Detection of Architectural Dysplastic Features from Histopathological Imagery of Oral Mucosa Using Neural Networks

Oral cancer is a serious illness, but it is potentially curable if early detection can be achieved successfully. Oral epithelial dysplasia (OED), which is a precursor to oral squamous cell carcinoma (OSCC), can provide abnormal characteristics to diagnose the risk of developing oral cancer. This paper proposes a neural network architecture for detecting dysplastic features of epithelial architecture, including irregular epithelial stratification and bulbous rete ridges. The different combinations of atrous convolution, batch normalization, global pooling, and dropout are discussed regarding their effects, along with an ablation study. A signature library containing image patches was constructed and utilized to train the models. The best-performing model in the validation set attained an average accuracy of 97.52%. The results of the blind test from the receiver operating characteristic (ROC) curves show that the best model reached the best probability of detection, 0.8571, for irregular epithelial stratifications and 0.8462 for the bulbous rete ridges.