Artificial Intelligence based Assessment of Oral Precancer to Aid Early Detection of Oral Cancer

Background: Oral epithelial dysplasia (OED) carries an increased risk of malignant transformation to oral squamous cell carcinoma (OSCC), which is amongst the leading cancers worldwide. The diagnostic gold standard for OED is histopathological assessment and grading, which is challenging with unreliable behaviour and progression prediction. Advancements in digital pathology and Artificial Intelligence (AI) provide opportunities to uncover novel data from wholeslide imaging (WSI) through automated detection, pattern recognition and quantitative analysis. Aims: This research uses a range of digital, computational, and quantitative approaches to reveal novel insights into OED progression and develops OSCC risk prediction models which are compared to existing clinical grading systems. Methods: Retrospective samples of OED and non-dysplastic WSI (with five-year clinical follow-up) were used to develop malignant transformation prediction models based on analysis of conventional architectural and cytological histological features (n=109), novel digital morphometric features (n=100) and mitotic features assessed on haematoxylin and eosin (H&E) WSI and immunohistochemistry for Phosphohistone H3 (n=68). Machine learning models were trained to detect dysplastic, immune, and stromal cells in OED (n=220-248). Deep learning neural networks were trained to segment OED epithelium (n=434) and validated on external unseen datasets. Prognostic relationships were explored, and spatial analysis conducted. Results: Six conventional histological features were significant for transformation (p<0.036) and recurrence (p<0.015). Significant differences in cytoplasmic eosin, nuclear eccentricity and circularity were seen in basal epithelial cells of OED (p<0.05). Nucleus circularity was associated with recurrence (p=0.018) and epithelial perimeter with malignancy (p=0.03). The developed models demonstrated better predictive strength for malignant transformation risk (AUROC:0.74- 0.81) compared to ‘gold-standard’ histological grading (AUROC:0.60-69) with superior performance maintained on unseen external datasets. Trained AI models segmented and classified OED epithelium, immune and stromal cells with good accuracy (F1 scores:0.80-0.87). Peri-epithelial lymphocyte count was associated with malignant transformation and reduced progression free survival (p<0.05). Conclusions: This novel research shows correlations between individual OED histological features, digital morphometric features, and prognosis for the first time on the largest digital multicentre cohort to date.