Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Master Class on Reproducibility: A Student Hackathon on Advanced MRI Reconstruction Methods

Lina Felsner, Sevgi G. Kafali, Hannah Eichhorn, Agnes A. J. Leth, Aidas Batvinskas, Andre Datchev, Fabian Klemm, Jan Aulich, Puntika Leepagorn, Ruben Klinger, Daniel Rueckert, Julia A. Schnabel

TL;DR

The study documents a student-driven reproducibility hackathon aimed at MRI reconstruction: MoDL, HUMUS-Net, and an untrained, physics-regularized method. By providing pre-trained weights and data, the teams tested in-domain and out-of-distribution performance, revealing that MoDL and HUMUS-Net can be reproduced under practical conditions while the Untrained+Physics method struggles without well-maintained code. The findings highlight the critical role of documentation, dependency management, and open workflows in achieving reproducibility, and they advocate applying FAIR and FUTURE-AI principles to enhance transparency and reuse in medical imaging research. Overall, the work demonstrates the value of education-focused reproducibility exercises for promoting credible, transferable MRI reconstruction advances.

Abstract

We report the design, protocol, and outcomes of a student reproducibility hackathon focused on replicating the results of three influential MRI reconstruction papers: (a) MoDL, an unrolled model-based network with learned denoising; (b) HUMUS-Net, a hybrid unrolled multiscale CNN+Transformer architecture; and (c) an untrained, physics-regularized dynamic MRI method that uses a quantitative MR model for early stopping. We describe the setup of the hackathon and present reproduction outcomes alongside additional experiments, and we detail fundamental practices for building reproducible codebases.

A Master Class on Reproducibility: A Student Hackathon on Advanced MRI Reconstruction Methods

TL;DR

The study documents a student-driven reproducibility hackathon aimed at MRI reconstruction: MoDL, HUMUS-Net, and an untrained, physics-regularized method. By providing pre-trained weights and data, the teams tested in-domain and out-of-distribution performance, revealing that MoDL and HUMUS-Net can be reproduced under practical conditions while the Untrained+Physics method struggles without well-maintained code. The findings highlight the critical role of documentation, dependency management, and open workflows in achieving reproducibility, and they advocate applying FAIR and FUTURE-AI principles to enhance transparency and reuse in medical imaging research. Overall, the work demonstrates the value of education-focused reproducibility exercises for promoting credible, transferable MRI reconstruction advances.

Abstract

We report the design, protocol, and outcomes of a student reproducibility hackathon focused on replicating the results of three influential MRI reconstruction papers: (a) MoDL, an unrolled model-based network with learned denoising; (b) HUMUS-Net, a hybrid unrolled multiscale CNN+Transformer architecture; and (c) an untrained, physics-regularized dynamic MRI method that uses a quantitative MR model for early stopping. We describe the setup of the hackathon and present reproduction outcomes alongside additional experiments, and we detail fundamental practices for building reproducible codebases.
Paper Structure (7 sections, 2 figures, 2 tables)

This paper contains 7 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Results
  4. Team A: MoDL
  5. Team B: HUMUS-Net
  6. Team C: Untrained+Physics
  7. Discussion & Conclusion

Figures (2)

  • Figure 1: Results for MoDL using in-domain and out-of-distribution (cross-domain) sampling masks. Image quality metrics for Acc. rate: acceleration rate. SSIM: structural similarity. PSNR: peak signal-to-noise ratio.
  • Figure 2: Results for HUMUS-Net using in-domain (knee) and out-of-distribution, i.e., cross-domain (brain) target organs. Image quality metrics. PSNR: peak signal-to-noise ratio. SSIM: structural similarity.