Lumbar Spine Computed Tomography to Magnetic Resonance Imaging Synthesis Using Generative Adversarial Network: Visual...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) are crucial for diagnosing spinal diseases.
• MRI offers superior resolution for spinal canal and intervertebral disc lesions compared to CT.
• CT may be used when MRI is contraindicated, potentially limiting diagnostic evaluation.

Purpose of the Study:

• To develop synthetic lumbar MRI from CT scans using Generative Adversarial Network (GAN) models.
• To evaluate the realism of synthetic MR images via Visual Turing Tests (VTTs).
• To compare the performance of unsupervised, semi-supervised, and supervised GAN training algorithms.

Main Methods:

• Trained GAN models on axial CT and T2-weighted axial MRI from 285 patients (≥40 years).
• Conducted VTTs with 59 patients, where four readers judged 600 axial images (150 true, 450 synthetic).
• Assessed image quality using Structural Similarity (SSIM) and Peak Signal to Noise Ratio (PSNR).

Main Results:

• Mean accuracy for identifying true images in VTT was 52.0% (first choice).
• Supervised GAN algorithm yielded the most frequently selected synthetic images (first and second choices combined).
• Supervised algorithm achieved the best image quality scores (PSNR: 15.987 ± 1.039, SSIM: 0.518 ± 0.042).

Conclusions:

• GANs can synthesize fairly realistic lumbar spine MR images from CT scans.
• The supervised training algorithm demonstrated superior performance in generating true-to-life MR images.
• This pilot study highlights GANs' potential to enhance spinal disease evaluation when MRI is limited.