Metody klasifikace obrazů u kolonoskopie na základě metod umělé inteligence s využitím anatomických orientačních značek

Abstract

Thesis is focused on Image classification using convolutional neural networks. Two pre-existing convolutional neural networks ResNet18 and GoogleNet were used for image classification. Both neural networks were trained and tested on the same endoscopic images. Experienced endoscopists classified images into classification classes and created eight classes. Image data were used as complete database then divided into anatomical landmarks and pathological findings. 60 % of Images were used for training, 40 % of images were used for validation. Learning processes were set to 5, 10, 20 and 30 epochs, learning rate was set to 0,001. The results were analyzed separately for each dataset and for each settings. Values of validation accuracy, sensitivity and specificity were criteria for comparison. GoogleNet had highest accuracy in cathegory pathological findings. ResNet18 was more successful in the rest of cathegories.