MileyCeph / README.md
Abddeus's picture
Upload 3 files
6000763 verified
|
Raw History Blame Contribute Delete
2.64 kB
metadata
title: OrthoCeph Pro
emoji: 🦷
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false

OrthoCeph Pro v3.4 β€” Accuracy-First Cephalometric Suite

Single-file, browser-based lateral cephalometric tracing. Everything runs client-side: patient radiographs never leave the browser; the only network calls are the ONNX Runtime library and the landmark models (browser-cached after the first download).

Models

  • Engine A β€” HRNet-W32 cepha29 (Aariz dataset), INT8, ~28 MB, loaded automatically from a public model repo. Published validation: MRE 1.24 mm, 83.1% of landmarks within 2 mm. All 29 output channels used.
  • Engine B (optional) β€” HRNet-W48 converted from MFouda128/Cephalometric-AI-Models, weight-INT8, 63 MB (orthoceph_engine_b_int8.onnx). The checkpoint's own recorded validation: MRE 1.10 mm, 87.6% within 2 mm. Its undocumented channel order was verified landmark-by-landmark against Engine A on two real films before integration (molar channels split by upper/lower position; R/LPM/UPM unused). To enable: upload orthoceph_engine_b_int8.onnx into this Space next to index.html β€” the app then finds it automatically. Or paste any direct .onnx URL in the app, or load the file per session. When loaded, Auto-Trace averages both engines per landmark and lowers confidence where they disagree (shown as A↔B Ξ” in the landmark list). Engine B expects grayscale input; the app grayscales automatically.
  • Steady mode (TTA) β€” Engine A runs normal + mirrored passes, averaged; disagreement lowers confidence.

Accuracy rules

  • No fake numbers β€” a measurement is shown only when every landmark it needs exists. Wits uses a real functional occlusal plane.
  • Calibration gate β€” mm values are flagged UNCALIBRATED until a 10 mm ruler calibration is done; calibration recalculates all mm values live.
  • Confidence + verification β€” per-landmark confidence, Review Queue (worst first), Verified workflow, JSON records storing AI-vs-final positions as a validation dataset.
  • Analyses β€” ABO Focus, Steiner, Downs, Tweed, McNamara, Ricketts, Holdaway, Jarabak (~48 measurements, 30 landmarks).
  • Print report β€” clean black-and-white report: patient header, diagnosis, traced film, all analysis tables.
  • T1/T2 superimposition β€” cranial base, maxillary, mandibular.
  • DICOM β€” uncompressed DICOM supported.

Note

Open this Space over its https address. Opening the file directly from a phone's file manager (content:// / file://) can restrict WebAssembly and model downloads β€” manual tracing still works there.