Nonlinear subspace reconstruction with denoising autoencoder improves DWI quality

A nonlinear subspace model using a denoising autoencoder jointly reconstructs k-q-space data, showing improved noise suppression and detail preservation versus MUSE and LLR in high b-value diffusion-weighted imaging, with minimal bias and higher precision.

• The method learns a latent subspace from biophysically simulated signals, then incorporates the decoder into the image reconstruction forward operator.
• Validation on a multi-shell brain scan showed improved noise suppression and detail compared to multiplexed sensitivity-encoding (MUSE) and locally low-rank (LLR) reconstructions.
• Fiber direction analysis demonstrated minimal bias but higher precision with the proposed method.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.