Anatomical Classification of the Gastrointestinal Tract Using Ensemble Transfer Learning
Date
2023-05-17
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0000-0001-9302-4930
Type
Thesis
Degree Level
Masters
Abstract
Endoscopy is a procedure used to visualize disorders of the gastrointestinal (GI) lumen. GI disorders can occur without symptoms, which is why gastroenterologists often recommend routine examinations of the GI tract. It allows a doctor to directly visualize the inside of the GI tract and identify the cause of symptoms, reducing the need for exploratory surgery or other invasive procedures. It can also detect the early stages of GI disorders, such as cancer, enabling prompt treatment that can improve outcomes. Endoscopic examinations generate significant numbers of GI images. Because of this vast amount of endoscopic image data, relying solely on human interpretation can be problematic. Artificial intelligence is gaining popularity in clinical medicine. Assist in medical image analysis and early detection of diseases, help with personalized treatment planning by analyzing a patient’s medical history and genomic data, and be used by surgical robots to improve precision and reduce invasiveness. It enables automated diagnosis, provides physicians with assistance, and may improve performance. One of the significant challenges is defining the specific anatomic locations of GI tract abnormalities. Clinicians can then determine appropriate treatment options, reducing the need for repetitive endoscopy. Due to the difficulty of collecting annotated data, very limited research has been conducted on the localization of anatomical locations by classification of endoscopy images. In this study, we present a classification of GI tract anatomical localization based on transfer learning and ensemble learning. Our approach involves the use of an autoencoder and the Xception model. The autoencoder was initially trained on thousands of unlabeled images, and the encoder then separated and used as a feature extractor. The Xception model was also used as a second model to extract features from the input images. The extracted feature vectors were then concatenated and fed into a Convolutional Neural Network for classification. This combination of models provides a powerful and versatile solution for image classification. By using the encoder as a feature extractor that can transfer the learned knowledge, it is possible to improve learning by allowing the model to focus on more relevant and useful data, which is extremely valuable when there are not enough appropriately labelled data. On the other hand, the Xception model provides additional feature extraction capabilities. Sometimes, one classifier is not enough in machine learning, as it depends on the problem we are trying to solve and the quality and quantity of data available. With ensemble learning, multiple learning networks can work together to create a stronger classifier. The final classification results are obtained by combining the information from both models through the CNN model. This approach demonstrates the potential for combining multiple models to improve the accuracy of image classification tasks in the medical domain. The HyperKvasir dataset is the main dataset used in this study. It contains 4,104 labelled and 99,417 unlabeled images taken at six different locations in the GI tract, including the cecum, ileum, pylorus, rectum, stomach, and Z line. After dataset preprocessing, which includes noise deduction and similarity removal, 871 labelled images remained for the purpose of this study. Our method was more accurate than state-of-the-art studies and had a higher F1 score while categorizing the input images into six different anatomical locations with less than a thousand labelled images. According to the results, feature extraction and ensemble learning increase accuracy by 5%, and a comparison with existing methods using the same dataset indicate improved performance and reduced cross entropy loss. The proposed method can therefore be used in the classification of endoscopy images.
Description
Keywords
Endoscopy, Classification, GI track anatomic locations, Transfer learning, Ensemble learning
Citation
Degree
Master of Engineering (M.Eng.)
Department
Electrical and Computer Engineering
Program
Electrical Engineering