Fusion of Selected Deep CNN and Handcrafted Features for Gastritis Detection from Wireless Capsule Endoscopy Images

The wireless capsule endoscope (WCE) is usually used in the detection of digestive tract diseases. It has the advantages of simple use, convenient inspection, no pain, and ability to inspect the small intestine, becoming one of the hot spots in the research field of medical device. At present, the examination of WCE images is mainly manual. Thousands of images need to be examined. Therefore, the accuracy of the diagnosis is closely related to the experience of the doctor and the state of the examination. This study intends to develop a model to assist doctors in the examination. First, we established a gastritis dataset based on the endoscopic data from 20 patients. Because of the small differences in the appearance of gastritis and normal images, we creatively proposed a method of combining manual features and depth features to characterize the gastritis images. The Least absolute shrinkage and selection operator (Lasso) feature filtering and Principal Component Analysis (PCA) dimensionality reduction methods were afterwards used to screen the informative features. Subsequently, the multiple machine learning methods such as K-Nearest Neighbor (KNN), Support Vector Machine (SVM), adaptive boostint (Adaboost) and RamdomFroest were used to model classifiers. The experimental results show that the best gastritis detection model using the selected features (7 handcrafted features and 13 deep features) and SVM realizes the accuracy, recall, precision and F1-score of 97.86\%\pm0.82\%, 97.32\%\pm2.28\%,\ 98.32\%\pm1.98\% and 0.98\pm 0.01 respectively.