A General Evaluation of Simple Convolutional Neural Networks and their Effectiveness for Automated Dental Diagnosis Using Intraoral Images: A Case Study with Dental Caries

Background: The incorporation of the growing field of Artificial Intelligence [AI] into medicine and healthcare has resulted in massive advancements in automated medical diagnostics. Specifically, in Deep Learning [DL], a subset of AI, the growth of Convolutional Neural Networks [CNNs], a DL architecture specialized in image classification and feature extraction, with simplistic, lightweight architectures offers computationally efficient alternatives to deeper, more complex Deep Learning architectures, increasing the accessibility of AI-based diagnostic solutions. With the lack of automated solutions available in dentistry, a subfield of medicine, this study aims to evaluate the effectiveness of simply structured Convolutional Neural Networks in dental diagnostics with a case study on dental caries classification given intraoral images, caries being a common oral disparity where it is of utmost importance for early intervention. Methods: A publicly available dataset of 2002 labeled intraoral images with 3 distinct classes [healthy enamel, early-stage caries, and advanced-stage caries] was utilized to train and test a lightweight Convolutional Neural Network with 4 Convolutional Layers, batch normalization, max pooling, and fully connected layers. Additionally, common data augmentation and preprocessing techniques, as well as cross-validation splits, the Adam optimizer, and Learning Rate Schedulers, were used to enhance model robustness and prevent overfitting from "memorizing" the training dataset. Certain hyperparameters were chosen, such as epoch count, batch size, cross-entropy loss function, etc., to maximize model performance. Key evaluation metrics, such as accuracy, precision, recall, and F1 score, were used to assess the model's performance on the unseen test dataset, with a confusion matrix being generated to highlight classifications and misclassifications. Results: The proposed simple-CNN achieved approximately 85% accuracy, reaching high levels of over 80% for precision, recall and F1 score as well, demonstrating excellent potential for the multi-class classification case study with dental caries by showing reduced risks of misclassifications and high accuracy in diagnosing the condition correctly, all while retaining computational efficiency. Conclusion: This study highlights the enormous potential of CNNs with simple architectures for automated dental diagnosis, evident with the tested lightweight CNN's strong performance in classifying dental caries, showing that they can perform well in diagnosing oral disparities all while balancing computational efficiency. Some applications of this study's results include the potential implementation of accessible automated solutions into dental clinics. Future research can involve and explore evaluating more complex DL architectures, multi-modality, explainable AI, etc.