Quantum-Inspired Explainable Deep Learning Framework for Early Enamel Caries Classification in Intraoral Photographs

Description

Early detection of enamel caries is crucial for preventive dentistry but remains challenging due to the subtle and subjective nature of visual examination. This study aims to develop and validate a quantum-inspired, explainable deep learning framework for the automated and interpretable classification of enamel caries from intraoral photographs. This study proposed a hybrid framework utilizing two deep learning models: a custom lightweight CNN named DentXCaries and a fine-tuned ResNet50 with squeeze-and-excitation attention. A novel quantum entanglement feature fusion technique was introduced to combine the deep features from both models. The fused features were classified using twelve machine learning classifiers. The model was developed on a public dataset of 2,000 intraoral images categorized into Early-Stage Enamel Caries, Advanced Enamel Caries, and No Enamel Caries. Explainable AI (Grad-CAM) provided visual explanations for predictions. Performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and statistical tests. The QNN-Caries framework demonstrates state-of-the-art accuracy for enamel caries classification while providing crucial visual interpretability. It represents a significant step towards a reliable, transparent, and clinically viable AI-assisted diagnostic tool for routine dental screenings.

Steps to reproduce

README – Enamel_Caries_Experiment 1. Project Overview This repository contains all experiment files for dental caries classification performed on Kaggle GPU (P100). You can load any folder to view results, extracted features, or trained models. 2. Folder Structure 📁 DentXCaries Features • Extracted feature vectors from DentXCaries model, Train, Test and Validation features and labels in NumPy. • Use these to train classifiers directly 📁 DentXCaries Model • Trained DentXCaries model weights • Can be loaded to reproduce model performance 📁 Modified ResNet50 Features • Feature embeddings generated from the modified ResNet50 network, Train, Test and Validation features and labels in NumPy. 📁 Modified ResNet50 Model • Trained model weights and logs of Modified ResNet50 📁 Quantum Entanglement Fused Features • Final fused features used to generate the best results, Fused Train, Test and Validation features and labels in NumPy. • 📁 Figures • All figures used in manuscript/poster • Confusion matrices, Grad-CAM visualizations, etc. 3. How to Use 1. Download or load the dataset in Kaggle 2. Load any folder (e.g., features or model) 3. Run training or evaluation to reproduce results 4. All experiments were conducted on Kaggle GPU P100 5. Loading features allows instant reproduction of accuracy results 4. Notes • All results are stored inside these folders • You can open each folder to review the outputs directly • No need to retrain models unless you want to

Institutions

• King Faisal University
• HITEC University
• Chulalongkorn University Faculty Of Dentistry

Categories

Licence