Endoscopy 2022; 54(S 01): S186-S187
DOI: 10.1055/s-0042-1745068
Abstracts | ESGE Days 2022
ESGE Days 2022 Digital poster exhibition

HIGH ACCURACY OF DEEP LEARNING BASED AUTOMATIC POLYP CHARACTERIZATION IN REAL-TIME DURING COLONOSCOPY

A. Benson
1   Hadassah Medical Center and Faculty of Medicine, Hebrew University of Jerusalem, Institute of Gastroenterology and Liver diseases, Department of Internal Medicine, Jerusalem, Israel
,
H. Jacob
1   Hadassah Medical Center and Faculty of Medicine, Hebrew University of Jerusalem, Institute of Gastroenterology and Liver diseases, Department of Internal Medicine, Jerusalem, Israel
,
L. Katz
1   Hadassah Medical Center and Faculty of Medicine, Hebrew University of Jerusalem, Institute of Gastroenterology and Liver diseases, Department of Internal Medicine, Jerusalem, Israel
,
H. Shirin
2   Shamir Medical Center, Institute of Gastroenterology, Zerifin, Israel
,
R. Hazzan
3   Assuta Centers, Haifa Gastroenterology Institute, Haifa, Israel
,
A. Kahloon
4   Erlanger Health System, Gastroenterology, Chattanooga, United States
,
P. Siersema
5   Radboud University Medical Center, Gastroenterology and Hepatology, Nijmegen, Netherlands
,
H. Neumann
6   University Medical Center Mainz, Interventional Endoscopy Center, Mainz, Germany
,
M. Landsman
7   MetroHealth System, Gastroenterology Division, Cleveland, United States
,
T. Berzin
8   Beth Israel Deaconess Medical Center, Division of Gastroenterology, Boston, United States
,
S. Ngamruengphon
9   Johns Hopkins Hospital, Division of Gastroenterology, Baltimore, United States
› Author Affiliations
› Further Information
Also available at
 

Aims Automatic polyp characterization (APC) during colonoscopy may enable endoscopists to determine histologic type of polyps more accurately. The MAGENTIQ-COLO is an artificial intelligence (AI) system that includes real-time polyp detection, size estimation, and APC. The purpose of this study was to assess polyp characterization performance.

Methods The APC categorizes polyps into two groups, neoplastic and non-neoplastic. It is implemented by a patch-wise convolutional neural network (CNN) with a backbone of ResNet50 CNN, and it was trained on 637,918 frames from 610 polyps with verified histopathology data. The APC results (neoplastic, non-neoplastic, or uncertain) are displayed in real-time. ([Figure 1]) In this study, 111,531 frames of 107 polyps taken from 88 colonoscopy videos were analyzed. To evaluate the performance of the APC, we measured accuracy (for the two groups), and precision and sensitivity of each group separately. The measurements were performed comparing the APC result to the histopathology results.

Zoom Image
Fig. 1

Results The APC accuracy on the dataset was 94.38%. The sensitivity was 99.95% for neoplastic polyps and 52.78% for non-neoplastic polyps. The precision was 94.04% for neoplastic polyps and 99.36% for non-neoplastic polyps. ([Table 1])

Table 1

Metric/Group

Neoplastic

Non-neoplastic

Sensitivity

99.95%

52.78%

Precision

94.04%

99.36%

Conclusions This newly introduced APC model has high accuracy and high precision for neoplastic and non-neoplastic polyps, and very high sensitivity for neoplastic polyps. The non-neoplastic sensitivity will improve as the system continues to train on a larger number of polyps (with emphasis on non-neoplastic polyps). By combining real-time APC and APSE, the AI system may support endoscopist decision-making in accurately diagnosing diminutive polyps, without the need for histologic confirmation.



Publication History

Article published online:
14 April 2022

© 2022. European Society of Gastrointestinal Endoscopy. All rights reserved.

Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany