MCMT-GAN: Multi-Task Coherent Modality Transferable GAN for 3D Brain Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Synthesizing multi-modality medical data is vital for computer-aided diagnosis and neuroscience.
• Acquiring diverse imaging data is challenging due to patient discomfort, cost, and scanner limitations.

Purpose of the Study:

• To develop an unsupervised method for synthesizing multi-modality brain MRI.
• To address limitations in acquiring diverse medical imaging data.

Main Methods:

• Proposed a multi-task coherent modality transferable Generative Adversarial Network (MCMT-GAN).
• Employed bidirectional adversarial loss, cycle-consistency loss, domain adaptation loss, and manifold regularization.
• Integrated discriminators with segmentors to ensure segmentation task utility.

Main Results:

• MCMT-GAN achieves robust, high-fidelity multi-modality brain image synthesis.
• Generated images are visually impressive and suitable for clinical post-processing.
• The method outperforms state-of-the-art techniques in cross-modality synthesis.

Conclusions:

• MCMT-GAN effectively synthesizes multi-modality brain MRI in an unsupervised manner.
• The approach enhances the utility of synthesized images for segmentation tasks.
• This method offers a viable solution for overcoming data acquisition challenges in medical imaging.