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

PURPOSE: To present a novel, nonlinear subspace modeling and joint k-q-space reconstruction technique for high-resolution, multi-band, multi-shell diffusion-weighted imaging (DWI). METHODS: High b-value (> 1000 s/mm2), high resolution DWI has the drawback of generally low signal-to-noise ratios (SNRs). We present an approach that leverages a denoising autoencoder (DAE) to learn a latent subspace from biophysically simulated diffusion-weighted signals. The decoder of this network is then used in the forward operator of the image reconstruction process. The decoded latent images are scaled by the b 0 image, phase is added and processed as regular DWI images with the forward operator for multi-shot, multicoil, multi-slice, k-q-undersampled acquisition schemes. The performance is investigated with a multi-shell, multi-direction imaging brain scan and compared to the results of the multiplexed sensitivity-encoding (MUSE) reconstruction and locally low-rank (LLR) regularized reconstruction. The results are further validated by a bias and precision analysis of reconstructed fiber directions. RESULTS: Comparing the reconstructed data using the proposed method shows improved noise suppression compared to MUSE and more details than LLR-reconstructed images. Specifically, in the higher b-value domain, the reconstruction results show improved detectability of small structures. This bias and precision analysis showed minimal bias introduction, but higher precision with the proposed method. CONCLUSION: Our method combines deep learning, latent signal modeling, joint k-q-space reconstruction, and biophysical simulation for diffusion data and shows strong noise suppression and a high degree of detail in reconstructed diffusion-weighted images.