AFIDs-Validator: An Open-Access AI-Guided Platform for Learning Anatomical Landmark Placement

Key results

1. ShownTrained-rater anatomical fiducial localization error (AFLE) varies roughly fourfold across the 32 landmarks. It is lowest at the commissures and highest at the temporal-horn and indusium griseum landmarks.

   From 492 rater files over 132 human subjects in four cohorts: median AFLE PC 0.37 mm and AC 0.38 mm vs RIAMTH 1.46 mm, LIAMTH 1.40 mm, LIGO 1.37 mm and RIGO 1.24 mm; global median 0.87 mm, with 57% of placements under 1 mm and 82% under 2 mm. · 3.1 A difficulty spectrum for landmark placement; 2.5 Rater-reliability prior

2. ShownRater errors are heavy-tailed at every landmark, so each placement is scored against that landmark's own distribution. The same millimetre error can fall outside the expert range at one landmark and at the median at another.

   A 1.2 mm error sits near the 4th precision percentile at the anterior commissure but about the 58th at the indusium griseum origin. Across landmarks, mean AFLE is 1.4 to 2.3 times the median (commissures: median about 0.37 mm, mean roughly 0.8 mm). · 3.2 Calibrated feedback relative to trained…; 3.1 A difficulty spectrum for landmark placement

3. ShownThe validator scores uploads against 21 expert-annotated templates in human and macaque spaces. It uses four error tiers tied to published multi-rater reliability, and learners practise in a five-step tutoring loop.

   15 human and 6 macaque templates, each with all 32 landmarks (480 human, 192 macaque, 672 total points); tiers excellent <1 mm, good 1 to 2 mm, fair 2 to 4 mm, needs work ≥4 mm; the 2 mm cutoff sits near the 82nd percentile of trained-rater errors; server-side scoring <50 ms. · 2.4 Reference template library; 2.3 Validation engine; 2.2 Guided learning mode

4. ShownIn a worked quality-control test, the engine located a template-space mismatch by region and by landmark. It also graded severity, telling a gross MNI305 offset apart from a milder MNI152Lin one.

   MNI305 vs MNI152NLin2009cAsym, aligned at AC only: mean error 2.9 mm; 13 of 32 landmarks beyond their rater 90th percentile; temporal 5.2 mm and ventricular 4.1 mm vs diencephalic floor 1.3 mm; PC 3.6 mm (4.1 times its 90th percentile). MNI152Lin mismatch: mean 1.6 mm, 23 of 32 within 2 mm. · 3.4 A quality-control example; Figure legends

5. ShownHow hard a landmark is to localize correlates with how much its position shifts across templates, but the two are distinct. AC-normalized positions differ across MNI templates enough that a reference template cannot be freely substituted at every landmark.

   Pearson r = 0.66 over 31 non-AC landmarks (RVOH varies across templates yet is rated reliably). Across eight MNI152/MNI305 templates, AC-normalized variability averages 1.20 mm, with a minimum of 0.41 mm at RMB and a maximum of 2.35 mm at LIAMTH; AC-PC distance 27.8 to 31.0 mm (MNI152Lin 30.3 mm, MNI305 31.0 mm). · 3.3 Two kinds of difficulty across templates; 2.6 Inter-template variability analysis; Figure legends