A fully automated and explainable algorithm for predicting malignant transformation in oral epithelial dysplasia

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Oral Cancer Research

Background:

• Oral epithelial dysplasia (OED) grading has high inter-/intra-observer variability.
• Current OED grading does not reliably predict malignant progression.
• This variability can lead to suboptimal clinical treatment decisions.

Purpose of the Study:

• To develop and validate an AI algorithm for predicting OED malignant transformation risk.
• To create an Oral Malignant Transformation (OMT) risk score using H&E whole slide images (WSIs).
• To assess the algorithm's performance against existing grading systems and its prognostic value.

Main Methods:

• Developed an AI pipeline using an in-house segmentation model for nuclei and epithelium detection.
• Utilized a shallow neural network with interpretable morphological and spatial features to predict progression.
• Validated the algorithm on internal (Sheffield; n=193) and external (Birmingham, Belfast; n=89) cohorts.

Main Results:

• The OMTscore achieved an AUROC of 0.75 (Recall=0.92) on external validation, outperforming binary grading (AUROC=0.72, Recall=0.85).
• Survival analyses demonstrated the OMTscore's prognostic value (C-index=0.60, p=0.02), comparable to WHO and binary grades.
• Identified peri-epithelial and intra-epithelial lymphocytes in predictive patches of transforming cases.

Conclusions:

• The developed AI algorithm provides an automated, explainable, and externally validated method for predicting OED transformation.
• The OMTscore demonstrates comparable-to-human performance, offering a promising solution for routine OED grading.
• This AI tool can aid in improving treatment decisions for patients with oral epithelial dysplasia.