DEEP LEARNING FOR ORAL EPITHELIAL DYSPLASIA GRADING

Abstract

Objective The present study proposes a Deep Learning model based on the binary system for grading oral epithelial dysplasia (OED) at the whole slide imaging level to eliminate inter-pathologist variability. Study Design A dataset of 99 whole slide images from three institutions were manually annotated, segmented, and fragmented into smaller patches of 299 × 299 pixels. A total of 40,893 images were sampled into 80%:10%:10% for training, validation, and independent test sets. An adaptation of ResNet50 with 2 hidden layers and 512 neurons in the fully connected layer (FC) was implemented with a low learning rate of 0.00001 for 200 epochs. Results The proposed ResNet50 reached 85.30% accuracy during training and 85.11% for validation, showing the potential of learning. The independent test showed an overall accuracy of 60%, with 61% sensitivity, 59% specificity, and 0.64AUROC, showing a lack of generalization ability for the present classification problem. Conclusion The proposed DL-based model presented a capacity for learning with the potential of achieving high accuracy but with relatively low generalization. Further work will encompass a 2-class problem (premalignant and malignant) to reinforce class separation and investigate the stability of accuracy and generalization of alternative DL models. The present study proposes a Deep Learning model based on the binary system for grading oral epithelial dysplasia (OED) at the whole slide imaging level to eliminate inter-pathologist variability. A dataset of 99 whole slide images from three institutions were manually annotated, segmented, and fragmented into smaller patches of 299 × 299 pixels. A total of 40,893 images were sampled into 80%:10%:10% for training, validation, and independent test sets. An adaptation of ResNet50 with 2 hidden layers and 512 neurons in the fully connected layer (FC) was implemented with a low learning rate of 0.00001 for 200 epochs. The proposed ResNet50 reached 85.30% accuracy during training and 85.11% for validation, showing the potential of learning. The independent test showed an overall accuracy of 60%, with 61% sensitivity, 59% specificity, and 0.64AUROC, showing a lack of generalization ability for the present classification problem. The proposed DL-based model presented a capacity for learning with the potential of achieving high accuracy but with relatively low generalization. Further work will encompass a 2-class problem (premalignant and malignant) to reinforce class separation and investigate the stability of accuracy and generalization of alternative DL models.