HEAR-MRI: helix-tragus-junction and nasion fiducials on structural MRI across nine EEG–MRI cohorts, with a cross-cohort detection benchmark

Version 1.0.1 supersedes 1.0.0. Users of 1.0.0 should migrate: it contains label sets whose subject identity could not be confirmed, and its benchmark splits permit leakage. What changed. The released collection goes from 361 to 342 label sets, one per distinct scan. Nineteen sets are withheld from the benchmark and released separately as labels_repeat.csv, each with its reason: five sat on another subject's scan, and fifteen NOD subjects are the same person's scan as fifteen HAD subjects (the two groups overlap by one subject). Every scan now appears in exactly one leave-one-cohort-out fold, which removes the leakage by construction — measured, not assumed: HAD's LOCO error rises from 5.52 mm with the duplicates present to 7.17 mm (s.d. 0.37 over five seeds) without them, about 4.5 times the between-seed spread. Baselines are regenerated and now seeded reproducibly; B2 is reported as mean ± s.d. over five seeds (LOCO 9.88 ± 0.33 mm, pooled 6.63 ± 0.23 mm). The intra-rater ICC is recomputed in an image-anchored frame, where the mean falls from 0.976 to 0.931 and the weakest landmark×axis combination is moderate rather than excellent when read from the lower confidence bound; the millimetre test–retest error (4.13 mm, n = 30) is the primary reliability result. Defacing statistics are corrected accordingly: the nasion is void in 99 of 342 subjects (29 %), 98 of 134 in LEMON, and none of the 684 released ears is flagged. labels_world_v1.0.0.csv retains the 1.0.0 table unchanged as the record. Full detail in CHANGELOG.md. The GFEX-EEG validation database (10.5281/zenodo.21032699), which released a four-cohort subset of these annotations, carries its own correction notice; the annotations released here supersede it. Manually placed helix-tragus-junction (HTJ; landmarks LHJ/RHJ) and nasion fiducials on 361 T1-weighted MRIs across nine publicly-archived OpenNeuro EEG-MRI cohorts, released as BIDS derivatives in native scanner-RAS millimetres and native voxel indices, with per-landmark defacing quality control, leave-one-cohort-out and pooled benchmark splits, a millimetre localization-error evaluator, and two reference baselines (multi-atlas registration transfer and a 3D CNN).