Early Diagnosis of Mouth Cancer Using Artificial Intelligence

Late diagnosis of oral cancer remains a major global health problem, especially in resource-limited primary care, where screening relies on subjective visual examination. Deep learning models have shown high accuracy in detecting oral cancer from images, but three gaps limit their clinical use: most operate as “black boxes” whose predictions cannot be checked by the user, most are developed and tested on a single dataset, and many require specialized equipment or computational infrastructure unavailable in primary care. This study aimed to develop an explainable, PC-based AI tool for screening oral cancer and potentially malignant disorders from clinical photographs. A retrospective diagnostic accuracy design was adopted using publicly available datasets (Kaggle, ORCHID, Sri Lanka and NDB-UFES) within a framework designed for multi-dataset evaluation. Images were resized to 224 × 224 pixels, normalized and augmented, and a convolutional neural network pre-trained on ImageNet was adapted through transfer learning for binary classification (normal versus suspicious). Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated to highlight the image regions driving each prediction. The model was packaged in an offline desktop application that shows, for every image, the predicted class, its confidence, the Grad-CAM heatmap and a referral-oriented recommendation at the point of care. The model achieved 100% sensitivity, 50% specificity, 80% accuracy, an 85.7% F1-score and an AUC of 1.00, and the heatmaps localized to lesion areas in suspicious cases. These results show that deep learning with integrated visual explanation is a promising basis for primary-care screening, though the small test set limits their reliability.Late diagnosis of oral cancer remains a major global health problem, especially in resource-limited primary care, where screening relies on subjective visual examination. Deep learning models have shown high accuracy in detecting oral cancer from images, but three gaps limit their clinical use: most operate as “black boxes” whose predictions cannot be checked by the user, most are developed and tested on a single dataset, and many require specialized equipment or computational infrastructure unavailable in primary care. This study aimed to develop an explainable, PC-based AI tool for screening oral cancer and potentially malignant disorders from clinical photographs. A retrospective diagnostic accuracy design was adopted using publicly available datasets (Kaggle, ORCHID, Sri Lanka and NDB-UFES) within a framework designed for multi-dataset evaluation. Images were resized to 224 × 224 pixels, normalized and augmented, and a convolutional neural network pre-trained on ImageNet was adapted through transfer learning for binary classification (normal versus suspicious). Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated to highlight the image regions driving each prediction. The model was packaged in an offline desktop application that shows, for every image, the predicted class, its confidence, the Grad-CAM heatmap and a referral-oriented recommendation at the point of care. The model achieved 100% sensitivity, 50% specificity, 80% accuracy, an 85.7% F1-score and an AUC of 1.00, and the heatmaps localized to lesion areas in suspicious cases. These results show that deep learning with integrated visual explanation is a promising basis for primary-care screening, though the small test set limits their reliability.