Optional burned-in text / PHI detection (phi-text-detector)

[luckfamousa]

What

Tag-based de-identification never touches pixel data, so PHI burned into images (name overlays, ultrasound banners, secondary-capture screenshots) survives. This adds an optional detector to flag such images.

New workspace crate phi-text-detector — a PaddleOCR DB text detector (detection only, no OCR) via ONNX Runtime (ort):

• onnx_detector — PP-OCR preprocessing (aspect-preserving resize to multiples of 32, ImageNet normalization, NCHW) + ONNX session (behind a Mutex so the detector is Send + Sync with a &self detect()).
• db_postprocess — pure-Rust probability-map → bitmap → connected components → scored/expanded/rescaled boxes (heavily unit-tested on synthetic maps).
• dicom_render — render first frame + extract screening metadata.
• image_prefilter — margin crops + a contrast gate.
• screening — Safe/Review/Unsafe from detection + metadata (BurnedInAnnotation, modality, secondary-capture SOP class); fail-closed.
• phi-screen CLI — screen images / DICOM / directories, JSON or JSONL.

Optional integration into the de-id CLI (the requested behavior)

Behind a text-detection feature (off by default, so the core library never pulls in ONNX Runtime):

• --detect-text off (default) — no detection
• --detect-text warn — flag on stderr, still writes the de-identified file
• --detect-text skip — writes no output for flagged images, exits 3

Model

Not committed. scripts/fetch-model.sh downloads and SHA-256-verifies ppocr_det.onnx (source + hash recorded in models/ppocr_det.metadata.json). CI fetches it.

Tests

• DB postprocessing, screening rules, prefilter — unit tests.
• Real ONNX detection on committed fixtures (positive_text.png → text; negative_no_text.png → none).
• De-id CLI policy (skip exits 3 + no output; warn writes + warns; clean image passes) via BurnedInAnnotation.
• CI runs the workspace + the text-detection integration; verified locally against rustc/clippy 1.95.

Notes

• This first increment renders the first frame and screens the full image (margin-first crops are available in image_prefilter but not yet wired as the default scan strategy).

🤖 Generated with Claude Code