KC-UNIT: Multi-kernel conversion using unpaired image-to-image translation with perceptual guidance in chest computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Computed tomography (CT) image reconstruction relies on convolution kernels, chosen based on anatomical structures and scan objectives.
• Limited storage in clinical settings necessitates discarding sinogram data, restricting kernel choices.
• Existing deep learning methods for CT kernel conversion struggle with unpaired data and preserving anatomical details.

Purpose of the Study:

• To develop a novel deep learning method for CT kernel conversion using unpaired image-to-image translation.
• To address the challenge of transferring kernel styles while preserving fine-grained anatomical structures in CT images.
• To improve semantic representation learning through discriminator regularization.

Main Methods:

• Proposed KC-UNIT, a novel kernel conversion method utilizing unpaired image-to-image translation.
• Implemented discriminator regularization using generator feature maps for enhanced semantic representation.
• Defined cosine similarity content and contrastive style losses between generator feature maps and discriminator semantic labels.

Main Results:

• KC-UNIT successfully preserved fine-grained anatomical structures during kernel transfer.
• The method demonstrated superior performance compared to existing generative adversarial network-based methods across multiple kernel domains.
• The approach effectively translates CT images between different kernel domains using unpaired data.

Conclusions:

• KC-UNIT offers an effective solution for CT kernel conversion with unpaired data.
• The method advances the field of medical image processing by enabling flexible kernel selection post-acquisition.
• The proposed technique preserves crucial anatomical details, enhancing the utility of CT imaging.