Cross-Vendor CT Image Data Harmonization Using CVH-CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiomics

Background:

• Computed Tomography (CT) imaging has advanced significantly, but harmonizing data from diverse scanners is crucial for large-scale radiomics research.
• Existing deep learning models face challenges with unpaired training data, hindering CT image harmonization efforts.
• Reducing radiation dose and enhancing image quality are primary focuses, leaving scanner variability an underexplored area.

Purpose of the Study:

• To develop a novel deep learning approach, CVH-CT, for harmonizing CT images acquired from different vendor scanners.
• To address the challenge of unpaired training data in deep learning models for CT image harmonization.
• To reduce scanner-related variability in CT images for improved radiomics analysis.

Main Methods:

• Proposed a novel deep learning framework named CVH-CT.
• The CVH-CT generator incorporates a self-attention mechanism to capture scanner-specific information.
• Introduced a VGG feature-based domain loss to extract texture properties from unpaired data, learning scanner-based texture distributions.

Main Results:

• CVH-CT demonstrated superior performance compared to baseline methods, attributed to the proposed domain loss.
• The method effectively minimized scanner-related variability in radiomic features.
• Experimental results confirmed the efficacy of CVH-CT in harmonizing CT images.

Conclusions:

• CVH-CT offers an effective solution for harmonizing CT images across different scanners, particularly in the absence of paired data.
• The proposed VGG feature-based domain loss is key to learning scanner-based texture distributions and improving harmonization.
• This approach significantly enhances the reliability and consistency of radiomic features in multi-center studies.