Diagnostic Accuracy of Deep Learning for Detection of Incipient and Advanced Radiographic Interproximal Caries in...

This study aimed to develop and compare deep-learning models for detecting proximal carious lesions of different radiographic severity on mixed-dentition periapical radiographs.

A retrospective diagnostic accuracy study included 1,838 digital periapical radiographs from pediatric patients in the mixed dentition stage. Two calibrated examiners independently annotated proximal radiographic lesions using bounding boxes (inter-examiner κ=0.89; intra-examiner κ=0.92), with disagreements resolved by consensus and specialist adjudication. Lesions were classified as incipient or advanced based on radiographic extent. Images were partitioned at the patient level into training (70%), validation (15%), and held-out internal test (15%) sets. Faster R-CNN, YOLOv11x, and transformer-based DEIMv2-X were evaluated using lesion-level precision, recall, F1-score, mAP, and AUPRC. Single-scale, multi-scale, and test-time augmentation (TTA) configurations were evaluated.

On the held-out internal test set, DEIMv2-X achieved the highest overall Macro-F1 (0.538) and mAP50 (0.486) using the multi-scale ensemble with TTA, whereas YOLOv11x achieved the highest overall recall (0.561) using the same configuration. Detection performance was consistently higher for advanced than incipient radiographic lesions. In binary lesion detection, DEIMv2-X multi-scale ensemble with TTA achieved the highest F1-score (0.630). External validation showed reduced performance across all architectures, with YOLOv11x demonstrating the highest overall Macro-F1 and precision.

Performance varied across architectures, inference configurations, and datasets. DEIMv2-X achieved the highest overall balance between precision and recall in the internal evaluation, whereas YOLOv11x achieved the highest overall Macro-F1 and precision on the independent external dataset. Incipient lesions remained challenging to detect, highlighting the need for further validation across diverse populations and clinical settings before clinical implementation.

Periapical radiographs acquired during routine dental care may provide additional diagnostic information regarding proximal carious lesions without additional radiation exposure. The observed differences among detection architectures and inference configurations provide a foundation for developing more accurate AI-assisted detection systems, although clinical examination remains essential for determining lesion activity, cavitation status, and treatment decisions.