Current evidence and future directions on improving the endoscopic recognition of early colorectal carcinoma using artificial intelligence – a scoping review

Aims Artificial intelligence (AI) has great potential to improve the endoscopic recognition of early stage colorectal carcinoma (CRC). As a result, AI may facilitate a higher rate of endoscopic resection of superficially invasive CRC. This scoping review aims to summarize current evidence regarding the use of AI for improving endoscopic recognition of early stage CRC, to identify knowledge gaps on this topic, and to provide an overview of the methodologies currently used.

Methods A systematic search was performed following the PRISMA-ScR guidelines. PubMed (including Medline), Scopus, Embase, IEEE Xplore, and ACM Digital Library were searched for relevant publications up to April 2023. Studies were suitable for inclusion when they were using AI for distinguishing (early stage) CRC from colorectal polyps on endoscopy or endocytoscopy imaging, using histopathology as gold standard. Sensitivity, specificity, or accuracy should be reported as outcome measures, and articles must be available in English. Study selection was performed by two reviewers independently.

Results Out of 4185 screened articles, 24 articles were included in this review. With the exception of three studies, all included studies were published in the past 5 years. None of the studies included reported CADx system testing in real-time. Convolutional neural network architectures were used in all studies except one, which employed a support vector machine and three studies that did not specify algorithm details. All studies that mentioned the endoscopy brand used Olympus endoscopy systems (n=21), six of which also used Fujifilm/Fujinon. CADx system classification categories ranged from two categories, such as lesions suitable or unsuitable for endoscopic resection, to five categories, such as hyperplastic polyp, sessile serrated lesion, adenoma, cancer, and other. CRC was classified varyingly within the classification categories, including diagnosis of CRC together with adenomas (n=11), diagnosis of CRC in a separate classification category (n=9), or estimation of CRC invasion depth (n=4). The number of images used in testing databases for the CADx systems varied from 69 to 48.391, the latter using dozens of images made of one lesion. The diagnostic performances have substantial variability, with sensitivities ranging from 55.0-98.1%, specificities from 67.5-100%, and accuracies from 74.7-94.9%.

Conclusions This scoping review highlights that the use of AI to improve endoscopic recognition of early stage CRC is an upcoming field of research. Diagnostic performances are promising, but large heterogeneity in the methodologies used, should be taken into account when interpreting the results. There is a knowledge gap regarding the real-time performance of CADx systems during multicenter external validation with sufficient amounts of original test data. To enhance the utility of CADx systems in clinical practice, future research should focus on the development of CADx systems that can differentiate CRC from premalignant lesions, while also providing an indication of submucosal invasion depth.

Conflicts of interest

Authors do not have any conflict of interest to disclose.

Publication History

Article published online:
15 April 2024