fidopt Interactive view
Heart, segmentation

LASC 2013 MRI

Whole-heart 3D bSSFP MRI at 1.5 T, with the left atrium, a short part of its appendage and the proximal sections of its four pulmonary veins segmented as one label (Left Atrial Segmentation Challenge 2013).

10 test casesheartimage MR
Look at the images
Read before comparing (9)
  • Scans. The challenge's 30 MRI scans, 10 for training and 20 for testing, all with four pulmonary veins and of mixed quality: 9 high, 10 moderate, 6 with local artefacts, 5 noisy. They come at 1.25 x 1.25 x 1.37 mm, but b001 and b006 at 0.75 x 0.75 x 0.85 and 0.75 mm, which the paper does not explain. The images stay as delivered.
  • Truth. Philips' model-based segmentation, corrected by an experienced observer in 2 to 6 hours a scan: each vein followed at least 10 mm from the body, the appendage cut close to it, the body reaching into the funnel of the mitral valve. The truth is that one label, gt_binary on a training scan, the union of gt_noclip's six labels on a test scan: not the atrium alone, hence its own name.
  • Scoring. The challenge cut every mask at the mitral plane and 10 mm along each vein, without the appendage, before scoring it, so long veins and leaks into the ventricle cost nothing. Here the map is scored against the reference as drawn: following a vein further, or reaching deeper below the valve, than the observer did costs. The challenge's numbers do not compare with these.
  • Orientation. Every header says axial, as do the Decathlon's, but the volumes are sagittal. prepare.py gives each image its true direction and its origin at zero, voxels untouched; the side was read on each scan from the descending aorta and the liver, and checked with TotalSegmentator on b002. A model trained on the Decathlon's files, reoriented by their headers, has seen this anatomy in another orientation.
  • Duplicate. Test scan b005 is training scan a006 voxel for voxel, under another reference: the two agree at a Dice of 0.927, most of the gap being the appendage. The scan is kept once, as development case a006. By image correlation, no other two of the 20 scans look alike.
  • References. Development cases carry training references, test cases test references; on the one scan with both, the training reference holds less of the appendage. The development labels measure 70 to 119 ml (93 on average), the test labels 81 to 137 ml (114).
  • Second observer. The challenge's second observer corrected the same automatic segmentations: against the reference on the ten test cases, a Dice of 0.89 (0.71 to 0.98), an HD95 of 6.3 mm (1.3 to 19.4) and a surface Dice at 1 mm of 0.59 (0.29 to 0.91). Independent tracings would differ more.
  • Room. Methods differ mostly where they cut the veins, the appendage and the valve; the 1 mm tolerance, under the voxel size, is hard even for the second observer.
  • Leakage. All 30 scans and their references have been public since 2015, and the Decathlon's Task02 Heart holds the same 30 images, 20 of them labelled: a model trained on either may have seen a test scan and its reference.

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

The first test case of LASC 2013 MRI
The first test caseClick to look at every case.
Test cases
10 test scans drawn with seed 2026 among the 19 of the challenge's test set (SET-B) that are not also in its training set: b005 is a006. The development cases are the challenge's 10 training scans (SET-A).
Truth
contours by experts the challenge's reference as one label, the atrial body with a short part of the appendage and the proximal sections of the four pulmonary veins: a model-based segmentation corrected by hand by an experienced observer, as drawn, before the challenge's clipping
Licence
CC BY 4.0
10 cases for tuning, never tested

a001, a002, a003, a004, a005, a006, a007, a008, a009, a010