QGLH-Net: quality-guided local-global hierarchical deep learning framework for oral disease classification using intraoral images

Dental diseases are a widespread yet under-recognized health problem worldwide, and oral health problems are associated with several systemic conditions including diabetes, cardiovascular disease and depression. Early diagnosis is crucial but depends on access to the trained dental professionals, which is difficult in underserved and resource-limited communities. This study proposes QGLH-Net (Quality-Guided Local-Global Hierarchical Network), a dual-stream deep learning framework that classifies oral diseases at six classes (gingivitis, caries, ulcer, calculus, tooth discoloration and hypodontia) from intraoral images while addressing the challenges of image quality, class imbalance and spatially localized lesion features. Experiments were conducted on publicly available Kaggle Oral Diseases dataset. Following quality filtering and deduplication, 3,076 images were retained and stratified into 2,153 training, 461 validation, and 462 test images. QGLH-Net framework introduces a dual-stream design combining ResNet-50 global branch selected because it achieved the highest macro-F1 among the tuned baselines, while for local branch a lightweight ResNet-18 is used operating on Grad-CAM-guided regions. Unlike post-hoc Grad-CAM visualization approaches, QGLH-Net uses Grad-CAM derived information to generate a complementary local input stream and integrate it with quality-aware curriculum and hierarchical supervision. Hierarchical supervision jointly optimizes six fine-grained disease classes and three coarse anatomical categories. Training further incorporates a two-stage quality-guided curriculum and a dynamic class-balanced focal loss in which inverse-frequency class weights are adapted using class-specific false-negative and false-positive rates, together with supervised contrastive learning. QGLH-Net achieved a best observed accuracy of 98.27% and macro F1-score of 0.9691, has a per-class F1-score ≥ 0.955. Paired statistical evaluation against the strongest ResNet-50 baseline yielded p = 0.152 for accuracy and p = 0.441 for macro-F1; therefore, the observed improvements are interpreted as performance gains rather than statistically confirmed superiority. Minority-class recall estimates showed wider 95% confidence intervals, particularly for caries (0.759–0.979) and tooth discoloration (0.750–0.978), reflecting their limited test support. End-to-end inference required approximately 117.1 ms per image (~ 8.5 images per second) on an NVIDIA Tesla T4 GPU. Beyond classification, we also explored how model outputs could be made more directly usable in a clinical workflow. As an additional proof-of-concept element, AI-assisted reporting module built with CrewAI and Google’s Gemini-1.5 Flash was implemented as a post-classification interface; it does not contribute to reported classification performance and has not undergone formal clinician-based evaluation. QGLH-Net is presented as an explainable screening and decision-support oral disease classification framework for intraoral-image assessment rather than an autonomous diagnostic system, with further clinical validation required before deployment.