Accuracy and applicability of the artificial intelligence integrated software in Z-line segmentation

Abstract

Conducted from April 2019 to July 2020, this study aimed to assess the software accuracy in assisting Z-line segmentation by comparing with medical doctors’ detection results and by recording doctors’ satisfaction in scale, time-of-implementation in interactive mode, and integrated mode. For the development of the Z-line detection algorithm, a dataset of 533 high-definition endoscopic WLI (white-light) images in diverse forms of Z-line were collected. The software was subsequently developed in 4 modes, including manual mode, interactive mode (using Superpixels-BPT), automatic mode (using AI algorithm), and integrated mode (the combination of BPT and U-Net). 30 endoscopic images were assigned to 2 groups of doctors (under 2-year experience and over 5-year experience) for the Z-line detection using the software in 4 modes. Time-of-implementation, number of mouse clicks, satisfaction in scales, and IoU (Interception over Union) metric with expert’s ground-truth are used for assessment. The results showed that IoU metrics of interactive and integrated modes in the experimental dataset was 88% with no statistical difference to the IoU value of manual mode, and mean IoU metrics from the results of 4 modes were high, from 86.7 to 90.8%. The mean values of time-of-implementation in interactive mode and integrated mode were not statistically different from manual mode. The median number of mouse-clicks each use in the interactive mode and the integrated mode were 24.5 and 15.5 times, respectively. The software received good feedbacks from the doctors, with the mean values of satisfaction scores of automatic mode, interactive mode and integrated mode are 7.2, 7.3, and 7.2 respectively. The development of the software for detecting endoscopic anatomy landmarks is a novel and feasible research direction in Vietnam. Further studies could focus on detecting some specific lesions classified according to anatomy landmarks.