work / cervical-lens

CervicalLens

Edge cytology screening plus genomic risk stratification for cervical cancer, with a clinician portal, an offline mobile app and an API.

Open cervicallens.hallelx2.com How it works The backend
Years
2026
Status
research
Live
cervicallens.hallelx2.com

What it is

Two models for two different moments. The first screens Pap-smear cell images on a low-end phone with no connection: a student model distilled from a larger teacher, exported to ONNX at 0.28 MB, 3.6 ms per image on CPU against 14.4 ms in PyTorch. It flags a slide when p(abnormal) ≥ 0.50. On the held-out set that catches 96% of abnormal slides and clears only 40% of normal ones (30 of 50 normals flagged), so a positive is a referral to a pathologist, not a diagnosis.

Read more

The second is for confirmed patients. Two regularised Cox models on TCGA-CESC data, a virulence axis and a metastasis/APOBEC axis, are each split at the median into a 2×2 matrix from LOW to CRITICAL. Combined C-index 0.655; CRITICAL against LOW has a hazard ratio of 2.2 (p = 0.010).

Both surface in a Next.js clinician portal, an offline-first Expo app and an API. Research use only.

Stack

  • Python
  • PyTorch
  • ONNX
  • lifelines
  • TCGA-CESC
  • TypeScript
  • Next.js
  • Expo

Counted

  • 0.28 MBmodel size
    ONNX export
  • 3.6 msCPU latency
    vs 14.4 ms PyTorch
  • 96% / 40%sensitivity / specificity
    held-out set
  • 0.655genomic C-index
    95% CI 0.588–0.713
  • 2.2hazard ratio, CRITICAL vs LOW
    p = 0.010

How it works

The system as a machine on a bench. Hover or use the arrow keys to drive it.

0.28 MB
Needle: the 0.28 MB student's p(abnormal) against the 0.50 markSlides: a dot on the label is abnormal on ground truth, a ring is normalCoin columns: TCGA-CESC patients per risk cell (91, 54, 56, 93)Research use only, not a clinical device: a distilled 0.28 MB student reads each slide in 3.6 ms on CPU and flags p ≥ 0.50, catching 96% of abnormal slides but sending 60% of normal ones to the pathologist too (specificity 0.40, AUC 0.84; the tour's slide readings are illustrative, its rates the held-out test set's). A confirmed case is then placed on two Cox axes from TCGA-CESC, virulence (C 0.591) and metastasis (C 0.645), split at their medians into LOW to CRITICAL (combined C 0.655, HR 2.2 CRITICAL vs LOW, p = 0.010).

The path through it

  1. CaptureA Pap-smear cell image on the phone, offline.
  2. ScreenThe distilled 0.28 MB ONNX model, 3.6 ms on CPU; p(abnormal) ≥ 0.50 flags the slide.
  3. ReferPositives go to a pathologist review queue.
  4. StratifyExpression, mutations and HPV genotype feed two Cox axes: virulence (C = 0.591) and metastasis/APOBEC (C = 0.645).
  5. Risk matrixMedian split on each axis → LOW, MODERATE, HIGH, CRITICAL.