Local Evidence Base
Grounded in Ugandan clinical realities with validated video-powered diagnostics.
Grounded in Ugandan Clinical Realities
A diagnostic decision-support system cannot be deployed on marketing claims; it must be backed by rigorous architectural validation. The UGDent platform was trained and evaluated on an extensive curated dataset of over 10,500 clinical dental images across six primary pathology categories (calculus, caries, gingivitis, mouth ulcers, tooth discoloration, and hypodontia).
Dataset Overview
Comprehensive Clinical Dental Imaging Dataset
Pathology Categories
Model Benchmarking
Evaluating Three Distinct Computational Architectures
EfficientNet-B3
91% Accuracy
42.4 MB • Highly sensitive to motion blur and shadows
BaselineCLIP Vision Transformer
87% Accuracy
600 MB • Robust to noise, occlusion, and lighting variations
ProductionMultimodal VLM
88% Accuracy
1.9 GB • 5s latency • Premium research tool
ResearchModel Performance Comparison
Accuracy vs. Memory footprint trade-offs across architectures
Implementation Methodology
How to Pilot UGDent
For tele-dentistry platforms, NGOs, and district health centers seeking to upgrade their screening capabilities, we provide a structured, low-risk Phased institutional pilot program.
Integration & Training
Connecting your clinical application to our FastAPI WebSocket backend. We provide localized training to community health workers on optimal smartphone camera sweeping techniques.
Shadow Screening
Health workers utilize the React Native mobile app in a shadow capacity. The CLIP ViT model processes video frames and generates diagnostic predictions, logged and compared against manual diagnoses.
XAI Clinical Auditing
Dentists review the Explainable AI attention maps to verify accurate lesion localization and rule out hallucinations based on lighting artifacts.
Full Deployment
Synthesis of pilot data into an impact report. The API transitions to active triage support, enabling confident referral of high-risk patients.
Understanding System Limitations
In clinical health-tech, overpromising is an ethical risk. We maintain total transparency regarding UGDent's current operational constraints.
Dataset Demographics & Imbalance
Class imbalance exists (e.g., Hypodontia under-represented). African intraoral demographics need expanded representation.
Multimodal Text Constraints
VLM constrained by absence of structured, clinically validated textual annotations in public datasets.
Hardware Dependencies
Sub-2-second inference currently requires GPU-accelerated cloud. Offline deployment requires aggressive quantization.
Immediate Next Steps
- Curate proprietary African intraoral dataset
- Apply Knowledge Distillation for edge deployment
- Expand multimodal clinical annotations
Ready to Pilot UGDent in Your Clinic?
Join our phased pilot program and experience video-powered dental diagnostics.