SeRL: Style-embedding representation learning for unsupervised CT images synthesis from unpaired MR images

Area of Science:

• Medical imaging
• Artificial intelligence
• Hepatocellular carcinoma (HCC) research

Background:

• Hepatocellular carcinoma (HCC) is a leading cause of cancer deaths globally and in the US.
• Computed tomography (CT) and magnetic resonance (MR) imaging are crucial for HCC diagnosis, but have drawbacks like radiation and cost.
• Detecting tumors in combined CT and MR data is challenging due to organ representation differences, and manual annotation is time-consuming.

Purpose of the Study:

• To develop an unsupervised and unpaired image translation method for abdomen CT and MR scans.
• To improve the quality of synthesized medical images by embedding style-representation information.
• To overcome challenges in detecting tumors from multimodal imaging data.

Main Methods:

• Proposed a style-embedding representation learning (SeRL) framework for unsupervised CT and MR image translation.
• Incorporated style-representation from real CT and MR images to enhance translation quality and avoid local minima.
• Utilized a patch-corrosion augmentation method for increased training data diversity.
• Integrated a self-attention module to reduce noise from low grayscale pixel values.

Main Results:

• SeRL successfully generated high-quality CT images from MR images for HCC patients.
• Evaluations using Frechet Inception Distance (FID), Sliced Wasserstein Distance (SWD), and liver segmentation dice score demonstrated superior performance.
• The method proved advantageous over existing state-of-the-art unsupervised medical image translation techniques.

Conclusions:

• The proposed SeRL method offers an effective solution for unsupervised medical image translation between CT and MR modalities.
• SeRL enhances image synthesis quality by leveraging style-representation and advanced augmentation techniques.
• This approach has the potential to reduce physician workload and improve diagnostic accuracy in HCC management.