Synthesizing lumbar computed tomography myelography from plain computed tomography images using a cycle-consistent...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Spinal Diagnostics
• Deep Learning for Medical Image Translation

Background:

• Magnetic resonance imaging (MRI) is standard for lumbar canal stenosis (LCS) diagnosis but has contraindications.
• Computed tomographic myelography (CTM) is an alternative but invasive.
• Need for noninvasive diagnostic imaging for LCS.

Purpose of the Study:

• Generate synthetic CTM images from plain CT using a modified Cycle GAN.
• Evaluate the diagnostic reliability of these synthetic CTM images for LCS grading.

Main Methods:

• Trained a modified Cycle GAN with Convolutional Block Attention Module and spectral normalization on 21,034 plain CT and 21,062 CTM images.
• Validated using an external dataset of 60 patients undergoing CT and MRI.
• Assessed quantitative performance (IoU, F1, SSIM, PSNR) and clinical reliability (inter/intrarater agreement).

Main Results:

• Modified Cycle GAN achieved high performance in spinal canal enhancement (IoU 0.70, F1 0.80) and bone structure delineation (IoU 0.89, F1 0.94).
• Synthetic CTM demonstrated substantial agreement with MRI for LCS grading.
• High intrarater (almost perfect) and inter-rater (substantial) reliability were observed.

Conclusions:

• Synthetic CTM generated from plain CT shows acceptable quantitative accuracy and clinical reliability for LCS grading.
• This AI-driven approach offers a potential noninvasive diagnostic alternative for patients ineligible for MRI.
• Further validation is needed for nerve-root depiction.