Deep Learning-Based Automated Detection of Oral Leukoplakia in Clinical Imaging.

OBJECTIVE

This study aims to develop and validate a deep learning-based system for automated identification of oral leukoplakia (OLK), addressing diagnostic challenges in clinical practice.

METHODS

We conducted a comparative analysis of 19 convolutional neural network (CNN) architectures using 446 clinical images of histopathologically confirmed oral leukoplakia cases. The dataset was augmented with 1,041 normal oral mucosa images for comparison. A fine-tuned EfficientNetB0 architecture was selected as the optimal model. Class Activation Mapping (CAM) visualized decision-making regions, with performance evaluated through area under the receiver operating characteristic curve (AUC-ROC) analysis and accuracy metrics.

RESULTS

The EfficientNetB0 model achieved 97.54% accuracy (95% confidence interval (CI): 95.2%-99.1%) with an AUC of 0.993 (95% CI: 0.981-0.998). Activation mapping demonstrated precise localization of leukoplakic lesions, correlating with clinical diagnostic criteria. The model maintained robust performance across varying illumination conditions and oral cavity locations.

CONCLUSION

This deep learning system demonstrates expert-level diagnostic capability for oral leukoplakia identification, showing potential for integration into clinical decision support systems. The model's high diagnostic accuracy and interpretability through activation mapping address critical needs in early oral cancer detection and screening programs.