| dc.description.abstract | Introduction: Regularly screening of the gastrointestinal tract for polyps is the an important measure for preventing colorectal cancer. Screening large population's gastrointestinal tract is with todays common methods too time consuming and expensive to accomplish. In this thesis explore the possibilities of using capsule video endoscopy (CVE) in combination with state of the art convolutional neural network (CNN) to create a computer aided diagnosis-system for automatic classification of diseases in the gastrointestinal tract.
Methods: First we create a dataset using images extracted from real CVE examinations of 19 patients. We use tensorflow framework to train and evaluate different state of the art CNNs compare with each other. We also propose a new CNN consisting of two state of the art CNNs in a parallel architecture. We also develop a Graphical User Interface (GUI) aimed at medical doctors for classifying VCE data.
Results: The dataset created contains 3267 labeled images with highlighted lesion annotation. We achieve F\textsubscript{Macro}-score, precision, recall and accuracy at 0.547, 0.553, 0.548 and 0.682 respectively on the best performing CNN. The GUI prototype is a simple application which classifies images based on predictions done by the trained network.
Discussion: The dataset have some balance issues as the largest class, normal images contains more than thousand samples, and five of the diseases contains less than hundred. The CNNs show promising preliminary results suggesting that it's a feasible approach. The GUI prototype suffers from severe performance issues and is only able classify about two images per second.
Conclusion: In this thesis we have shown that using CNN in combination with CVE is a a feasible approach. The results are promising but more research is required. The big challenge in image classification continues to be having a large and reliable dataset. | |
