Optimiranje neuronske mreže evolucijskim računanjem za detekciju tumora u MR slikama mozga

master's thesis

• Title (english): Evolutionary computing optimization of neural network for tumor detection in brain MR images
• Author: Jan Đurinec
• Mentor: Jelena Božek (mentor)
• Committee member: Jelena Božek (predsjednik povjerenstva)
• Committee member: Sonja Grgić (član povjerenstva)
• Committee member: Filip Turčinović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2025-02-28, Croatia

Abstract

Rad istražuje primjenu tehnika evolucijskog računanja (ER) za optimizaciju konvolucijskih neuronskih mreža (CNN) u detekciji tumora mozga na MRI snimkama. ER metode korištene su za obradu slika na ulazu modela, podešavanje hiperparametara i optimizaciju funkcije gubitka. Na skupu podataka od 3509 MRI snimaka, početna F1 mjera CNN-a iznosila je 71% (točnost: 63%). Testirana su tri ER algoritma: PSO, MFO i NLCMFO – pri čemu je NLCMFO povećao F1 mjeru na 97%. ER metoda uklanjanja lubanje nije pokazala značajnije rezultate. Konačna usporedba optimiziranog CNN-a s unaprijed treniranim VGG-16 modelom pokazala je slične performanse. Navedeni rezultati naglašavaju potencijal ER algoritama za optimiranje CNN modela kako bi postigli rezultate usporedive sa state-of-the-art modelima. Unatoč računalnim troškovima i izazovima vezanim uz podatke, ER pruža čvrste temelje za daljnja istraživanja u medicinskoj analizi slika i općenito.

Abstract (english)

This thesis explores the application of evolutionary computing (EC) techniques to optimize convolutional neural networks (CNNs) for brain tumor detection in MRI scans. EC techniques were used for image preprocessing, hyperparameter tuning, and loss function optimization. Using a dataset of 3509 MRI scans, initial CNN F1 score was 71% (accuracy: 63%). Three EC algorithms—PSO, MFO, and NLCMFO—were tested, with NLCMFO boosting the F1 score to 97%. While EC-assisted skull stripping showed mixed results, simpler preprocessing proved more reliable. A final comparison between the optimized CNN and the pre-trained VGG-16 model showed similar performance. This result highlights the potential of evolutionary algorithms to enable custom CNNs to perform comparably to state-of-the-art pre-trained models in medical image analysis. Despite computational costs and dataset challenges, EC shows promise for enhancing medical image analysis, warranting further research.

• Keywords: Detekcija tumora mozga; konvolucijske neuronske mreže; evolucijsko računarstvo
• Keywords (english): Brain tumor detection; Convolutional neural networks; Evolutionary computing
• Language: english
• URN:NBN: urn:nbn:hr:168:222588
• 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: Open access
• Created on: 2025-03-12 11:07:48