MMCMR-427K cine-CMR (CineMA 网格预处理版)

# MMCMR-427K cine-CMR, preprocessed onto the CineMA grid Cine cardiac MRI reconstructed from the raw multi-coil k-space of CardioMM / MMCMR-427K and preprocessed to match the layout of ACDC and M&Ms, so it can be used directly with CineMA-based pipelines. The source release ships **raw k-space, not images**. Nothing in it is viewable until it is reconstructed. This dataset is the reconstructed and gridded result. ## Provenance and licence Derived from **CardioMM / MMCMR-427K** (CMRxRecon), released under **CC BY-NC 4.0**: research and non-commercial use only, attribution required. This derived dataset carries the same licence and the same restriction. If you use it, cite the source: - Wang et al., *CardioMM / MMCMR-427K*, arXiv:2512.21652 - plus the four CMRxRecon challenge papers Source: https://huggingface.co/datasets/CMRxRecon/MMCMR-427K ## Contents 1,159 patients across three splits, 1,442 imaging series. | split | patients | |---|---| | train | 705 | | test | 216 | | analysis | 238 | | view | series | |---|---| | short axis (`sax`) | 1,083 | | four chamber (`lax_4c`) | 359 | Per series: the full cardiac cycle, the end-diastolic and end-systolic frames, and CineMA-predicted segmentation masks for both. 7,210 NIfTI files, 3.9 GB. ``` processed/ <split>/<pid>/ <pid>_sax_t.nii.gz full cycle, 12 cardiac phases <pid>_sax_ed.nii.gz end diastole <pid>_sax_es.nii.gz end systole <pid>_sax_ed_pred.nii.gz predicted mask (RV=1, MYO=2, LV=3) <pid>_sax_es_pred.nii.gz <pid>_lax_4c_*.nii.gz same set where the view exists train_metadata.csv test_metadata.csv analysis_metadata.csv ``` `<pid>` is `C<centre>-<scanner>-<set>-P<patient>`, for example `C001-UIH30Tumr780-Test-P007`. All four parts are needed: patient numbers restart per centre and per scanner in the source archive. ## Format Identical to ACDC and M&Ms in this project: **192x192, uint8, 1.0 mm in-plane**, with **through-plane spacing preserved at its real value** (3-10 mm, varying by scan). Verified against ACDC file-for-file. ## Metadata columns `pid, pathology, sex, age, height, weight, edv, esv, ef, ed_frame, es_frame, usable_for_volumetry, source_flags` `pathology` uses a 7-class vocabulary; 349 of 1,159 patients carry one: | label | patients | |---|---| | HCM | 136 | | DCM | 76 | | NOR | 73 | | Ischemic | 47 | | HHD | 17 | | (unlabelled) | 810 | Multi-etiology patients are deliberately left unlabelled rather than assigned an arbitrary winner; the source diagnoses are preserved in the catalog. ## How it was produced 1. **One sequential pass** over the 2.8 TB source archive. The 362 parts are a single gzip stream split at 8 GiB, so there is no random access and no resume point. 2. **Reconstruction**: inverse FFT, root-sum-of-squares coil combination, and a crop of the two-fold readout oversampling. Both MATLAB v7 and v7.3 (HDF5) containers are handled; cine exceeds v7's 2 GB variable limit, so it is necessarily v7.3. 3. **Gridding**: in-plane resample to 1.0 mm, centre crop or pad to 192x192, uint8 by 1st-99th percentile windowing. Through-plane untouched. 4. **ED/ES**: every cardiac phase segmented with CineMA (`mnms_sax` head, chosen because M&Ms is multi-vendor and multi-centre like this data); end diastole is the phase of maximum left-ventricular cavity volume, end systole the minimum. Reconstruction fidelity was checked by reconstructing scans directly from retained raw k-space and comparing against the pipeline output: agreement to 7.6e-06, the uint16 quantisation floor. ## Known limitations Read these before using the volumes. **211 patients have `ef` blank, and blank does not mean zero.** 132 of them have a placeholder 1.0 mm through-plane spacing because their source sidecar left `SliceThickness` empty or NaN; the true value is unrecoverable, and through-plane spacing multiplies directly into any volume, so a number here would be wrong by 5-8x. The rest are patients with a four-chamber view but no short-axis stack. `usable_for_volumetry` marks them. **ED and ES are derived, not annotated.** The source provides no indices. Against ACDC's ground truth, the same method gets ED exactly right 12/30 times (63% within one phase) and ES 15/30 (90% within one phase). This data has 12 cardiac phases against ACDC's 30, so each phase spans more of the cycle and the estimate is coarser. Anything sensitive to exact phase should treat these as estimates. **No demographics at all.** The source release contains no age, sex, height or weight, so no body-surface-area normalisation is possible. **About 3% of scans segment poorly** — an implausible end-diastolic volume or an empty mask at some phase. `source_flags` carries per-scan warnings. **Intensity distribution differs from ACDC.** These images retain more lung and air, so median grey is lower (25 against 51). CineMA min-max renormalises on load, and a controlled comparison found no degradation in segmentation quality: ejection fraction median 60.2% here against 55.3% on an ACDC control run through the identical code path. **37 source scans could not be reconstructed** and are absent: one 0.55 T scanner ships no acquisition sidecars, so in-plane spacing cannot be derived. ## What this is not Not a segmentation ground-truth dataset. The masks are CineMA predictions, not expert annotations. The source provides no manual contours.