PhD Scientific Days 2025

Poster Session I. - I: Theoretical and Translational Medicine

A Versatile Toolbox for High-Speed MRI Image Reconstruction

Name of the presenter

Hakkel Tamás

Institute/workplace of the presenter

Semmelweis University, Department of Biophysics and Radiation Biology

Authors

Hakkel Tamás¹

1: Semmelweis University, Department of Biophysics and Radiation Biology

Text of the abstract

Introduction
Magnetic resonance imaging (MRI) is a complex task that requires advanced software solutions to compensate for the limitation in data acquisition speed. The two most common approaches to accelerate measurement are to increase the amount of data collected simultaneously (parallel imaging) and undersampling the measurement space (partial Fourier technique and compressed sensing). These are typically addressed with iterative algorithms, which demand expert knowledge and additional optimization effort. Many such methods are difficult to implement or reproduce, resulting in long reconstruction times that hinder routine application.
Aims
We present an MRI reconstruction toolbox designed specifically for iterative methods, especially compressed sensing. This toolbox emphasizes flexibility and ease of use, enabling rapid experimentation and reproducibility of published methods, while avoiding the drawbacks of slow prototyping environments and delivering high performance.
Method
The toolbox is implemented in Julia, which combines the speed of C/C++ with the simplicity of Python or MATLAB. Its modular design supports the integration of custom logic at various levels. Users can implement new optimization algorithms, mix existing regularization terms, or define new ones. To demonstrate its capabilities, we benchmarked it against the Berkeley Advanced Reconstruction Toolbox (BART) and MRIReco.jl across several compressed sensing tasks, including reconstructions using Cartesian and non-Cartesian (radial) data, applying Total Variation, Wavelet regularization, temporal Fourier transforms, and low-rank plus sparse models (robust principal component analysis – rPCA). Performance tests were conducted on both a Windows desktop and a Linux-based supercomputer.
Results
Our software offers superior customization compared to existing alternatives. Benchmark results show a reconstruction speedup of 10–50%, depending on the method used.
Conclusion
This toolbox delivers state-of-the-art reconstruction speed and deep customization options. As demonstrated, it is suitable for both rapid prototyping and deployment in production environments.
Funding
This work was supported by the Cooperative Doctoral Program (EKÖP-KDP 2024) and the Governmental Agency for IT Development, through access to the Komondor supercomputer.