Subchondral bone length in knee osteoarthritis: A deep learning derived imaging measure and its association with radiographic and clinical outcomes

Gary H. Chang, Lisa K. Park, Nina A. Le, Ray Jhun, Tejus Surendran, Joseph H. C. Lai, Henry Hojoon Seo, Nuwapa Promchotichai, Grace H. Yoon, Jonathan Scalera, Terence D. Capellini, David T. Felson, Vijaya B. KolachalamamedRxiv 2021 · DOI

Introduction

Knee osteoarthritis (OA) is a leading cause of chronic pain and physical disability worldwide. Despite its prevalence, tracking the progression of the disease remains a significant clinical challenge. The current gold standard for assessment relies on radiographic Joint Space Narrowing (JSN), a metric that is often insensitive to early pathological changes and provides a limited view of the joint's three-dimensional health. While Magnetic Resonance Imaging (MRI) offers a more comprehensive view of soft tissues and bone morphology, the manual segmentation required to extract meaningful data is too labor-intensive for large-scale clinical application.

Furthermore, existing automated "bone shape" metrics often include marginal osteophytes (bone spurs). While these spurs are characteristic of OA, they do not always correlate well with a patient’s actual pain levels or the severity of cartilage loss. This study introduces a novel imaging biomarker—Subchondral Bone Length (SBL)—designed to focus specifically on the bone-cartilage interface, providing a more pathologically relevant measure of joint degeneration.

Methodology

To automate the measurement of SBL, the researchers developed a multi-stage deep learning pipeline applied to sagittal dual-echo steady state (DESS) MRI sequences from the Osteoarthritis Initiative (OAI) dataset.

The AI Pipeline

  1. Localization: A Fast Region-based Convolutional Neural Network (Fast R-CNN) was utilized to automatically identify and crop the knee joint area from the MRI slices, ensuring the subsequent models focused only on the relevant anatomy.
  2. Segmentation: A 2D U-Net architecture was trained to segment the femur, tibia, and their respective cartilage plates. To overcome the scarcity of expert-annotated data, the team employed an efficient strategy using Distance Regularized Level Set Evolution (DRLSE) to generate initial masks, which were then refined by experts.
  3. SBL Calculation: SBL is defined as the arc length of the cortical bone surface that is in direct contact with the overlying cartilage. Using a distance transform and skeletonization process, the team extracted this one-pixel-wide interface to calculate its precise length.

Validation and Application

The model was validated against an independent test set of 27 knees and then applied to a massive longitudinal cohort of 9,434 knees from the OAI. The researchers examined the relationship between SBL changes and radiographic JSN grades, as well as patient-reported outcomes via Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) scores.

Key Insights: The "Flattening" Signal

SBL is a dynamic measure influenced by two competing pathological processes: cartilage loss (which can shorten the measured interface) and bone flattening (which increases the surface area and lengthens the interface). The study’s findings revealed that as OA progresses, bone flattening is the dominant effect, leading to a net increase in SBL in the affected compartments.

By specifically measuring the bone-cartilage interface and excluding marginal osteophytes, SBL captures the "signature pathologic feature" of OA—the mechanical remodeling and deformation of the subchondral bone that occurs as the protective cartilage layer denudes. Furthermore, by mapping SBL across the medial-to-lateral axis, the researchers could create a localized "heat map" of joint degeneration, showing that SBL variations were most significant in the specific compartment where JSN was occurring.

Results and Clinical Significance

The automated system demonstrated high technical performance and strong clinical predictive power:

  • Technical Accuracy: The U-Net achieved Dice coefficients of approximately 0.90 for both cartilage and bone. Automated SBL measurements showed a high correlation ($R^2 > 0.90$) with manual radiologist annotations.
  • Predicting Surgical Outcomes: SBL proved to be a potent predictor of future total knee replacement (TKR). Over an 8-year follow-up, patients in the highest quartile of SBL change had significantly higher odds of undergoing surgery:
  • Medial JSN: Odds Ratio of 5.68 (95% CI: [3.90, 8.27])
    • Lateral JSN: Odds Ratio of 7.19 (95% CI: [3.71, 13.95])
  • Clinical Correlation: Increased SBL values were consistently associated with higher WOMAC scores for both pain and physical disability. For instance, the odds of experiencing severe pain increased roughly four-fold for patients in the highest SBL quartile.

Conclusion and Future Directions

Subchondral Bone Length represents a promising new imaging biomarker that quantifies OA status based on structural changes that matter to patients. By focusing on the bone-cartilage interface rather than total bone shape, SBL provides a clearer signal of disease severity and a more accurate prediction of clinical progression.

While the current 2D U-Net architecture is highly effective, future work may explore 3D architectures to capture the entire volume of the subchondral bone. Additionally, further research is needed to determine if SBL can detect "pre-radiographic" OA—identifying structural changes before they become visible on standard X-rays—potentially opening a window for earlier clinical intervention.

Summary updated

Abstract

ABSTRACT Objective Develop a bone shape measure that reflects the extent of cartilage loss and bone flattening in knee osteoarthritis (OA) and test it against estimates of disease severity. Methods A fast region-based convolutional neural network was trained to crop the knee joints in sagittal dual-echo steady state MRI sequences obtained from the Osteoarthritis Initiative (OAI). Publicly available annotations of the cartilage and menisci were used as references to annotate the tibia and the femur in 61 knees. Another deep neural network (U-Net) was developed to learn these annotations. Model predictions were compared with radiologist-driven annotations on an independent test set (27 knees). The U-Net was applied to automatically extract the knee joint structures on the larger OAI dataset (9,434 knees). We defined subchondral bone length (SBL), a novel shape measure characterizing the extent of overlying cartilage and bone flattening, and examined its relationship with radiographic joint space narrowing (JSN), concurrent WOMAC pain and disability as well as subsequent partial or total knee replacement (KR). Odds ratios for each outcome were estimated using relative changes in SBL on the OAI dataset into quartiles. Result Mean SBL values for knees with JSN were consistently different from knees without JSN. Greater changes of SBL from baseline were associated with greater pain and disability. For knees with medial or lateral JSN, the odds ratios between lowest and highest quartiles corresponding to SBL changes for future KR were 5.68 (95% CI:[3.90,8.27]) and 7.19 (95% CI:[3.71,13.95]), respectively. Conclusion SBL quantified OA status based on JSN severity. It has promise as an imaging marker in predicting clinical and structural OA outcomes.
BibTeX
@misc{chang2021subchondral,
  author = {Gary H. Chang and Lisa K. Park and Nina A. Le and Ray Jhun and Tejus Surendran and Joseph H. C. Lai and Henry Hojoon Seo and Nuwapa Promchotichai and Grace H. Yoon and Jonathan Scalera and Terence D. Capellini and David T. Felson and Vijaya B. Kolachalama},
  title = {{Subchondral bone length in knee osteoarthritis: A deep learning derived imaging measure and its association with radiographic and clinical outcomes}},
  year = {2021},
  doi = {10.1101/2021.04.28.21256271},
  howpublished = {medRxiv},
}
Figure