Artificial Intelligence for Automatic Measurement of Sagittal Vertical Axis Using ResUNet Framework.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Sagittal vertical axis (SVA) is a critical parameter for assessing spinal alignment.
• Accurate SVA measurement is essential for diagnosing and managing spinal deformities and degenerative conditions.
• Manual SVA measurement can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop and validate an automated method for measuring SVA from lateral whole spine radiography.
• To assess the accuracy, robustness, and consistency of the automated method compared to manual measurements by physicians.

Main Methods:

• Utilized a convolutional neural network (ResUNet) for keypoint detection to automate SVA measurement.
• Trained and evaluated the ResUNet model on 990 standing lateral radiographs.
• Employed an improved localization method for enhanced accuracy in keypoint detection.

Main Results:

• Achieved a median absolute error of 1.183 ± 0.166 mm for SVA measurement.
• Demonstrated high detection rates: 91% for C7 body and 87% for sacrum within 5 mm.
• The algorithm processes each image in approximately 0.2 seconds.
• Intra-class correlation coefficients ranged from 0.946 to 0.993, indicating excellent consistency with physicians.

Conclusions:

• The automated SVA measurement method is accurate, robust, and efficient.
• The algorithm shows excellent consistency with measurements made by physicians of varying experience levels.
• This AI-powered tool has significant potential for clinical application in routine SVA assessment.