Related Experiment Video
Updated: Jan 7, 2026
06:17
Augmented Reality Navigation-Guided Core Decompression for Osteonecrosis of Femoral Head
Published on: April 12, 2022
4.2K
Classification of Osteonecrosis of the Femoral Head Stage on Radiographic Images Using Deep Learning Techniques
Hyun Hee Lee1, Joeun Jeong2,3, Taehoon Shin2,3
1Department of Orthopedic Surgery, International St. Mary's Hospital, Catholic Kwandong University College of Medicine, Incheon 22711, Republic of Korea.
Bioengineering (Basel, Switzerland)
|December 30, 2025
Summary
This study introduces a deep learning method using X-rays to diagnose osteonecrosis of the femoral head (ONFH). The AI model effectively differentiates ONFH stages, offering an accessible alternative to MRI for early detection and disease management.
Area of Science:
- Orthopedics
- Radiology
- Artificial Intelligence
Background:
- Magnetic resonance imaging (MRI) is sensitive for early osteonecrosis of the femoral head (ONFH) but costly and inaccessible.
- Radiography is accessible but lacks sensitivity for early ONFH diagnosis.
- Accurate ONFH staging is crucial for timely intervention and treatment planning.
Purpose of the Study:
- To develop and validate a deep learning approach for classifying ONFH stages using radiographic images.
- To provide a more accessible and cost-effective method for early ONFH diagnosis and stage differentiation.
- To assess the performance of a radiograph-based normative modeling approach compared to MRI.
Main Methods:
- A dataset of 909 hip radiographs was utilized, with femoral heads segmented using a U-Net model (DSC of 0.977).
- A variational autoencoder (VAE) was trained on healthy femoral head images to establish a normative latent distribution.
- Osteonecrosis of the femoral head (ONFH) images were projected into the latent space to analyze Mahalanobis distance distributions across different disease grades.
Main Results:
- The deep learning model successfully differentiated healthy femoral heads from those with ONFH.
- Significant differences in Mahalanobis distance were observed between grade 0 and grades 2-4 ONFH.
- No significant difference was found between grade 0 and grade 1 ONFH, reflecting the subtlety of early radiographic changes.
Conclusions:
- The proposed radiograph-based deep learning method offers an accessible alternative for ONFH stage differentiation, especially in resource-limited settings.
- The approach effectively captures gradewise structural progression of ONFH within the latent space.
- While early-stage differentiation remains a challenge, the system shows potential for improving diagnostic efficiency and supporting automated ONFH staging.
