NIMARC-MRI: Abdominal HASTE Dataset for Motion Correction in LMIC Settings

In plain language

NIMARC-MRI is an abdominal magnetic resonance imaging dataset established using a 1.5 Tesla scanner at a Nigerian tertiary medical centre. Created specifically to benchmark vendor-neutral, image-domain deep learning methods for respiratory motion correction, it targets clinical environments operating without manufacturer motion-correction licences. The resource comprises three non-overlapping patient groups totalling 243 individuals and 322 volumetric files. These cohorts include 139 cases featuring synthetic three-dimensional respiratory motion for supervised algorithmic training, 25 clinical cases exhibiting native motion degradation for validation, and 79 paired single-shot and respiratory-triggered acquisitions designed to assess cross-sequence generalisation. All scans are standardised, resampled to isotropic resolution, and distributed under a controlled non-commercial data use agreement.

Key takeaways

NIMARC-MRI provides the first public abdominal magnetic resonance imaging dataset sourced from West Africa.
The collection contains 322 preprocessed volumes across 243 patients, divided into synthetic motion, native motion degradation, and paired acquisition cohorts.
Scans were acquired on a 1.5 Tesla scanner lacking integrated commercial motion-correction software licences.
The resource is designed to benchmark vendor-neutral deep learning algorithms for respiratory motion correction in resource-constrained healthcare settings.

Why it matters

Abdominal magnetic resonance imaging often suffers from image blur caused by patient breathing, which typically requires expensive proprietary software to fix. By offering openly benchmarkable imaging data from an operating clinical scanner in Nigeria, this resource enables researchers to build and validate software solutions tailored to healthcare facilities that cannot afford high-cost vendor licences.

Commercialisation angle

This dataset is aimed at medical software developers and artificial intelligence researchers seeking to create vendor-neutral, post-processing motion correction tools for magnetic resonance imaging. Because the data is released under a non-commercial licence to benchmark algorithms, it sits at an early, pre-commercial research stage, though algorithms trained upon it could eventually inform low-cost diagnostic software solutions for low-resource clinics.

AI-generated from the published abstract. Always read the original work before citing.

Abstract

Overview NIMARC-MRI is the first public abdominal MRI dataset from West Africa, collected at a Nigerian tertiary centre operating a 1.5 T Siemens Magnetom Essenza scanner without integrated motion-correction licenses. The dataset was created to benchmark vendor-neutral, image-domain deep learning for respiratory motion correction in resource-constrained settings. Cohort Composition The dataset comprises three mutually exclusive patient cohorts: Table Cohort Description Patients Files Format Synthetic 3D-consistent synthetic respiratory motion for supervised training 139 139 .npz (train/val/test split) Real Motion Native motion-degraded HASTE for clinical validation 25 25 .npz Paired HASTE-TRIGGER Paired acquisitions for cross-sequence generalisation 79 158 .npz (79 HASTE + 79 TRIGGER) Total 243 322 Note: Patient IDs are non-identifiable codes. No individual appears in more than one cohort. Directory Structure plain nimarc-mri-v1.0/ ├── README.md ├── LICENSE.txt ├── DATA_USE_AGREEMENT.pdf ├── Synthetic/ │ ├── train/ │ │ ├── ORIGINAL_pt001_synthetic.npz │ │ ├── ORIGINAL_pt002_synthetic.npz │ │ └── ... (100 patients) │ ├── val/ │ │ ├── ORIGINAL_pt101_synthetic.npz │ │ └── ... (20 patients) │ └── test/ │ ├── ORIGINAL_pt121_synthetic.npz │ └── ... (19 patients) ├── real_motion_25/ │ ├── motion_metadata.csv │ ├── MOTION_pt001_haste.npz │ ├── MOTION_pt002_haste.npz │ └── ... (25 patients) └── Paired_79/ ├── paired_metadata.csv ├── PAIRED_pt001_haste.npz ├── PAIRED_pt001_trigger.npz ├── PAIRED_pt002_haste.npz ├── PAIRED_pt002_trigger.npz └── ... (79 patients x 2 sequences) File Naming Convention Table Cohort Pattern Example Synthetic (train/val/test) ORIGINAL_pt{001}_synthetic.npz ORIGINAL_pt131_synthetic.npz Real Motion MOTION_pt{002}_haste.npz MOTION_pt025_haste.npz Paired HASTE PAIRED_pt{003}_haste.npz PAIRED_pt079_haste.npz Paired TRIGGER PAIRED_pt{004}_trigger.npz PAIRED_pt079_trigger.npz Data Format All imaging data are stored as NumPy .npz archives with the following keys: data — the imaging volume (slices x height x width) affine — the NIfTI affine matrix for spatial orientation header — acquisition parameters extracted from DICOM Volumes were converted from DICOM via dcm2niix (v6.0.0), resampled to isotropic 1.0 mm, intensity-normalised to [0, 1] via min-max scaling with 3.5 SD outlier clipping, and serialised as .npz. Acquisition Parameters Table Parameter T2 HASTE (Synthetic & Real Motion) T2 TSE-TRIGGER (Paired) Scanner Siemens Magnetom Essenza 1.5 T Siemens Magnetom Essenza 1.5 T Sequence Single-shot turbo spin-echo, axial Multi-shot respiratory-triggered TSE, axial TR (ms) 4,574 ± 890 3,500–5,500 TE (ms) 80–150 85–120 Echo train length 77–96 12–26 (mean 19) Slice thickness (mm) 2.5–8.5 4.0–6.0 Matrix 256 × 256 256 × 256 Fat suppression SPAIR / None SPAIR Breathing Breath-hold (single-shot) Free-breathing, belt triggering Acquisition time ~18–22 s per slab ~3–5 min per slab Metadata Files motion_metadata.csv (real_motion_25/) Table Column Description patient_id Non-identifiable code (e.g., MOTION_pt025) file_name Corresponding .npz file tr_ms Repetition time te_ms Echo time slice_thickness_mm Slice thickness motion_severity Mild / Moderate / Severe (radiologist-assessed) likert_original Blinded Likert score (original) likert_corrected Blinded Likert score (model-corrected) notes Additional observations paired_metadata.csv (Paired_79/) Table Column Description patient_id Non-identifiable code (e.g., PAIRED_pt079) haste_file HASTE .npz filename trigger_file TRIGGER .npz filename tr_haste_ms HASTE TR te_haste_ms HASTE TE tr_trigger_ms TRIGGER TR te_trigger_ms TRIGGER TE slice_thickness_haste_mm HASTE slice thickness slice_thickness_trigger_mm TRIGGER slice thickness acquisition_date Date of scan (YYYY-MM-DD) Demographics Age and sex metadata were removed during institutional DICOM anonymisation per National Health Research Ethics Committee requirements (NHREC/17/03/2018: SHREC/2026/7923). The released dataset contains imaging data and acquisition parameters only. Data Use Agreement Full patient-level data are available under a Controlled Data Use Agreement requiring: Citation of the accompanying paper. Institutional ethics clearance or IRB approval for the proposed research. No attempt to re-identify patients. Non-commercial use (CC-BY-NC 4.0). To request access to the full dataset, send email it to abdulrazaq.zubair@fuhsa.edu.ng, aazubair01@gmail.com, +2348189817182 whatsapp Sample Subset A 10-patient sample subset is available for immediate download: 5 synthetic (2 train, 1 val, 2 test) 3 real motion 2 paired (4 files: 2 HASTE + 2 TRIGGER) See sample_subset/ for details. Citation If you use this dataset, please cite: bibtex {nimarc2026, title={NIMARC-MRI: Abdominal HASTE Dataset and a Baseline U-Net Exposing the Synthetic-to-Real Gap in Low-Resource Motion Correction}, author={Zubair, Abdulrazaq A. and Muhammad, Nafiu Musa and Krah, Simeon and Ibrahim, Ummasalma Usman and Abdumumin, Yusuf Tijjani and Tijjani, Ismail Ismail and Ibrahim, Ummulkhairi and Anas, Alyasaa and Yakubu, Mubaraq and Muhammad, Abbas Rabiu}, {MIRASOL Workshop, MICCAI}, year={2026} } Contact For questions regarding the dataset or DUA, contact:Abdulrazaq A. Zubair — abdulrazaq.zubair@fuhsa.edu.ng, aazubair01@gmail.com, +2348189817182 Acknowledgements This work was conducted at Medserve Kano Diagnostic Center, we also acknowledge African Institute for Research Advancement & Innovation (AIRA Africa) and SPARK academy for their tremendous support.