Every 105 minutes, a woman in Ethiopia dies of a cancer that is almost entirely preventable. We built this because waiting for a pathologist shouldn't be part of the diagnosis.
CerviScan AI is a cervical cancer screening intelligence platform — part live data observatory, part working diagnostic model. It tracks the epidemic through WHO and IARC data, reads Africa's news and research as it happens, and runs a trained Pap-smear classifier entirely in your browser.
No servers. No uploads. No patient image ever leaves your device.
In Ethiopia, cervical cancer is the second most common cancer in women: ~8,159 diagnosed and ~5,007 die every year, 80%+ found at an advanced stage, fewer than 1 in 25 ever screened. Africa carries ~1/4 of global cases but ~1/3 of global deaths.
We can't ship pathologists. But a 4.3M-parameter model that runs in a $50 phone's browser fits in a clinic tomorrow.
| Feature | What it does | |
|---|---|---|
| 🩺 | Live Demo | Upload a Pap smear cell — or twenty in batch mode — and get a classification, confidence score, and triage action in milliseconds. Non-cytology images are automatically rejected. |
| 📤 | Exports | Batch results as CSV / JSON — file, prediction, confidence, gate status, action, and rejection audit. |
| 📊 | Data Hub | Curated WHO / IARC epidemiology: the human clock, the screening gap, the 90·70·90 elimination tracker — every figure linked to its primary source. |
| And Many More.., |
EfficientNet-B0 (timm) + custom head · 4.3M params · trained on SIPaKMeD (4,049 cells, 80/20 stratified split) · serving as fp16 ONNX, 8.3 MB.
| Metric | Value |
|---|---|
| 5-class accuracy | 98.40% · macro F1 0.9840 |
| Binary triage accuracy | 99.51% |
| Abnormal sensitivity (triage) | 99.38% — 3 of 487 missed |
| Normal specificity | 99.69% |
| Best class | Parabasal — 157/157 perfect |
The number we're most proud of isn't the accuracy. At the decision that actually matters in a clinic — normal vs. needs-a-human — the system misses 3 in 487. And every prediction below 70% confidence is routed to mandatory human review, by design. The workflow was built around the model's weaknesses, not just its strengths.
- ✅ Export verified on the full validation set — fp16 ONNX is numerically identical to the PyTorch model (max diff
1.3e-07). - 🗂 The checkpoint describes itself — class order, normalization stats, and image size are embedded. No tensor archaeology for the next person.
- 🔒 Privacy is architectural, not a policy — inference runs in WASM on the visitor's machine. There is no backend to trust, because there is no backend.
- ⛔ Non-medical images are rejected, not guessed — two-layer validation (image statistics + model-confidence floor).
git clone https://github.com/ethioel/cerviscan-ai.git
cd cerviscan-ai
python -m http.server 8000Then open http://localhost:8000 — first demo visit downloads the 8.3 MB model from Hugging Face; after that it's browser-cached.
The full pipeline — data ingest, augmentation, training, early stopping, Grad-CAM, verified ONNX export — lives in the Kaggle notebook. Edit the dataset a little bit, enable GPU, Save & Run All: reproducible end to end on free hardware.
- Data sources: WHO · IARC GLOBOCAN · ICO/IARC HPV Centre · WHO AFRO · FMOH Ethiopia · NCBI PubMed
- Dataset: SIPaKMeD — with gratitude to its creators
- Runtime: onnxruntime-web · Chart.js
This is a research prototype, not a medical device. The model was trained and validated on isolated cells from one public dataset — it has not seen whole slides, other scanners, or real clinical populations. Its confidence gate and not-a-cell rejection layer exist precisely because of that. Epidemiological figures are approximations from the cited primary sources; verify them there before citing them anywhere that matters.
If this work speeds up one referral, or keeps one woman from being diagnosed late — that's the whole point.
If this project helps you, a star helps it reach someone else.
