betaFeedback

WorksPhysics and Astronomy

Scan-specific k-space interpolation using deep Koopman autoencoders

Abstract

Purpose: To develop a scan-specific, interpretable k-space interpolation technique for the reconstruction of accelerated MRI data. Methods: A deep Koopman autoencoder for nonlinear k-space interpolation was developed consisting of an encoder, a linear layer for interpolation and a decoder. The resulting network is implemented using mean square error loss function on k-space neighbourhoods derived from the autocalibration signals (ACS), combining a prediction loss and an autoencoding loss. The approach was compared to GRAPPA and RAKI in 281 retrospectively undersampled brain datasets, using quantitative reconstruction metrics. In a subset of 52 randomly selected images, a clinical reader study compared the proposed approach and RAKI in a forced choice scenario and using a five-point ordinal scoring, assessing sharpness, SNR, aliasing and overall image quality. Additionally, the effect of varying parameters such as the number of autocalibration signals, the contribution of the autoencoding loss, the kernel size, and the latent space dimension for the proposed technique was studied. Results: The proposed approach showed improved quantitative metrics compared to GRAPPA and RAKI. Visually, the method showed higher resilience to noise amplification and improved preservation of sharp details. In the clinical reader study, the proposed approach was preferred overall to RAKI and yielded higher scores than GRAPPA and RAKI. The proposed method showed robustness to parameter variations. Conclusion: Deep Learning based k-space interpolation can be achieved using a Koopman autoencoder for k-space neighbourhoods. Keywords: Parallel imaging, Koopman autoencoder, nonlinear mapping, k-space interpolation, deep learning reconstruction

Paper details

vlnc:W_22m3au9r
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC-BY-NC-SA-4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…