A Deep Learning System to Predict Epithelial Dysplasia in Oral Leukoplakia

Area of Science:

• Oral pathology and oncology
• Artificial intelligence in medicine
• Medical imaging analysis

Background:

• Oral leukoplakia (OL) carries a significant risk of malignant transformation into oral cancer.
• Assessing epithelial dysplasia (OED) in OL is crucial but typically requires invasive biopsy procedures.
• There is a need for non-invasive methods to predict OED and guide diagnostic decisions.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting OED probability using oral photographs.
• To compare the DL model's performance against clinicians' assessments for OED diagnosis.
• To evaluate the DL model's utility as a decision support tool for biopsy selection in OL patients.

Main Methods:

• A DL model with an EfficientNet-B2 backbone was trained on 2,073 retrospective oral leukoplakia images.
• Histopathology served as the gold standard for OED status determination.
• The model's performance was evaluated through internal validation, temporal testing, and external validation with human rater comparisons.

Main Results:

• The DL model demonstrated strong performance with high area under the curve (0.882) and balanced accuracy (81.8%) during testing.
• External validation confirmed good performance (AUC 0.828, balanced accuracy 76.4%) and outperformed 92.3% of human raters in OED classification.
• The model showed significant potential net benefit in guiding biopsy decisions, potentially reducing unnecessary procedures.

Conclusions:

• A photograph-based DL model can accurately predict OED probability and status in oral leukoplakia.
• The model shows promise as a decision support tool, assisting in selecting patients for biopsy and potentially enabling self-monitoring.
• This AI approach offers a non-invasive adjunct to traditional methods for managing oral leukoplakia and its malignant potential.