<b>AI-Powered Dental Image Screening for Early Detection of Oral Conditions: A Comparative Evaluation of Lightweight CNN Architectures for Low-Resource Settings in Nigeria</b>

Dental health remains a critical yet underserved aspect of healthcare in low-resource settings, particularly in Nigeria, where access to timely diagnosis is limited by cost, distance, and lack of awareness. This study presents Oral Sight, an AI-powered dental image screening prototype that evaluates the feasibility of lightweight convolutional neural network (CNN) architectures for classifying oral health conditions from patient-uploaded images. Three models were trained and compared on a publicly available oral lesions dataset (n=261): a custom baseline CNN, MobileNetV2, and ResNet50. MobileNetV2 achieved the highest macro F1-score of 0.54 on the held-out test set (n=40), outperforming the baseline CNN (F1: 0.24) and ResNet50 (F1: 0.18), demonstrating its superior trade-off between accuracy and computational efficiency for low-resource deployment.

Grad-CAM visualizations indicated that the best-performing model appeared to focus on clinically relevant oral tissue regions during classification. Results indicate that while transfer learning significantly improves performance over baseline approaches, dataset size remains the primary bottleneck, highlighting the critical need for larger, locally sourced dental image datasets in Nigerian clinical settings. All code, notebooks, and reproducibility documentation are publicly available at: github.com

Note that comparative model selection was conducted on the validation set, while the headline F1 of 0.54 reflects final evaluation of the best-performing model on a strictly held-out test set.

This work was conducted under the supervision of Winner Emeto, DataraFlow (info@dataraflow.com) as part of the DataraFlow Internship Programme. Intern ID: DF2025-074