MVG_extract — pre-extracted 2D slice tiles for the MVG_Multimodal medical baseline

Canonical 256×256 2D slice tiles for the four datasets used by the multimodal MVG_Multimodal in-context image→image model (cross-modal translation + binary segmentation): BraTS2023, SynthRAD2023, AMOS22, CHAOS.

These are the exact model inputs — extracting them once removes the per-sample full-volume load (AMOS volumes are ~200 MB) so training/eval is compute-bound, and they are the canonical bytes the fair-evaluation contract hashes (so any method consuming them is byte-comparable).

Contents

Everything unzips to a single MVG_extract/ directory:

MVG_extract/{dataset}/{split}/{subject}__{img|mask}__{key}.npz

• {dataset} ∈ brats, synthrad, amos, chaos
• {split} ∈ train, val, test (BraTS has train/val only — its test uses the nnUNet/paired-PNG pools)
• {key} = modality (t1n,t1c,t2w,t2f / mr,ct / ct,mri / t1dual,t2spir) or seg/liver mask
• each .npz holds zs (int32, n slice indices) + tiles (uint8, (n, 256, 256))

dataset	train	val	test	modalities	task
brats	5752	307	—	t1n,t1c,t2w,t2f (+seg)	translation + tumor seg
synthrad	576	32	112	mr, ct (brain/pelvis)	MR↔CT translation
amos	580	32	108	ct, mri (+liver)	liver seg
chaos	90	12	18	ct, t1dual, t2spir (+liver)	liver seg

(counts are .npz stacks; ~7,619 total, 9.3 GB.)

Usage

import numpy as np
d = np.load("MVG_extract/synthrad/test/brain__BA211__img__mr.npz")
zs, tiles = d["zs"], d["tiles"]           # (n,), (n,256,256) uint8
img01 = tiles[0].astype(np.float32) / 255.0   # image  -> [0,1]
# mask stacks (…__mask__…) are already {0,1}: gt = tiles[0].astype(np.float32)

Point the loader at the unzipped root: MedicalProvider(name, split, 256, extract_root="…/MVG_extract").

Provenance

Rendered by mm_extract.py via the vendored mm_preprocess recipe (LPS reorient; per-volume body bbox +15% → pad-square → resize 256; MR non-zero (1,99)-pct → CT HU window (−1000,1000); MR & CT share the CT bbox). Frozen subject splits from latent-diffusion_Multimodal/ldm/data/medical_splits (seed 0). uint8 8-bit quantization is byte-consistent with the medical-imaging PNG pipeline.

Code + fair-evaluation protocol: https://github.com/ChicagoPark/MVG_Multimodal