fidopt Interactive view
Brain tumours, four MR sequences, segmentation

UTSW-Glioma, adult glioma before surgery

Brain MR of an adult with a glioma before any resection: T1, T1Gd, T2 and FLAIR as one four-channel image, registered to SRI24 and skull-stripped, with necrotic core, oedema and enhancing tumour labelled (UTSW-Glioma, TCIA).

21 test casesbrainimage MR
Look at the images
Read before comparing (10)
  • Images. One examination a subject before surgery, on scanners of five makers at 0.3 to 3 T; one hospital, so the draw has no strata. The FeTS tool registered each scan to SRI24, resampled it to 1 mm and skull-stripped it, on BraTS 2021's grid. The channels T1, T1Gd, T2, FLAIR keep the provider's values, whose scales differ between scanners. No mask: the brain is the foreground.
  • Skull stripping. Ragged on some subjects, it cuts into tumours at the brain's surface, leaving label voxels where all four sequences are zero: 10,756 of test case BT0694's 58,356, 5,813 of BT0143's 60,818, 4,585 of development case BT0285's, 824 or fewer on the other test cases. No method sees anything there.
  • Truth. Trained research staff refined the FeTS tool's map by hand, and neuroradiologists sent maps back until they accepted them; one map a subject, no second reader. Refined and automatic maps agree at a median Dice of 0.988 for the whole tumour and 0.79 for the necrosis, so a method close to FeTS's networks may be favoured. Graded semi-automatic.
  • Labels. 4 is what enhances on the T1Gd, 1 the necrotic core, 2 the FLAIR hyperintensity, non-enhancing tumour included. The core lies mostly within the enhancing tumour but not wholly: it touches the oedema on 15 of the 17 test maps that hold it, and test case BT0233's 101 voxels of 1 lie in the oedema with no 4 at all. A glioma without enhancement is label 2 alone, as test cases BT0652, BT0694 and BT0757 are. Each label is scored on its own, not as BraTS's nested regions, so challenge numbers do not compare.
  • Left out. On most maps with a 3, the 3 rings every boundary, as if the map had been resampled as an image; the 24 shell maps have label 1 on the tumour's outer surface. The rule that left out BT0848 was checked only on the subjects prepared. The maps kept are those edited least: 72 % glioblastomas against 62 % of the collection.
  • Absent labels. Every pool map holds the oedema, 176 the enhancing tumour, 180 the necrosis. A label a case lacks is not scored there unless a method writes it, a false positive at 0. In test cases BT0191 and BT0716 the necrosis is a few voxels.
  • Room. Moved by one voxel, the first ten test labels keep a Dice of 0.82, 0.91 and 0.89 for the necrosis, the oedema and the enhancing tumour. The surface Dice is at 1 mm, as the provider sets no tolerance.
  • Outside numbers. None on these cases as scored here. The label maps are public, so a network trained on this collection may have seen these subjects.
  • BraTS. Not BraTS 2021's patients as far as its sources tell: its crosswalk, TCGA site codes and contributors name no UT Southwestern, though some of its hospitals are unknown. On the 20 cases built first the T1Gd correlates at most 0.48 with any of the 20 BraTS2021 cases then built, where a subject moved by a voxel keeps 0.91.
  • Overlap. No other dataset of the corpus holds these patients. The registration screening lists this collection: a dataset built from it must keep these test subjects among its test cases and these development subjects out.
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBImages
1MONAI BraTS SegResNetGPUapp0.615 ±0.1130.629 ±0.1158.02 ±3.46145.23.8

The order averages each method's ranks on Dice, Surface Dice, HD95 mm, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 21 cases. Two neighbours are tied when a paired Wilcoxon test on the shared cases does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Workstation.

1 result not ranked: an older run, cases still missing, or a failure on every case
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBMachineImages
MONAI BraTS SegResNetGPUappolder run0.615 ±0.1130.629 ±0.1158.02 ±3.4793.53.1Laptop

Dice: how much the organ contours overlap, 1 is perfect. Surface Dice: the share of the contours' surfaces within a tolerance of each other, 1 is perfect. HD95: distance between the contours' surfaces.

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 UTSW-Glioma, adult glioma before surgery
The first test caseClick to look at every case.
Test cases
21 test subjects drawn with seed 2026 among the 197 outside development, then 19 development subjects, the last of the same seeded order. The pool of 216 is the 362 subjects with a label map refined by hand, less 121 whose map holds a value the provider does not describe (3 or 5), 24 whose label 1 is a shell around the tumour, and BT0848, its map off the brain.
Truth
contours by a program, verified by people the dataset's three labels, 1 the necrotic tumour core, 2 the peritumoral oedema and non-enhancing infiltrated tissue, 4 the enhancing tumour: the FeTS tool's automatic map refined in 3D Slicer by trained research staff given the four sequences, reviewed by neuroradiologists and returned for revision until they accepted it
Licence
CC BY 4.0
19 cases for tuning, never tested

BT0285, BT0196, BT0276, BT0460, BT0226, BT0602, BT0501, BT0083, BT0632, BT0491, BT0818, BT0324, BT0643, BT0361, BT0235, BT0292, BT0772, BT0835, BT0802