Deep learning-based micro-CT grayscale analysis for early detection and staging of osteoporosis in rats.
Authors
Affiliations (5)
Affiliations (5)
- Emergency Department, 731 Hospital of China Aerospace Science and Industry Group, Beijing, 100074, China; Department of Orthopaedic Trauma, Peking University People's Hospital, Peking University, Beijing, 100083, China.
- Department of Orthopaedic Trauma, Peking University People's Hospital, Peking University, Beijing, 100083, China.
- Trauma Center, Peking University People's Hospital, Beijing, 100083, China; Key Laboratory of Trauma and Neural Regeneration, Peking University, Ministry of Education, Beijing, 100083, China; National Trauma Medical Center, Beijing, 100083, China.
- Changzhi Medical College,046013, China.
- Department of Orthopaedic Trauma, Peking University People's Hospital, Peking University, Beijing, 100083, China; National Trauma Medical Center, Beijing, 100083, China; School of Medicine, Shenzhen University, Shenzhen, Guangdong, 518060, China. Electronic address: [email protected].
Abstract
Rat models are widely used in preclinical osteoporosis research to study disease mechanisms and evaluate therapies. Current Micro-CT studies mostly rely on cross-sectional comparisons at a single time point, and there is a lack of standardized reference data across multiple time points. To address this gap, the present study provides standardized reference data from multiple time points using a deep learning-based Micro-CT grayscale analysis, enabling early detection and precise staging of osteoporosis. A standardized osteoporosis model was established in ovariectomized Sprague-Dawley rats (n = 32) with a sham-operated group (n = 32). Femurs were harvested at 4, 8, 16, and 24 weeks post-surgery. The proximal 0-250 slice region adjacent to the growth plate was defined as the region of interest (ROI), and six representative slices per femur were analyzed. Voxels within each ROI were classified into four grayscale regions: 0-50 (non-bone), 51-100 (bone-nonbone transition), 101-150 (defined bone), and 151-255 (highly mineralized bone). The percentage areas of the four regions across the six slices (4 × 6 input) were used to train a custom deep learning model. Diagnostic performance for early osteoporosis detection and staging was compared with conventional trabecular parameters. Both the grayscale-based algorithm and conventional Micro-CT parameters distinguished Sham and OVX rats at 4 weeks, enabling early detection, whereas DXA only detected differences at 16 weeks. In osteoporosis staging within the OVX group, the grayscale-based model achieved higher accuracy (88.4 % ± 6.4 %) than conventional parameters (55.9 % ± 8.4 %, p < 0.05). For single-time-point osteoporosis diagnosis, the grayscale-based algorithm (98.3 % ± 3.4 %) also outperformed conventional parameters (85.3 % ± 3.4 %, p < 0.05). The grayscale-based deep learning method allows sensitive early detection and more accurate staging of osteoporosis, providing a robust quantitative tool for assessment of osteoporotic progression in OVX rats.