Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

MSLesSeg-2024 — preprocessed for segmentation (fixed v2)

A training-ready derivative of the MSLesSeg-2024 multiple sclerosis lesion segmentation dataset (T1, T2, FLAIR + expert lesion masks, MNI space). The original release is unchanged in content; this version repairs and standardises its preprocessing so the volumes can be fed to a segmentation model directly.

Please cite the original MSLesSeg dataset and follow its license terms when using this data.

Layout

Same layout as the original release, so it is a drop-in replacement:

train/P{n}/T{k}/P{n}_T{k}_{FLAIR,T1,T2,MASK,BRAIN}.nii.gz   + P{n}_T{k}_session.json
test/P{n}/P{n}_{FLAIR,T1,T2,MASK,BRAIN}.nii.gz              + P{n}_session.json
covariates.csv   per-session clinical + acquisition covariates
splits.json      5-fold patient-grouped CV splits + held-out test list
manifest.json    build configuration and per-session statistics

T{k} is a timepoint (longitudinal session), not a modality: P1_T2_T1.nii.gz is patient 1, timepoint 2, T1-weighted image. The test split has no timepoint level.

File Contents
FLAIR, T1, T2 N4 bias-corrected, z-score normalised per image inside the brain mask
MASK Binary expert lesion mask (drawn on the registered FLAIR)
BRAIN Brain mask used for normalisation
session.json Shape, affine, crop origin, lesion count / sizes / centroids, QC info

Sessions

110 sessions from 75 patients: 88 train (53 patients) and 22 test (22 patients). All volumes are cropped to a common box of 176 × 208 × 176 at 1 mm.

The original release has 115 sessions. Five were dropped because their lesion labels sit over missing or collapsed image data, which no preprocessing can fix: P12_T1, P12_T3, P12_T4, P49_T2, P53_T1.

What differs from the original release

  • N4 bias-field correction and per-image z-score normalisation (the original has neither, and mixes integer-rounded and full-float intensity scales).
  • Brain mask repair: FLAIR brain extraction failed (skull retained) on some sessions; those use the intersection of the FLAIR, T1 and T2 masks instead.
  • Common crop removing empty background, padded to a multiple of 16.
  • Patient-grouped CV splits: a patient's timepoints never straddle a fold boundary.

Caveats

  • Lesion voxel counts are in MNI space and inflated by a patient-dependent registration scaling (1.05–4.1×). Use lesion_volume_native_mm3 in covariates.csv to compare lesion load across patients.
  • T1 and T2 were registered to MNI independently of FLAIR and sit ~2 mm off it.
  • Acquisition is heterogeneous (~100 Philips / 9 GE / 1 Siemens, 9 scanner models). Treat scanner as a confounder.
Downloads last month
843