A Vision Transformer that diagnoses dental conditions at 92.4% accuracy, in under a second.
Dental diagnostics are subjective, expensive, and inaccessible in underserved regions. A dentist's assessment of caries or gingivitis from an intraoral photo varies by experience and fatigue. DentaVision applies a pretrained Vision Transformer (ViT) to standardize and accelerate dental diagnosis, providing consistent, explainable classifications from dental images with sub-second inference via a clinical-ready API. The ViT architecture was chosen over CNNs for its global attention mechanism, which better captures the distributed visual patterns of dental pathology.
A pretrained ViT-Base/16 model is fine-tuned on a curated dental image dataset. The attention mechanism allows the model to focus on distributed caries and gingivitis indicators across the entire image.
The ViT backbone is frozen; only the classification head and final transformer blocks are fine-tuned. Data augmentation (rotation, brightness, contrast jitter) expands the training distribution.
A FastAPI service wraps the trained model, accepting base64-encoded images and returning classification predictions with confidence scores in JSON. Designed for clinical EMR system integration.
A React interface allows clinicians to upload dental images and receive instant diagnosis with confidence breakdown, designed for low-friction clinical workflow integration.
92.4% classification accuracy on caries and gingivitis detection
Sub-second inference latency suitable for real-time clinical use
FastAPI backend designed for EMR/EHR system integration
ViT attention maps provide interpretable visual explanations for clinicians