DC-cycleGAN: Bidirectional CT-to-MR synthesis from unpaired data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic Resonance (MR) and Computer Tomography (CT) imaging provide complementary diagnostic information.
• Acquiring both MR and CT images can be challenging due to cost, radiation exposure, and equipment availability.
• Medical image synthesis offers a potential solution to overcome these limitations.

Purpose of the Study:

• To propose a novel bidirectional learning model, dual contrast cycleGAN (DC-cycleGAN), for synthesizing medical images from unpaired datasets.
• To enhance the quality and clinical utility of synthesized medical images.

Main Methods:

• Developed a DC-cycleGAN model incorporating a dual contrast loss function within discriminators.
• Utilized source domain samples as negative examples to enforce distinctness between real and synthetic images.
• Integrated cross-entropy and Structural Similarity Index (SSIM) to preserve luminance and structural integrity during synthesis.

Main Results:

• DC-cycleGAN demonstrated superior performance in medical image synthesis compared to existing cycleGAN-based methods (cycleGAN, RegGAN, DualGAN, NiceGAN).
• The proposed dual contrast loss effectively constrained the generation of realistic synthetic images.
• Integration of cross-entropy and SSIM improved the fidelity of synthesized images.

Conclusions:

• DC-cycleGAN effectively synthesizes high-quality medical images from unpaired data.
• The model offers a promising approach to address limitations in acquiring multimodal medical imaging.
• The developed technique has the potential to improve clinical diagnosis and treatment planning.