A novel AI device for real-time optical characterization of colorectal polyps

Abstract

Accurate in-vivo optical characterization of colorectal polyps is key to select the optimal treatment regimen during colonoscopy. However, reported accuracies vary widely among endoscopists. We developed a novel intelligent medical device able to seamlessly operate in real-time using conventional white light (WL) endoscopy video stream without virtual chromoendoscopy (blue light, BL). In this work, we evaluated the standalone performance of this computer-aided diagnosis device (CADx) on a prospectively acquired dataset of unaltered colonoscopy videos. An international group of endoscopists performed optical characterization of each polyp acquired in a prospective study, blinded to both histology and CADx result, by means of an online platform enabling careful video assessment. Colorectal polyps were categorized by reviewers, subdivided into 10 experts and 11 non-experts endoscopists, and by the CADx as either "adenoma" or "non-adenoma". A total of 513 polyps from 165 patients were assessed. CADx accuracy in WL was found comparable to the accuracy of expert endoscopists (CADx_WL/Exp; OR 1.211 [0.766-1.915]) using histopathology as the reference standard. Moreover, CADx accuracy in WL was found superior to the accuracy of non-expert endoscopists (CADx_WL/NonExp; OR 1.875 [1.191-2.953]), and CADx accuracy in BL was found comparable to it (CADx_BL/CADx_WL; OR 0.886 [0.612-1.282]). The proposed intelligent device shows the potential to support non-expert endoscopists in systematically reaching the performances of expert endoscopists in optical characterization.

Fig. 1. AI device intended use workflow.

The endoscopy video stream flows through the device with no modification or delay (<1.5 μs). In real time (50–60 ms), the device augments the video stream by adding overlay markers to surround areas of interest, such that they can be further inspected by the endoscopist. During colon navigation (1), the endoscopist is focused on exposing the mucosa appropriately in order to facilitate the task of polyp detection. The AI intelligent device activates when a polyp is detected (2, CADe). At this stage, the endoscopist examines the mucosa in order to characterize the lesion and decide on clinical action: if a polyp is framed consistently by the endoscopist, the CADx activates automatically (3) and the histology prediction is added to the green box. After possible endoscopic resection of the polyp (4), navigation is resumed (1) and CADx automatically disengages. These tasks are iterated as a loop until the end of the procedure.

Fig. 2. ROC curves calculated using the bootstrap method.

From left to right: CADx white light vs non-expert endoscopists, CADx white light vs expert endoscopists, and CADx white light vs CADx blue light. The light band around the ROC curves represents 95% confidence intervals derived by bootstrapping. Black dots represent individual reviewers' performances.

Fig. 3. Graphical visualization of agreement between reviewer predictions and histology.

Results for polyps with non-adenomatous and adenomatous histology are reported in left and right plots, respectively. In each box, the rightmost pair of bins reports polyps for which 70–100% of the endoscopists' predictions are in agreement with histology, while the leftmost pair of bins reports polyps for which less than 30% of the endoscopists' predictions are in agreement with histology. Each bin pair reports non-experts on the left and experts on the right. The distribution suggests that most of CADx FPs and FNs are polyps where reviewers disagree with each other, or are unanimously in disagreement with histology.

Fig. 4. Exemplar images of polyps with disputed histology.

Video frames of polyps for which both CADx and all reviewer predictions are in contrast with the histology ground truth. These polyps are considered false negatives (top row—histology: adenoma, CADx prediction: non-adenoma) or false positives (bottom row—histology: non-adenoma, CADx prediction: adenoma) in the reported accuracies.

Fig. 5. Colonoscopy videos of 165 patients screened in the CHANGE study were considered for the Standalone CADx study.

A total of 544 polyps videoclips from 130 patients was obtained after discarding dropout patients and patients with no polyps. This number was further reduced to 513 polyp videoclips (198 adenomas, 315 non-adenomas) after polyps with no recorded histology or missing videoclips were discarded by scientific annotation experts.