"GastroAid": A Semi-Supervised Learning Approach to Gastrointestinal Tract Image Classification

Abstract

The human gastrointestinal (GI) tract is subjected to a range of diseases and endoscopy can be considered the global standard for the examination of the GI tract. However, the potential to detect abnormalities in the GI tract highly depends on the gastroenterologist’s experience and the ability to detect during the endoscopic procedure. “GastroAid” is a computer-aided GI pathology detection system which can be integrated to the endoscopy device as an assistant mechanism for medical experts. The model used for detection was created using a semi-supervised learning approach as a solution for lack of annotated medical data. Here a novel hybrid ensemble deep learning architecture was used for training along with the use of both labelled and raw data. The proposed approach succeeded in obtaining a 5% overall F1 score difference compared to the initial and final models and more than 80% in all evaluation metrics which outperforms most of existing approaches in respect to overall performance and generalizability.