Functional-Consistent CycleGAN for CT to Iodine Perfusion Map Translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generative adversarial neural networks (GANs) are effective for medical image-to-image translation.
• CycleGAN enables unpaired image translation using cycle-consistency loss.
• CycleGAN assumes invertible mappings, which is not always true in medical applications, e.g., inferring function from structure.

Purpose of the Study:

• To address the limitations of traditional CycleGAN in non-invertible medical image translation tasks.
• To propose a novel functional-consistent CycleGAN (fc-CycleGAN) framework.
• To enhance the learning of fundamental characteristics in image translation using a shared proxy domain.

Main Methods:

• Developed a functional-consistent CycleGAN by incorporating a proxy structural image in a shared third domain.
• The proxy domain facilitates learning fundamental characteristics while maintaining cycle consistency.
• Applied the method to estimate iodine perfusion maps from contrast CT scans.

Main Results:

• The proposed fc-CycleGAN demonstrates improved performance in non-invertible image translation tasks.
• Comparison with traditional CycleGAN shows the effectiveness of the functional-consistent approach.
• Successful estimation of iodine perfusion maps from contrast CT scans.

Conclusions:

• The functional-consistent CycleGAN is a viable solution for medical image translation problems with non-invertible mappings.
• Leveraging a proxy domain enhances the robustness and accuracy of the translation process.
• This approach holds promise for applications like functional activity estimation from structural medical images.