When It Comes to Polyp Diagnosis With CADx, Location Matters

VIENNA — The effectiveness of computer-aided diagnosis (CADx) in differentiating neoplastic from non-neoplastic polyps depends on the region of the colon examined, according to a systematic review and meta-analysis.

In particular, the diagnostic performance of CADx for polyps showed significantly lower specificity in the proximal colon than in the distal colon.

“While current CADx systems are suitable for use in the distal colon, they should not be employed for diagnosing polyps in the proximal colon until new, higher performing systems are developed specifically for these lesions,” said study lead Tommy Rizkala, MD, Endoscopy Unit, IRCCS Humanitas Clinical and Research Center, Rozzano, Italy.

The “main strength” of the review is that we contacted each study author for more specific information and were therefore able to divide the data into the proximal colon and the rectosigmoid colon, he explained.

“This is the first paper that has really collected these data. Most papers provide data for the entire colon or just for the rectosigmoid colon,” said Rizkala, who presented the findings at the United European Gastroenterology (UEG) Week 2024.

The study was also recently published in Clinical Gastroenterology and Hepatology.

Optical diagnosis enables real-time histologic predictions of polyps 5 mm or smaller during colonoscopy, offering potential clinical and cost-saving benefits. Two optical diagnostic strategies are used for polyps in this size range based on location: A leave-in-situ strategy (applied only in the rectosigmoid colon when there is high confidence of non-neoplastic polyps) and a resect-and-discard strategy (applied only in the whole colon when there is high confidence of neoplastic polyps upon optical diagnosis).

Rizkala carried out a review of studies that evaluated the performance of real-time CADx alone — independent of endoscopist judgement — for predicting the histology of colorectal polyps 5 mm or smaller. The primary endpoints were CADx sensitivity and specificity in the proximal colon (the portion extending from the descending colon to the cecum) and the distal colon (limited to the rectosigmoid region). Secondary outcomes were the negative predictive value (NPV), positive predictive value (PPV), and accuracy of the CADx alone in the proximal colon and the distal colon.

Lower Specificity in the Proximal Colon

An analysis of data based on 7782 polyps ≤ 5 mm from 11 studies found specificity values of 0.62 (95% CI, 0.52-0.71) and 0.85 (95% CI, 0.75-0.92) for the proximal and distal regions of the colon, respectively, with a risk ratio (RR) of 0.74 (95% CI, 0.72-0.84), meaning that CADx accuracy was significantly lower in the proximal colon than in the distal colon.

“According to the optical diagnosis strategy, we can use the leave-in-situ approach for the distal colon because the performance is adequate, but for the rest of the colon, CADx requires further enhancement,” Rizkala said.

Sensitivity values were 0.89 (95% CI, 0.83-0.93) and 0.87 (95% CI, 0.80-0.92) for the proximal and distal regions, respectively, with an RR of 1.00 (95% CI, 0.97-1.03).

Regarding the secondary outcomes, the NPV was 0.64 vs 0.93 for the proximal vs distal colon, with an RR of 0.71 (95% CI, 0.64-0.79), and accuracy was 0.81 vs 0.86, with an RR of 0.95 (95% CI, 0.91-0.99).

With the higher prevalence of neoplastic lesions in the proximal colon than in the distal colon, a lower NPV was observed in the proximal colon, Rizkala noted.

The PPV was 0.87 vs 0.76 for the proximal vs distal colon, with an RR of 1.11 (95% CI, 1.06-1.17), so the two parts of the colon were comparable, he reported.

In the future, CADx systems should focus on using lesions from the proximal colon to train more accurately because currently CADx systems are trained on the available endoscopic data in which most of those polyps are from the rectosigmoid colon, Rizkala said.

We would also “like manufacturers of CADx systems to provide public access to data balanced between the proximal and distal regions of the colon,” he added.

Diagnosis More Challenging Than Detection With CADx

Commenting on the study, co-moderator David G. Graham, MD, consultant gastroenterologist at University College London Hospital, London, England, remarked that, “The key questions here relate to why are these systems underperforming in the proximal colon, and how can we improve this?”

Are these results “due to the very different appearance of adenomas in the distal colon vs the proximal colon on CADx (which is not what we see as endoscopists but seems to be what the systems are seeing), or is it due to a different characterization of polyps,” that is, more sessile serrated lesions in the proximal colon than in the distal colon? he speculated.

Also commenting on the study was Raf Bisschops, MD, head of endoscopy at KU Leuven, Leuven, Belgium. He remarked that the review underscores the fact that optical diagnosis by artificial intelligence is a more challenging task than detection.

It is “not entirely clear” what would explain the difference in performance of CADx between the distal colon and proximal colon, he said. It can’t be excluded that the inclusion of different CADx systems, some of which clearly underperformed, may account for the difference.

He went on to suggest that the differences might be down to location beyond proximal and distal.

“The difference in performance between the right and left colon is also interesting, since recent insights in the molecular and morphological features of hyperplastic polyps indicates that there are different classes with more goblet cell-rich hyperplastic polyps in the right colon, and more microvesicular hyperplastic polyps in the left.”

These have “distinct microscopic and endoscopic appearances” that could account for a difference in performance of a CADx system if not included in the training and validation sets, he explained.

Rizkala and Graham reported no relevant disclosures. Bisschops reported receiving research grants and speaker fees from Medtronic, Fujifilm, and Pentax.