Toward automatic and reliable evaluation of human gastric motility using magnetically controlled capsule endoscope and deep learning.

In this paper, we develop a combination of algorithms, including camera motion detector (CMD), deep learning models, class activation mapping (CAM), and periodical feature detector for the purpose of evaluating human gastric motility by detecting the presence of gastric peristalsis and measuring the period of gastric peristalsis. Moreover, we use visual interpretations provided by CAM to improve the sensitivity of the detection results. We evaluate the performance of detecting peristalsis and measuring period by calculating accuracy, F1, and area under curve (AUC) scores. Also, we evaluate the performance of the periodical feature detector using the error rate. We perform extensive experiments on the magnetically controlled capsule endoscope (MCCE) dataset with more than 100,000 frames (100,055 specifically). We have achieved high accuracy (0.8882), F1 (0.8192), and AUC scores (0.9400) for detecting human gastric peristalsis, and low error rate (8.36%) in measuring peristalsis periods from the clinical dataset. The proposed combination of algorithms has demonstrated the feasibility of assisting in the evaluation of human gastric motility.