OraScan - AI Oral Screening Platform
The product definition behind OraScan's 94.7% accuracy across 11 disease categories.
Accuracy Target Set & Met
Training Images Curated
Disease Classes Defined
Tools & Methods
The problem
Most people in smaller Indian cities skip regular dental checkups because specialists aren't nearby and consultations cost too much. The goal was to make AI-powered oral screening available on a simple kiosk - so any clinic, even without a specialist on staff, could flag problems before they became serious.
The challenge
Oral diseases affect more than 3.5 billion people globally but often go undiagnosed in low-resource settings. The product needed to be clinically credible and deployable on kiosk hardware without a GPU. That meant setting accuracy targets for each of the 11 disease categories, defining dataset curation rules for a reproducible training corpus, and choosing a kiosk-first launch with a REST API for future integrations.
Product Requirements Document
PRODUCT REQUIREMENTS DOCUMENT
OraScan - AI Oral Disease Detection
Varun Cumbanungam · AI Product Manager · Oralens HealthCare (2023)
Doc ID
ORS-PRD-V1
Status
Approved
Owner
Varun C.
Date
2023
Version
1.0
Problem Statement
Oral diseases affect 3.5B people globally but go undiagnosed. The product must classify 11 disease classes at ≥94% accuracy on GPU-less kiosk hardware in under 200ms.
Disease Classes & Dataset
- Caries, Calculus, Gingivitis - high-prevalence classes
- Periodontal disease, Oral cancer - critical recall required
- Hypodontia, Mucocele, Ulcer, Fluorosis + more
- 78,058 labelled images (DENTEX + SMART-OM datasets)
- Class imbalance: weighted sampling + CutMix augmentation
Model & Deployment Strategy
- EfficientNet-B0 backbone - 4.67M parameters
- PyTorch + AMD GPU (DirectML) training pipeline
- ONNX INT8 export - edge kiosk CPU deployment
- 23 training iterations tracked in Edge Impulse
Key Acceptance Criteria
- AC-1Overall test accuracy ≥94% across all 11 classes
- AC-2Oral cancer recall ≥97% - no false negatives
- AC-3ONNX INT8 inference on CPU under 200ms
- AC-4Model size under 20MB post-quantisation
- AC-5FastAPI inference server p95 under 300ms
- AC-6Accuracy drop post-INT8 quantisation under 1%
Key Risks
- HIGH
Oral cancer false negative - missed diagnosis
Recall ≥97% gate + clinical review threshold
- HIGH
INT8 accuracy regression on kiosk hardware
Per-class evaluation before deployment sign-off
- MED
Dataset class imbalance skews model
Weighted sampling + confusion matrix gate
Product Artefacts Delivered
- PRD V1 - disease scope, dataset strategy, ACs
- Model evaluation - per-class metrics, confusion matrix
- ONNX deployment spec - INT8 quantisation runbook
- Kiosk integration guide - FastAPI + hardware setup
CONFIDENTIAL · OraScan PRD · Property of Oralens HealthCare
PRD · Model Evaluation · ONNX Deployment Spec · Kiosk Guide
Full PRD and supporting artefacts available upon request
Results
The model shipped meeting the 94.7% test accuracy target across all 11 disease classes. ONNX INT8 quantisation met the <200 ms kiosk inference latency requirement without meaningful accuracy regression. The product was integrated into two workflows - automated scanning (MediaPipe FaceMesh-triggered) and manual desktop scanning - enabling OraScan to serve both high-throughput kiosk and assisted-diagnosis use cases from a single model artefact.
Gallery & Demos
AI Analysis Results Screen
Kiosk interface displaying disease classification output across all 11 oral condition categories with confidence scores.
Web Portal - Clinical Dashboard
Clinician view showing patient scans, historical analysis, and disease progression tracking.
Click any image or video to expand · ← → keys navigate
More from OraLens Healthcare Pvt. Ltd.
OraScan - Oral Disease Detection
An oral disease classifier built with a custom EfficientNet-B0 model trained on 78,000+ dental images. It reaches 94.7% accuracy across 11 disease categories and runs through ONNX on kiosk hardware.
OraScan H - Halitosis IoT Device
Full-stack IoT device for halitosis (bad breath) detection: a Raspberry Pi Zero 2W with H2S gas sensors communicates via BLE to a Flutter mobile app, with on-device TFLite AI classification and PHP backend.
OraScan H - Connected Halitosis Device
Product ownership of OraScan H, a consumer IoT halitosis detector built with a Raspberry Pi Zero 2W, H2S gas sensors, and a Flutter mobile app. Defined the BLE pairing flow and sprint milestones through 92% software completion.
ArogyaLens - Unified Health Platform
A multilingual healthcare platform connecting patient records, doctor workflows, hospital operations, appointments, pharmacy, labs, telemedicine, and oral-health screening.
ArogyaLens - Unified Health Platform
Product ownership for a health platform spanning 15+ modules, 13 languages, telemedicine, and oral-health screening. Authored the PRD, SRS, and technology specification.
LuxySmile Oral Care - Product Website & Dashboard
Full product suite for LuxySmile Oral Care (Oralens Healthcare): a React + Vite marketing site targeting B2B dental wellness programs for schools and corporates, plus an internal case management dashboard with direct-to-S3 uploads and CSV bulk import.
LuxySmile - Oral Care Brand & Operations
Product and go-to-market ownership for LuxySmile Oral Care - OraLens Healthcare's B2B wellness brand. Defined the dual B2B/B2C positioning strategy, school and corporate wellness program inquiry flow, and the AI kiosk upsell path on a premium product website.
Interested in this work?
I can walk through the architecture and code during an interview.