Manas Dave reflects on topics in our sister journal Evidence-Based Dentistry.

figure 1

©NEMES LASZLO/SCIENCE PHOTO LIBRARY/Getty Images Plus

Introduction

‘Application and performance of artificial intelligence technology in oral cancer diagnosis and prediction of prognosis: A systematic review', by S. B. Khanagar et al. was published in Diagnostics in 2021.1 A summary of this article was published in Evidence-Based Dentistry in 2022.2

Background

Mouth cancer encompasses cancer affecting different parts of the mouth (including the lips, gums and soft tissues) whilst oropharyngeal cancer starts in the throat (pharynx) behind the mouth including tonsil cancer and cancer of the back of the tongue. In 2016, 3,744 people in the UK were diagnosed with mouth cancer and 2,977 diagnosed with oropharyngeal cancer. Whilst survival rates are improving they are still limited with 56% of patients with mouth cancer and 66% of patients with oropharyngeal cancer surviving for five years or more.3 Amongst oral cancers, squamous cell carcinomas are the most common encompassing 90% of all oral cavity cancers.1

Risk factors for developing oral cancer include oral potentially malignant disorders (OPMDs) which encompass a number of disease entities from chronic candidiasis to actinic keratosis.4 Other factors include oral epithelial dysplasia which describe a series of abnormal changes to the oral epithelium. Modifiable risk factors include those environmentally exposed such as smoking, alcohol and nutrition. With artificial intelligence (AI) providing a novel approach to current challenges, the aim of this systematic review was to report on the application and role of AI in the diagnosis and prediction of oral cancer occurrence.

Methods

An electronic database search of PubMed, Scopus, Embase, Cochrane, Web of Science, Saudi Digital Library and Google Scholar were conducted. From January 2000-March 2021. Original research reporting on AI technology and data used in evaluating these AI-models were included with no restrictions on study design. QUAmDAS-2 was used to assess the quality of the studies reporting on diagnostic accuracy.

Results

  • Studies included: 16 articles were included in this review encompassing 14 cross-sectional, one cohort and one retrospective study. The number of photographs ranged from 10-33,065 and the modalities included datasets, clinical imaging, recorded spectra, auto-fluorescent imaging, hyperspectral imaging, histological imaging and high-resolution cytology imaging

  • Role of AI in oral cancer: AI technology was used to differentiate between normal, premalignant and malignant conditions predicting the likelihood of oral cancer incidence. Additionally, AI was used for prognosis, early detection, prediction of risk recurrence, disease development from potential malignant lesions and predicting the survival of patients. Nine studies reported using convolutional neural networks (CNNs) and another seven studies used artificial neural networks (ANNs)

  • AI accuracy: The accuracy of AI diagnosing oral cancer ranged from 59.9-99.7%, sensitivity 41.98-98.6% and specificity 64.7-10%

  • Quality assessment: Most studies used photographic data as input into the CNNS and ANNs, hence 76.5% reported a low risk of bias for the patient-selection domain. Studies which used histopathology as the reference standard were also graded as low risk. In some studies, patient selection and index test domains were unclear. There were no domains identified as high risk of bias.

Conclusions

The authors stated:

‘…AI-based algorithms showed more accurate results in predicting oral cancer occurrence.'

Evidence-based AI applications could greatly contribute to public health and aid clinicians in diagnostics.

Commentary

This systematic review highlighted a range of different AI tools that can identify patterns in a wide range of images to diagnose oral cancer. Whilst the topic was interesting, the systematic review had limited detail on its protocol (there was missing information such as a clear search strategy), no title or abstract screening and limited detail on the data extraction process. The included studies were heterogeneous in their assessed variables which precluded quantitative analysis. Evidence-based AI applications could greatly contribute to public health and aid clinicians in diagnostics however more high-quality research is needed in this area.