Poster 242: Deep Learning for Identifying Patellofemoral Measurements Associated with Cartilage Lesions on MRI

OBJECTIVES: Magnetic resonance imaging (MRI) enables patellofemoral joint (PFJ) geometric measurements that may guide metrics of operative treatments, such as anteromedialization osteotomy, offloading stress on PFJs with chondral defects. However, utilization of MRI is limited due to expense required in manual image annotation. Preliminary work in small cohorts has suggested differences in PFJ measurements involving the anterior tibia and trochlea among symptomatic and control patients. Recently, deep learning (DL) has been used to perform accurate, automated segmentation of knee MRI. Herein, our aim was to develop a DL algorithm to segment the bony structures in knee MRIs and determine the relationship between PFJ geometries and cartilage lesions in a large, longitudinal cohort. METHODS: Subjects: We obtained data from the Osteoarthritis Initiative (OAI), a multi-institutional study conducted between 2005 and 2006, consisting of 4796 participants aged 45 to 80 years (NCT#00080171, on ClinicalTrials.gov). Eligible participants had osteoarthritis (OA) or elevated risk of OA in at least one knee at baseline. Each participant was assessed yearly with questionnaires, radiographs, and MRI. MRI sequences: MRIs were obtained using 3T scanners (Siemens Trio, Germany) on both right and left knees. From the OAI database, we accessed 3D sagittal double echo steady-state volumes with the following parameters: resolution=0.365×0.456×0.7mm, field of view=14cm, repetition time/echo time=16.2/4.7ms, matrix=384×307×160, bandwidth=62.5kHz. Radiologist grading: A subset of knee MRIs from OAI were graded according to MRI Osteoarthritis Knee Score (MOAKS) as part of several previous studies and shared publicly. A centralized group performed the grading through direct supervision of two musculoskeletal radiologists with more than nine years of training in grading knee OA with semi-quantitative scoring systems. Clinical data and case-control status were not made available to the radiologists during grading. In total, 2653 unique participants received imaging at either or both of two visits (baseline and 4 years), resulting in gradings for 4413 knee MRIs from 3117 unique knees. Automated segmentation: We built a DL segmentation model to determine the locations of patella, femur, and tibia within the double echo steady-state MRIs. The model utilized a 3D V-Net architecture, and training was performed on 40 manually annotated MRIs. We then used the trained model to segment femur, tibia, and patella bone from the radiologist MOAKS-graded subset of the OAI. Geometric measurements: We measured the tibial tubercle-trochlear groove (TTTG) length, sagittal tibial tubercle-trochlear groove (sTTTG) length, trochlear sulcus angle, trochlear dysplasia, Caton-Deschamps index, and flexion angle using the bone segmentations of the MOAKS-graded cohort. The trochlea was chosen using the axial slice of the segmentation with the largest femur width. The tibial tubercle was chosen using the anterior-most point of the axial cut 13.8mm superior to the inferior-most axial cut of the segmentation. This axial cut of the tibia was chosen using the average axial location of the tibial tubercle in 15 randomly selected segmentations. Statistical analysis: Dice score determined accuracy of the segmentation model on the holdout set of manually segmented images. Kruskal-Wallis H-test compared differences in demographics, radiographic findings, and geometric measurements between subjects without and with PFJ OA. PFJ OA was defined as having a MOAKS score greater than 2 in one or more of the following articular surfaces of the PFJ: anterior medial femur, anterior lateral femur, medial patella, lateral patella. Controls with no PFJ OA were classified as MOAKS score less than or equal to 2 in all four PFJ articular surfaces. Two-tailed p-values less than 0.05 were considered statistically significant. Statistical analyses were performed in Python (version 3.6.5; Python Software Foundation, Beaverton, Ore); important packages included numpy, pandas, and scipy. RESULTS: Baseline demographics for all subjects were as follows: women=1868, men=1243, age (mean(SD))=62.7(8.9), BMI=28.7(4.8). Kellgren-Lawrence (KL) grades of the knees were KL0=1106, KL1=590, KL2=695, KL3=493, KL4=194. Dice scores of the model on the holdout test set were 97.2% (95% confidence interval (CI): 96.6-97.7%), 97.3% (95% CI: 96.6-97.9%), and 96.0% (95% CI: 95.2-96.7%) for femur, tibia, and patella, respectively. A total of 2051 knees did not have PFJ OA, while 1060 knees possessed PFJ OA (Table 1). Subjects with PFJ OA were older, had greater female predominance, greater BMI, greater KL scores, greater pain levels, and greater proportion of total knee replacement at 8 years. Knees without PFJ OA had a greater sTTTG (-8.2(5.7) mm) compared to subjects with PFJ OA (-9.8(5.8) mm) (p<0.001), indicating a more anteriorly-positioned tibial tubercle for those subjects without PFJ OA. There were no differences in TTTG (9.2(4.2) mm v. 9.4(4.9) mm, p=0.12) and sulcus angle (145.7(9.1) degrees v. 146.0(9.6) degrees, p=0.41) among knees without and with PFJ OA. CONCLUSIONS: We used DL to segment bony structures within knee MRIs and compare PFJ geometric measurements with presence of OA. Knees with PFJ OA were associated with a more posteriorly positioned tibial tubercle; however, we found no differences in TTTG and sulcus angle in knees with and without OA. This relationship of the tibial tubercle and trochlea in the sagittal plane is consistent with prior smaller case series, though the methodology of this study allows for evaluation of these measurements on a much broader scale in this large cohort. Understanding the geometric measurements of knees without chondral defects may serve as metrics for correctional chondral deformity operations.