Automatic ROI placement: validation

Striatum Quant · hand atlas v1 · Toolkit v2.18. Every number below is reproduced by training/striatumHandAtlas-v1/ in the source project.

137 µmmedian distance of arrow ends from the hand placement, each animal held out

9 of 9control–knockout differences reproduced (strain × arrow)

0.046median difference from the hand measurement per arrow (normalized; controls ≈ 0.75)

249hand-quantified sections from 20 animals

How placement works

There are no anatomy rules. The ROIs (DV, ML and MV-DL arrows and three noise boxes) are carried over from the most similar of 249 TH sections whose ROIs were placed by hand.

The section is resampled to 80 µm per pixel and contrast-normalized; its tissue is found.
Each library section, and its mirror image for the other hemisphere, is located by normalized cross-correlation and aligned by enhanced-correlation-coefficient registration (shift, rotation, scale), then refit over its striatum.
The 8 best matches carry their arrows over; each arrow end is their correlation-weighted median. Four ends that sit on the striatum's edge in the hand placements (DV start, both ML ends, MV-DL end) are moved to the nearest strong edge.
A striatum clearly paler than its partner that also matches worse (a lesion) takes the other side's arrows, mirrored.
Noise boxes are chosen in tissue clear of both striata, at the background level the matches' hand boxes measured.
Both hemispheres are placed; "Auto" takes the one with the lower expected error, and the other is shown faintly, one click away.

Placement accuracy

Leave-one-animal-out: each of the 249 sections was placed as a new image, with every section from its own animal removed from the library, on the hemisphere the person had quantified.

Measure	Value
Median arrow-end distance, all sections	137 µm (mean 181 µm)
Whole 10x sections (204)	129 µm; 3.2% of ends over 500 µm
Sections cropped to the striata (45)	166 µm; 10.7% of ends over 500 µm
Previous rule-based engine, whole sections	373 µm; 37.1% over 500 µm
Noise-box background / hand noise-box background	0.99 (IQR 0.90–1.08)

Arrow end	Median	90th percentile
DV start	90 µm	201 µm
DV end	190 µm	416 µm
ML start	121 µm	494 µm
ML end	137 µm	339 µm
MV-DL start	183 µm	384 µm
MV-DL end	122 µm	313 µm

For scale, the hand-drawn arrows are about 3.0 mm (DV), 2.1 mm (ML) and 2.7 mm (MV-DL) long. Mean error by group: crhr cKO 170 µm, crhr control 126 µm, dat cKO 211 µm, dat control 198 µm, ntsr cKO 247 µm, ntsr control 159 µm.

Effect on measurements

All 249 sections were measured at full resolution with the hand ROIs and with the held-out automatic ROIs, both at line width 2 as in Fiji, and both went through Striatum Quant's own profile analysis (noise-subtracted, 30 bins, normalized to the maximum of the mean control curve of the same strain, staining round and arrow). Re-measuring the hand ROIs reproduced the original Fiji CSVs to within 2e-05, so the comparison isolates ROI placement.

Control minus knockout, averaged over the 30 bins (bins with p < 0.05, Welch's t-test per bin):

Strain, arrow	Control / cKO	Hand ROIs	Auto, hand hemisphere	Auto, Auto hemisphere
Crhr1 DV	71 / 41	0.72 (30/30)	0.74 (30/30)	0.74 (30/30)
Crhr1 ML	71 / 41	0.62 (30/30)	0.72 (30/30)	0.72 (30/30)
Crhr1 MV-DL	71 / 41	0.72 (30/30)	0.76 (30/30)	0.73 (30/30)
Dat DV	23 / 42	0.75 (30/30)	0.72 (30/30)	0.75 (30/30)
Dat ML	23 / 42	0.63 (30/30)	0.66 (30/30)	0.67 (30/30)
Dat MV-DL	23 / 42	0.76 (30/30)	0.71 (30/30)	0.72 (30/30)
Ntsr1 DV	25 / 45	0.61 (30/30)	0.48 (30/30)	0.40 (30/30)
Ntsr1 ML	25 / 45	0.38 (27/30)	0.49 (30/30)	0.40 (30/30)
Ntsr1 MV-DL	25 / 45	0.69 (30/30)	0.60 (30/30)	0.47 (30/30)

Group mean profiles along each arrow for controls and knockouts of three strains, with hand ROIs and automatic ROIs nearly overlapping except for Ntsr1 knockouts. — Group mean profiles: hand ROIs solid, automatic ROIs dashed (hand hemisphere, each animal held out).

Per section and arrow, the automatic ROIs' whole-arrow mean lies a median 0.046 from the hand value (90th percentile 0.23); with the Auto hemisphere, 0.057 (0.31). The arrows and the noise boxes contribute about equally: automatic arrows with the hand boxes are within 0.024, hand arrows with the automatic boxes within 0.031. The bin-by-bin shape of a single section does not reproduce (median correlation 0.51), because moving an arrow ~140 µm crosses different TH patches; group curves average this out.

Ntsr1 differs most: its knockouts read more residual TH with automatic ROIs, because the hand noise boxes in those sections sit on brighter tissue than in any other group, which the automatic boxes do not reproduce.

Hemisphere

Which hemisphere a person quantifies could not be predicted from the image, and both striata are usually intact. "Auto" therefore takes the side whose placement is expected to be more accurate, from how well its matches fit and how far they disagree (rank correlation with the actual error 0.46). Where both striata are intact, the two sides' expected errors usually differ by a few micrometres. Auto agreed with the person's choice in 152 of 249 sections.

Sections from other laboratories

Eleven DAB-stained sections from three open-access papers (CC BY 4.0: Sci Rep 2024, Front Pharmacol 2022, Front Aging Neurosci 2026), inverted to look like fluorescence, both hemispheres each, judged by eye by the person whose placements form the library: 11 of 22 hemispheres placed properly at first, 17 of 22 after the striatum refit and mirroring, and judged ready to publish after edge snapping. A caudal section and a heavily damaged hemisphere remain known limits.

Limitations

The library's sections are mid-striatal; at caudal levels the ventral arrow ends can run below the striatum.
Built from far-red or CY3 TH sections of four FloxTH lines. Other stains, magnifications or heavily damaged sections may match poorly; the match level shown says so.
Mirroring assumes a section cut close to symmetrically.
Placement imitates one person's style. It is a starting point to review, not a measurement.

Suggested methods text

ROIs (dorsoventral, mediolateral and medioventral-to-dorsolateral line profiles and three background boxes) were placed automatically with Striatum Quant (Toolkit v2.18), which carries ROIs over from the most similar of 249 coronal TH sections whose ROIs were placed by hand (20 animals): sections are matched by normalized cross-correlation and aligned by enhanced-correlation-coefficient registration, the matches' ROIs are combined by correlation-weighted median, and arrow ends are adjusted to the striatal edge. Held out by animal, arrow endpoints fell a median 137 µm from the hand placements, and the resulting normalized profiles reproduced every control–knockout difference obtained with hand-placed ROIs. Every automatic placement was reviewed and corrected where needed before measurement.iEdit this last sentence to say what was actually done: for example, that every placement was reviewed, or how many were moved by hand before measuring.

Reproducing

The source project's training/striatumHandAtlas-v1/ builds the library from the hand ROI sets and ND2 files (research data, not included) and reproduces every number here; re-run from an empty folder, it rebuilt the shipped library exactly. The model card (striatumHandAtlas-v1.card.md) and the full numbers (validation.json) ship with the site.