Assessment of AlexNet for Oral Epithelial Dysplasia Classification

The progression of potentially malignant lesions into oral squamous cell carcinoma (OSCC) is assessed based on the dysplasia grading, in which several cytological and architectural changes drive pathologists to assign a malignization risk for proper patient management. The present study implemented AlexNet, a Deep Learning (DL) model based on the binary system for grading oral epithelial dysplasia (OED), the only prognostic hallmark of oral potentially malignant disorders (OPMD). A TOTAL of 63 digital slides from two institutions were used. Annotated regions of interest (ROI) were segmented and fragmented into 56,406 smaller patches of 220 × 220 pixels. After data separation into training, validation, and testing sets (80%, 10%, and 10%), data augmentation was conducted in training/validation images. After carrying out the test, the trained AlexNet reached 91.57% accuracy, with 93.55% sensitivity, and 87.57% specificity. F1-score was 0.93857 and AUC was 0.9771. Consequently, showing potential for the proposed application.