Research and development of artificial intelligence (AI) technology has been remarkable, especially in the field of endoscopy, resulting in its increasing use.1,2 Most studies that have employed AI technology related to colonoscopy have focused on computer-aided detection, which is used to detect polyps and lesions and alert the examiner, and also on computer-aided diagnosis, which can qualitatively determine whether lesions on an examination screen are neoplastic.3,4 Computed-aided detection functions have been used in clinical practice, and many studies that have employed AI technology have shown increased adenoma detection rates and adenomas per colonoscopy, thus highlighting its usefulness.4,5 Moreover, other studies that have employed AI technology have measured the endoscope withdrawal time to ensure the quality of examinations.6 However, only a few studies based on AI technology have contributed towards reducing the workload of physicians by assisting with the generation of endoscopy reports.7 This study was approved by the Institutional Review Board of the University of Fukui (approval number: 20210090).
Although endoscopists are required to write an endoscopy report immediately after performing a colonoscopy, it is difficult to recollect and describe all the colorectal polyps found at multiple locations and the numerous biopsies that have been performed. Additionally, writing an endoscopy report immediately after completion of a colonoscopy is not always possible because of other tasks that have to be performed and unforeseen incidents that may occur, thus obscuring memory. Furthermore, performing several colonoscopies and accurately recording the locations of lesions and all related findings during the course of a single day is physically and mentally draining and time-consuming.
AR-C1 (FUJIFILM Corp.) is an AI-based endoscopy report generation support system that automatically recognizes the use of instruments such as snares and biopsy forceps as well as insertion and removal of the endoscope during procedures (Fig. 1A). This system is supported by AI technology that was developed using machine learning and a convolutional neural network. AR-C1 is trained to recognize more than 100,000 pieces of data related to biopsy forceps, snares, hemostatic clips, and other treatment tools. Real-world snare recognition during 189 procedures exhibited 100% sensitivity, 96.9% specificity, and 97.4% accuracy. Additionally, real-world biopsy recognition during 218 procedures exhibited 100% sensitivity, 99.5% specificity, and 99.5% accuracy. AR-C1 automatically recognizes snares and biopsy forceps and temporarily registers a picture of the procedure as well as the specimen number and location on a report card (Fig. 1B). After completion of the examination, the examiner confirms the report card, and the findings are reflected in the endoscopy report (Fig. 1C). Thus, detection of polyps performed by the CAD-EYE (FUJIFILM Corp.), execution of procedures, specimen number, cecal intubation time, and total examination time are automatically registered by AI during colonoscopy. Manual actions, which include specifying the exact location and describing detailed procedures, such as the resection technique used, are input using a footswitch during colonoscopy. Lesion sizes are also input manually while generating the report after completion of the colonoscopy. If the AI system misidentifies any of the instruments, the examiner can make corrections using the footswitch during the examination.
In this study, we retrospectively determined whether AR-C1 could reduce the burden of generating reports related to colorectal endoscopic mucosal resection in clinical practice. The study was conducted in clinical practice in real time and included 40 endoscopic mucosal resections performed by a single examiner between September 2023 and March 2024; 20 procedures each with and without using AR-C1 were performed. For the control group, a report was generated by selecting the location and size of lesions, procedure, and specimen number from the available selection items and then selecting the desired pictures. For the AR-C1 group, pictures and examination details were inserted by finalizing the temporarily registered report card, and missing information was selected from the same selection items as those used for the control group. Additionally, for both the groups, comments regarding any other necessary details were added as needed. The report generation times were examined to determine whether there were any differences between the two groups. Each report generation time was automatically measured by the program from the time the report generation screen was opened until the time the report was finalized. These times were disclosed and analyzed only after all the 40 reports were generated.
There were no significant differences between the two groups of patients in terms of sex, age, number of polyps, polyp location, polyp size, polyp shape, and number of procedures performed, including non-polypectomy procedures (Table 1). The average report generation times per procedure were 258.4 seconds for procedures performed without using AR-C1 and 186.6 seconds for those performed using AR-C1, indicating that the generation time for the procedures performed using AR-C1 was significantly shorter by approximately one minute (Fig. 2A). The report generation times per 100 characters were 59.8 seconds for procedures performed without using AR-C1 and 42.4 seconds for those performed using AR-C1, thus indicating that the generation time was significantly shorter for procedures performed using AR-C1 (Fig. 2B).
In the present study, the report generation time using AR-C1 was significantly shorter than that without using AR-C1. Furthermore, footswitch operation during the examination did not increase the total examination time, which included the procedure and withdrawal times (Table 1). We believe that the shorter report generation time was attributable to the reduced number of tasks that required completion by the examiner. For example, the number and location of polyps did not require recall, and the locations of polyps and procedures did not require description by the endoscopist because AR-C1 performed these tasks.
Nevertheless, our study has several limitations. First, this was a single-center study conducted by a single examiner. Second, the sample size was small. Third, the endoscopy reports generated in this study were written in Japanese. Finally, although the report generation times were automatically measured by the program for each case and were not revealed until the time of analysis, it was impossible to blind the examiner as to whether AR-C1 was used; therefore, bias associated with measurement related to the generation time of the endoscopy report by the examiner may have possibly occurred. Therefore, a multicenter multilingual study with well-matched patient groups is required to validate our findings.
With the current AR-C1 system, details of the location of lesions and procedure performed must be entered using a footswitch, which is somewhat cumbersome; however, this does not affect the examination time. A device that uses voice recognition and a microphone instead of a footswitch is currently being developed, and is expected to improve operability. Further advances in AI technology may enable systems to automatically reflect other information, such as the location of lesion and shape as well as details of the procedure, in endoscopy reports. Regarding the recognition of locations of lesions, previous studies have reported that some AI technologies can discriminate between the cecum or appendix and other areas, while others can approximately divide the lesion into different areas.8,9 However, current AI technologies are not yet capable of identifying each location of lesions in the cecum to rectum with high accuracy. We initially attempted to use AI technology to distinguish each location in the colon; however, misrecognitions that required human correction occurred, thus increasing the time and effort required to generate reports. However, in the present study, we obtained good results by using a footswitch to specify the location but not to indicate endoscope insertion and removal, which were recognized by AI technology. Efficacy of computer-aided diagnosis has been demonstrated in clinical practice. If AI technology continues to evolve at the current rate, technology that combines various functions can be established in the near future.10 As a result, reports can be accurately generated at the completion of examinations and treatments.
In conclusion, the use of an AI-based endoscopy report support system can reduce the time required to generate reports. These findings can be applied to the development of future AI technology systems.
Conflicts of Interest
The authors declare that they have received research grants from FUJIFILM Corp. to support this study. No other potential conflicts of interest relevant to this article were reported.
Funding
None.
Author Contributions
Conceptualization: TNa, YN; Data curation: TNa; Formal analysis: TNa; Investigation: all authors; Methodology: TNa, YN; Project administration: TNa, YN; Supervision: YN; Visualization: TNa; Writing–original draft: TNa; Writing–review & editing: all authors.
Fig. 1.Flow chart of the AR-C1 system (FUJIFILM Corp.) used during colonoscopy. (A) AR-C1 automatically recognizes insertion and removal of the endoscope and use of instruments such as snares and biopsy forceps during the examination. (B) At the completion of the examination, details of the procedure recognized by AR-C1, and pictures of the procedure as well as the specimen number and site are temporarily registered on the report card. (C) Contents of the report card.
Fig. 2.(A) Report generation time per case without using AR-C1 (without artificial intelligence [AI]) and using AR-C1 (with AI). (B) Report generation time per 100 characters without using AR-C1 (without AI) and using AR-C1 (with AI).
Table 1.Background characteristics
|
Characteristic |
Without AI (n=20) |
With AI (n=20) |
p-value |
|
Sex (F/M) |
5/15 |
8/12 |
0.50 |
|
Median age (yr, IQR) |
69 (39–89) |
72 (36–87) |
0.32 |
|
Mean no. of polyps |
3.3 |
3.6 |
0.69 |
|
Polyp location (rectum/colon) |
9/57 |
7/64 |
0.60 |
|
Polyp size (<10 mm/≥10 mm) |
24/42 |
22/49 |
0.59 |
|
Polyp shape (flat/protruded) |
33/33 |
29/42 |
0.31 |
|
Mean no. of procedures |
3.4 |
3.6 |
0.70 |
|
Mean character count of the reports |
492.3 |
480.4 |
0.88 |
|
Total examination time (s) |
1,410.9 |
1,494.0 |
0.71 |
REFERENCES
- 1. Choi J, Shin K, Jung J, et al. Convolutional neural network technology in endoscopic imaging: artificial intelligence for endoscopy. Clin Endosc 2020;53:117–126.ArticlePubMedPMCPDF
- 2. Hassan C, Misawa M, Rizkala T, et al. Computer-aided diagnosis for leaving colorectal polyps in situ : a systematic review and meta-analysis. Ann Intern Med 2024;177:919–928.ArticlePubMedPDF
- 3. Djinbachian R, Haumesser C, Taghiakbari M, et al. Autonomous artificial intelligence vs artificial intelligence-assisted human optical diagnosis of colorectal polyps: a randomized controlled trial. Gastroenterology 2024;167:392–399.ArticlePubMed
- 4. Park JB, Bae JH. Effectiveness of a novel artificial intelligence-assisted colonoscopy system for adenoma detection: a prospective, propensity score-matched, non-randomized controlled study in Korea. Clin Endosc 2025;58:112–120.ArticlePubMed
- 5. Lou S, Du F, Song W, et al. Artificial intelligence for colorectal neoplasia detection during colonoscopy: a systematic review and meta-analysis of randomized clinical trials. EClinicalMedicine 2023;66:102341.ArticlePubMedPMC
- 6. Lux TJ, Saßmannshausen Z, Kafetzis I, et al. Assisted documentation as a new focus for artificial intelligence in endoscopy: the precedent of reliable withdrawal time and image reporting. Endoscopy 2023;55:1118–1123.ArticlePubMedPMC
- 7. Rey JF. As how artificial intelligence is revolutionizing endoscopy. Clin Endosc 2024;57:302–308.ArticlePubMedPMCPDF
- 8. Taghiakbari M, Hamidi Ghalehjegh S, Jehanno E, et al. Automated detection of anatomical landmarks during colonoscopy using a deep learning model. J Can Assoc Gastroenterol 2023;6:145–151.ArticlePubMedPMCPDF
- 9. Kim BS, Cho M, Chung GE, et al. Density clustering-based automatic anatomical section recognition in colonoscopy video using deep learning. Sci Rep 2024;14:872.ArticlePubMedPMCPDF
- 10. Gong EJ, Bang CS. Revolutionizing gastrointestinal endoscopy: the emerging role of large language models. Clin Endosc 2024;57:759–762.ArticlePubMed
Citations
Citations to this article as recorded by
KSGE
, Takuto Nosaka
PubReader
ePub Link
Cite
XML Download
