Oral AI

Inspiration

Oral cancer is one of the most survivable cancers if caught early — but in India and much of Asia, it isn't. 60–80% of cases are diagnosed at Stage III/IV, when 5-year survival drops to 20–30%, versus over 80% for Stage I/II. India and Asia account for a disproportionate share of the global burden (66% of the world's cases), yet rural screening coverage is under 1% (NFHS-5), even as tobacco and areca nut use remain high in exactly those high-risk states. Team Codeine wanted to close that gap using something almost everyone already has: a smartphone.

What it does

OralAI is a multimodal AI screening tool for early oral cancer detection. A health worker captures a smartphone photo of an oral lesion along with a short clinical questionnaire (demographics, tobacco/areca nut/alcohol/HPV risk, lesion characteristics, prior history). The model fuses the image and clinical context to output:

A malignancy-risk score (e.g., 0.87 High Risk)
A suggested action along a referral pathway — routine review, refer for biopsy, or urgent referral
An automatically generated referral slip and logged case for follow-up/reporting

It's built to deliver specialist-level triage in rural hands — a clear next step, not just a label.

How we built it

Inputs: point-of-care oral images + structured clinical context
Model architecture: an image branch (CNN / Vision Transformer — MobileNet/EfficientNet) and a clinical branch (Random Forest / XGBoost / small feed-forward NN), combined in a fusion layer to produce the final malignancy probability/class
Stack: Python, TensorFlow/PyTorch, scikit-learn, FastAPI backend, with an Android/Web front end for health workers
Workflow: capture photo + questionnaire → guided image-quality check → multimodal AI inference (on-device or cloud) → risk classification + suggested action → automatic referral slip → case logging
Deployment: offline-first with on-device inference for low-connectivity settings, and cloud fallback for sync, model updates, and analytics

Challenges we ran into

Variable image quality across different phones and lighting conditions
Limited, imbalanced annotated datasets for training a reliable model
Patchy rural connectivity, which shaped the offline-first architecture decision
Training/adoption burden on frontline health workers who'd actually use the tool
High clinical stakes of false negatives and false positives, requiring a conservative, referral-based design rather than an autonomous diagnostic one

Accomplishments that we're proud of

Designing around proven real-world feasibility: similar CNN-based mHealth triage has been field-tested with ASHA workers at ~95% sensitivity across 5,025 subjects and 32,128 images
Anchoring the solution to India's existing 1M+ ASHA rural health worker network and 66% smartphone penetration, so deployment doesn't require new infrastructure
Keeping the system scoped to screening/triage + referral, not autonomous diagnosis, with strict specialist sign-off built into the pathway
Modeling the real clinical upside: shifting cases from Stage III/IV (20–30% survival) toward Stage I/II (80%+ survival) — potentially closing the gap between the 34.1% rural and 48.5% urban 5-year survival rates cited from JAMA Network Open (2025)

What we learned

Early detection tools for underserved populations succeed or fail on deployment realism, not just model accuracy — offline-first design and guided-capture UI mattered as much as the fusion architecture
Partnering the AI output with a human referral pathway builds trust and safety into a tool making high-stakes suggestions
Frontline health worker networks (like ASHAs) are an existing, underused channel for scaling AI health tools without needing new clinical infrastructure

What's next for Oral AI

Partnerships for annotated data and clinical validation (e.g., with Hack2Heal, IIFR)
Piloting the guided-capture + triage workflow with ASHA-style frontline health workers
Feeding structured screening data into ICMR-NCRP registries to support public health planning
Extending the multimodal fusion approach to other visually detectable conditions beyond oral cancer
Scaling reach with the rallying call: *"Let's end oral cancer. Partner. Fund. Scale impact."