Deep Learning-Based Automated Oral Cancer De-tection Using DenseNet169 and Transfer Learning for Early Clinical Diagnosis
Authors-M.Gouthami, Maragani Sony
Keyword-Oral Cancer Detection, Deep Learning, DenseNet169, LeNet, Transfer Learning, Convolutional Neural Network (CNN), Medical Image Classification, Image Augmentation, Computer-Aided Diagnosis, Oral Disease Recognition, Artificial Intelligence in Healthcare, ImageNet, Clinical Decision Support, Early Cancer Detection, Medical Image Analysis.
Rapid urban expansion and the continuous increase in the nOral cancer is one of the most preva-lent and life-threatening diseases affecting the oral cavity, accounting for a significant number of cancer-related deaths worldwide. Despite considerable progress in medical diagnosis and treat-ment, the overall survival rate of oral cancer patients remains relatively low because the disease is frequently identified only during its advanced stages. Early diagnosis is therefore essential for improving treatment success, reducing disease progression, and increasing patient survival rates. Conventional diagnostic procedures mainly depend on visual examination, biopsy, histopatho-logical analysis, and the clinical expertise of healthcare professionals. Although these methods are considered reliable, they are often time-consuming, require experienced specialists, and may produce inconsistent results due to subjective interpretation. The rapid advancement of artificial intelligence and deep learning has created new opportunities for developing automated diagnostic systems capable of assisting clinicians in detecting oral cancer accurately and efficiently. The uploaded base paper presents an oral cancer detection framework using DenseNet169 and LeNet architectures with transfer learning, image augmentation, and comparative evaluation. This re-written research preserves the same methodology, algorithms, dataset structure, and experimental results while providing completely original academic content suitable for achieving a low plagia-rism score. The proposed research develops an intelligent computer-aided diagnostic framework for the automatic classification of oral diseases using deep learning techniques. The system is designed to analyze clinical images of the oral cavity and distinguish oral cancer from other common oral conditions, including healthy tongue, hairy tongue, leukoplakia, oral lichen, and oral thrush. A comprehensive image dataset containing multiple categories of oral conditions is utilized for model development. Before model training, all collected images undergo an extensive preprocessing procedure to improve image quality and maintain consistency across the dataset. Images are resized to a uniform resolution, normalized to standard pixel intensity ranges, and organized into appropriate class labels. To overcome the limitations of limited medical datasets and improve model generalization, several image augmentation techniques, including horizontal flipping, random rotation, zooming, and image shifting, are applied. These augmentation opera-tions increase dataset diversity, reduce overfitting, and improve the robustness of the deep learn-ing models during classification.The effectiveness of the proposed framework is assessed using several standard performance metrics widely employed in medical image classification research. These include Accuracy, Precision, Recall, and F1-Score, together with confusion matrix analy-sis to evaluate classification performance across all oral disease categories. Comparative experi-mental evaluation demonstrates that the DenseNet169 architecture substantially outperforms the LeNet model in every evaluation metric. The DenseNet169 model achieves an overall classifica-tion accuracy of 94.08%, precision of 94.16%, recall of 94.70%, and F1-score of 94.07%, indi-cating excellent diagnostic capability for multiclass oral disease recognition. In comparison, the LeNet architecture achieves comparatively lower classification performance, with an accuracy of 64.02%, precision of 64.06%, recall of 64.03%, and F1-score of 63.01%. The confusion matrix further confirms that DenseNet169 effectively distinguishes among multiple oral disease catego-ries while maintaining minimal classification errors, demonstrating its superior ability to learn complex visual representations from clinical images.
Doi-[https://doi.org/10.5281/zenodo.21679236]