Continuous Conversion of CT Kernel Using Switchable CycleGAN With AdaIN.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology

Background:

• X-ray computed tomography (CT) utilizes various filter kernels to enhance specific anatomical structures.
• Re-scanning patients is often necessary for generating CT images with different kernels due to the deletion of raw sinogram data.

Purpose of the Study:

• To develop a novel unsupervised method for post-hoc conversion of CT images between different filter kernels without compromising image quality.
• To enable image translation between kernel domains and along interpolation paths using a cycle-consistent generative adversarial network (cycleGAN).

Main Methods:

• Proposed an unsupervised continuous kernel conversion method leveraging cycle-consistent generative adversarial networks (cycleGAN) with adaptive instance normalization (AdaIN).
• The method allows for image translation between distinct kernel domains without requiring paired training data.
• Explored the generation of intermediate kernel domain images through interpolation.

Main Results:

• The proposed cycleGAN-based method successfully translates CT images between different kernels accurately, achieving quality comparable to supervised methods.
• The network can generate images along the interpolation path between kernel domains, creating unseen intermediate kernel images.
• Availability of intermediate kernel domain images further enhances the quality of the generated images.

Conclusions:

• The developed unsupervised method offers an effective solution for post-hoc CT image kernel conversion, eliminating the need for repeat patient scans.
• Generated intermediate kernel images demonstrate utility in diagnostic tasks, such as hypopharyngeal cancer diagnosis.
• This approach advances the flexibility and efficiency of medical image processing in radiology.