A method for anomaly classification of endoscopic images from the entire gastrointestinal tract

Abstract

The gastrointestinal tract is part of the digestive system and is essential for digestion. Digestive problems can be symptoms of chronic diseases such as cancer and should be treated seriously. Endoscopic examinations of the tract enable the detection of these diseases in their early stages, allowing for effective treatment. Although they are the gold standard for GI tract analysis, variations in operator performance limit their usefulness. Support systems for experts to detect and diagnose such pathologies are desired. The proposed method aims to develop a method capable of classifying endoscopic images as normal or with anomalies. The proposed method aims to develop a classification method capable of distinguishing between healthy and anomalous endoscopic images, identifying specific anomalies within the gastrointestinal tract, and classifying such pathologies in a three-step process. A Convolutional Neural Network, specifically EfficientNetV2M, is used for the initial step, which is responsible for distinguishing between healthy and anomalous images. A Deep Learning architecture based on MambaVision was used for the second and third steps of the proposed method to classify GI tract anomalies. The second stage is responsible for categorizing pathologies into groups to forward images to specific binary classification models in the third stage, which are trained to distinguish images between pathologies within each of these groups. This work uses a rarely used database, the ERS database, containing 89,195 labeled images of the entire length of the gastrointestinal tract with more than 100 labels. The results obtained for the first stage, achieved using a model based on the EfficientNetV2 architecture, yielded an average F1-Score of 88.15%. The MambaVision architecture model used for the proposed method’s second stage obtained an average F1-Score of 76.10%. In contrast, the models for the last stage, responsible for classifying Cancer and Ulcer, Polyp and Other Pathologies, were 82.07% and 75.08%, respectively. When evaluating the proposed method end-to-end, an average F1-Score of 57.35% was obtained.