WaveViT-YOLO: A Hybrid Architecture for Dental Caries Detection in Intraoral Photographs
Applied Sciences · 19.08.2026 · 16 (16) · 8257
DOI: 10.3390/app16168257 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Sciences, online erschienen am 19.08.2026. Verfasst von Ines Neji und 2 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Princess Nourah bint Abdulrahman University.
Abstract (Originalsprache)
Dental caries remains one of the most prevalent oral health problems worldwide, yet automated detection in intraoral photographs is challenging because of variable lighting, specular reflections, saliva, restoration margins, and subtle early demineralization. We propose WaveViT-YOLO, a hybrid architecture built on a YOLOv9m backbone and integrating (i) a discrete wavelet transform (DWT) preprocessing stage, (ii) learnable WaveletAttention modules at the feature-pyramid scales, and (iii) ViT-based MultiScaleCrossAttention fusion. On the publicly available Annotated Intraoral Image Dataset (6313images; patient-level 70/15/15 split; three independent seeds), YOLOv9m is the strongest standalone YOLO model by mAP@50 (mAP@50 = 0.807±0.003; mAP@50–95 = 0.642±0.004). WaveViT-YOLO achieves the highest measured mAP@50 (0.814±0.005; 0.007 absolute and +0.87% relative), mAP@50–95 (0.647±0.004), F1 (0.831±0.005), and PR-AUC (0.845) among the evaluated models. The model contains 26.3 M parameters, a 30.8% increase over the 20.1 M YOLOv9m baseline. Model-only inference is 27.3±1.6 ms on an NVIDIA T4 GPU, while the current CPU DWT stage adds 235.2±8.0 ms, giving approximately 262.5 ms/image end-to-end; therefore, the current pipeline is not real-time end-to-end. Using the displayed seed-averaged mAP@50 values, the isolated relative changes are +0.62% for DWT and +0.37% for either WaveletAttention or ViT fusion, whereas the full configuration reaches +0.87%. The paired three-seed comparison against YOLOv9m yields t(2)=6.06, p=0.026, and Cohen’s dz=3.50; because n=3, this analysis is treated as exploratory. Small lesions (
Förderung
- Princess Nourah bint Abdulrahman University
Zitieren
APA
Neji, I., Filali, I., & Ejbali, R. (2026). WaveViT-YOLO: A Hybrid Architecture for Dental Caries Detection in Intraoral Photographs. Applied Sciences, 16(16), 8257. https://doi.org/10.3390/app16168257
Vancouver
Neji I, Filali I, Ejbali R. WaveViT-YOLO: A Hybrid Architecture for Dental Caries Detection in Intraoral Photographs. Applied Sciences. 2026;16(16):8257. doi:10.3390/app16168257
BibTeX
@article{Neji2026,
author = {Neji, Ines and Filali, Imen and Ejbali, Ridha},
title = {WaveViT-YOLO: A Hybrid Architecture for Dental Caries Detection in Intraoral Photographs},
journal = {Applied Sciences},
year = {2026},
volume = {16},
number = {16},
pages = {8257},
doi = {10.3390/app16168257}
}
EUVIMED