fidoptYork Cardiac MRI
Short-axis cardiac cine MR of a patient under 18 at end diastole (fixed) and end systole (moving), with the left ventricular cavity and myocardium drawn by hand on both (York University's cardiac MRI dataset).
Read before comparing (9)
- Truth. One person, the first author, then a graduate student in computer science, drew every contour by hand where both the endocardium and the epicardium were visible. No clinician check, second observer or agreement is reported. The papillary muscles and trabeculae lie inside the cavity label.
- Labels. prepare.py fills each contour on its slice, the closed cubic spline through its 32 points that the dataset's viewer draws: 1 inside the endocardium, 2 between it and the epicardium. Read from 1 at a pixel's centre, the contours follow the image's edges within half a pixel.
- Slices. Only slices where both contours were seen carry labels. On 9 of the 10 test pairs end systole has one or two labelled slices fewer, and a slice gained or lost weighs a whole slab in the Dice and 6 to 10 mm in the HD95. The myocardium label changes volume by -12 % (pat17) to +30 % (pat2) between the frames.
- Images. 33 subjects from the Hospital for Sick Children, Toronto, on a GE Genesis Signa with the FIESTA protocol, 20 frames a cycle; on the test pairs 10 to 15 slices of 1.25 to 1.64 mm pixels, 6 to 10 mm apart. The files hold only pixels: prepare.py stacks each frame with the identity direction at the origin, values as stored.
- Phases. The dataset names none: the fixed frame has the largest cavity, the moving frame the smallest. The pair comes from one cine on one grid and is used as delivered.
- Scored. On six of the ten test subjects the diagnosis concerns the right ventricle or its outflow tract, which no label covers: the board measures the left ventricle only. The cavity loses 58 % (pat1, pat2) to 88 % (pat8) of its volume between the frames.
- Baseline. With no registration the test pairs start at a mean Dice of 0.43 (0.31 to 0.52; cavity 0.47, myocardium 0.39) and an HD95 of 12.8 mm (10.5 to 16.8).
- Truth bar. It passes the main board's three checks but is weaker than ACDC, which leaves 0.56 unregistered: one annotator, not said to be a clinician, and no right ventricle.
- Outside numbers. None: the paper measures segmentation, not registration.
No method has a result on every pair of this dataset yet.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 mm | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| No registration | 0.426 ±0.050 | 12.84 ±1.55 | 0.000 | 0.00 | ||||||
Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 10 test pairs drawn with seed 2026 among the 23 subjects outside development, and 10 development pairs, the last ten of the same seeded order. One source, one hospital and one scanner model, and no diagnosis groups: the draw has no strata.
- Truth
- contours by experts the left ventricular cavity (1) and the myocardium (2), inside the endocardial and epicardial contours drawn by hand on each slice of the two frames where both were visible
- Source
- jtl.lassonde.yorku.ca/software/datasets
data.nvision.eecs.yorku.ca/datasets/MRI/metadata.txt
data.nvision.eecs.yorku.ca/datasets/MRI/show_segs.zip
doi.org/10.1016/j.media.2007.12.003 - Licence
- no licence: use is free of charge, for research purposes only, and the authors ask that Andreopoulos and Tsotsos, Medical Image Analysis 2008, be cited
10 pairs for tuning, never tested
pat32, pat7, pat31, pat9, pat10, pat3, pat18, pat19, pat24, pat27