Xu Xie

Thesis Topic: Neural Implicit Representation (NIR) for non Cartesian image reconstruction

Supervisors: Prof Dr. Florian Knoll

Description

The goal of this project is to explore neural implicit methods for the reconstruction of undersampled cardiac MRI data. Recent advances in the field have shown that neural implicit representations can achieve high-quality reconstructions without requiring large datasets or binning, which are typically necessary in traditional methods. Huang et al. [1] propose a novel neural implicit k-space representation for binning-free non-Cartesian cardiac MRI imaging. Their approach leverages a multi-layer perceptron (MLP) to map k-space coordinates directly to signal intensities, bypassing the need for density compensation and non-uniform Fourier transforms. This method demonstrated improved artifact removal and spatial-temporal resolution in high-undersampling scenarios.
In a similar vein, Catalán et al. [2] present an unsupervised method, NF-cMRI, that utilizes neural fields to reconstruct accelerated cardiac cine MRI. Their approach is based on implicit regularization using spatio-temporal Fourier features, achieving high-quality reconstructions from undersampled radial k-space data. The method avoids the need for large training datasets and eliminates the complexity associated with NUFFT, offering a simpler and more efficient framework for cardiac MRI. Both methods are designed to address the limitations of conventional MRI reconstruction techniques by exploiting the power of implicit neural networks.

References

[1] Huang, W., Li, H. B., Pan, J., Cruz, G., Rueckert, D., & Hammernik, K. (2023). Neural Implicit k-Space for Binning-free Non-Cartesian Cardiac MR Imaging. arXiv preprint arXiv:2212.08479.
[2] Catalán, T., Courdurier, M., Osses, A., Botnar, R., Sahli Costabal, F., & Prieto, C. (2023). Unsupervised Reconstruction of Accelerated Cardiac Cine MRI using Neural Fields. arXiv preprint arXiv:2307.14363.