Data crime detectives
A session about data crimes in MR recon
| Session duration: | 60m |
| Intended audience: | Physicists, Radiologists |
| Moderators: | Moritz Zaiss, Martin Uecker (Confirmed: MZ confirmed) |
| First Speaker: | Reinhard Heckel (15 min) ( (Confirmed: No) |
| Second Speaker: | Efrat Shimron (15 min) (Confirmed: No) |
| Third Speaker: | Moritz Zaiss (Confirmed: No) |
Session description
This session will be about Data Crimes in MRI reconstructon and how simmulation can help.
Retrospective subsampling, commonly used in developing reconstruction algorithms, introduces challenges that can bias results and hinder reproducibility. When k-space data synthesized from reconstructed images, such as those from public repositories, are subsampled, hidden preprocessing pipelines (e.g., lossy compression) often enhance data sparsity. This artificially boosts algorithm performance during evaluation, leading to overly optimistic results that fail to generalize to raw data (https://doi.org/10.1073/pnas.2117203119). Another challenge emerges in transient sequences, such as MP-RAGE or FLAIR, where signal evolution depends on the history of k-space acquisitions itself. As demonstrated in recent studies (https://doi.org/10.1007/978-3-031-16446-0_67, https://doi.org/10.1007/978-3-031-72104-5_23), retrospective subsampling of these sequences fails to account for altered transient magnetization dynamics. For example, subsampling k-space lines effectively removes excitation or refocusing pulses from the prospective sequence, leading to k-space data corresponding to sharper reconstructed images than those of the original sequence. If not performed with care, this approach can cause data domain shifts or data distribution shifts (https://arxiv.org/pdf/2204.07204) between retrospectively subsampled and prospectively acquired data. To raise awareness of such "data crimes" we want to dedicate this session playing detective in public datasets, papers and hands-on simulations.
The program will be the following:
- Introduction (10 min)
- Presentation by Reinhard Heckel (15 min) (Confirmed:No)
- Presentation by Efrat Shimron (15 min) (Confirmed: Virtual Yes)
- Presentation by Moritz Zaiss on Simulations (15 min) (Confirmed:Yes)
- Questions and Answers (15 min)