Development and validation of a CNN model for grading oral epithelial dysplasia

Diagnosis of oral epithelial dysplasia (OED), particularly grading, is challenging as it shows inter- and intra-observer variability. New technologies to assist pathologists are needed to decrease variability and enhance reproducibility. AI has brought about a new wave of increased precision in oncological pathology. High accuracy and objectivity in OED grading may now be attainable in histopathology because of AI tools such as deep learning.

A deep learning model, Densenet121, was developed for automated and reliable grading of OED according to the 2017 WHO classification (mild, moderate, and severe). The dataset comprised 1500 images, and slides were digitised at 4×, 10× and 40× magnification. We developed and trained a classification model.

Overall performance of the Densenet 121 model was good, with accuracy above 80% across all magnifications. However, the 10× magnification yielded a good accuracy of 92.7%, precision of 92.8% and recall of 92.5%. For classification into mild, moderate, and severe categories, accuracy reached 91%, with good precision and recall.

DenseNet121 achieves high accuracy making it computationally efficient. This model was able to distinguish between normal epithelium and different grades of OED. These findings support the potential role of AI-based tools in reducing variability in grading dysplasia.