AI · Healthcare · Medical Imaging

Dental AI
Inference Pipeline

We built a proprietary AI pipeline that analyses intraoral photographs to detect individual teeth, classify dental anomalies, and produce structured per-tooth health assessments for dentist review, deployed end-to-end on cloud infrastructure.

The problem

Dental examination is inherently subjective and time-consuming. A dentist reviewing intraoral photographs must mentally track every visible tooth, assess multiple anomaly types simultaneously, and form a holistic view of oral health, all under time pressure. Early-stage issues like incipient caries or gum inflammation are easily missed when the examination relies entirely on visual inspection of individual images.

Our client needed a system that could analyse photographs programmatically, detecting and tracking individual teeth across multiple frames, classifying the health status of each, and presenting findings in a structured format that supports rather than replaces clinical decision-making.

The AI inference pipeline

We designed and iterated on a multi-stage inference pipeline that turns raw intraoral photographs into reliable, structured clinical outputs. The architecture is organised around a few core phases, each addressing a distinct challenge in moving from unprocessed imagery to a usable per-tooth assessment.

Input preparation

Incoming images are normalised and quality-filtered to ensure a consistent baseline for downstream analysis.

Detection and analysis

Individual teeth are located within each frame and assessed against the relevant clinical anomaly categories.

Aggregation and quality control

Per-frame results are consolidated across the full image set, with validation steps that filter noise and stabilise the final per-tooth verdict.

Structured reporting

Findings are synthesised into a holistic oral health assessment and serialised into a structured format ready for downstream consumption.

Cloud deployment

We led the end-to-end cloud deployment. The inference service is containerised for reproducible, environment-independent rollout and runs on serverless infrastructure, scaling automatically with demand and avoiding idle compute costs. Encryption and access controls are in place to meet the data security expectations of a healthcare context.

A CI/CD pipeline automates the full build and release cycle, so every code change reaches production through a consistent, tested path with no manual deployment steps.

Path to model independence

The pipeline is designed with a clear path toward progressively replacing external model dependencies with in-house trained alternatives, reducing costs, latency, and third-party exposure as annotation data accumulates and performance thresholds are met.

The phasing prioritises the highest-cost and highest-risk dependencies first, sequenced around the pace at which labelled clinical data can be produced.

Clinical collaboration and data quality

AI performance in medical imaging is only as good as the training data, and training data quality depends on annotation consistency. We worked directly with clinical stakeholders to establish labelling standards for each anomaly category, resolving edge cases, defining boundary conditions, and producing alignment materials including decision flowcharts and classification guides. These materials reduce annotator disagreement and ensure the model trains on labels that reflect genuine clinical consensus rather than individual interpretation.