Automatic hip osteoarthritis grading with uncertainty estimation from computed tomography using digitally-reconstructed radiographs

Masachika Masuda, Mazen Soufi, Yoshito Otake, Keisuke Uemura, Sotaro Kono, Kazuma Takashima, Hidetoshi Hamada, Yi Gu, Masaki Takao, Seiji Okada, Nobuhiko Sugano, Yoshinobu Sato

Introduction

Diagnosing the severity of hip osteoarthritis (OA) is a nuanced clinical task essential for determining surgical interventions, such as total hip arthroplasty. Traditionally, orthopedic surgeons rely on two primary grading systems: the Kellgren-Lawrence (KL) grade, which measures joint space narrowing and osteophyte formation, and the Crowe grade, which assesses the degree of femoral head dislocation. Despite their widespread use, these assessments are notoriously subjective, leading to significant inter-observer variability.

While deep learning has demonstrated potential in medical imaging, many existing models simplify the diagnostic process into a binary "diseased vs. healthy" classification. Such binary outputs lack the granularity required for complex surgical planning. A recent study by Masuda et al. (2023) addresses this by introducing an automated pipeline that grades the full spectrum of hip OA progression and provides an estimate of the model's confidence in its predictions.

Methodology

Digitally Reconstructed Radiographs (DRRs)

To bridge the gap between 3D CT data and 2D radiographic grading standards, the researchers developed a workflow to generate Digitally Reconstructed Radiographs (DRRs). By automatically detecting femoral head centers within 3D CT volumes, the system projects these regions into 2D images. This approach allows the model to benefit from the high-resolution structural information of CT scans while utilizing the familiar visual features used in standard KL and Crowe grading.

Model Architectures and Grading Schemes

The study benchmarked three prominent computer vision architectures: VGG16, DenseNet161, and the Vision Transformer (ViT). The researchers explored two distinct grading strategies:

  1. Combined: A single 7-class ordinal output representing the continuous progression of the disease.
  2. Separated: Two independent prediction heads for Crowe and KL grades.

Uncertainty Estimation

To address the "black box" nature of deep learning, the authors implemented Monte-Carlo (MC) Dropout to quantify epistemic uncertainty. By keeping dropout layers active during inference and running the model 50 times for each image, the system calculates the variance of the predictions. High variance serves as a proxy for model uncertainty, potentially flagging difficult cases for human review.

Experimental Setup and Datasets

Data Sources

The models were trained and evaluated using an internal dataset of 394 unilateral DRR images from 197 patients at Osaka University Hospital. For independent validation, an external dataset of 104 DRRs from 52 patients was utilized.

Ground Truth

Ground truth labels were provided by an orthopedic surgeon with six years of experience. The 7-class scale was constructed by mapping the biological progression of the disease, combining Crowe (1–4) and KL (1–4) grades into a logical ordinal sequence ranging from mild joint narrowing to severe dislocation.

Key Results

Performance Metrics

The models achieved an Exact Class Accuracy (ECA) of approximately 0.65. However, the One-Neighbor Class Accuracy (ONCA) reached approximately 0.95. This indicates that even when the model misclassified an image, the prediction was almost always within one grade of the expert's assessment, suggesting high clinical relevance.

Architecture Comparison

The Vision Transformer (ViT) emerged as the superior architecture, yielding the lowest uncertainty levels and the smallest regression error (0.383). The authors attribute this to the ViT’s attention mechanism, which effectively captures long-range spatial dependencies—such as the relationship between the pelvis and the femoral head—more effectively than standard convolutional layers.

Uncertainty as a Safety Mechanism

A critical finding was the statistical correlation between model uncertainty and prediction accuracy ($P < 6e-3$). Misclassified cases consistently exhibited higher uncertainty scores. In a clinical workflow, this allows the AI to act as a triage tool: automatically grading high-confidence cases while alerting surgeons to "uncertain" cases that require closer inspection.

Discussion and Future Work

While the results are promising, the study identifies several areas for improvement. The dataset exhibited a natural imbalance, with fewer cases of severe dislocation (Crowe 4), leading to lower accuracy in high-severity classes. Additionally, while DRRs provide a "clean" representation derived from CT, future validation is required on conventional X-ray radiographs, which remain the clinical gold standard due to lower cost and radiation exposure.

The authors plan to expand their database to over 2,000 cases to improve the grading of rare deformities and further refine the uncertainty thresholds. By moving beyond binary classification and incorporating self-knowledge of error, this framework represents a significant step toward trustworthy AI in orthopedic surgery.

Summary updated

BibTeX
@article{masuda2024automatic,
  author = {Masachika Masuda and Mazen Soufi and Yoshito Otake and Keisuke Uemura and Sotaro Kono and Kazuma Takashima and Hidetoshi Hamada and Yi Gu and Masaki Takao and Seiji Okada and Nobuhiko Sugano and Yoshinobu Sato},
  title = {{Automatic hip osteoarthritis grading with uncertainty estimation from computed tomography using digitally-reconstructed radiographs}},
  journal = {Int. J. Comput. Assist. Radiol. Surg.},
  year = {2024},
  doi = {10.1007/s11548-024-03087-1},
}
Figure