CISC498-Johnson-Project

MRE Non-Cartesian MRI Reconstruction (MATLAB → PyTorch Conversion)

This project aims to convert the Johnson Lab’s non-Cartesian MRI reconstruction pipeline from MATLAB into a modern, modular, and GPU-accelerated Python/PyTorch framework.

Background

Magnetic Resonance Elastography (MRE) measures the mechanical stiffness of tissue.
The Johnson Lab develops high-resolution spiral k-space imaging methods for brain imaging applications (e.g., Alzheimer’s research).
However, current MATLAB implementations are difficult to maintain and slow to run, especially for iterative reconstruction and multi-coil processing.

The goal of this project is to implement a clean, open Python/PyTorch pipeline for NUFFT-based non-Cartesian MRI reconstruction to improve usability, speed, and research flexibility.
:contentReference[oaicite:0]{index=0}

Objectives

• Replace MATLAB code with a fully PyTorch-based implementation
• Support GPU acceleration (NVIDIA CUDA + Apple MPS + CPU fallback)
• Use NUFFT operators (not gridding) for forward/adjoint image models
• Implement iterative reconstruction solvers:
  • Conjugate Gradient (CG)
  • PWLS / Tikhonov / Total Variation regularizers
  • Optionally: FISTA / ADMM / Learned Unrolled Networks
• Build a modular pipeline allowing rapid research modifications
• Provide Jupyter/Colab demos for reproducibility

Key Components to Port from MATLAB

MATLAB Module	Python Equivalent (to build or adopt)
NUFFT.m, Gdft.m	Use torchkbnufft or custom PyTorch NUFFT ops
SENSE.m (coil combination)	Torch-based sensitivity multiplication w/ broadcasting
TimeSegmentation.m, int_tim_seg.m	Port to PyTorch (phase correction kernels)
solve_pwls_pcg.m	Rewrite using Torch autograd + custom linear operator classes
C_sparse.m, diag_sp.m, Robj.m	Implement modular regularizers in PyTorch
Spiral trajectory generation (genkspace.m, genspi*)	Port trajectory and sampling operator code

Repository Layout (Recommended)

Dependencies

• Python 3.9+
• PyTorch (CUDA or MPS recommended)
• torchkbnufft or tfkbnufft (we will evaluate performance)
• NumPy, SciPy, Matplotlib
• JupyterLab / Google Colab

Getting Started (Colab Friendly)

git clone https://github.com/<YOUR_USERNAME>/mre-recon.git
cd mre-recon
pip install -r requirements.txt