Cargando…

Automated Disease Detection in Gastroscopy Videos Using Convolutional Neural Networks

A large percentage of the world's population is affected by gastric diseases ranging from erosion and ulcer to serious ailments such as gastric cancer, which is mainly caused by Helicobacter pylori(H.pylori) infection. While most erosions and ulcers are benign, severe cases of gastric diseases...

Descripción completa

Detalles Bibliográficos
Autores principales: Zhang, Chenxi, Xiong, Zinan, Chen, Shuijiao, Ding, Alex, Cao, Yu, Liu, Benyuan, Liu, Xiaowei
Formato: Online Artículo Texto
Lenguaje:English
Publicado: Frontiers Media S.A. 2022
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9041416/
https://www.ncbi.nlm.nih.gov/pubmed/35492307
http://dx.doi.org/10.3389/fmed.2022.846024
Descripción
Sumario:A large percentage of the world's population is affected by gastric diseases ranging from erosion and ulcer to serious ailments such as gastric cancer, which is mainly caused by Helicobacter pylori(H.pylori) infection. While most erosions and ulcers are benign, severe cases of gastric diseases can still develop into cancer. Thus, early screening and treatment of all gastric diseases are of great importance. Upper gastroscopy is one such common screening procedure that visualizes the patient's upper digestive system by inserting a camera attached to a rubber tube down the patient's digestive tracts, but since the procedure requires manual inspection of the video feed, it is prone to human errors. To improve the sensitivity and specificity of gastroscopies, we applied deep learning methods to develop an automated gastric disease detection system that detects frames of the video feed showing signs of gastric diseases. To this end, we collected data from images in anonymous patient case reports and gastroscopy videos to train and evaluate a convolutional neural network (CNN), and we used sliding window to improve the stability of our model's video performance. Our CNN model achieved 84.92% sensitivity, 88.26% specificity, and 85.2% F1-score on the test set, as well as 97% true positive rate and 16.2% false positive rate on a separate video test set.