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

Background: Magnetic resonance imaging (MRI) is the gold standard for diagnosing lumbar canal stenosis (LCS), but it may be contraindicated in patients with implants or claustrophobia. Computed tomographic myelography (CTM) is an alternative imaging modality but requires an invasive procedure. This study aimed to generate synthetic CTM images from plain CT using a modified Cycle GAN and evaluate their diagnostic reliability. Methods: The training dataset included 111 patients from 3 hospitals, comprising 21,034 plain CT and 21,062 CTM images. The external test dataset included 60 patients from another hospital who underwent plain lumbar CT and MRI, including 20 patients each with LCS, lumbar disc herniation, and no spinal disease. A modified Cycle GAN incorporating a Convolutional Block Attention Module and spectral normalization was trained to translate plain CT into synthetic CTM. Quantitative performance was assessed using Intersection over Union, F1 score, Structural Similarity Index Measure, and Peak Signal-to-Noise Ratio. Two spine surgeons independently graded stenosis on synthetic CTM and MRI using the Lee staging system. Intrarater reliability, inter-rater reliability, and agreement between synthetic CTM and MRI were evaluated using weighted kappa coefficients. Results: For spinal canal enhancement, the modified Cycle GAN achieved a mean Intersection over Union of 0.70 and F1 score of 0.80. Bone structures showed a mean Intersection over Union of 0.89 and F1 score of 0.94. Surrounding tissues demonstrated a Structural Similarity Index Measure of 0.91 and Peak Signal-to-Noise Ratio of 30 dB. Synthetic CTM showed almost perfect intrarater reliability and substantial inter-rater reliability. Agreement between synthetic CTM and MRI was substantial. Conclusions: Synthetic CTM generated from plain CT demonstrated acceptable quantitative accuracy and clinical reliability for grading LCS. Although spatial overlap was imperfect and nerve-root depiction was not validated, this method may offer a noninvasive diagnostic alternative for patients ineligible for MRI.