Exploring the feasibility of AI-based analysis of histopathological variability in salivary gland tumours

This study uses artificial intelligence (AI) for differentiation between salivary gland tumours (SGT) using digitised Haematoxylin and Eosin stained whole-slide images (WSI). Machine learning (ML) classifiers were developed and tested using 320 scanned WSI. These included a benign versus malignant classifier (BvM) for automated identification of benign and malignant tumours, a malignant sub-typing (MST) classifier for subtyping four most common malignant SGT and a third classifier for malignant tumour grading. ML results were also compared with deep learning models. All ML classifiers showed an excellent accuracy. An F1 score of 0.95 was seen for benign vs. malignant and malignant subtyping tasks and 0.87 for automated grading. In comparison, the best performing DL models showed F1 scores of 0.80, 0.60 and 0.70 for the same tasks respectively. External validation on an independent cohort demonstrated good accuracy, with an F1 score of 0.87 for both the benign vs. malignant and grading classifiers. A notable association between cellularity, nuclear haematoxylin, cytoplasmic eosin, and nucleus/cell ratio (p < 0.01) were seen between tumours. Our novel findings show that AI can be used for automated differentiation between SGT. Analysis of larger multicentre cohorts is required to establish the significance and clinical usefulness of these findings.