Q-Space Guided Multi-Modal Translation Network for Diffusion-Weighted Image Synthesis.

Area of Science:

• Medical Imaging
• Neuroimaging
• Computational Neuroscience

Background:

• Diffusion-weighted imaging (DWI) is crucial for non-invasive tissue microstructure analysis.
• Acquiring densely sampled q-space data for DWI is clinically impractical due to time constraints.
• Current deep learning methods for DWI are limited by fixed q-space sampling requirements.

Purpose of the Study:

• To develop a flexible deep learning framework for synthesizing multi-shell, high-angular resolution DWI (MS-HARDI) from variable q-space sampling.
• To leverage readily available structural MRI data (T1- and T2-weighted) for DWI synthesis.
• To overcome the limitations of fixed sampling schemes in existing deep learning approaches for DWI.

Main Methods:

• Proposed Q-space Guided Multi-Modal Translation Network (Q-MMTN) integrating a hybrid encoder and multi-modal attention fusion.
• Employed a flexible q-space-aware embedding for dynamic feature modulation.
• Incorporated adversarial, reconstruction, and anatomical consistency losses to ensure signal realism and fidelity.

Main Results:

• Q-MMTN successfully synthesized MS-HARDI from flexible q-space sampling, outperforming existing methods.
• The network demonstrated superior performance in estimating parameter maps and fiber tracts with fine-grained anatomical details across diverse lifespan datasets.
• Q-MMTN showed adaptability to variable sampling scenarios, unlike fixed-sampling deep learning models.

Conclusions:

• Q-MMTN offers a promising solution for efficient and flexible DWI data acquisition in clinical and research settings.
• The method enhances the practical utility of DWI by reducing acquisition time and improving data quality.
• The network's ability to handle variable sampling makes it a valuable tool for advancing neuroimaging studies.