Is the use of artificial intelligence the main stage for detecting polyps during colonoscopy?

Abstract

Colorectal cancer (CRC) is the third most frequently diagnosed cancer and the second leading cause of cancer death worldwide. In this regard, CRC screening is one of the most important issues in modern preventive medicine. Colorectal polyps are potential predictors of CRC, and therefore represent one of the leading targets for screening colonoscopy. The difficulty of analyzing the information obtained during colonoscopy, including the size, location, shape, type of polyps, the need to standardize morphological data, determines that recently a number of works have promoted the opinion on the advisability of using various artificial intelligence (AI) methods to improve the effectiveness of endoscopic screening for CRC. At the same time, they point to a number of errors and methodological problems in the use of AI systems for the diagnosis of colorectal polyps. In this regard, the interpretation of the work of Shi et al, devoted to the use of a machine learning-based predictive model for monitoring the results of colorectal polypectomy, is undoubtedly interesting. In our opinion, the prospects for using AI to assess endoscopic screening for CRC look certainly positive, but the road to its widespread use will not be easy.

Key Words: Artificial intelligence; Colorectal cancer; Colorectal polyps; Colonoscopy; Diagnostics

Core Tip: Endoscopic screening of colorectal polyps is one of the most relevant methods for preventing colorectal cancer. Recently, various artificial intelligence (AI) systems have been used extremely actively to improve the efficiency of polyp diagnostics during colonoscopy. However, until now along with increasing the efficiency of polyp detection and reducing the number of errors during colonoscopy, a number of modern studies note the presence of significant limitations and the possibility of false diagnostics when using AI in real clinical practice. The prospects for using AI for this issue are undeniable, but long-term efforts are required along this road.

TO THE EDITOR

We are delighted to read the high-quality article by Shi et al[1], published in the World Journal of Gastroenterology. The authors examined 1694 patients who underwent colorectal polypectomy. To assess the recurrence of colorectal polyps, patients were followed up for one year. A predictive model was constructed using machine learning to identify risk factors for polyp recurrence. Multivariate logistic regression analysis identified eight independent risk factors for the recurrence of colorectal polyps after one year. The most significant among these risk factors were smoking duration, family history, and age. We consider it important to draw attention to the fact that the listed risk factors were associated with the recurrence of colorectal polyps and not with the detection of initial polyps. The authors concluded that the XGBoost model is highly effective in optimizing the surveillance of patients who underwent colorectal polypectomy.

Colorectal cancer (CRC) is the third most frequently diagnosed cancer and the second leading cause of cancer death worldwide. In this regard, CRC screening is one of the most important issues in modern preventive medicine[2]. Colorectal polyps are potential predictors of CRC, and therefore represent one of the leading targets for screening colonoscopy[3,4]. Modern studies confirm the advisability of using colonoscopy for CRC screening. Knudsen et al[5] observed 195453 people for 12 years. Of 81151 patients of them had a screening colonoscopy with no pathological changes (the first group). Of 114302 people did not have an endoscopic examination (the second group). 10 years after the first examination, CRC was detected on repeat endoscopy in 0.5% of patients in the first group and 1.95% of people in the second group. The frequency of deaths from CRC was 0.2% in the first group and 0.6% in the second group. The authors concluded that colonoscopic screening is highly effective in reducing the incidence and mortality of CRC[5].

Given the complexity of the information obtained during colonoscopy, including the analysis of the size, location, shape, type of polyps, and the need to standardize morphological data, a number of recent studies have promoted the view on the advisability of using various artificial intelligence (AI) methods to improve the effectiveness of endoscopic screening for CRC[6-9]. However, some authors have expressed doubts about this point of view[10]. A recent systematic review of recommendations in different regions of the world for surveillance of patients with colorectal polyps found significant variations and limitations[11].

New studies generally support the idea of the complexity and discussion ability of using AI in colonoscopy. Comparison of the results of analysis of 77 colorectal polyps in 59 patients revealed higher accuracy, sensitivity and specificity of polyp detection using AI compared to assessment by five board-certified specialists[12]. Ozawa et al[13] trained a convolutional neural network (CNN) using 16418 images of 4752 colorectal polyps and 4013 images of normal colorectal tissues, and subsequently validated the performance of the trained CNN on 7077 colonoscopy images, including 1172 images from 309 various types of colorectal polyps. In summary, the CNN showed promising results in the ability to detect and classify colorectal polyps using endoscopic images. A modern meta-analysis that pooled the results of six randomized trials (1718 patients) found a significantly lower rate of errors in the diagnosis of colorectal adenomas (P < 0.001) and colorectal polyps (P < 0.001) using computer-aided colonoscopy compared with standard white-light colonoscopy[14]. However, some studies have raised doubts about the effectiveness of using AI to improve the efficiency of colonoscopy. A number of studies using computer-aided quality improvement are characterized by a low number of endoscopies performed, which is insufficient to assess the quality of the method[15]. There is a risk of the AI itself changing the settings when using a small sample of patients for machine learning, which limits the possibilities of AI training in clinical practice[16]. Computer-aided detection (CAD) in a number of endoscopic studies did not demonstrate an increase in the detection rate of advanced or large (> 10 mm) adenomas, sessile serrated lesions and CRC[17,18]. CAD can increase the frequency of false-positive results in colonoscopy[19], which increases psychological distraction of endoscopists and prolongs the time of endoscopic examination[20]. Troya et al[21] assessed the visual reaction time for polyp detection in an endoscopist with and without CAD. The use of CAD did not improve human reaction time and made it difficult to assess the picture of normal intestinal mucosa. The authors of a review article on the role of AI in colonoscopy, to be published in the journal Digestion in 2025, analyzed the effectiveness of various AI systems in improving the diagnosis and characterization of lesions on colonoscopy. It was concluded that AI can increase the detection rate of colorectal adenomas and reduce the error rate in diagnosing adenomas, but may increase the rate of false positive results and lead to an increase in unnecessary polypectomies. CAD of the characteristics of formations during colonoscopy gives ambiguous results and may distort the accuracy of differentiation of polyp types. The authors of the review think that the AI use can lead to a decrease in the qualifications of endoscopists and can worsen the cost-effectiveness of screening. The authors summarize their work by noting that although there are positive results of using AI for the diagnosis of colorectal polyps, potential disadvantages of the new technique may hinder its further implementation in real clinical practice[22].

We think that to better assess the effectiveness of AI in detecting colorectal polyps, it would be useful to conduct new, larger randomized control trials comparing AI-assisted colonoscopy with traditional methods. Different methods of standardizing information obtained from patients using AI should be compared with each other. Such studies could include several groups using AI-assisted colonoscopy with different data standardization methods that would be compared with traditional colonoscopy. The study should include data such as polyp detection rate, polyp location and type, procedure time, all of which would enhance the reliability of the evidence regarding the effectiveness of AI in clinical practice.

CONCLUSION

In our opinion, the prospects for using AI to assess endoscopic CRC screening look certainly positive. When thinking about the reasons for the difficulties in implementing AI systems in assessing colonoscopy results, a “simple” question arises: “Who should think: Computer or human?” The likely misconception is that we overestimate the capabilities of the computer and do not sufficiently understand the specificity and standardization of the information that AI must receive. It can be assumed that the capabilities of associative analysis in a computer may differ significantly from human synthetic perception. This difference requires new data structures adapted to AI processing methods, such as pre-processed video streams that improve pattern recognition. Future efforts should focus on improving the technology to better detect clinically significant lesions, reduce the risk of potential deskilling of endoscopists, and assess the long-term consequences for patient outcomes. By addressing this issue, AI can strengthen its role as a transformative tool in the fight against CRC, complementing and enhancing the endoscopist’s expertise rather than replacing it[23]. Let us repeat that the prospects for the use of AI in image processing and, in practical terms, in colonoscopy, are absolutely positive, but the road to its widespread use may not be easy and may be quite long. The work of Shi et al[1] on the application of a machine learning-based predictive model for prolonged monitoring of the results of colorectal polypectomy is devoted to the issue of the AI use in medicine and therefore deserves attention.

ACKNOWLEDGEMENTS

We thank the reviewers for their comments that helped to improve the manuscript.