Supervisors info:
Περαντώνης Σταύρος, Διευθυντής έρευνας, κέντρο πληροφορικής και τηλεπικοινωνιών, ΕΚΕΦΕ "Δημόκριτος"
Summary:
Wireless endoscopy is a relatively new imaging method, which highly contributes to the diagnosis of many digestive system diseases. The main problem of this method, concerns the huge number of images that the clinician has to examine, in order to end up with the right diagnosis. This procedure may be very tiresome for the clinician, because it requires many hours of uninterrupted attention. This master thesis, offers a solution to the aforementioned problem, by presenting the bag of words classification method, which can classify those images in categories, according to the disease that they depict. This method is based on the detection of regions of interests in images. Once regions of interest are found, features will be extracted from those regions. Those features are the abbreviated representation of an image. Two major datasets were used for the experiments conduction. The first is KID dataset 2, which contains 8 categories and the second is the Normal-Abnormal set which contains only 2 categories. The experiments were conducted in MATLAB programming environment. During the experiments, many feature extraction methods were used and their results were evaluated. Those methods were: SIFT, RGB-SIFT HSV-SIFT, Opponent-SIFT and HSV&Opponent SIFT (Combination). Apart from those methods, two more variables were constantly changing their values between the experiments, in order to conclude if they had any impact on the results. HSV-SIFT extraction method presented the highest results, as it has reached 81.46% classification percentage in Normal-Abnormal set, and 47.92% in KID dataset 2.
Keywords:
Wireless endoscopy, image classification, bag of words model, feature extraction, SIFT, RGB-SIFT, HSV-SIFT, Opponent-SIFT
