Explainable Deep Learning Framework for Automated Classi-fication of Enamel Caries: Bridging Artificial Intelligence and Dental Public Health
Description
This study proposes an explainable and interpretable deep learning framework for automated enamel caries classification, emphasizing diagnostic transparency and reliability for clinical use. The dual-model framework demonstrated high diagnostic precision and transparency, highlighting its potential for integration into clinical and community-level dental screening. Although dataset diversity remains a limitation, future research will focus on multi-modal and federated learning approaches to ensure broader generalization and population-level applicability.
Files
Steps to reproduce
Setup Instructions 1. Download and Extract • Download the file EnamelCaries_Experiment-II.zip from this repository. • Extract all folders to your working directory (local or Kaggle environment). 2. Load the Code • Open the folder Experiment Code/. • Upload EnamelCaries_Experiment-II.py to your Kaggle Notebook or run locally using Jupyter/VS Code. 3. Connect the Dataset • Link or upload the intraoral enamel caries dataset. • Update dataset paths in the script (train, validation, and test directories). • The models expect images resized to 224 × 224 × 3 format. 4. Run the Models Execute the notebook sequentially to perform: • Training of ExplainableDentalNet and Interpretable ResNet50-SE • Evaluation using accuracy, precision, recall, and F1-score • Generation of confusion matrices and statistical reports • Grad-CAM heatmap visualization for explainability 5. View Results • Navigate to: o Classwise Classification Results/ → for per-class statistics (.xlsx, .png) o Confusion Matrix/ → for visual confusion matrices and accuracy tables o Grad-CAM/ → to inspect lesion localization visualizations Requirements Environment • Kaggle Notebook (recommended) • Python ≥ 3.8 Libraries • TensorFlow / Keras • NumPy, Pandas, Scikit-learn • Matplotlib, Seaborn Hardware • GPU: NVIDIA Tesla P100 or higher (recommended for full training) Reproducibility and Citation This repository ensures full methodological transparency and reproducibility following CLAIM 2024 and STARD 2015 reporting guidelines.
Institutions
- King Faisal University