Deep Feature Learning Framework for Automated Oral Cancer Detection

This work presents a deep learning–based framework for the automatic identification of oral cancer from clinical tongue images. The study utilizes a multi-class oral image dataset comprising healthy tongue samples along with pathological conditions such as oral cancer, leukoplakia, oral lichen planus, thrush, and hairy tongue. Images collected from both affected and non-affected individuals were used to train and evaluate the proposed model. A pre-trained DenseNet169 network was adopted as the backbone architecture and fine-tuned using transfer learning, with additional fully connected layers introduced to enhance class discrimination. To reduce overfitting and improve generalization, extensive image augmentation techniques were applied during training. The effectiveness of the proposed approach was validated through a comparative analysis with a classical LeNet-based convolutional network. Experimental results indicate that the DenseNet-based model achieved superior performance, recording an accuracy of 94.08%, precision of 94.16%, recall of 94.70%, and an F1-score of 94.70%. In contrast, the LeNet model produced significantly lower results, with accuracy, precision, recall, and F1-score values close to 64%. The findings emphasize the importance of appropriate model architecture selection, robust data preprocessing, and systematic evaluation in medical image classification tasks. Further optimization and large-scale validation could strengthen the applicability of the proposed system for real-time clinical oral cancer screening.