Deep Learning-Based Oral Cancer Detection Using Clinical Images
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Oral cancer is a significant health concern where early detection greatly improves patient outcomes. This study develops and evaluates a deep learning model to automatically detect oral cancer from clinical photographic images. A convolutional neural network (CNN) was trained on a dataset of 750 oral lesion images sourced from Kaggle, using transfer learning with EfficientNet-B0 to compensate for limited sample size. The model's performance was validated on a held-out test set and assessed with accuracy, precision, recall, and receiver operating characteristic (ROC) curve analysis. Results indicate high diagnostic performance: the model achieved approximately 95% overall accuracy in distinguishing cancerous lesions from non-cancerous oral tissue. The CNN demonstrated a sensitivity of about 94% for identifying oral cancers and a specificity of about 95% for correctly recognizing non-cancer cases. The area under the ROC curve was 0.97, indicating excellent discriminative ability. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations were used as an exploratory interpretability method to identify image regions contributing to model predictions. These heatmaps may suggest overlap between model attention and visually relevant lesion regions, but they do not confirm that the model used clinically meaningful features. These findings suggest that deep learning could serve as a valuable tool in assisting dental professionals with early oral cancer detection, though furt
Abstract
Oral cancer is a significant health concern where early detection greatly improves patient outcomes. This study develops and evaluates a deep learning model to automatically detect oral cancer from clinical photographic images. A convolutional neural network (CNN) was trained on a dataset of 750 oral lesion images sourced from Kaggle, using transfer learning with EfficientNet-B0 to compensate for limited sample size. The model's performance was validated on a held-out test set and assessed with accuracy, precision, recall, and receiver operating characteristic (ROC) curve analysis. Results indicate high diagnostic performance: the model achieved approximately 95% overall accuracy in distinguishing cancerous lesions from non-cancerous oral tissue. The CNN demonstrated a sensitivity of about 94% for identifying oral cancers and a specificity of about 95% for correctly recognizing non-cancer cases. The area under the ROC curve was 0.97, indicating excellent discriminative ability. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations were used as an exploratory interpretability method to identify image regions contributing to model predictions. These heatmaps may suggest overlap between model attention and visually relevant lesion regions, but they do not confirm that the model used clinically meaningful features. These findings suggest that deep learning could serve as a valuable tool in assisting dental professionals with early oral cancer detection, though further clinical validation is required before adoption.