Computer-Aided Diagnosis for Determining Sagittal Spinal Curvatures Using Deep Learning and Radiography.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Spinal Biomechanics

Background:

• Manual analysis of spinal curvatures is subjective and time-consuming.
• Clinician variability in measurements can impact diagnostic accuracy.
• Accurate assessment of sagittal spinal alignment is crucial for diagnosing various spinal conditions.

Purpose of the Study:

• To develop and evaluate a deep learning-based computer-aided diagnosis (CAD) tool for automated measurement of spinal sagittal alignment.
• To assess the accuracy and reliability of the CAD system in measuring thoracic kyphosis and lumbar lordosis angles from X-ray images.
• To reduce the time and observer variability associated with manual spinal curvature analysis.

Main Methods:

• A U-Net based deep learning model was employed for lateral spine segmentation.
• Parameter equations and trigonometric functions were used to develop algorithms for angle measurement.
• The system was trained on 322 datasets with data augmentation and evaluated using fivefold cross-validation.
• Performance was validated using Pearson correlation, Bland-Altman analysis, and intra-class correlation coefficient (ICC).

Main Results:

• The segmentation model achieved high performance with precision (90.53%), sensitivity (89.53%), and Dice similarity coefficient (90.22%).
• The automated angle measurement algorithm demonstrated high similarity and reliability in verification analyses.
• The CAD system effectively measured thoracic kyphosis (T4-T12) and lumbar lordosis (L1-S1, L1-L5) angles.

Conclusions:

• The developed deep learning-based CAD tool accurately measures spinal sagittal alignment.
• This automated system can assist clinicians in diagnosing spinal conditions by providing objective and reliable measurements.
• The CAD tool has the potential to significantly reduce diagnostic time and inter-observer variability in spinal curvature analysis.