Generating synthetic CTs from magnetic resonance images using generative adversarial networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• MR-only treatment planning streamlines workflows but requires efficient synthetic CT (synCT) generation.
• Automated synCT generation in the brain is crucial for near real-time planning.

Purpose of the Study:

• To develop and evaluate a novel generative adversarial network (GAN) for brain synCT generation.
• To compare the performance of the GAN method against a deep convolutional neural network (CNN).

Main Methods:

• A GAN model was developed using T1-weighted MRI as input, with a ResNet generator and a CNN discriminator.
• Retrospective analysis of 15 brain cancer patients' T1-weighted MRI and CT-SIM images.
• Performance evaluation using Mean Absolute Error (MAE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR).

Main Results:

• GAN model achieved a testing time of ~5.7 seconds per case.
• GAN synCTs showed lower MAE (41.9 HU) compared to CNN (102.4 HU) across tissues.
• GAN demonstrated superior performance in detail preservation and representation of abnormal anatomy (PSNR: 26.6, SSIM: 0.83).

Conclusions:

• A validated GAN model effectively generates high-quality brain synCTs from T1-weighted MRI in seconds.
• This method significantly enhances the potential for near real-time MR-only radiotherapy planning.
• The GAN approach offers a robust solution for efficient and automated synCT generation.