Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Selective extraction and identification of malicious traces in imaging data of endoscopic capsule, by using machine learning

Barpagiannis Christos

Full record

Year 2019
Type of Item Diploma Work
Bibliographic Citation Christos Barpagiannis, "Selective extraction and identification of malicious traces in imaging data of endoscopic capsule, by using machine learning", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2019
Appears in Collections


Nowadays, the small bowel capsule endoscopy (SBCE) is the most reliable way to diagnose patients suffering from small bowel bleeding (SBB). The patient swallows a specially designed electronic capsule, which moves through the gastrointestinal (GI) tract, photographing the environment, as it moves on. The time the capsule needs to traverse the small intestine ranges from one to five hours, depending on the situation. The photos are transferred to a computer in order to be reviewed by the doctor, in a form of a video stream, for further diagnosis. Visualization of a massive number of images is a very time consuming and tedious task for gastroenterologists. Decreasing the time of reading the video during the diagnostic phase by the medical staff can be achieved automatically by using available software that employs algorithms, mainly based on the deletion of similar images and images that do not contain the red color (indicating blood). Machine Learning (ML) and Artificial Neural Networks (ANNs) have a crucial importance as supporting tools for doctorsIn the present Thesis we have developed three different Deep Learning Convolutional NN models, which achieve an automatic diagnosis of vasculature and bleeding from CE image data and classify them into two classes, healthy and unhealthy. The three different CNN models contain Conv2D, Max Pooling layers. All models at the end have a classifier consisting of a flatten layer and fully connected layers. In the last fully connected layer we used the sigmoid function for activation as opposed to all other layers we used with the ReLU function. These three models were gradually applied to a combination of different techniques to evaluate their effectiveness in improving our CNN models. We have implemented methods like Augmentation that increase the data with images it produces from the set by rotating and shifting them, and methods that change the structure of the models. We implemented Dropout which temporarily removes neuronal model units along with their connections and the method Transfer Learning using the already pre-trained models versions VGG16 and VGG19 (Visual Geometry Group) from University of Oxford and the ResNet which are trained in a large set of images e.g. the ImageNet dataset. In order to save computational time, different data feed techniques were also tested on the CNN models.This Thesis took in account data available in the Thesis of Mr. A. Polydorou, at the ECE Department of TUC in 2018. These data are extracted from 171 CE videos of different patients from five Greek hospitals. In collaboration with a gastrointestinal surgeon doctor, a total of 3800 images were selected from different forms of angioectasia and different haemorrhages, from different sites of the GI (gastrointestinal), which include mainly difficult cases of diagnosis of lesions.No more than one image from each lesion was included. After successive trials of different numbers and combination of images, the best training results were achieved by using a balanced training set of at least 1000 images, consisting of 500 healthy and 500 bleeding or with angioectasia. The automatic diagnostic experiments were performed on selected images from 33 videos that were not used during the training/validation procedure. The testing set consists of 100 images, including 50 normal and 50 bleeding or angioectasia. Our data are the values of the RGB components for each pixel.The total versions of our three CNN models are 22. Performance metrics such as accuracy (Acc), sensitivity (Sen), and specificity (Spe) were computed to evaluate the effectiveness of the proposed models. We find out that the methods Dropout and Augmentation improve the performance of our models but in the case of Transfer Learning the usage of VGGNet, ResNet we had the opposite results. The best performance of the model based on unseen data achieved 90% Sen, 92% Spe, 91.8% Precision, 8% FPR, 10% FNR. Compared to other research teams that have developed CNN for the same problem, better results have been reported, but this does not allow the algorithms to be compared because of their control over different data collection. The results of the present work compared to the work developed by Mr. A. Polydorou Thesis (diagnosis based on handcrafted color metrics of the HSV color space), the color metrics results were better than those of our CNN method (99% Sen, 96.1% Acc, 93.2% Spe, 83,3% Precision, 20% FPR, 0% FNR).

Available Files