZIB for providing the data. CORRESPONDENCE ADDRESS: mitra.daneshmand@oulu.fi