fidopt Interactive view
Brain tumours, four MR sequences, segmentation

BraTS 2023 meningioma

Pre-operative brain MR of a patient with a meningioma, T1, T1Gd, T2 and FLAIR as one four-channel image, with the non-enhancing tumour core, the surrounding FLAIR hyperintensity and the enhancing tumour labelled (ASNR-MICCAI BraTS 2023 meningioma training set, Synapse).

30 test casesbrainimage MRresearch only
Look at the images
Read before comparing (12)
  • Access. The archive is given only to a Synapse account that accepts the terms, so prepare.py reads a copy kept on the reference machine, checked against Synapse's MD5 of version 3 of the training archive.
  • Pool. 944 patients with a meningioma imaged before treatment at six US centres (Duke, Thomas Jefferson, UCSF, Yale, Penn and Missouri) on varied scanners and protocols; a patient with any other intracranial tumour was excluded. Sixteen patients have two to eight examinations: only the first enters, the 56 others stay out. The test patients are 18 to 86 years old; the table grades 13 of them 1, five 2, two 3 (one an anaplastic haemangiopericytoma) and ten not.
  • Images. One image of four channels a voxel, T1, T1Gd, T2, FLAIR (the files t1n, t1c, t2w, t2f), as delivered: registered to the SRI24 atlas, resampled to 1 mm and skull-stripped by the provider, 240 x 240 x 155 voxels on every case. No mask is given.
  • Values. Each scanner's units, untouched: the T1Gd's median in the brain runs from 68 to 447,000 over the test cases, three of them above 300,000.
  • Truth. An nnU-Net pre-segmentation corrected in ITK-SNAP by annotators of every experience, each map reviewed by one board-certified neuroradiologist and returned until accepted. One map a patient, no second reader and no agreement measure: graded semi-automatic.
  • Labels. 1 the non-enhancing tumour core, 2 the surrounding non-enhancing FLAIR hyperintensity (oedema or other change, named peritumoral_oedema), 3 the enhancing tumour, values as delivered. Each label is scored on its own, not as the challenge's nested regions, so the numbers do not compare with the challenge's. A label a case lacks scores only if a method writes it.
  • Tumours. The 30 test tumours hold 0.3 to 209 ml (median 11.9 ml); all have an enhancing tumour, 16 the FLAIR hyperintensity, 12 a non-enhancing core, nine all three.
  • Start. The whole brain given as each label scores a mean Dice of 0.025 for the enhancing tumour on the 30 test cases, 0.067 for the FLAIR hyperintensity on the 16 that hold it and 0.001 for the non-enhancing core on its 12; counting the cases without the label, where it is a false positive at 0, 0.035 and 0.000.
  • Room. The test labels moved by one voxel keep a mean Dice of 0.66 for the non-enhancing core, 0.84 for the FLAIR hyperintensity and 0.91 for the enhancing tumour, means of six moves; the surface Dice is at 1 mm, BraTS setting no tolerance.
  • Version. The FIX-V4 archive beside it is older and not applied: its 46 label maps are in version 3 byte for byte, its images differ by rounding only, and its one other map, BraTS-MEN-00364's, was revised again in version 3.
  • Exposure. The training labels are public: a network trained on BraTS data may have seen these patients.
  • Overlap. The other BraTS datasets of the corpus (BraTS2021, BraTS2021_Syn, BraTSAfrica, BraTSPEDs, BraTSReg), GliomaUTSWMR and UCSDGliomaPostMR hold gliomas, BrainMetsMR Yale's metastases and MUGliomaPostMR Missouri's gliomas, which this cohort's criteria exclude, and ExVivoMTL post-mortem lobes from Penn; PairedHeadMRCT_Syn's meningiomas are from Sun Yat-sen University and ReMINDTumourMR's tumours from Brigham and Women's, neither a centre of this cohort. On the T1Gd inside the brain, each drawn case correlates at most 0.54 with the 356 SRI24 images of those sets, where one examination in its two releases (FIX-V4 and version 3) reads 0.9998 and a case moved by a voxel 0.92; another examination of one patient reads -0.06 to 0.91, so only the cohorts tell those apart.
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBImages
1MONAI BraTS SegResNetGPUapp0.362 ±0.0930.348 ±0.08934.76 ±16.42145.34.9

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 30 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.

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

Test cases
30 test patients drawn with seed 2026 among the 944 patients of the 1,000 training examinations, then 30 development patients, the last of the same seeded order. A patient enters by their first examination; the clinical table names no centre, so the draw is not stratified, and the challenge's validation and test labels are not public.
Truth
contours by a program, verified by people the dataset's three labels, 1 the non-enhancing tumour core, 2 the surrounding non-enhancing FLAIR hyperintensity, 3 the enhancing tumour: an nnU-Net pre-segmentation corrected in ITK-SNAP by annotators from medical students to neuroradiologists, every map reviewed by one board-certified neuroradiologist and returned until accepted
Licence
CC BY-NC 4.0 (BraTS 2023 post-challenge terms)
30 cases for tuning, never tested

BraTS-MEN-01309-000, BraTS-MEN-00047-000, BraTS-MEN-00367-000, BraTS-MEN-00199-000, BraTS-MEN-00671-000, BraTS-MEN-00941-000, BraTS-MEN-01214-000, BraTS-MEN-01183-000, BraTS-MEN-01268-000, BraTS-MEN-01085-000, BraTS-MEN-00417-000, BraTS-MEN-00704-000, BraTS-MEN-00404-000, BraTS-MEN-00977-000, BraTS-MEN-00056-000, BraTS-MEN-00596-000, BraTS-MEN-01300-000, BraTS-MEN-01321-000, BraTS-MEN-01294-000, BraTS-MEN-01216-000, BraTS-MEN-00412-000, BraTS-MEN-01051-000, BraTS-MEN-00159-000, BraTS-MEN-00161-000, BraTS-MEN-00057-000, BraTS-MEN-01042-000, BraTS-MEN-00209-000, BraTS-MEN-01366-000, BraTS-MEN-00741-000, BraTS-MEN-00946-000