Adaptive local contrast
A pixel must differ from its own neighbourhood inside the calibrated, unmasked iris annulus.
IRISPECTRAUpload photos →IRIS COMPUTER VISION · PROTOCOL 0.2
Measure the object before naming the pattern.
A professional detector must know where a structure begins, whether its boundary closes, what it touches and how certain it is. Regional darkness alone cannot distinguish a crypt, lacuna, pigment mark or shadow.
Open iris measurement00 / CURRENT LIVE BASELINE
Object candidates are now explicit.
Connected components
Each surviving region becomes a separate object with its own centre, area and contour proxy.
Geometry vector
Elongation, circularity, radial alignment and collarette distance support a provisional shape family.
Confidence ceiling
Confidence is capped and weak evidence remains unresolved because pigment, shadow and tissue depth are not fully separated.
LIVE MODEL 0.5
This is an interpretable classical computer-vision baseline, not a trained clinical classifier. Its purpose is to create inspectable candidate masks and measurable features for the expert-labelled dataset that the later instance model will require.
01 / PROCESS
Eight validation gates.
Acquisition gate
Focus, glare, visible annulus, scale, camera metadata and repeat-image consistency.
Anatomy segmentation
Separate pupil, limbus, eyelids, eyelashes and specular reflections before texture analysis.
Coordinate model
Create a polar iris strip while preserving a reversible map to the original pixels.
Collarette contour
Estimate the irregular collarette independently rather than assuming a perfect circle.
Pattern instances
Give each crypt, lacuna, furrow and pigment candidate its own mask and confidence.
Shape + topology
Measure closure, area, circularity, elongation, skeleton branches, adjacency and collarette attachment.
Calibrated inference
Report class probabilities, model version and an unresolved state for weak or unfamiliar evidence.
Locked validation
Evaluate on unseen participants, cameras and acquisition conditions before releasing a label.
02 / PHENOTYPE TAXONOMY
Shape is measured. Meaning is separated.
observed
Pattern names describe visible form. Interpretations inherited from historical iris maps remain a separate historical claim layer; they are not diagnoses or validated organ findings.
| Candidate | Operational definition | Required evidence |
|---|---|---|
| Closed lacuna | One opening whose contour converges and closes. | closure · contour continuity · hole count |
| Asparagus-like | Solitary, collarette-attached and elongated with an outward-pointing tip. | attachment · tip direction · elongation |
| Three-adjacent configuration | Three distinct neighbouring crypt or small-lacuna instances at the collarette edge. | instance count · adjacency graph · collarette distance |
| Double lacuna | Two attached closed instances with comparable size. | count · attachment · closure · size ratio |
| Leaf-like | Usually closed and collarette-attached with a leaf contour and possible internal branches. | shape · attachment · internal skeleton |
| Circular lacuna | A small encapsulated near-circular opening associated with the collarette. | circularity · closure · scale · attachment |
| Contraction furrow | A curvilinear or annular groove with tangential continuity. | arc length · skeleton · radius consistency |
| Unresolved | Quality is insufficient, evidence conflicts, or the phenotype is outside training data. | quality gate · OOD score · uncertainty |
03 / MODEL STACK
Interpretable before impressive.
open-iris + independent NIST benchmark
Iris, pupil and occlusion geometry; visible-light domain validation is required.
CVAT + SAM 2 assistance
Expert masks, attributes, review states and adjudicated corrections.
nnU-Net + instance segmentation
Semantic benchmark plus separate crypt, lacuna, furrow and pigment instances.
OpenCV + scikit-image
Contours, region properties, skeletons, holes and adjacency graphs.
04 / RELEASE CRITERIA
A label is not released because it looks plausible.
Iris / pupil / collaretteDice · IoU · boundary distance
Crypt / lacuna / furrow instancesmask AP by class · precision · recall · count MAE
Double / triple groupingexact group accuracy · graph precision / recall
Expert reliabilityCohen/Fleiss κ · adjudication rate
Probability qualitycalibration error · selective-risk curve
Repeatabilitysame-eye repeat ICC · localisation error
Domain robustnesscamera · iris colour · glare · occlusion strata
METHOD BOUNDARY
A two-dimensional photograph can provide a photometric depth proxy, not true anatomical depth. Confident pattern subtypes require a locked participant-level test set, repeat-image reliability and calibrated uncertainty. The system abstains when those conditions are not met.