Unit:
Κατεύθυνση ΒιοπληροφορικήΠληροφορική
Author:
Giannoulis Georgios
Supervisors info:
Ιωάννης Καλατζής, Αναπληρωτής Καθηγητής, Τμήμα Μηχανικών Βιοϊατρικής, Πανεπιστήμιο Δυτικής Αττικής
Original Title:
Ανάπτυξη και αξιολόγηση συνελικτικών νευρωνικών δικτύων για την κατηγοριοποίηση ιατρικών εικόνων
Translated title:
Development and evaluation of Convolutional Neural Networks for classification of medical images
Summary:
In this diploma thesis we present basic elements of neural network theory with emphasis on convolutional neural networks (CNN) with the main aim of studying the categorization of medical images. In particular, breast cancer images of the Invasive Ductal Carcinoma (IDC). This form of breast cancer consists of 80% of the cases of breast cancer. We aim to separate given images of tissue regarding to being positive or negative on IDC.
The implementation of the algorithms used, is based on the technology of CNN that exhibit a greater depth than used to be applied. CNN technology makes it possible to take a picture as input and discover specific information or features from the data the network deems important in order to separate the images into pre-specified categories that the images belong without any human input as to what differentiates the categories.
CNNs come in very different architectures, with different depth, different plurality of convolutions, different activation functions as well as many other technologies and techniques. In recent years, therefore, a great deal of research has been made to discover the most suitable architectures. In the present study we qualitatively and quantitative explore 9 modifications to an award-winning neural network architecture, the VGG16. Our aim is to evaluate various structural and quality parameters of a CNN in relation with how should we train such networks for better and faster results. Also, we focus on the transfer of knowledge from already trained networks which solved a different set of problems.
Main subject category:
Technology - Computer science
Keywords:
medical imaging, convolutional networks, neural networks, diagnostic methods, breast cancer
