What question did this study set out to answer?

To evaluate hybrid deep learning strategies for classifying oral conditions using small and imbalanced datasets.

May 13, 2026BMC Oral Health0 citationsOpen Access

Research on the efficacy of hybrid deep learning models for image-based classification of common oral conditions

YXY X XiaLishui Central Hospital WLWeisong LiuLishui Central Hospital ZLZhongda LiuLishui Central Hospital

Key Points

To evaluate hybrid deep learning strategies for classifying oral conditions using small and imbalanced datasets.
Analyzed 10,735 oral images using hybrid and end-to-end deep learning models.
Compared models including ResNet18, EfficientNetV2-S, and ConvNeXt-Tiny with various classifiers.
Employed 5-fold cross-validation, class-imbalance ablations, and statistical tests to validate results.
Strongest end-to-end model was ConvNeXt-Tiny with 93.22% accuracy.
Hybrid model (ConvNeXt-Tiny + RBF-SVM) achieved 93.96% accuracy and a macro-F1 of 0.9242.
In periodontal evaluation, ConvNeXt-Tiny + linear SVM yielded 88.89% accuracy.

Abstract

To compare tuned end-to-end and hybrid deep learning strategies for image-based classification of common oral conditions under small and imbalanced data conditions, and to determine whether the same pattern persists in a clinically focused periodontal subset. A total of 10,735 oral images were analyzed. We compared tuned end-to-end and frozen-feature hybrid models across ResNet18, EfficientNetV2-S, and ConvNeXt-Tiny backbones. Post-hoc classifiers included K-nearest neighbors (KNN), random forest (RF), linear support vector machine (SVM), and radial basis function (RBF)-SVM. In addition to the original holdout setting, stratified 5-fold cross-validation, class-imbalance-oriented ablations, paired McNemar tests, Grad-CAM, and ROC/PR analyses were performed. A periodontal 3-class subtask (Calculus/Caries/Gingivitis) was also evaluated. On the full holdout task, the strongest end-to-end model was ConvNeXt-Tiny at 224 × 224 (accuracy 0.9322, macro-F1 0.9177), whereas the strongest overall holdout model was ConvNeXt-Tiny + RBF-SVM (accuracy 0.9396, macro-F1 0.9242). On the periodontal holdout task, the strongest model was ConvNeXt-Tiny + linear SVM (accuracy 0.8889, macro-F1 0.8727). Under pooled 5-fold cross-validation, hybrid gains remained statistically supported for several backbone-specific comparisons, including full-task ResNet18 end-to-end versus RBF-SVM (macro-F1 0.8984 vs. 0.9130; McNemar p = 5.41e-08). The dominant error pattern remained mutual confusion between Calculus and Gingivitis. In the task of oral image classification, the hybrid model strategy demonstrates improved performance metrics over standard end-to-end models, particularly within the clinically focused periodontal subset. While further refinement is needed to address specific overlaps between closely related classes (such as Calculus and Gingivitis), this approach proves to be an effective optimization strategy. Ultimately, it provides valuable methodological options and foundational insights for the future development of intelligent periodontal auxiliary diagnostic tools.

Demander à l'IA

Bookmark

View Full Paper