Performance comparison of deep learning models for automated dental caries classification with Grad-CAM visualization
TL;DR
This study is among the first few reproducible benchmark studies for CNNs on the Caries_Spectra dataset that involves repeated cross-validation, multi-seed testing, statistical comparison, computational profiling, and explainability studies.
Abstract
Dental caries is a common oral disorder, and detection of this disease at an early stage is difficult because of the inconsistencies in the clinical assessment of the disease. While deep learning has shown remarkable results in analyzing dental images, the inconsistencies in the datasets, validation process, and evaluation methods make it difficult to compare the existing methods. In this paper, we perform a systematic benchmarking test on five deep learning models including the ResNet18, MobileNetV2, EfficientNet-B0, DenseNet121 and Custom CNN for enamel caries classification using the publicly available Caries_Spectra dataset of 2000 RGB intraoral images. All models have been analyzed under the same experimental setup consisting of stratified 5-fold cross-validation with two independent random seeds, early stopping, and the same training parameters. Evaluation measures included accuracy, precision, recall, F1 score, confidence interval calculation, and statistical testing, alongside computational complexity analysis and interpretability using Grad-CAM. EfficientNet-B0 model had the best quantitative performance with 98.18 ± 0.82 mean accuracy, 0.985 ± 0.006 F1-score, 0.985 ± 0.006 precision and 0.985 ± 0.006 recall, with DenseNet121 also showing similar results. It was found that the pretrained models outperformed the custom CNN models significantly, but there was no statistical difference between the pretrained models. Grad-CAM visualization was done to qualitatively analyze the model's activations for enamel caries detection. As far as we are aware, this study is among the first few reproducible benchmark studies for CNNs on the Caries_Spectra dataset that involves repeated cross-validation, multi-seed testing, statistical comparison, computational profiling, and explainability studies. The designed benchmark framework serves as a standard reference for future advancements in the area of artificial intelligence for enamel caries classification systems.