Head Lice Checker Methodology

Our screening flow is intended to reduce uncertainty, not to replace a clinician. The main purpose is to help parents decide whether a professional confirmation should be prioritized and to reduce delayed action when multiple risk indicators are visible.

The system works best on close-up scalp photos where hair is parted and lighting is strong. It is not optimized for distant portraits, heavy motion blur, or obstructed views. In low-quality conditions we keep language conservative and encourage a better re-scan.

Outputs are therefore written as indicative guidance. We avoid definitive medical claims and direct users toward qualified clinical support whenever confidence is low, symptoms persist, or findings appear significant.

When a user uploads an image, the file is checked for validity and minimum dimensions before screening. This protects model stability and avoids confidence distortions caused by very small or corrupted inputs.

The same prepared image variant used for inference is also used in result rendering. This alignment matters because detection overlays are projected using the source coordinate system, and visual trust depends on marker placement matching the displayed frame.

Preprocessing is intentionally minimal: we do not apply aggressive enhancement pipelines that could alter visual artifacts. Users are encouraged to provide better source captures rather than relying on synthetic sharpening.

Provider outputs can include nested prediction objects, class aliases, and mixed confidence keys. We normalize these structures into a stable internal detection array so frontend behavior remains consistent as long as providers return valid bounding metadata.

Class aliases are mapped into standardized labels such as lice and nits. We reject malformed detections, enforce numeric coordinate validation, and apply a minimum confidence floor before items are exposed in UI summaries.

For each accepted detection we retain center coordinates, width, height, raw confidence, and tier assignment. This normalized object powers evidence overlays, counts, summary chips, and analytics instrumentation.

Confidence tiers are grouped for parent-facing clarity. High confidence indicates strong model evidence, medium confidence indicates useful but less certain evidence, and low confidence signals that image quality or visual ambiguity may affect interpretation.

When valid detections exist, top-level result labels are determined by the strongest detection class. When detections are absent, fallback logic returns a clear-screening outcome with supportive language and re-check guidance if symptoms continue.

This policy avoids overpromising while still giving clear direction. Users can inspect marker overlays and summaries before deciding whether to seek clinic confirmation.

Result cards prioritize evidence-first communication. We surface the uploaded photo, detection markers, total indicator count, and confidence tier in a compact hierarchy so users can understand why a suggestion is being shown.

Guidance text then translates output into practical next actions: monitor, rescan with better lighting, or request clinic follow-up. We deliberately avoid alarmist language and avoid diagnostic wording.

By combining visible evidence with calm copy, the methodology supports trust while keeping decisions grounded in professional confirmation when appropriate.

We track anonymized interaction events to improve reliability and user comprehension. Examples include overlay toggles, legend filtering, scan retries, and clinic CTA engagement.

These signals help us evaluate whether users understand results and whether next-step guidance is effective. Improvements are prioritized when we detect confusion patterns such as immediate repeated rescans after strong positive signals.

Method updates are reviewed before release and documented in content timestamps so users and partners can see when guidance standards were last updated.

Method changes follow a release checklist that validates parser stability, confidence tier mapping, and user-facing language consistency before deployment. This helps prevent regressions where back-end detection structure changes could silently affect frontend interpretation.

Each release includes deterministic tests around nested prediction extraction, alias mapping, and malformed detection rejection. We also verify overlay coordinate rendering against the exact image variant used for inference so trust is preserved in evidence view.

When major behavior changes are introduced, content guidance and FAQ references are updated in the same cycle so user education remains aligned with product behavior.