Procjena stadija dijabetičke retinopatije u slikama očne pozadine pomoću dubokih neuronskih mreža

master's thesis

• Title (english): Diabetic Retinopathy Stage Estimation from Retinal Photographs Using Deep Neural Networks
• Author: Andrea Lukač
• Mentor: Marko Subašić (mentor)
• Committee member: Marko Subašić (predsjednik povjerenstva)
• Committee member: Sven Lončarić (član povjerenstva)
• Committee member: Tomislav Petković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2017-07-07, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U ovom su radu primijenjene konvolucijske neuronske mreže za analizu slika očne
pozadine. Postupak analize ovakvih slika ključan je kako bi se pravovremeno detek-
tirala dijabetička retinopatija čiji se simptomi manifestiraju na očnoj pozadini. Za
problem segmentacije krvnih žila korištene su potpuno konvolucijske mreže. Model je
treniran i testiran na DRiDB [26] bazi slika te su izmjereni odlični rezultati. U svrhu
procjene stadija dijabetičke retinopatije konstruirana je konvolucijska mreža za klasi-
fikaciju. Dobiveni model je treniran na bazi slika iz [29]. Testni rezultati pokazali su
se obećavajućima za problem koji nije jednostavan. Dodatno je uvedena rezidualnost
u arhitekturu što je dovelo do malo boljih rezultata. Pokazano je da se za dva različita
problema može koristiti slična arhitektura uz odred̄ene modifikacije.

Abstract (english)

In this paper, convolutional neural networks were exploited for retinal image analysis.
Procedure of analyzing this kind of images is crucial in order to promptly detect diabe-
tic retinopathy, whose symptoms manifest on retina. For the blood vessel segmentation
problem, fully convolutional networks were used. The model was trained and tested on
DRiDB [26] image dataset and great results were measured. In purpose of assessment
of diabetic retinopathy stage we constructed a convolutional network for classification.
Resulting model was trained on the dataset from [29]. Test results have shown as pro-
mising for the problem that is not simple. Additionally, residuality was introduced in
the architecture which led to slightly better results. It has been shown that a similar
architecture with certain modifications can be used for two different problems.

• Keywords: segmentacija; dijabetička retinopatija; klasifikacija; duboke neuronske mreže; očna pozadina
• Keywords (english): segmentation; diabetic retinopathy; classification; deep neural networks; retina
• Language: croatian
• URN:NBN: urn:nbn:hr:168:654033
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-03-13 22:24:18