Pearl Dental Paris: Dentaline

CASE STUDY · HEALTHCARE & LIFE SCIENCES

Deep Learning at the service of dental diagnosis.

Automated identification of teeth, anomalies and implants on panoramic X-rays: IT Nights 2022 Silver Medal, HDS and GDPR compliant.

Client

Pearl Dental Paris

Expertise

Computer Vision · Deep Learning · Product co-construction

Tech Stack

TensorFlow · Python · Next.js · GraphQL · Azure · MongoDB

Engagement

5 experts · 14 months

CONTEXT

Project Context

Pearl Dental Paris wanted to improve its diagnostic efficiency in the face of a growing patient volume. The challenge: identify teeth, anomalies and implants quickly and precisely on panoramic radiographs, while offering patients clear visualisations of their oral health and reducing practitioners' analysis time.

Manual periodontal charting took up to 2 minutes per patient and generated 35% of data-entry errors: time lost, and clinical information weakened over the long term. With a growing activity, this friction was becoming a direct obstacle to care quality and practice capacity.

The stake: design a Deep Learning solution that is reliable, fast and HDS / GDPR compliant, and that fits naturally into the existing practice workflow, without replacing the medical decision but enhancing the precision, traceability and peace of mind of the practitioner.

Strategic Objectives

OBJECTIVES

(01)

Automated analysis

Design a Deep Learning system able to analyse panoramic radiographs and precisely identify the 32 dental elements (FDI), anomalies (caries, lesions, fractures) and implants, with a clinical accuracy compatible with practice use.

(02)

Diagnostic support

Provide a detailed charting report to assist practitioners in their diagnosis and treatment plan, with high accuracy, native explainability and viewable bounding boxes, without ever replacing the medical decision.

(03)

Patient and practice experience

Improve the patient experience through clear visualisations of the oral situation and give dentists a significant time saving (from 2 minutes to 5 seconds per charting), a massive release of clinical capacity.

Solutions Delivered by Adservio

SOLUTIONS

Adservio co-built Dentaline with Pearl Dental Paris (Data Scientists, ML Engineers, Product Designers, HDS expert) over 14 months.

(01)

Data processing and labelling

Labelling and classification of 3,250 panoramic radiographs by expert practitioners, with precise clinical tagging (FDI, anomalies, implants). An inter-rater agreement protocol was put in place to strengthen the reliability of the training set.

(02)

Deep Learning architecture

Training of ResNet-101 + Faster R-CNN models in supervised and semi-supervised mode. Optimisation of confidence thresholds, NMS parameters and calibration to produce a robust initial model across the morphological diversity of patients.

(03)

Clinical evaluation cycles

Iterative cycles of testing, validation and improvement until a reliable final model was reached (F1-score 0.94, accuracy 96%). Continuous clinical feedback with Pearl Dental practitioners on true positives, false positives and edge cases.

(04)

Dentaline application and UX

Design of a Next.js + GraphQL web application with interactive clinical visualisations: bounding boxes overlaid on the radiograph, auto-generated FDI charting, exportable PDF report, integration into the practice EHR, all without adding friction to the workflow.

(05)

HDS and GDPR compliance

Certified Health Data Hosting (HDS), native anonymisation of PII before inference, AES-256 encryption at rest and in transit, immutable audit log of every diagnosis. GDPR compliance by design and aligned with the European regulation on AI in healthcare (MDR class IIa).

From a panoramic X-ray to a full charting.
In 5 seconds. F1-score 0.94.

The Faster R-CNN model identifies the 32 teeth (FDI), anomalies (caries, lesions) and implants in a single pass. Bounding boxes, confidence scores, structured charting, encrypted audit log: a complete diagnosis, HDS and GDPR secured, ready to be validated by the practitioner.

Results

MEASURED FIGURES

−95%

charting time

Recording a periodontal charting cut from 2 minutes to 5 seconds per patient.

96%

model accuracy

Clinical accuracy of the final model, for a 0.94 F1-score.

35% → 1%

error rate

Data entry errors on charting collapsed, making the medical record more reliable.

3,250

annotated X-rays

Panoramic radiographs labelled by expert practitioners.

dental elements

Teeth (FDI), anomalies and implants identified in a single pass.

IT Nights 2022

Silver Medal

Winner of the Co-construction Challenge, with a CAC 40 CIO jury.

Impact

WHAT IT CHANGES

Accelerated periodontal charting

−95% in time to record a periodontal charting, from 2 minutes to 5 seconds per patient. The practitioner validates what the AI proposes, rather than entering data by hand.

Augmented accuracy

The data-entry error rate drops from 35% to 1%, a major qualitative leap in the longitudinal follow-up of patients and the reliability of the medical record.

IT Nights 2022 Silver Medal

Winner in the "Co-construction Challenge" category, with a jury made up of the CIOs of Vinci, L'Oréal, LVMH, Stellantis, BNP Paribas and EDF, an industry recognition of the consultancy/practice partnership approach.

Reliable visual diagnosis

Complete and detailed detection of teeth, anomalies and implants in seconds, with an actionable report for the practitioner and bounding boxes that overlay on the original radiograph for diagnostic traceability.

Native HDS and GDPR compliance

Certified HDS hosting, anonymisation of patient data before inference, immutable audit log. Full compliance with the French and European healthcare regulatory framework, validated during the DPO audit in 2023.

Clinical capacity freed up

The time saved on charting is reinvested in care: chair productivity rose without pressure on practitioners, and the patient experience is noticeably smoother.