muqsitamir/abide-harmonization-audit

ABIDE I Harmonization Audit Results

Results behind Frozen Site-Probe Accuracy Is Not a Standalone Harmonization Score: An Audit of Evaluation Practice for Multi-Site MRI (Muhammad Muqsit Islam, Gloria García Cuenco, José M. Martínez, Pierrick Coupé; arXiv:2610.14140, submitted to IEEE ISBI 2027).

The audit asks whether a common way of scoring MRI harmonization, the balanced accuracy of a frozen site classifier on harmonized scans, says what it is taken to say. It uses ten output sets from seven harmonization methods on ABIDE I (1,112 T1-weighted scans, 17 sites) as test cases, not as a leaderboard, with controls for chance, head shape, intensity, preprocessing and target alignment.

No images are included. These are metrics, probe predictions summarised as balanced accuracies, and bootstrap intervals. Interactive view: Harmonization Audit Explorer.

Subsets

Config	Rows	What it holds
paper_table (default)	12	Paper Table 1: image change (PSNR, XCorr, W), target alignment (ΔW, ΔKL to NYU), and site-probe balanced accuracy for the frozen probe, the reseeded probes and the converged probes, per output
probe_seeds	110	One row per probe seed and output: balanced accuracy on the 90 source test subjects with its 95% bootstrap interval
decodability	32	Paper Fig. 2: probes trained on raw scans vs. on the harmonized scans, on whole-head and brain-only slices, for image and intensity-histogram probes
all_runs	~36K	Every evaluation run in long format (run, probe family, training source, seed, checkpoint, method, subject group, metric, estimate, 95% interval)

Key columns

• key / method: output identifier (cyclegan_tuned, stargan_aggressive, dlest_1000, diffusion_20k_redraw, haca3, ...). label is the name used in the paper.
• probe (in probe_seeds): reseeded = the frozen probe's training recipe rerun with five seeds (best-validation checkpoint); converged = five probes trained four times longer (last checkpoint).
• ba, estimate: balanced accuracy over the 16 source sites (uniform guessing: 0.0625; shuffled-label probes: 0.03–0.08).
• ci_low, ci_high: 95% percentile bootstrap interval, 2,000 resamples of test subjects stratified by site.
• group (in all_runs): source_non_nyu (the 90 subjects the paper reports), target_nyu, or all.
• delta_w_nyu, delta_kl_nyu: output's distance to NYU intensities minus the input's; negative means closer to the target site.

Main findings, in numbers

• Rerunning the frozen probe's recipe with five seeds gives the same CycleGAN outputs balanced accuracies from 0.13 to 0.59.
• A Gaussian blur (σ = 2) that moves intensities away from NYU brings the frozen probe to 0.07.
• Probes trained on harmonized whole-head slices recover the site at 0.87–0.91 (histogram matching, CycleGAN, diffusion); for HACA3, the frozen probe scores 0.25, a probe trained on its outputs 0.75, and an intensity-histogram probe 0.71.

Provenance

Built with build_audit_data.py (in this repository) from results/isbi2027/ in github.com/muqsitamir/HarmonIt at tag isbi2027-submission. The build checks every per-seed value against the paper's Table 1 and the probe-difference files before writing. The written protocol, with amendments dated before each outcome was seen, is in documentation/ISBI2027_PROTOCOL.md of that repository.

License and data terms

The underlying images are from ABIDE I (Di Martino et al., 2014), shared by the International Neuroimaging Data-sharing Initiative under non-commercial, share-alike terms. These derived results are released under CC BY-NC-SA 4.0 to match. The code that produced them is Apache-2.0 (see the HarmonIt repository).

Citation

The paper is under review at IEEE ISBI 2027. Please cite the arXiv version:

@misc{islam2026frozenprobe,
  title         = {Frozen Site-Probe Accuracy Is Not a Standalone Harmonization Score:
                   An Audit of Evaluation Practice for Multi-Site MRI},
  author        = {Islam, Muhammad Muqsit and Garc{\'i}a Cuenco, Gloria and
                   Mart{\'i}nez, Jos{\'e} M. and Coup{\'e}, Pierrick},
  year          = {2026},
  eprint        = {2610.14140},
  archivePrefix = {arXiv},
  primaryClass  = {eess.IV},
  note          = {Submitted to IEEE ISBI 2027},
  url           = {https://arxiv.org/abs/2610.14140}
}