Semantička segmentacija slika primjenom dubokog učenja

master's thesis

• Title (english): Semantic segmentation of images using deep learning
• Author: Marko Damjanić
• Mentor: Tomislav Hrkać (mentor)
• Committee member: Tomislav Hrkać (predsjednik povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Committee member: Alan Jović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-11, Croatia

Abstract

Ovaj rad evaluira tri modela za semantičku segmentaciju gastrointestinalnih polipa
na slikama kolonoskopije: U-Net, PraNet i DUCK-Net. Započinje se s objašnjavanjem os-
nova dubokih neuronskih mreža i konvolucijskih neuronskih mreža, nakon čega slijedi
detaljna analiza arhitekture i prednosti svakog modela. Javno dostupni skupovi poda-
taka standardizirani su i pripremljeni za trening i testiranje, s naglaskom na evaluaciju
performansi modela korištenjem uobičajenih metrika segmentacije. U-Net je poslužio
kao robusni bazni model, dok je PraNet pokazao izvanredne performanse uz efikasne
zahtjeve za treniranje. DUCK-Net, iako kompleksan i resursno zahtjevan, nije posti-
gao rezultate kakve se očekivalo. Rezultati sugeriraju da U-Net i PraNet imaju značajan
potencijal za poboljšanje s modernim tehnikama, dok kompleksnost DUCK-Net-a pred-
stavlja izazove. Dobiveni rezultati ističu potencijal modela za semantičku segmentaciju
u pomoći medicinskim dijagnozama, pri čemu se PraNet posebno ističe kao obećavajući
za aplikacije u stvarnom vremenu.

Abstract (english)

This thesis evaluates three models for the semantic segmentation of gastrointesti-
nal polyps in colonoscopy images: U-Net, PraNet, and DUCK-Net. The study begins by
explaining the basics of deep neural networks and convolutional neural networks, fol-
lowed by a detailed analysis of the architecture and advantages of each model. Publicly
available datasets were standardized and prepared for training and testing, with a focus
on evaluating model performance using common segmentation metrics. U-Net served
as a robust baseline model, while PraNet demonstrated superior performance with effi-
cient training requirements. DUCK-Net, although complex and resource-intensive, did
not perform as expected. The findings suggest that U-Net and PraNet have significant
potential for improvement with modern techniques, whereas DUCK-Net’s complexity
poses challenges. The results underscore the potential of semantic segmentation mod-
els in aiding medical diagnoses, with PraNet being particularly promising for real-time
applications.

• Keywords: semantička segmentacija; gastrointestinalni polipi; U-Net; PraNet; DUCK-Net; duboko učenje; konvolucijske neuronske mreže; analiza medicinskih slika
• Keywords (english): semantic segmentation; gastrointestinal polyps; U-Net; PraNet; DUCK- Net; deep learning; convolutional neural networks; medical image analysis
• Language: croatian
• URN:NBN: urn:nbn:hr:168:765015
• 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: 2024-10-12 20:24:05