December's Paper of the Month report looks at how computer-aided diagnosis (CAD) can be used to improve the performance of diagnostic techniques for smaller polyps and to decrease the incidence of colorectal cancer.

Real-time use of artificial intelligence in identification of diminutive polyps during colonoscopy. Ann Intern Med, 2018
Authors: Mori Y, Kudo S, Misawa M, Saito Y, Ikematsu H, Hotta K, et al.

What is known on the subject?

Colonoscopy is considered the gold standard to decrease colorectal cancer incidence and mortality with polypectomy recommended for all polyps over 5mm. For smaller (diminutive) polyps in the rectosigmoid, narrow band imaging or staining can be used to accurately differentiate neoplastic and benign polyps. According to the PIVI-2 declaration, an advanced endoscopic diagnosis with a negative predictive value of 90% can be used to choose a 'diagnose-and-leave' strategy for these polyps. However, only experts have been shown to achieve this level of performance and thus standardisation of optical diagnosis is an unmet need. Computer-aided diagnosis (CAD) methods have been tested to overcome this limitation but so far none has been prospectively tested.

What this study adds

Study design

Single group, open label, prospective study.

Primary endpoint

Whether CAD-stained analysis produced ≥ 90% negative predictive value (NPV) for diagnosing diminutive rectosigmoid adenomas.

Secondary endpoints

The ability of CAD-NBI to distinguish diminutive rectosigmoid polyps and technical success rate of performing CAD.


Consecutive patients scheduled to undergo routine colonoscopy at an academic centre in Japan, between June and December 2017, were recruited.


Patients were submitted to staining with 1.0% methylene blue and narrow band imaging with contact light microscopy. A CAD system was connected to the endoscope and provided a prediction of the pathologic status in real time. The outputs of the CAD prediction were compared against the goldstandard of formal pathologic assessment of the excised polyps. Results were calculated based both on worst-case scenario, where polyps lacking either CAD or pathology were treated as false-positive or -negative; and best-case scenario, where they were treated as true-positive or -negative.


Overall, 466 diminutive (including 250 rectosigmoid) polyps from 325 patients were assessed by CAD, with a pathologic prediction rate of 98.1% (457 of 466). Colonoscopies were performed by novices and experts. The indications were screening in 42 patients, surveillance after polypectomy in 121 patients, the presence of symptoms in 91 patients, planned treatment of other polyps (≥6 mm) in 69 patients, and other reasons in 2 patients. The negative predictive value of CAD for diminutive rectosigmoid adenomas was 96.4% (95% CI: 91.8% to 98.8%) (best-case scenario) and 93.7% (CI: 88.3% to 97.1%) (worst-case scenario) with stained mode and 96.5% (CI: 92.1% to 98.9%) (best-case scenario) and 95.2% (CI: 90.3% to 98.0%) (worst-case scenario) with narrow-band imaging.

Author interpretation

Real-time use of the fully automated CAD system designed for endocytoscopes can meet the clinical threshold required for the “diagnose-and-leave” strategy for diminutive, nonneoplastic rectosigmoid polyps, which may help improve the cost- effectiveness of colonoscopy.

Implications for colorectal practice

This study has two main implications:

  1. for endoscopists
  2. for the role of artificial intelligence in current practice.

Implications for endoscopy

This method presents a way to standardise the assessment of diminutive polyps with a CAD system that provides real time decision support recommendations. This has been shown to achieve the benchmark for use in the clinical setting according to the PIVI-2 declaration. Moreover, the results were similar for staining and for NBI and the performance was similar for novices and experts. If it is possible to diagnose the hyperplastic polyps with a high degree of accuracy, then it is possible to leave them in place and avoid complications, decrease procedural time and decrease costs.

Implications for AI in clinical practice

Decision support tools using artificial intelligence have been gaining momentum and there are great expectations for this kind of technology. However, as the field is very novel, the methods to test and implement it are not standardised. This paper represents a prospective open-label study that takes real world data as the input and not datasets prepared and analysed separate from clinical practice. Therefore, the great efforts of the authors should be congratulated. It also informs us on how the scrutiny of AI tools should be implemented i.e. through a strict evidence-based framework that is the best assurance that ultimately, patients will indeed benefit from this new technology.