fidopt Interactive view
Lesions and tumours, segmentation

BONBID-HIE neonatal lesions

Skull-stripped ADC map of the brain of a term neonate with hypoxic-ischaemic encephalopathy, the injured tissue outlined by experts (BONBID-HIE, Massachusetts General Hospital).

10 test casesbrainimage MRresearch only
Look at the images
Read before comparing (10)
  • Cohort. Term neonates with clinically confirmed HIE, 0 to 14 days old at the scan, at Massachusetts General Hospital, 2001 to 2018: a GE Signa 1.5 T (33 of the pool) or a Siemens TrioTim, Prisma fit or Skyra 3 T (56). One examination a patient.
  • Images. A case is the provider's skull-stripped ADC map, made by the scanner and stripped with a multi-atlas method, zero outside the brain, 32-bit floats in 10^-6 mm^2/s as delivered; 24 of the 89 hold a few negative voxels. Grids range from 0.63 to 2 mm in plane and 2 to 6 mm between slices. The provider's Z_ADC map, the ADC compared voxel by voxel with a normal neonatal atlas, is not given, nor is a mask.
  • Truth. One annotator, a clinical fellow, drew every mask from the radiology report; three paediatric neuroradiologists settled 27 patients by majority. No agreement between readers is published. Graded expert with that limit.
  • Lesions. Often diffuse and small: 0.11 to 347 mL on the 10 test cases (median 4.9 mL), four under 1.2 mL; MGHNICU_072's covers 65% of the brain. The masks equal the clinical table's lesion volumes to the voxel on all 20 cases, and on each test case the lesion's median ADC is 0.47 to 0.91 times the rest of the brain's, as an acute injury reads.
  • Room. The ten test masks moved one voxel in plane keep a Dice of 0.21 to 0.96, along the slices 0.05 to 0.88: on thick slices a small lesion is a few voxels and the surface Dice at 1 mm is strict.
  • Start. Labelling the whole brain (the non-zero voxels) scores a mean Dice of 0.16 on the test cases (0.0006 to 0.79) and a surface Dice of 0.01; Otsu's dark voxels 0.04. The provider's own baseline, Z_ADC below -2, reaches a mean Dice of 0.54 on these ten (its clinical table), from a map methods are not given here. TotalSegmentator has no lesion task for the neonatal brain.
  • Held out. Seven training patients have an empty mask, no injury seen: MGHNICU_209, 215, 231, 329, 432, 442 and 458. A Dice needs a lesion.
  • Grids. Five masks (MGHNICU_234, 331, 336, 358, 359) carry the ADC's header rounded, at most 1 µm and 7e-5 in the direction off; prepare.py writes every mask under its ADC's header.
  • Overlap. No other dataset of the corpus holds neonates from MGH or names the collection; no built image has the grid of any of the 20 cases (12,304 stores compared by size and spacing, a check that finds AbdomenDIRQA's case20 in MSD_Pancreas). M-CRIB's neonates are from Melbourne.
  • Outside numbers. The challenges ranked on the 44 locked test patients, with Z_ADC: none compares with these.

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
10 test patients drawn with seed 2026 among the 72 outside development, the two scanner families in proportion (4 GE, 6 Siemens), and 10 development patients, the last of the same seeded order. The pool is the 89 patients of the released training and validation archives less seven with an empty mask (draw.held_out); the provider's 44 test patients sit in an archive locked by a password the organisers have not published.
Truth
contours by experts the dataset's mask of the HIE lesions on the ADC map: drawn in MRIcroN by a clinical fellow trained in neonatal neuroradiology, on the regions the clinical radiology report names, starting axially and corrected in the coronal and sagittal planes; on the 27 patients where the fellow and the report disagreed, settled by majority among three paediatric neuroradiologists (Bao et al. 2025)
Licence
CC BY-NC-ND 2.5
10 cases for tuning, never tested

MGHNICU_446, MGHNICU_338, MGHNICU_290, MGHNICU_174, MGHNICU_028, MGHNICU_439, MGHNICU_062, MGHNICU_437, MGHNICU_309, MGHNICU_071