Improved CT-based Osteoporosis Assessment with a Fully Automated Deep Learning Tool.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Quantitative CT (qCT) is crucial for bone mineral density (BMD) assessment.
• Previous feature-based image processing algorithms for CT-based BMD have limitations.
• Deep learning (DL) offers potential for enhanced image analysis in medical diagnostics.

Purpose of the Study:

• To develop and validate a novel deep learning (DL) tool for BMD assessment using CT images.
• To compare the performance of the DL tool against a prior feature-based image processing algorithm.
• To evaluate the DL tool's accuracy relative to manual L1 trabecular Hounsfield unit measurements as a reference standard.

Main Methods:

• Retrospective analysis of 11,035 abdominal CT scans.
• Manual L1 trabecular Hounsfield unit measurements served as the reference standard.
• Automated region of interest (ROI) placement using both a feature-based tool and a new DL tool.

Main Results:

• The DL tool achieved a significantly higher overall technical success rate (99.3%) than the older algorithm (89.4%).
• Median Hounsfield unit values from the DL tool were within 10% of the manual standard in 75.6% of seven-slice vertebral ROIs.
• Osteoporosis assessment showed trade-offs: single-slice (sensitivity 39.4%, specificity 98.3%) vs. seven-slice (sensitivity 71.3%, specificity 94.6%).

Conclusions:

• The deep learning BMD tool demonstrates superior performance and a higher success rate compared to previous methods.
• The DL tool's outputs can be optimized for either higher sensitivity or specificity in osteoporosis detection.
• This DL approach represents a significant advancement in automated quantitative CT for BMD assessment.