fidopt2018 Left Atrium Segmentation Challenge
3D late gadolinium-enhanced MRI of the left atrium of a patient with atrial fibrillation, before or after ablation, the atrial cavity traced by hand (2018 Atrial Segmentation Challenge, University of Utah).
54 test casesheartimage MRresearch only
Read before comparing (8)
- Cohort. Patients with atrial fibrillation imaged before and after ablation, mostly from the University of Utah (NIH/NIGMS CIBC), the rest from other institutes the page does not name: 154 scans of 60 patients, 1.5 T and 3 T, a respiratory-navigated 3D inversion-recovery gradient echo 20 to 25 minutes after gadolinium.
- Patients. A patient has several scans and the files say whose they are nowhere. Matched to cDEMRIS (below), the test set holds both scans of one patient (DN9MLO9S6NWF7GYPC8ZD and AY8LY0Z1KBIZ6LMDOAZO) and one test patient's other scan is a training scan (X730E6GWUSAFPYEAH15H and LJSDNMND9SHKM7Q4IRHJ); development case DYXSCIWHLSUOZIDDSZ40's other scan is a training scan too. Other links stay unknown: test cases are not independent, and development may share patients with test.
- Geometry. 640 x 640 x 88 or 576 x 576 x 88 voxels, 8-bit, every header 1 mm in every direction at origin zero. The pages give 0.625 mm isotropic, but cDEMRIS's copies of 14 of these scans hold 44 slices 2.5 mm apart on the same in-plane grid, so the 88 slices are most likely 1.25 mm apart. The header is left as delivered: distances and volumes here are in file voxels, not millimetres.
- Truth. One consensus tracing a scan by three observers; no agreement between readers is given. Graded expert. The masks hold 134,558 to 443,607 voxels on the 64 cases (median 204,030); the cavity reads 2.8 to 10 times the median of the rest of the image's non-zero voxels on the first ten test scans.
- Room. The first ten test masks moved one voxel keep a Dice of 0.95 to 0.98; the surface Dice at 1 file voxel decides.
- Start. Labelling the whole image (its non-zero voxels) scores a mean Dice of 0.02 on the first ten test cases; Otsu's bright voxels 0.06. TotalSegmentator has no task for LGE-MRI of the atrium.
- Overlap. cDEMRIS (ISBI 2012 scar challenge, figshare 4214532, not in the corpus): 14 of its 20 scans on Utah's 640 or 576 grid are scans of this set, correlation 0.997 to 0.999 once the 88 slices are averaged two by two, against 0.76 at most for any other pair; each cDEMRIS scan rebuilt as these files would hold it scores 0.997 to 0.999, the check's positive control. Pre and post scans of its patients p6, p8, p13, p15, p16, p20 and p24 are among them: a dataset later built from cDEMRIS must leave those scans out. LASC2013 is the 2013 challenge's bSSFP MRI, other images.
- Outside numbers. The challenge's best entry reached a Dice of 0.932 on these 54 test scans (Xiong et al. 2021), scored on the same masks.
No method has a result on every case of this dataset yet.
Case by case
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- The challenge's test set taken whole: its 54 scans, whose cavity masks are now in the archive. 10 development scans, the last of the seeded order over its 100 training scans. The scans are not one a patient: 154 scans of 60 patients, and the files carry no patient key (see the notes).
- Truth
- contours by experts the left atrium with its appendage and the pulmonary vein antra, not the atrium alone (hence atrium_pulmonary_veins_left, as LASC2013): the cavity traced by hand on every slice of the LGE-MRI in consensus by three trained observers (Corview): the blood pool inside the endocardial surface, with the mitral valve, the left atrial appendage and the pulmonary vein antra up to where the veins stop narrowing, under 10 mm on average (Xiong et al. 2021)
- Source
- www.cardiacatlas.org/atriaseg2018-challenge
www.cardiacatlas.org/atriaseg2018-challenge/atria-seg-data
doi.org/10.1016/j.media.2020.101832
arxiv.org/abs/2004.12314
www.cardiacatlas.org/policy-documents - Licence
- none stated
10 cases for tuning, never tested
9DHWWP5Y66VDMPXISZ13, DYXSCIWHLSUOZIDDSZ40, PVNXUK681N9BY14K4Z86, A5RNNK0A891WUSC2V624, IDWWHGWJ5STOQXSDT6GU, 1MHBF3G6DCPWHSKG7XCP, CCGAKN4EDT72KC8TTJ76, CMPXO4J23G58J53Q98SZ, E2ZMO66WGS74UKXTZPPQ, A11O45O3NAXWM7T2H8CH