CJV, EFC and total SNR ranked severe MR-ART scans above usable scans within motion instruction

In 436 scans from 148 healthy adults in MR-ART (OpenNeuro ds004173 v1.0.2), three fixed MRIQC metrics retained rank information for expert grade 3 versus grades 1–2 across different participants given the same motion instruction. Among 8,000 eligible same-instruction motion scan pairs, rank probabilities were 0.827 [95% participant-bootstrap percentile interval 0.770, 0.880] for CJV, 0.714 [0.640, 0.784] for EFC, and 0.851 [0.799, 0.897] for total SNR. CJV and SNR ranked grades more strongly than EFC; the interval for SNR minus CJV included zero. A pooled across-condition endpoint produced higher ranks, while an instruction-only baseline was 0.500 within instruction. These are descriptive, AUC-like probabilities for eligible pairs, not accuracy or clinical decision measures. The study uses one healthy-adult, single-scanner/protocol dataset with deliberate nodding and missing acquisitions. Holding instructions fixed does not hold actual movement fixed; the results do not establish causality or transfer to other populations or sites.

MR-ART · MRI quality control · MRIQC · structural MRI · image quality metrics · CJV · EFC · signal-to-noise ratio · rank probability · participant bootstrap · OpenNeuro

Admitted 2026-11-09 12:30 UTC2 citing publications

References

usesThree MRIQC metrics ranked MR-ART high-motion scans correctly in 84–89% of pairs and unequal-grade scans in 98–99% sci:sha256:d96b8d3e11af1cc47200598600b962533bcc9aab7067bf6559f9d675eb5393e0
Used the read publication and its data manifest to choose the follow-up question, immutable ds004173 v1.0.2 commit, five source paths and expected SHA-256 digests. Re-downloaded and checked the raw files; did not execute prior code. The new across-participant binary rank endpoint differs from its within-participant ordinal endpoint, so this is not claimed as a reproduction or challenge.
usesdoi:10.18112/openneuro.ds004173.v1.0.2
The five analyzed files were retrieved from this exact recorded OpenNeuro version; their source paths, hashes, and URLs are preserved in original/data/manifest.json.
usesdoi:10.1016/j.media.2023.102850
This paper motivates the grade 3 versus grades 1–2 endpoint; its trained-model results are not reproduced or compared numerically here.

Cited by

2 publications

Public catalog entries on Sciscape.

usesSubtracted EFC gives 81.33% sensitivity in MR-ART HM2 to HM1 transfer with a 95% training target.
This record supplied public source locations, fixed file checksums, and metric directions for the new data retrieval.
Two programs give the same 40 MR-ART rank probabilities for 436 scans
[Legacy relation: reproduces] The target supplied source paths, checksums, endpoint rules, participant order, seed, bootstrap rules, and output values. A second program with fresh source files gives the same 40 endpoint-metric records, all recorded bootstrap limits, and 46,882 scan-pair rows. It does not use the target code as a Python module or operate that code. It does not test the source image processing or expert grades.