CervicalLens

AI Screening & Prognosis System · Low-Resource Settings

DATICAN AI-in-Medicine Competition Entry

Re-engineering Cervical Cancer Triage for Low-Resource Environments.

An end-to-end clinical workflow integrating offline cytology screening at the edge with prognostic genomic risk stratification to connect field workers, pathologists, and clinicians.

Edge Inference

3.6 ms

Offline CPU latency on low-end mobile devices (vs 14.4 ms PyTorch).

Model Footprint

0.28 MB

Distilled MobileNetV3-Small ONNX model running natively on-device.

Screening Filter

96.0%

Sensitivity achieved on held-out test cytology sets for high-security triage.

Prognostic Survival

2.2 HR

Hazard Ratio of CRITICAL vs LOW risk categories (Cox model, p = 0.010).

Pillar 01

Primary Screening at the Edge

Cytology screening at the point of care is throttled by laboratory shortage. CervicalLens deploys a distilled student MobileNetV3 model directly onto native field-worker devices, allowing community health workers to perform real-time offline classifications of Pap-smear slide tiles without an active internet connection.

Distillation Loss Formula

L = α T² D_KL(softmax(Z_s/T) || softmax(Z_t/T)) + (1-α) L_CE(y, Z_s)

Knowledge distilled from pathology foundation models (UNI, CTransPath) into a highly optimized edge binary.

Pillar 02

Prognostic Genomic Risk Stratification

Once cervical cancer is confirmed, patient clinical outcomes are driven by host-viral genetics. CervicalLens computes Progression-Free Interval (PFI) risk using a 2-axis survival prognostic model evaluated on the real TCGA-CESC patient cohort.

2D Risk Classification Matrix

Metastasis Low
Metastasis High
Virulence Low
LOW RISKSurveillance
MODERATEMild hazard
Virulence High
HIGH RISKHigh hazard
CRITICALHR = 2.2
Genomic Sensitivity Check

Disproving the APOBEC3B Survival Contradiction

A published clinical study (PMC10076974) claimed that high expression of the APOBEC3B enzyme led to worse overall survival. CervicalLens ran robust sensitivity audits on the same TCGA-CESC cohort and successfully disproved this claim, identifying a critical label inversion bug in their codebase.

01. FIGO Adjusted

Adjusting for age and FIGO stage reveals APOBEC3B is actually highly protective (HR 0.66–0.72, p=0.0008).

02. Quantile Scan

A complete scan of expression quantiles (10% to 90%) failed to yield a single threshold predicting worse survival.

03. Group Inversion

Flipping the high/low group labels reproduced their published KM curves and log-rank statistics exactly.

Hosted System Architecture

Turborepo workspaces deployed and active for NACOS × DATICAN evaluation.

Clinician Console

Next.js clinician management dashboard tracking patient codes, pap-smear tiles, and genomic risks.

Open Web Portal →

Hono API Backend

Serverless REST API (Vercel Node.js 22) conducting live model inference in under 1 microsecond.

Verify API Health →

Field Worker App

Expo React Native client (v0.1.0) with local file system caching, camera specimen uploads, and offline support.

Download v0.1.0 APK Archive →