Automated Measurements of Spinal Parameters for Scoliosis Using Deep Learning.

Area of Science:

• Radiology
• Artificial Intelligence
• Orthopedics

Background:

• Scoliosis diagnosis relies on accurate measurements like the Cobb angle, which are traditionally time-consuming and prone to interobserver variability.
• Existing automated tools often lack comprehensive measurement capabilities or require manual input.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated measurement of key scoliosis parameters.
• To assess the accuracy and efficiency of the CNN in diagnosing scoliosis compared to manual measurements.

Main Methods:

• A retrospective study using 1682 scoliosis patient radiographs (AP and lateral).
• A CNN model incorporating coarse segmentation, landmark localization, and fine segmentation was developed.
• Performance was evaluated using Dice coefficient, Mean Absolute Error (MAE), and Percentage of Correct Key-points (PCK).

Main Results:

• The CNN achieved high accuracy with Dice coefficients >0.90 and PCK of 89.7%-93.7%.
• Automated measurements demonstrated acceptable agreement with manual measurements, with significantly reduced measurement time.
• MAE for T5-T12 Cobb angle and sagittal balance was low in adolescents and acceptable in older adults.

Conclusions:

• The developed deep learning system provides rapid, accurate, and reliable scoliosis measurements.
• This automated approach can enhance clinical workflow efficiency for scoliosis diagnosis and treatment planning.