Use of artificial intelligence in diagnosis of head and neck precancerous and cancerous lesions: A systematic review.

Abstract

This systematic review analyses and describes the application and diagnostic accuracy of Artificial Intelligence (AI) methods used for detection and grading of potentially malignant (pre-cancerous) and cancerous head and neck lesions using whole slide images (WSI) of human tissue slides. Electronic databases MEDLINE via OVID, Scopus and Web of Science were searched between October 2009 - April 2020. Tailored search-strings were developed using database-specific terms. Studies were selected using a strict inclusion criterion following PRISMA Guidelines. Risk of bias assessment was conducted using a tailored QUADAS-2 tool. Out of 315 records, 11 fulfilled the inclusion criteria. AI-based methods were employed for analysis of specific histological features for oral epithelial dysplasia (n = 1), oral submucous fibrosis (n = 5), oral squamous cell carcinoma (n = 4) and oropharyngeal squamous cell carcinoma (n = 1). A combination of heuristics, supervised and unsupervised learning methods were employed, including more than 10 different classification and segmentation techniques. Most studies used uni-centric datasets (range 40-270 images) comprising small sub-images within WSI with accuracy between 79 and 100%. This review provides early evidence to support the potential application of supervised machine learning methods as a diagnostic aid for some oral potentially malignant and malignant lesions; however, there is a paucity of evidence using AI for diagnosis of other head and neck pathologies. Overall, the quality of evidence is low, with most studies showing a high risk of bias which is likely to have overestimated accuracy rates. This review highlights the need for development of state-of-the-art deep learning techniques in future head and neck research.