Deep learning based automated detection and grading of oral epithelial dysplasia: A comparative histopathological image analysis.

OBJECTIVE

To develop and comparatively evaluate multiple deep learning architectures for automated detection and grading of oral epithelial dysplasia in hematoxylin and eosin (H&E)-stained histopathological images.

METHODS

A total of 2080 digitized H&E-stained histopathological images representing mild, moderate and severe grades of OED, along with normal mucosa, were acquired at 100x and 400x magnifications. Following standardized preprocessing and augmentation, nine deep learning architectures, Vision Transformer (ViT), VGG16, YOLOv11, ResNet-50, YOLOv8, MobileNet-v3, Inception-v3, EfficientNet-B0, and DenseNet-121 were trained and evaluated. Model performance was assessed using accuracy, precision, recall, and F1- score metrics.

RESULTS

Among these nine architectures, DenseNet-121 achieved the highest accuracy of 95.47% with a macro-average F1-score of 0.94, outperforming both transformer-based and conventional convolutional models. It demonstrated strong generalization and minimal overfitting, with class-wise F1-scores exceeding 0.95 for normal, mild, and severe grades of dysplasia.

CONCLUSION

The DenseNet-121 model provides a reliable and generalizable framework for automated grading of oral epithelial dysplasia, offering enhanced diagnostic precision, reduced observer bias, and potential integration into digital pathology workflows for early detection and risk stratification of oral precancerous lesions.

CLINICAL SIGNIFICANCE: Deep learning based automated grading of OED can enhance diagnostic accuracy and reproducibility in histopathological evaluation. Integration of such AI-assisted systems into digital pathology may support early identification of high-risk lesions and assist clinicians in making timely and evidence-based treatment decisions.