Double U-Net CycleGAN for 3D MR to CT image synthesis.

Area of Science:

• Medical imaging
• Artificial intelligence
• Image synthesis

Background:

• Generative Adversarial Networks (GANs), including CycleGAN, are used for medical image synthesis from unpaired data.
• Current methods often process 2D slices and concatenate them, leading to spatial inconsistencies in 3D reconstructions.
• Synthesizing high-quality 3D medical images with spatial integrity remains a challenge.

Purpose of the Study:

• To develop a novel CycleGAN-based model for high-quality conversion of magnetic resonance (MR) to computed tomography (CT) images.
• To address and resolve spatial inconsistencies in 3D medical image synthesis.
• To improve the accuracy and efficiency of medical image synthesis using unpaired data.

Main Methods:

• A 2.5D input representation was created by reorganizing adjacent 3 slices to maintain spatial consistency and avoid heavy 3D convolutions.
• A U-Net discriminator network was employed for enhanced local and global feature perception.
• Content-Aware ReAssembly of Features (CARAFE) upsampling was utilized for improved content awareness and a larger receptive field.

Main Results:

• The proposed double U-Net CycleGAN achieved a Mean Absolute Error (MAE) of 74.56±10.02, Peak-Signal-to-Noise Ratio (PSNR) of 27.12±0.71, and Structural Similarity Index Measure (SSIM) of 0.84±0.03.
• The method demonstrated superior performance compared to existing state-of-the-art techniques.
• Quantitative metrics indicate successful and accurate 3D image synthesis.

Conclusions:

• The developed method effectively converts MR to CT images using unpaired data, outperforming current state-of-the-art approaches.
• The model synthesizes superior 3D CT images with reduced computational and memory requirements compared to 3D CycleGAN.
• This approach offers a promising solution for generating spatially consistent 3D medical images.