Colonoscopy with computer-aided diagnosis may allow better assessment of diminutive polyps

The study aimed to determine whether computer-aided diagnosis had a negative predictive value of 90% or more for identifying diminutive rectosigmoid adenomas.


Colonoscopy using computer-aided diagnosis (CAD) may allow clinicians to better distinguish diminutive polyps that require resection from those that do not, a study found.

Researchers in Japan performed a single-group, open-label, prospective study of consecutive patients undergoing colonoscopy to determine the performance of real-time CAD with endocytoscopes in narrow-band imaging and methylene blue staining modes. Pathologic diagnosis of detected diminutive polyps, defined as those 5 mm or smaller, by real-time CAD outputs was compared with pathologic diagnosis of the resected specimen.

The main objective of the study was to determine whether CAD had a negative predictive value (NPV) of 90% or more for identifying diminutive rectosigmoid adenomas, which is the threshold at which clinicians can “diagnose and leave” non-neoplastic polyps. The researchers used a best-case scenario in which it was assumed that polyps lacking CAD or pathologic diagnosis were considered true-positive or true-negative, respectively, and a worst-case scenario in which it was assumed that polyps lacking CAD or pathologic diagnosis were considered false-positive or false-negative, respectively.

The study results were published Aug. 14 by Annals of Internal Medicine.

CAD was used to assess a total of 466 diminutive polyps from 325 patients (median age, 67 years; 72.3% men). Two hundred eighty-seven diminutive polyps were neoplastic, and 175 were non-neoplastic. Two hundred fifty diminutive polyps were in the rectosigmoid colon and were included in evaluation of the study's primary end point. The pathologic prediction rate was 98.1% (457 of 466 polyps). The NPV for the stained mode of CAD in assessing diminutive rectosigmoid adenomas was 96.4% (95% CI, 91.8% to 98.8%) in the best-case scenario and 93.7% (95% CI, 88.3% to 97.1%) in the worst-case scenario. For CAD with narrow-band imaging, the NPVs for the best- and worst-case scenarios were 96.5% (95% CI, 92.1% to 98.9%) and 95.2% (95% CI, 90.3% to 98.0%).

The researchers noted that two-thirds of the colonoscopies were done by experts, that 186 polyps that were not assessed by CAD were excluded, and that no comparative data were available, among other limitations. However, they concluded that real-time CAD as designed for endocytoscopy performs well enough to meet the clinical threshold for the “diagnose-and-leave” strategy in diminutive non-neoplastic rectosigmoid polyps and indicated that such technology might improve colonoscopy's cost-effectiveness.

An accompanying editorial pointed out several issues that should be resolved before this use of CAD becomes common in clinical practice. For example, CAD could not distinguish between hyperplastic and adenomatous polyps proximal to the sigmoid colon and does a poor job of assessing sessile serrated polyps, a newly recognized precursor to cancer, the editorialists noted.

However, they said the study is important and provides a glimpse of the future of the field, in which collaboration among engineers, computer scientists, and clinicians may help improve health care quality. “The study also represents another example of the benefit of big data, a prerequisite for the deep-learning technology used to train the CAD system,” the editorialists wrote. “To err is human, but CAD may help us reduce the frequency of human errors.”