Virtual magnetic resonance lumbar spine images generated from computed tomography images using conditional generative...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed Tomography (CT) and Magnetic Resonance (MR) are crucial diagnostic imaging modalities.
• Generating MR images from CT data can enhance diagnostic capabilities and reduce patient exposure to radiation.
• Conditional generative adversarial networks (cGANs) offer a novel approach for image synthesis.

Purpose of the Study:

• To develop and evaluate a cGAN model for generating virtual MR lumbar spine images from existing CT scans.
• To assess the quality and feasibility of virtual MR images generated by cGANs compared to actual MR images.

Main Methods:

• A cGAN model was trained using CT and corresponding MR lumbar spine examinations from 22 adult patients.
• The model generated virtual MR images from CT data, with 4 examinations reserved for testing.
• Image quality was quantitatively assessed using Mean Square Error (MSE) and Structural Similarity Index (SSIM), alongside qualitative evaluation by radiologists.

Main Results:

• The Mean Square Error (MSE) between generated virtual MR images and actual MR images decreased significantly with increased training epochs.
• Quantitative metrics showed improved similarity to actual MR images as the cGAN training progressed.
• Radiologists found the virtual MR images comparable to original MR images, with minor exceptions in spinal canal structure.

Conclusions:

• Conditional generative adversarial networks (cGANs) provide a feasible method for generating high-quality virtual MR lumbar spine images from CT data.
• This technique can potentially provide MR-like imaging for evaluating vertebral bodies and intervertebral discs without requiring a separate MR examination.
• The generated virtual MR images hold potential for applications such as attenuation correction in PET imaging and radiotherapy dose planning.