In a randomized multicenter study, an international research team in Milan found that computer image analysis significantly increased the detection rate of adenomas and the number of adenomas detected per colonoscopy.1 With a comparable withdrawal time, the results of digitally assisted colonoscopy were also particularly robust and reliable. Given the relevance played by the most complete detection to date for later degeneration risk, this procedure can be a valuable aid even for experienced diagnosticians.
The study included data from 685 participants in three Italian centers who were hospitalized for cancer screening or underwent colonoscopy after polypectomy or positive fecal immunochemical tests (FIT). For the computer-aided detection (CADe) of polyps, novel devices were used, which allowed a real-time superimposition of suspected findings of the software with the image of the conventional endoscopy. The procedure of adenoma detection based on trained artificial intelligence (AI) thus facilitated the comparison of findings with those of the examiners.
The conventional colonoscopies were performed exclusively by experienced specialists in endoscopy (> 2,000 colonoscopies each).
The study was statistically designed to detect at least a 10% difference in results between endoscopists and AI. It was found that CADe, at 54.8%, showed a significantly higher adenoma detection rate (ADR) than conventional endoscopy by experienced diagnosticians (40.4%).
CADe also found significantly more adenomas per endoscopy with an average of 1.07 than the human control group with 0.71. This higher detection rate was particularly evident for adenomas of the size 6-9 mm. CADe found these tumors in 10.6% of the participants, the classical endoscopy only in 5.8%. Even the smallest adenomas (≤ 5 mm) were detected significantly more frequently by computer than by specialists (33.7% vs. 26.5%).
The average duration of endoscope retraction was about the same for both procedures, about 7 minutes each. The rate of non-neoplastic lesion dissection (26.0% vs. 28.7%) was also consistent. Therefore, the authors of the study do not assume that there was a systematic difference between the study groups, e.g. with regard to attention during the examination.
The authors also emphasize that the treating endoscopists and CADe each had simultaneous access to the same information from the test subjects - in contrast to previous studies, in which the digitized data could not be shown in real-time and not in superposition with the current endoscopy image. In addition, the endoscopists often confirmed the accuracy of the CADe diagnosis on close examination when the region of the CADe's findings was highlighted in green on the endoscopy image.
The study authors also consider the presented results to be particularly robust, as they were independent of the localization and morphology of the individual findings. As was to be expected with a machine evaluation, the results did not vary with these parameters. In contrast, the diagnostic reliability of conventional endoscopy decreased depending on the difficulty of the endoscopy.
Source:
Repici A, et al. Efficacy of Real-Time Computer-Aided Detection of Colorectal Neoplasia in a Randomized Trial. Gastroenterology 2020, e-Pub ahead of print. Doi: https://doi.org/10.1053/j.gastro.2020.04.062.