Multimodal MRI synthesis using unified generative adversarial networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Acquiring multiple magnetic resonance image (MRI) contrasts is time-consuming and expensive.
• Medical image synthesis offers an effective alternative for obtaining complementary diagnostic information.
• Multimodal MRIs are crucial for comprehensive disease assessment, diagnosis, and treatment planning.

Purpose of the Study:

• To develop a unified framework for multimodal MRI synthesis.
• To create a single generative adversarial network capable of learning mappings between different MRI modalities.
• To enable the synthesis of various MRI contrasts from a single input image.

Main Methods:

• A unified generative adversarial network (GAN) with a single generator and discriminator was developed.
• The generator takes an MRI and its modality label to synthesize images in target modalities.
• The network was trained and tested on multimodal brain MRI data (T1-weighted, T1-weighted contrast-enhanced, T2-weighted, and FLAIR).

Main Results:

• The proposed GAN model demonstrated high accuracy and robustness in synthesizing arbitrary MRI modalities.
• Quantitative assessments showed favorable results for synthesized images across different metrics (NMAE, PSNR, SSIM, VIF, NIQE).
• For instance, using T1 as input, generated T1c, T2, and Flair images achieved high SSIM values (e.g., 0.974 ± 0.059 for T1c).

Conclusions:

• A novel multimodal MRI synthesis method using a unified GAN was successfully proposed.
• The framework synthesizes multimodal MR images efficiently in a single forward pass.
• The method accurately generates multimodal MR images from a single input MRI scan.