Stacking-Based Hybrid Deep Learning Framework for Dental Cavity Detection

Abstract

Dental cavity detection plays a vital role in preventive dentistry and early diagnosis of oral diseases. Manual identification of caries from dental images is often subjective and prone to human error, which necessitates the development of automated and intelligent detection systems. In this work, a hybrid stacking-based framework is proposed that combines deep convolutional features from EfficientNetB0 with handcrafted descriptors such as Histogram of Oriented Gradients (HOG) and color moments. The extracted features are normalized, reduced through Principal Component Analysis (PCA), and classified using a multi-level stacking ensemble integrating XG Boost, Light GBM, Random For-est, and Support Vector Machine (SVM) base learners, with a logistic regression meta-classifier for final decision fusion. The model was evaluated on the publicly available Kaggle Dental Cavity dataset, achieving an Area Under the ROC Curve (AUC) of 0.938 and an Average Precision (AP) score of 0.9287, indicating excellent discriminative performance. Class-wise evaluation demonstrated a high recall of 95% for the cavity class and a precision of 93% for the no-cavity class, reflecting the model’s robustness in minimizing both false negatives and false positives. These results confirm that the proposed hybrid stacking approach effectively leverages complementary feature representations, providing a reliable and clinically meaningful solution for automated dental cavity detection.