Supremacy of attention-based transformer in oral cancer classification using histopathology images

Oral cancer has emerged as one of the ubiquitous malignant tumors globally. Timely detection and treatment reduces the mortality rate of oral cancer. This study utilizes a vision transformer (ViT) framework to classify oral squamous cell carcinoma (OSCC) and healthy oral histopathology images. The proposed approach is implemented on a public database consisting of 4946 oral histopathology images. Although ViT architectures have been extensively used in the medical imaging field, they have not yet been explored in oral cancer detection. Though transformer architecture needs large dataset to attain better performance, our modified architecture accomplishes an accuracy, specificity and sensitivity of 97.78%, 96.72%, and 98.80%, respectively, on a relatively smaller medical dataset. The evaluation metrics of the proposed method have also been compared with eight pre-trained deep learning models, namely Xception, Resnet50, InceptionV3, InceptionResnetV2, Densenet121, Densenet169, Densenet201 and EfficientNetB7. It is observed that the modified ViT model performs better than the deep learning models, demonstrating the ability to extract various features from the histopathology images for the classification. The results of the proposed approach would aid the clinical community for detection of oral cancer in patients of diverse origin.