Background: Salivary gland tumours (SGT) are a heterogeneous group of neoplasms with morphological diversity and overlapping features. SGT can be diagnostically challenging due to a large number of entities and markedly similar features but different clinical behaviour. Recently, artificial intelligence (AI) has been found to be precise for histological diagnosis and prognosis prediction including identification of sub-visual features. However, its application to SGT has not been reported to date. The epidemiological data for SGT is also somewhat outdated and limited to single centre reports.
Objectives: This study aims first, to mitigate epidemiological data shortcomings by analysing information including demographic, anatomical location and histological diagnoses of SGT from multiple centres across the world. Second, to examine if AI can be used to differentiate between different SGT subtypes and grades based on the analysis of digitised whole slide images (WSIs) of Haematoxylin and Eosin (H&E) stained slides.
Methods: All primary benign and malignant SGT demographical data including age, gender, location and histological diagnosis from fifteen centres covering the majority of the world health organisation (WHO) geographical regions between 2006 and 2019 were included in the study. A total of 240 scanned H&E WSIs were obtained. An open-source bioimage analysis software (QuPath) was used for training and analysis of features on representative regions-of-interest (ROIs). The first machine learning (ML) classifier was trained and tested to differentiate between benign and malignant (BvM) SGT (n=120 each). The second ML classifier was used for malignant SGT subtyping (MST) (n=120). A third ML classifier was used for automated grade prediction (TG) (n=120) in two most common gradable SGT and results compared with deep learning (DL) methods using multiple state-of-the-art DL convolutional neural networks (CNNs). Finally, a quantitative analysis of geometrical and morphometrical features and their correlation with tumour type, grade, and molecular status (MAML2 rearrangement in mucoepidermoid carcinoma) was carried out.
Results: A total of 5739 cases were analysed including 65% benign and 35% malignant tumours. The most common benign tumour was pleomorphic adenoma, while the most prevalent malignant tumour was mucoepidermoid carcinoma. Our novel (ML) classifiers results achieved excellent performance with F1 score of 0.90, 0.92 and 0.87, for benign vs malignant, malignant subtyping and grading, respectively. Significant differences between cellularity, cytoplasmic eosin and nucleus/cell ratio (p<0.05) were seen for all experiments, potentially signifying important diagnostic features. Most of the DL CNNs also achieved high F1 scores for benign versus malignant differentiation (>0.80), with EfficientNet-B0 giving the best performance (F1= 0.87) but with inferior accuracy than the ML classifier for malignant subtyping (highest F1=0.60 with ResNet-18 and ResNet-50) and tumour grading (highest F1 score=0.70 with EfficientNet-B0). The quantitative analysis showed a statistically significant difference between nuclear eccentricity and nucleus/cell ratio (p<0.01) and nuclear circularity and eccentricity (p<0.05) between low and high grades in AdCC and PAC respectively, as well as nuclear area and perimeter (p<0.01) between MAML2 fusion positive and negative cases.
Conclusion: SGT are rare, but have shown a gradual increasing incidence over the last decade and a half. We have reported the largest multicentre investigation of SGT to date but more extensive studies of SGT need to be conducted to understand and update the epidemiological landscape of these tumours. Our novel results report the successful application of ML and DL for histological analysis, subtyping and grading of SGT on H&E images for the first time. These findings will aid in pathological diagnosis and clinical decision making, but a larger multicentre cohort needs to be analysed to determine the true significance and clinical usefulness.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)