A Deep Nonlinear Subspace Modeling and Reconstruction for Diffusion-Weighted Imaging Using Denoising Auto-Encoder

Area of Science:

• Medical Imaging
• Biophysics
• Machine Learning

Background:

• High b-value, high-resolution diffusion-weighted imaging (DWI) often suffers from low signal-to-noise ratios (SNRs).
• Existing reconstruction methods like MUSE and LLR have limitations in noise suppression and detail preservation.

Purpose of the Study:

• To develop and present a novel nonlinear subspace modeling and joint k-q-space reconstruction technique for high-resolution, multi-band, multi-shell DWI.
• To address the SNR limitations in high-quality DWI acquisition.

Main Methods:

• Leveraged a denoising autoencoder (DAE) to learn a latent subspace from biophysically simulated diffusion signals.
• Integrated the DAE decoder into the image reconstruction process, utilizing it within a joint k-q-space framework for undersampled acquisitions.
• Validated the method on a multi-shell, multi-direction brain scan, comparing it against MUSE and LLR reconstructions.

Main Results:

• The proposed method demonstrated superior noise suppression compared to MUSE.
• It provided more detailed images than LLR reconstruction, particularly in the higher b-value domain.
• Bias and precision analysis indicated minimal bias introduction and higher precision in reconstructed fiber directions.

Conclusions:

• The novel technique effectively combines deep learning, latent signal modeling, and joint k-q-space reconstruction with biophysical simulations.
• It achieves strong noise suppression and a high degree of detail in reconstructed diffusion-weighted images.
• This approach offers a significant advancement for high-resolution DWI acquisition and analysis.