NCT03761771

Artificial Intelligence Identifying Polyps in Real-world Colonoscopy

Validating the Performance of Artificial Intelligence in Identifying Polyps in Real-world Colonoscopy

• Status: Completed
• Phase: —
• Study type: Observational
• Enrollment: 209 (actual)
• Sponsor: Zhaoshen Li · Academic / Other
• Sex: All
• Age: 18 Years – 75 Years
• Healthy volunteers: Not accepted

Summary

Recently, artificial intelligence (AI) assisted image recognition has made remarkable breakthroughs in various medical fields with the developing of deep learning and conventional neural networks (CNNs). However, all current AI assisted-diagnosis systems (ADSs) were established and validated on endoscopic images or selected videos, while its actual assisted-diagnosis performance in real-world colonoscopy is up to now unknown. Therefore, we validated the performance of an ADS in real-world colonoscopy, which is based on deep learning algorithm and CNNs, trained and tested in multicenter datasets of 20 endoscopy centers.

Detailed description

The ADS were established in changhai digestive endoscopy center to assess its efficacy in clinical practice. The ADS automatically initiated once the ileocecal valve was pictured by the colonoscopist or the colonoscopist recorded any image of colon during the insertion. When colonoscopists withdrew the colonoscopies and inspect the colons, the video streaming of colonoscopies was real-time switched to the ADS, which made it feasible to identify and classify lesions in real time. Colonoscopists were invited to respond if they doubted potential polyps in the screen, and the ADS also made a voice when identifying potential polyps, followed by repeatedly inspecting to confirm the existence of lesions. The voice of ADS could be real-time heard by colonoscopists, while the screen of ADS was placed right behind colonoscopists, where polyps identified by ADS could be seen after the colonoscopists' turning but not simultaneously. The lesion detection by ADS or colonoscopists were determined as follow: A. polyps only identified by ADS, which was considered to be missed by colonoscopists: polyps were reported by the ADS and the colonoscopists did not know the location of polyps without reminder of the ADS until the polyps disappeared from the view; B. polyps first identified by ADS: polyps were first reported by the ADS and the colonoscopists also later knew the location of polyps by themselves; C. polyps simultaneously identified by the ADS and colonoscopists: the time of reporting polyps was closely synchronal (within 1 second); D. polyps first reported by colonoscopists: polyps were first reported by the colonoscopists and the ADS also later identified the location of polyps before the colonoscopists unfolded and pictured the polyps; E. polyps only reported by colonoscopists, which was considered to be missed by the ADS: polyps were reported by the colonoscopists and the ADS did not identify the location of polyps until colonoscopists unfolded and pictured the polyps. Besides, the false-positives of real-world ADS were also reported with potential causes analyzed by colonoscopists.

Conditions

• Sensitivity of the ADS in Identifying Polyps in Real-world Colonoscopy
• Mean Number of Polyps Per Colonoscopy for Colonoscopists and Colonoscopists + ADS

Interventions

Type	Name	Description
DEVICE	colonoscopy withdrawal with the ADS monitoring	During the testing of trained ADS, when the system doubts colonic lesions from the input data of the test images, a rectangular frame was displayed in the endoscopic image to surround the lesion. If the system confirmed it as the colonic lesions, a sound of reminder will be played and the types of lesions (non-adenomatous polyps, adenomatous polyps and colorectal cancers) will be classified by the system. We adopted several standards to define the identification and classification of colonic lesions: 1) when the system identified and confirmed any lesion in the images of no polyps or cancers, the results were judged to be false-positive. 2) when the system both confirmed and correctly localized the lesions in images (IoU \> 0.3), the results were judged to be true-positive. 3) when the system did not confirm or correctly localize the lesions, the results were judged as false-negative. 4) when system confirmed no lesions in the normal images, the results were judged to be true-negative.

Timeline

• Start date: 2018-11-01
• Primary completion: 2018-12-10
• Completion: 2018-12-10
• First posted: 2018-12-03
• Last updated: 2018-12-17

Locations

2 sites across 1 country: China