Generalizability of Deep Learning Classification of Spinal Osteoporotic Compression Fractures on Radiographs Using an
Qifei Dong1, Gang Luo1, Nancy E Lane2
1Department of Biomedical Informatics and Medical Education, University of Washington, Seattle, Washington (Q.D., G.L., B.C.C.).
Academic Radiology
|July 12, 2023
Summary
A deep learning model accurately identifies spinal osteoporotic compression fractures (OCFs) on radiographs. This tool shows promise for opportunistic osteoporosis screening, aiding early diagnosis and treatment.
Area of Science:
- Radiology
- Artificial Intelligence
- Osteoporosis Research
Background:
- Spinal osteoporotic compression fractures (OCFs) are early osteoporosis indicators but often missed.
- Early diagnosis and treatment are crucial for managing osteoporosis.
Purpose of the Study:
- To develop a deep learning vertebral body classifier for OCFs.
- To create a component for an automated opportunistic screening tool for osteoporosis.
Main Methods:
- Retrospective analysis of local (1790 subjects) and Osteoporotic Fractures in Men (MrOS) Study (5994 men) datasets.
- Training five deep learning algorithms to classify vertebral bodies on spine radiographs.
- Comparing model performance using AUC-ROC, sensitivity, specificity, and PPV.
Main Results:
- The best ensemble averaging model achieved AUC-ROC of 0.948 (local) and 0.936 (MrOS).
- Optimized for PPV, the model demonstrated high specificity (99.7% and 99.6%) and PPV (89.8% and 94.8%).
Conclusions:
- The deep learning model shows strong performance (AUC-ROC > 0.90) on independent datasets.
- The model exhibits generalizability for real-world clinical use in opportunistic osteoporosis screening.
