# Examination of new clinical dental caries in school children using real intra oral photos with artificial intelligence model YOLO-V8x

## TL;DR

This study evaluates an AI app using YOLOv8x to detect dental caries in children's photos, showing performance comparable to dentists but with faster diagnosis.

## Contribution

The novel use of YOLOv8x in a mobile app for dental caries detection in school children, with performance benchmarked against human dentists.

## Key findings

- YOLOv8x achieved 45.8% mAP, 72.6% precision, and 41.1% recall in detecting dental caries.
- AI diagnosis was four times faster than manual diagnosis by dentists.
- Sensitivity and specificity for caries detection ranged from 82.31% to 99.33% across ICDAS categories.

## Abstract

Dental caries is a common chronic condition among school children, and in Indonesia, rates have been higher than 80%. Traditional diagnostic techniques tend to be long and dependent upon the availability of healthcare professionals and hence promote development of alternatives involving artificial intelligence (AI). This work aimed to evaluate the effectiveness of the application titled HI Bogi that embeds the model known as YOLOv8 in distinguishing dental caries among primary school pupils in Cimahi, Indonesia.

A cross-sectional analytic model with prospectively collected data was adopted. A dataset consisting of 3,221 JPG photographs was created and labeled based on the International Caries Detection and Assessment System (ICDAS, D0 - D6) using Roboflow software. The photographs were resized to a size of 640 × 640 pixels and distributed into training (2,266 photographs), validation (635 photographs), and testing (320 photographs) sets. A YOLOv8x algorithm was used for application in deep learning tasks, and its performance was determined through mean Average Precision (mAP), Intersection over Union (IoU) measurements, and precision and recall metrics. Additionally, a Mann Whitney statistical test was carried out to contrast classification efficiency between the AI system and dental practitioners’ classification efficacy, while both methods’ diagnostic speed was tested.

mAP was 45.8%, and precision was 72.6%, and recall was 41.1% for YOLOv8x model. Comparative testing across ICDAS categories did not vary significantly between dentists and AI (p > 0.05). Detection of caries (D1–D6) was observed to have a sensitivity ranged from 82.31% to 96.45% and specificity ranged from 77.10% to 99.33% on new data. Additionally, testing was completed approximately four times faster with AI compared to manual testing (p = 0.000).

The YOLOv8x model in the HI Bogi app demonstrated diagnostic performance equal to dentists and a drastically shortened examination time, validating its use in school-based dental health programs. But class-specific precision variation suggests further refinement is needed. Further studies ought to increase datasets and characterize advanced architecture, such as transformer-based architecture, to achieve higher specificity and detection rates in rare lesion categories.