Automated osteoporosis classification and T-score prediction using hip radiographs via deep learning algorithm
Yu-Pin Chen1,2, Wing P Chan3,4, Han-Wei Zhang5,6,7,8
1Department of Orthopedics, Wan Fang Hospital, Taipei Medical University, Taipei City, Taiwan.
Therapeutic Advances in Musculoskeletal Disease
|April 26, 2024
Summary
A new AI model accurately detects osteoporosis and predicts T-scores from hip X-rays. This offers a low-cost, adaptable screening tool for wider population use.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Radiology and Imaging
- Osteoporosis Diagnosis
Background:
- Dual-energy X-ray absorptiometry (DXA) is the gold standard for osteoporosis diagnosis but is underutilized.
- There is a need for accessible and efficient osteoporosis screening methods.
Purpose of the Study:
- To develop and validate a convolutional neural network (CNN) model for osteoporosis classification and T-score prediction using simple hip radiographs.
- To evaluate the performance of the proposed CNN model against existing methods.
Main Methods:
- A retrospective study using a dataset of 3460 unilateral hip images from 1730 patients (age ≥50 years).
- A fully automated CNN model (X1AI-Osteo) with a controllable feature layer and preprocessing algorithm was developed.
- The model was trained on 2473 images and tested on 497 images, with DXA T-scores for validation.
Main Results:
- The proposed CNN model achieved high performance in predicting osteoporosis (sensitivity: 97.2%; specificity: 95.6%; AUC: 0.96).
- The model demonstrated high consistency with DXA T-scores (r=0.996, p<0.001) when incorporating age, BMI, and sex.
- Performance surpassed open-sourced CNN models.
Conclusions:
- The CNN model accurately identifies osteoporosis and predicts T-scores from simple hip radiographs.
- This technology holds potential for cost-effective, population-based opportunistic osteoporosis screening.
- The model's adaptability makes it suitable for a broader population at risk.
