Deep Learning-Based Referral Triage for Childhood Caries Screening Using Smartphone Photographs

Introduction and aims Childhood caries is the most prevalent chronic childhood disease worldwide, and early detection reduces treatment complexity and cost. To address the shortage of paediatric dentists, particularly in underserved settings, this study developed and internally evaluated an end-to-end deep learning system for automatic childhood caries screening using smartphone photographs taken under non-standardized conditions. Methods From an online platform, 463 images (5,279 teeth) from children aged 2 to 12 (2022-2023) were annotated pixel-wise for caries. A multi-scale path aggregation network (MS-PANet) with robust data augmentation was evaluated on an independent internal hold-out test set of 50 images collected from the same platform during a subsequent period (January-June 2024). Its performance was compared with several state-of-the-art object-detection frameworks and with the caries-identification accuracy of parents, dental students, and continuing education program dentists (CEPD) specializing in paediatric dentistry. Results On the development validation set, MS-PANet achieved an mAP@50 of 88.1%, and an mAP@50 of 83.6% on low-quality validation images. On the independent internal hold-out test set, MS-PANet achieved an mAP@50 of 85.4%, with specificity, sensitivity, and AUC of 91.0% (95% CI: 87.9%-94.0%), 80.9% (95% CI: 75.9%-86.0%), and 0.932 (95% CI: 0.890-0.960), respectively; mAP@50 on the low-quality subset of this hold-out set was 80.9%. In an exploratory small-rater comparison, AI showed higher accuracy than parents and dental students, while no statistically significant difference was observed between AI and CEPD raters. Conclusion The AI-assisted, smartphone-based system shows promise as a scalable referral-triage tool for childhood caries screening in underserved and remote paediatric populations, prompting timely referral for definitive diagnosis. Clinical Relevance This smartphone-based system may help caregivers and non-specialist personnel identify photograph-visible childhood caries outside clinical settings. Its performance showed no statistically significant difference from CEPD raters in a small exploratory comparison; larger external and clinical validation is required before implementation.