1. Introduction
The proposed challenge aims to advance the development of robust and high-quality reconstruction algorithms tailored for magnetic particle imaging (MPI) under low tracer concentration conditions. MPI is an emerging noninvasive imaging modality that enablesreal-time visualization of superparamagnetic iron oxide nanoparticles, offering significant potential for applications in oncology, cardiovascular imaging, and stem cell tracking. However, the reconstruction process in MPI is inherently ill-conditioned, particularly at low concentrations, where signal-to-noise ratios degrade, leading to artifacts, reduced resolution, and inaccurate quantification. This challenge focuses on field free line MPI (FFL-MPI) setups, which have been increasingly adopted for human-scale systems and clinical translation due to their ability to provide potential high sensitivity. By soliciting innovative signal processing and machine learning-based solutions, the challenge seeks to address critical barriers in clinical MPI adoption, such as reducing biological toxicity and meeting the demands of cellular tracking applications. Participants will be tasked with reconstructing high-fidelity images from simulated and real lowconcentration FFL MPI datasets, evaluating metrics like structural similarity, peak signal-to-noise ratio, and quantitative accuracy.
2. Organizers
-
Jie Tiantian@ieee.orgBeihang University
-
Yu Anyuan1989@buaa.edu.cnBeihang University
-
Hui Huihui.hui@ia.ac.cnChinese academy of science
-
Xin Fengxin.feng@ia.ac.cnChinese academy of science
3. Timeline
Submit your challenge paper on the EDAS platform. The review period remains open for two weeks, while method evaluations continue in parallel during this time (i.e. challenge is still running).
Upload the final paper on EDAS. Ensure your method artifacts (e.g., code/models) are prepared for possible verification. Organizers finalize the ranking.
Final dates follow the official ISBI 2026 schedule. Please refer to the ISBI Challenge page for any updates.
ISBI Challenge page: https://biomedicalimaging.org/2026/challenges/
4. Datasets
December 7th, 2025: Dataset Release
Go To Download5. Evaluation Metrics
| Metric | Description | Purpose |
|---|---|---|
| Structural Similarity Index (SSIM) | Measures structure similarity between reconstructed and reference high-concentration images | Fidelity and structure preservation |
| Peak Signal-to-Noise Ratio (PSNR) | Quantifies signal quality relative to reconstruction noise | Noise suppression and contrast quality |
| Normalized Root Mean Square Error (NRMSE) | Measures normalized reconstruction error relative to the ground truth | Global reconstruction accuracy |
Primary and Secondary Ranking Rules
The Structural Similarity Index (SSIM) is designated as the primary determinant of the final ranking.
- Teams will first be ranked in descending order of SSIM (higher is better).
- This reflects the challenge's primary goal of achieving high-fidelity image reconstruction under low tracer concentrations.
If two or more teams achieve SSIM scores within 0.005 of each other, the following hierarchical tie-breakers will be applied sequentially:
- PSNR - higher value preferred
- NRMSE - lower error value preferred
While only SSIM determines the official ranking, a weighted composite score will be reported for analysis and publication:
\[ S_{\text{composite}} = 0.5 \times SSIM + 0.3 \frac{PSNR}{PSNR_{\max}} + 0.2 \times \left(1 - \frac{NRMSE}{NRMSE_{\max}}\right) \]
The computation rule is that for each team, compute the average PSNR across all test images (in dB) and defined as:
\[ PSNR_{\max} = \max_i (PSNR_i) \]
NRMSEmax is defined as the largest NRMSE value observed among all participant submissions:
\[ NRMSE_{\max} = \max_i (NRMSE_i) \]
\[ NRMSE_i = \frac{ \sqrt{\frac{1}{N} \sum_{k=1}^{N} (I_{i,k} - I_{H,k})^2} }{ \max(I_H) - \min(I_H) } \]
Note: This composite score is supplementary and will not affect the official ranking order.