Richard Osuala
Universitat de Barcelona, Spain
Challenge Co-LeadContext
Dynamic contrast-enhanced MRI (DCE-MRI) plays a central role in breast cancer management, but its reliance on gadolinium-based contrast agents raises various concerns. MAMA-SYNTH introduces a standardized, clinically informed benchmark for evaluating generative models, with the goal of advancing the development of contrast-reduced and contrast-free breast MRI protocols.
Why?64
Gd
Gadolinium
Side Effects
Gadolinium deposits, even from chelated agents, can lead to long-term accumulation, potential neurotoxicity and trigger nephrogenic systemic fibrosis.
π§
Contamination
Gadolinium is detected in drinking water supplies worldwide and beverages, raising concerns about long-term exposure and environmental health impacts.
π°
Accessibility
Gadolinium contrast agents significantly increase the cost of MRI examinations, limiting accessibility in resource-constrained settings.
Task
The task of the challenge is to synthesize single-timepoint 2-dimensional post-contrast breast DCE-MRI slices from corresponding pre-contrast T1-weighted MRI inputs using paired clinical data. Participating algorithms operate on pre-contrast images and generate synthetic peak enhancement post-contrast output.
Pre-contrast
π€
β
Peak-enhancement
Data
The challenge utilizes diverse datasets to ensure algorithmic generalizability across different scanners and populations:
Training Cohort: MAMA-MIA Dataset
This dataset contains pre-treatment DCE-MRI from 1,506 patients from 25 + centers across the United States. The MAMA-MIA dataset was utilized as the training set in the first edition of MAMA Challenges for Primary Tumor Segmentation and Pathologic Complete Response Prediction. Learn more about the MAMA-MIA benchmark at the MAMA-MIA Challenge 2025 and access the complete dataset at MAMA-MIA Dataset with key properties shown below:
Acquisition Plane
Axial
84.4%
Sagittal
15.6%
Magnetic Field Strength
1.5T
72.1%
3T
27.9%
Scanner Manufacturers
GE
64.1%
Siemens
27.3%
Philips
8.6%
We note that participants are allowed to train their models on any further dataset, as long as said dataset is publicly available. To ensure fair evaluation across participating teams, the usage of private data is not allowed in this challenge. We further note that by participating in this challenge, participants agree to comply with EO 14117, 28 CFR Part 202, and Guide Notice NOT-OD-25-083 and acknowledge that the usage of NIH Controlled-access Data Repositories (CADRs) is prohibited in this challenge.
Testing Cohorts
The test data were acquired from two external centers located in the Netherlands and Argentina. Each test case refers to a 2D slice extracted from a patientβs DCE scan.
For each patient, the slice containing the largest malignant tumor area is selected from the peak enhancement phase.
The peak-enhancement phase is defined as the time point with the highest signal intensity within the tumor region.
All images are fat-suppressed and acquired in the axial plane. The main statistics are summarized below:
| Field | Radboud UMC The Netherlands |
Instituto Alexander Fleming Argentina |
|---|---|---|
| Number of Cases | 200 | 100 |
| Image Dimension | 416 Γ 416 px | 512 Γ 512 px |
| Contrast Agent | DOTAREM (99%), GADOVIST (0.5%) | DOTAREM, GADOVIST |
| Manufacturer | Siemens | GE |
| Magnetic Field Strength | 3T | 1.5T |
| Molecular Subtype | ||
| β³ Luminal | 165 (85.7 %) | 37 (37 %) |
| β³ Triple Negative | 23 (9.4 %) | 30 (30 %) |
| β³ Other | 12 (4.9 %) | 20 (20 %) |
Evaluation Framework
Submissions are evaluated across four metric groups spanning pixel-level fidelity, perceptual realism, diagnostic classification performance, and segmentation accuracy.
Metric Group 1
Image-to-Image Comparison
Pixel Similarity
MSE Mean Squared Error
MSE measures the average squared difference between each pixel in the synthesized output and its corresponding pixel in the ground-truth post-contrast image. Lower values indicate greater pixel-level fidelity.
Perceptual Similarity
LPIPS Learned Perceptual Image Patch Similarity
LPIPS computes perceptual distance between images using deep network feature activations, capturing texture and structural similarity closer to human perception than pixel-wise metrics. Lower values indicate more realistic synthesis.
Metric Group 2
ROI-to-ROI Comparison
Tumor Texture Similarity
SSIM Structural Similarity Index
SSIM evaluates image quality by jointly measuring luminance, contrast, and structural similarity within local patches. Applied to the tumor ROI, it captures how well the synthesized enhancement texture matches the reference. Values range from 0β1; higher is better.
Distribution Realism
FRD FrΓ©chet Radiomics Distance
FRD adapts the FID framework to radiomic feature space, measuring the FrΓ©chet distance between the feature distributions of the synthesized ROI patches and real post-contrast patches. Lower values indicate that the synthesized tumors are more statistically indistinguishable from real enhancement patterns.
Metric Group 3
Downstream Task: Classification
Binary Classification
AUROC Luminal
Measures the classifier's ability to distinguish Luminal A/B tumors from all other subtypes on synthesized images. A score of 1.0 indicates perfect separation; 0.5 is chance.
Binary Classification
AUROC TNBC
Measures the classifier's ability to distinguish Triple-Negative tumors from all other subtypes on synthesized images. A score of 1.0 indicates perfect separation; 0.5 is chance.
Metric Group 4
Downstream Task: Segmentation
Overlap Accuracy
DICE SΓΈrensenβDice Coefficient
The Dice coefficient measures voxel-level overlap between the predicted segmentation mask on synthesized post-contrast and the ground-truth mask. It is the harmonic mean of precision and recall over the segmented region. Values range from 0β1; higher values indicate better spatial overlap.
Boundary Accuracy
Hausdorff Distance 95th Percentile
The 95th-percentile Hausdorff Distance (HD95) measures the worst-case boundary deviation between the predicted segmentation mask on synthesized post-contrast and reference segmentation contours, excluding the top 5% of outlier distances for robustness. Lower values indicate more precise boundary delineation.
Ranking Scheme
Timeline
| May 1 | Validation Phase Opens | |
| June 15 | Test Phase Opens | |
| June 30 | Last Submission Deadline | |
| August 1 | Official Results Release | |
| October 8 | Winners Announcement at Deep-Breath Workshop (MICCAI 2026) | |
Organization Committee
Smriti Joshi
Universitat de Barcelona, Spain
Challenge Co-Lead
Jarek van Dijk
Radboud University Medical Centre, Netherlands
Challenge Co-Lead
Luyi Han
Radboud University Medical Centre, Netherlands
Lidia Garrucho
Universitat de Barcelona, Spain
Oliver Diaz
Universitat de Barcelona & CVC, Spain
Maria Laura Cosaka
Instituto Alexander Fleming, Argentina
Antonio Portaluri
The Netherlands Cancer Institute (NKI), Netherlands
Advisory Committee
Simone Balocco
Universitat de Barcelona & CVC, Spain
Karim Lekadir
Universitat de Barcelona & ICREA, Spain
Yaofei Duan
Radboud University Medical Centre, Netherlands
Tianyu Zhang
Radboud University Medical Centre, Netherlands
Ritse Mann
Radboud University Medical Centre, Netherlands
Daniel Mysler
Instituto Alexander Fleming, Argentina
Contact
For inquiries about the MAMA-SYNTH challenge, feel free to reach out to Richard Osuala (richard.osuala[at]ub.edu), and Smriti Joshi (smriti.joshi[at]ub.edu).
Partners & Institutions
