Automated detection of premalignant oral lesions on whole slide images using convolutional neural networks

•We successfully trained a deep learning CNN model to recognize suspicious for high-grade epithelial dysplasia of the tongue and floor of mouth based on histopathologic features.•The model can process a digital whole-slide image and mark the suspected regions of high-grade oral dysplasia for pathologists to review.•The model recognized all regions of severe epithelial dysplasia across our test set and achieved an overall accuracy of 90.9 percent.