Automated Classification of Oral Cancer Histopathology images using Convolutional Neural Network

Abstract

In latest years, convolutional neural networks (CNNs) have accomplished state-of-the-art performance in many computer vision tasks such as classification of images, object detection, instance segmentation etc. CNN has a robust learning capacity and can enhance the use of datasets for feature extraction. Identifying the key visual features from the oral squamous cells are the significant and compulsory task for the clinicians to detect the different stages of oral cancer. The computer-aided instrument performing the same identifying job would provide clinicians with a vital guidance during diagnosis for evaluating histological images. In this study, we suggest the use of 4-layered (5X5X3) patches of convolutional neural networks (CNNs) for feature extraction and classification from oral cancer images. To prevent overfitting the images were augmented by rotating, inverting and flipping. The proposed model has achieved 96.77 % accuracy, with 10 fold cross validation, which is at par with the accuracy of cytotechnologists and pathologists. Therefore, this model is helpful in classifying oral cancer microscopic images.