Screening opportunistic osteoporosis through multimodal techniques of hip joint CT images: exploring 2D and 3D deep learning, radiomics, clinical data, and their integration.
Authors
Affiliations (4)
Affiliations (4)
- Department of Sports Medicine, The Second Affiliated Hospital of Fujian Medical University, Quanzhou, China.
- Liming Vocational University, Quanzhou, China.
- Department of Orthopaedic Surgery, Jinjiang General Hospital, Quanzhou, China.
- Department of Orthopaedic Surgery, The First Affiliated Hospital of Fujian Medical University, Fuzhou, China.
Abstract
Osteoporosis is a systemic disorder that is characterized by reduced bone mass and density, resulting in heightened bone fragility and vulnerability to fractures. The opportunistic screening of osteoporosis through hip CT holds significant clinical significance. This study explores methods for osteoporosis detection using hip CT data analysis, integrating 2D/3D deep learning, radiomics, and clinical data to improve diagnostic accuracy and reliability. This study enrolled a total of 567 patients with hip joint CT images. Clinical data, including age, gender, and other relevant factors, were utilized to establish an opportunistic osteoporosis screening clinical model. Regions of interest on the patients' hip joint CT scans were outlined and assessed for opportunistic osteoporosis screening through radiomic techniques, 2D deep learning technology, and 3D deep learning technology. A Nomogram model for opportunistic osteoporosis screening based on hip joint CT was established after integrating the radiomic model with the clinical model. The efficacy of each model was compared to identify the optimal model for opportunistic osteoporosis screening based on hip joint CT. Among all Nomogram models (radiomics model + clinical model), the GradientBoosting machine learning algorithm demonstrated the best performance in the validation group. The accuracy of this model in screening for opportunistic osteoporosis in the validation group was 0.849, with an AUC of 0.911. Among all two-dimensional deep learning models, densenet201 exhibited the best performance in the validation group. The accuracy of this model in screening for opportunistic osteoporosis in the validation group was 0.817, with an AUC of 0.884. In three-dimensional deep learning models, the three-dimensional ResNet34 showed the best performance in the validation group, with an accuracy of 0.806 and an AUC of 0.889. Our study showcases the potential of employing radiomics, Nomogram techniques and deep learning techniques to analyze CT images of patients' hips, facilitating the evaluation of osteoporosis.