Globoko učenje na neslikovnih medicinskih podatkih

Pavlin, Jan (author), Kukar, Matjaž (mentor)

V zadnjem desetletju se vse bolj množično uporabljajo tehnike globokega učenja. Zaslugo za to lahko pripišemo razvoju tehnologije in zlasti znižanju cen grafičnih kartic, ki omogočajo hitro učenje. Globoko učenje je področje strojnega učenja in se je uveljavilo predvsem na področjih računalniškega vida, prepoznavanja govora, prepoznavanja slik, analize teksta. Uporaba tega področja počasi prodira tudi v medicino. V diplomskem delu bom preučil področja globokega učenja, upravljanja s pomanjkljivimi podatki in upravljanja z neuravnoteženimi podatki. Zgradil bom modele različnih topologij globokih nevronskih mrež in med njimi primerjal dosežene rezultate na podatkovni množici medicinskih podatkov. Analiziral bom tudi uporabo grafične kartice in procesorja za učenje ter uporabo nevronskih mrež. Cilj diplomske naloge je preizkusiti in analizirati uporabo globokega učenja na medicinskih podatkih z uporabo različnih pristopov k reševanju problema pomanjkljivih podatkov. Določiti je potrebno, katera topologija in katera metoda predprocesiranja podatkov se najbolje obneseta, kakšne rezultate dosežeta in koliko časa je potrebnega za učenje te nevronske mreže.

• Language: Slovenian
• Keywords: globoko učenje, medicinski podatki, nevronske mreže
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-105046
• Publication date in RUL: 23.10.2018

Secondary language

• Language: English
• Title: Deep learning on non-image medical data

Abstract

Use of deep learning is increasing in the last decade. This is mostly because advancement of technology and also price reduction of various graphical processors, which allow quick learning. Deep learning is an area of machine learning and has been used for computer vision, speech recognition, image classification and other. Usage of deep learning is being slowly used in medicine. In the diploma thesis, I will examine the areas of deep learning, management of insufficient data and management of unbalanced data. I will build models of different topologies of deep neural networks and compare the results achieved on the data set of medical data among them. I will also analyse the learning and use of a graphics card and a processor for learning and using neural networks. The goal of the thesis is to test and analyse the use of deep learning on medical data using different approaches to solving the problem of poor data. It is necessary to specify which topology and which methods of data preprocessing are best managed, what results they achieve and how much time is needed to learn this neural network.

• Keywords: deep learning, medical data, neural networks