MRI-only based synthetic CT generation using dense cycle consistent generative adversarial networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Current radiation therapy planning often requires both MRI and CT scans.
• CT simulation is a time-consuming step in the patient treatment workflow.
• MRI-only based treatment planning simplifies workflow and reduces patient burden.

Purpose of the Study:

• To develop an automated method for synthetic CT (sCT) generation from MRI images.
• To enable MRI-only radiation therapy planning.
• To investigate the use of dense cycle-consistent generative adversarial networks (cycle GAN) for sCT generation.

Main Methods:

• A novel method using dense cycle GAN was developed for patch-based sCT image generation.
• The cycle GAN model simultaneously learns MRI-to-CT and CT-to-MRI transformations.
• Network optimization involved gradient difference (GD) loss and a novel distance loss metric.

Main Results:

• The model was validated using leave-one-out cross-validation.
• Quantitative metrics included Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), and Normalized Cross Correlation (NCC).
• Mean MAE was 55.7 HU (brain) and 50.8 HU (pelvis); mean PSNR/NCC were 26.6 dB/0.963 (brain) and 24.5 dB/0.929 (pelvis).

Conclusions:

• A novel learning-based approach effectively generates CT images from MRIs using dense cycle GAN.
• The method produces robust, high-quality sCT images rapidly.
• This approach shows significant potential for supporting near real-time MRI-only treatment planning in brain and pelvis.