aluk4/iabi-mri-brain

IABI MRI Brain Pack

Clean ground-truth brain MR slices for a university course on inverse problems in biomedical imaging. Derived from the IXI Dataset.

This pack contains images only — no k-space, no raw data. Undersampled measurements are simulated from these images at training time, which is what makes the acceleration factor a free experimental knob.

Contents

Split	Slices	Shape	dtype
train.npz	1500	160×160	float16
val.npz	200	160×160	float16
test.npz	200	160×160	float16

Each file holds one array under the key images.

import numpy as np
from huggingface_hub import hf_hub_download
path = hf_hub_download("aluk4/iabi-mri-brain", "train.npz", repo_type="dataset")
images = np.load(path)["images"]   # (1500, 160, 160) float16 in [0, 1]

How it was built

T2-weighted volumes from IXI-T2.tar. Six slices are taken from the central 30% of each volume — the ends of a head scan are mostly air — then centre-cropped to square, resized to 160×160 by area-averaging, and normalised per slice to [0, 1].

Splits are disjoint by subject. 578 IXI subjects were assigned to splits before any slices were extracted: 326 train, 45 val, 45 test. This matters. Adjacent slices of one subject are nearly the same image, so a slice-level random split would put a near-duplicate of almost every test slice into the training set and the resulting test score would not measure generalisation. Verified empirically: maximum cosine similarity between a test slice and any training slice is 0.849, against 0.882 within the training set.

Limitations

These are magnitude images from a single coil. Real MRI is complex-valued and multi-coil. Reconstruction results on simulated single-coil k-space are not comparable to published multi-coil benchmarks such as fastMRI.

Licence and attribution

CC BY-SA 3.0, inherited from IXI — share-alike propagates, so this derived pack carries the same licence.

The IXI data was collected at three London hospitals (Hammersmith Hospital 3T Philips, Guy's Hospital 1.5T Philips, Institute of Psychiatry 1.5T GE) and is made available by the IXI project under CC BY-SA 3.0. Please credit the IXI project in any work using this pack.