teooi's picture
Add 19 files
f642e77 verified
|
Raw History Blame Contribute Delete
3.18 kB
metadata
license: other
license_name: nsd-terms-and-conditions
license_link: https://cvnlab.slite.page/p/IB6BSeW_7o/Terms-and-Conditions
pretty_name: CortexDecoder NSD cortical flatmaps
tags:
  - fmri
  - neuroscience
  - brain-decoding
  - natural-scenes-dataset
extra_gated_prompt: >-
  This dataset is derived from the Natural Scenes Dataset (NSD). To access it,
  you agree to the NSD Terms and Conditions
  (https://cvnlab.slite.page/p/IB6BSeW_7o/Terms-and-Conditions).
extra_gated_fields:
  I agree to the NSD Terms and Conditions: checkbox

CortexDecoder NSD cortical flatmaps

Cortical flatmaps and reconstructions for NSD Subject 1, from CortexDecoder. Code is on GitHub and the model weights are in teooi/cortexdecoder.

Files

flatmaps/subj01/
  subj01_sessionNN.npy               (750, 1, 224, 560) float32 flatmaps, one per trial (39 sessions, session 11 excluded)
  subj01_sessionNN_image_ids.npy     NSD image index (0-based, 73k set) of each trial
  subj01_sessionNN.json              per-session record
  valid_mask.npy                     224x560 cortex mask
  manifest.json                      preprocessing record with the SHA-256 of every file
roi/subj01_roi_overlap.json          visual-cortex ROI on the CortexMAE patch grid
reconstructions/subj01/
  final.zip                          998 final reconstructions (PNG, 256 px, one diffusion sample per image)
  lowlevel.zip                       998 low-level Deep Image Prior reconstructions (PNG, 256 px)
  image_ids.csv                      file name and NSD image index of each reconstruction

Flatmaps

NSD single-trial betas are resampled from fsaverage to fsLR-32k and projected onto CortexMAE's 224x560 flatmap. Each flatmap location is standardized with statistics fitted on training images only (seed 42, 5% of non-test images held out for validation). Pixels outside valid_mask.npy are not cortex.

The flatmaps cover the whole cortex. The visual-cortex ROI used by the final model is a mask over CortexMAE's 364 patches (16x16 pixels each). roi/subj01_roi_overlap.json gives each patch's grid cell (cells_row_col) and the fraction of it inside NSD's nsdgeneral mask (coverage). The ROI is the 50 patches with coverage of at least 0.5.

import numpy as np
flatmaps = np.load("flatmaps/subj01/subj01_session01.npy")          # (750, 1, 224, 560)
image_ids = np.load("flatmaps/subj01/subj01_session01_image_ids.npy")

Reconstructions

The 998 shared test images that Subject 1 viewed. Files are named <row>_img<NSD image index>.png. final.zip matches the CortexDecoder row of Tables 1 and 2 in the paper, and lowlevel.zip the CortexDecoder (low-level) row.

Citation

@article{imototar2026cortexdecoder,
  title   = {CortexDecoder: fMRI-to-image Reconstruction from Cortical Flatmaps},
  author  = {Imoto-Tar, Teo and Fung, Maxwell and Gurha, Arnav},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.10.09.757766}
}

Please also cite the Natural Scenes Dataset (Allen et al., 2022).