Attention-Based Q-Space Deep Learning Generalized for Accelerated Diffusion Magnetic Resonance Imaging.

Area of Science:

• Medical Imaging
• Neuroscience
• Machine Learning

Background:

• Diffusion MRI (dMRI) quantifies tissue microstructure but requires lengthy acquisition times due to dense q-space sampling.
• Accelerating dMRI acquisition often involves undersampling q-space data and using deep learning for reconstruction.
• Existing deep learning methods are limited by predetermined q-space sampling strategies.

Purpose of the Study:

• To develop a novel deep learning model, attention-based q-space deep learning (aqDL), for dMRI reconstruction.
• To enable accurate parameter reconstruction using variable q-space sampling strategies.
• To improve the efficiency and generalizability of dMRI data reconstruction.

Main Methods:

• Proposed the attention-based q-space deep learning (aqDL) model.
• Utilized Transformer encoders to map dMRI data from diverse scanning strategies to a common feature space.
• Employed a multilayer perceptron for reconstructing dMRI parameters from latent features.
• Validated the model on Human Connectome Project datasets and two independent datasets with varying undersampling ratios.

Main Results:

• The aqDL model achieved the highest reconstruction accuracy across various undersampling numbers.
• Performance remained high irrespective of whether variable or predetermined q-space scanning strategies were used.
• The model demonstrated consistent accuracy on independent datasets, confirming its generalizability.

Conclusions:

• aqDL effectively reconstructs dMRI parameters from undersampled q-space data with variable sampling strategies.
• The model offers superior accuracy and generalizability compared to traditional deep learning approaches.
• aqDL shows significant potential for application in general clinical dMRI datasets, improving efficiency.