Gaze-resolved 3D MRI of the eye. We acquire a T1-weighted 3D radial-spiral k-space (pulseq LIBRE) while an MR-compatible eye tracker (SR Research EyeLink) records where the participant looks. Every k-space readout is labelled with the gaze direction at the moment it was acquired, the readouts are binned by direction, and each bin is reconstructed on its own. The result is four motion-resolved images per subject: up, down, left and right.
The full dataset (15 subjects: raw k-space, eye tracking, gaze labels and all reconstructions, in BIDS format) is on
OpenNeuro: ds008973, DOI
10.18112/openneuro.ds008973.v1.0.0. Its README explains the
file naming. To rebuild it from data/study, run analysis/conversion/build_bids_dataset.py.
| step | folder | what it does |
|---|---|---|
| S1 | recon/1-CoilSensitivity |
coil sensitivity maps from the body-coil and head-coil prescans |
| S2 | recon/2-Binning |
sliding window over the synchronised eye-tracking signal gives one gaze label per readout |
| S3 | recon/3-Mitosius |
splits the k-space into one subset per gaze direction |
| S4 | recon/4-Recon |
gridding (bmMathilda) and compressed-sensing reconstruction (bmSteva, 20 iterations) |
The eye tracker and the scanner have separate clocks, so a per-subject offset and rate correction is applied
(model/sync_table.py, stored in sync_corrections.mat). Use the eMask_th0.75_winLen3_sync labels. The readout
indexing convention is described in CLAUDE.md and is asserted in model/dataset.py.
| folder | content |
|---|---|
recon/ |
MATLAB reconstruction pipeline (S1 to S4), coil compression (ROI-PCA, ROVir), cluster job files |
eye_tracker/ |
EyeLink EDF to BIDS conversion and the gaze mask generation used by S2 |
visual_stimuli/ |
PsychoPy fixation paradigms shown in the scanner |
analysis/ |
NIfTI conversion, image-quality metrics, undersampling study, orientation checks |
model/ |
investigation of whether gaze can be decoded from individual readouts (read model/README.md) |
docs/ |
README figures (python docs/make_readme_figures.py) |
You need MATLAB with the monalisa and pulseq libraries on the path. For one subject:
subject_num = 5;
mask_type = 'clean'; % 'clean' | 'clean_0.50' | 'clean_0.75' | 'clean_0.95'
region_idx = 0; % 0 = up, 1 = down, 2 = left, 3 = right
run('recon/1-CoilSensitivity/S1_CalCoilSensitivity_jb_pulseq.m')
run('recon/2-Binning/S2_eyeMask_t1_binning_pulseq.m')
run('recon/3-Mitosius/S3_mitosius_binning_pulseq.m')
run('recon/4-Recon/S4_recon_debi_4fr_pulseq.m')All subjects at once: bash recon/4-Recon/run_resume_recon_logged.sh (skips outputs that already exist). Convert
the .mat results to NIfTI with mat2nii_single (analysis/conversion/). Paths, variables and the pipeline
internals are documented in CLAUDE.md.
CC0 1.0, see LICENSE. eye_tracker/hcph-sops-fork is a fork of the HCPh SOPs and keeps its original Apache 2.0
license.