chicagoypark/Inv_BraSyn

Inv-BraSyn — a frozen 800-slice benchmark for medical inverse problems

800 slices (8 classes × 100) with ground truth, the exact degraded measurement, and the naive reconstruction already materialised. Load it with numpy alone — no source datasets, no loaders, no index caches, no preprocessing to re-derive.

The point of freezing it: reproducing this set from raw data requires the original volumes, the exact crop/resize pipeline, the frozen index caches, and the seed rule — and the seed rule is the trap (see below). Getting it subtly wrong shifts results by 0.5–2 dB in a way that looks entirely plausible. Shipping the arrays removes every one of those failure modes.

Contents

file	class	operator	n
test100_brats_t1n.npz	BraTS-GLI T1n	MRI ×4	100
test100_brats_t1c.npz	BraTS-GLI T1c	MRI ×4	100
test100_brats_t2w.npz	BraTS-GLI T2w	MRI ×4	100
test100_brats_t2f.npz	BraTS-GLI T2f	MRI ×4	100
test100_synthrad_brain_mri.npz	SynthRAD brain MR	MRI ×4	100
test100_synthrad_brain_ct.npz	SynthRAD brain CT	sparse-view CT	100
test100_synthrad_pelvis_mri.npz	SynthRAD pelvis MR	MRI ×4	100
test100_synthrad_pelvis_ct.npz	SynthRAD pelvis CT	sparse-view CT	100

Each .npz holds:

key	dtype	shape	meaning
clean	float32	(100, 256, 256)	ground truth, range [0, 1]
y	complex64 (MRI) / float32 (CT)	(100, 256, 256) / (100, 256, 60)	the measurement
naive	float32	(100, 256, 256)	zero-filled (MRI) / FBP (CT), scale-matched to clean
case_tag, case_key, z, seed_si	str/int32	(100,)	row identity, joins to the manifest
input_psnr, input_ssim, input_lpips	float64	(100,)	locked metrics of naive vs clean

Plus test100_cases.json (the frozen manifest, sha256 c9fefd282b012de5ec1db51faef3db2541831c4f75e4bb64900c22b1d872961b), operators.json, and verify_bundle.py.

Usage

import numpy as np
d = np.load("test100_synthrad_brain_ct.npz")
clean, y, naive = d["clean"], d["y"], d["naive"]      # (100,256,256), (100,256,60), (100,256,256)
print(d["case_tag"][0], d["input_psnr"][0])

Verify before you trust it

python verify_bundle.py     # needs numpy + scikit-image; no GPU, no model, no source data

Recomputes clean_sha256 and meas_sha256 for all 800 slices and compares them byte-for-byte against the frozen manifest, then re-derives input_psnr/input_ssim and checks them at the protocol tolerances (1e-5 / 1e-5). This is what makes the bundle provably the same data the published numbers were computed on rather than a lookalike re-export.

The operators

	MRI (6 classes)	CT (2 classes)
forward model	single-coil Cartesian undersampling	sparse-view parallel-beam Radon
geometry	accel=4, acs=20, randomly_cartesian	n_angles=60, arc=180°
noise	σ=0.02 absolute	σ=0.02 relative (effective 0.68–1.20)

⚠ The seed rule. The measurement seed is manifest["seed"] + case["seed_si"] — not the row position. 512 of the 800 slices carry offsets from a larger run so their measurements stay bit-identical to it; the other 288 use 1000 + k. This is already baked into the shipped y, so it only matters if you regenerate measurements yourself.

Two reporting rules

1. Never average all 8 classes. MRI and CT are not at matched SNR — the CT noise is relative, giving an effective σ of 0.68–1.20 against MRI's 0.02. Report a 6-class MRI mean and a 2-class CT mean separately.
2. Quote n_cases beside n_slices. The four SynthRAD classes draw 3–5 slices from each of 28 cases, so their effective sample size is 28, not 100. The BraTS classes are 1 slice from each of 100 subjects.

Reference numbers

Published results on this exact set, for sanity-checking a harness (PSNR, dB):

class	naive input	best reported
brats_t1n	25.18	35.63
brats_t1c	26.61	35.69
brats_t2w	25.04	36.61
brats_t2f	26.07	35.04
synthrad_brain_mri	22.27	31.31
synthrad_brain_ct	22.58	32.64
synthrad_pelvis_mri	24.26	29.49
synthrad_pelvis_ct	24.23	34.30

If your naive column does not reproduce to ~1e-5, your loader or metric convention differs from the protocol's — verify_bundle.py will tell you which.

Provenance, licensing and citation

This is a derived dataset: 256×256 slices extracted, windowed and resized from two source datasets, then degraded by the operators above. No original volumes are redistributed here.

Both sources are distributed through registration-gated challenge portals with their own terms:

• BraTS 2023 (ASNR-MICCAI, Glioma segmentation) — via Synapse. The split used here is the challenge's ValidationData cohort.
• SynthRAD2023, Task 1 (MR→CT) — via grand-challenge.org / Zenodo.

You should obtain your own access to both source challenges and comply with their terms. If you are a rights-holder for either dataset and object to this derived release, please open a discussion on this repo and it will be taken down.

Please cite both original datasets alongside this benchmark:

@article{baid2021rsna,
  title={The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification},
  author={Baid, Ujjwal and others}, journal={arXiv:2107.02314}, year={2021}}

@article{thummerer2023synthrad,
  title={SynthRAD2023 Grand Challenge dataset: Generating synthetic CT for radiotherapy},
  author={Thummerer, Adrian and others}, journal={Medical Physics}, volume={50}, number={7}, year={2023}}