Supremacy of attention based convolution neural network in classification of oral cancer using histopathological images

Introduction: Oral cancer has grown to be one of the most prevalent malignant tumours and one of the deadliest diseases in emerging and low-to-middle income nations. The mortality rate can be significantly reduced if oral cancer is detected early and treated effectively. Objectives: This study proposes an effective histopathological image classification model for oral cancer diagnosis using Vision Transformer deep learning based on multi-head attention mechanism. Methods: The oral histopathological image dataset used in the study consists of 4946 images, which were categorized into 2435 images of healthy oral mucosa and 2511 images of oral squamous cell carcinoma (OSCC). In our proposed approach, along with Vision Transformer model eight pre-trained deep learning models known as Xception, Resnet50, InceptionV3, InceptionResnetV2, Densenet121, Densenet169, Densenet201 and EfficientNetB7 have been used for the comparative analysis. 90% of the images are used for training the models while the rest 10% of the images are used for testing purposes. Results: Vision Transformer model achieved the highest classification accuracy of 97.78% in comparison to other considered deep learning models. Specificity, sensitivity and ROC AUC score are recorded as 96.88%, 98.74% and 97.74% respectively. Conclusion: We found that our proposed Vision Transformer model outperforms compared to other pre-trained deep learning models, demonstrating a stronger transfer ability of the learning in histopathological image classification from the analysis of the obtained results. This method considerably lowers the cost of diagnostic testing while increasing the diagnostic effectiveness, and accuracy for oral cancer detection in patients of diverse origin.