Artificial intelligence empowered quality control for magnetic resonance imaging

IVANOVSKI, MARKO (author), Demšar, Jure (mentor)

Magnetic resonance imaging is a popular noninvasive way for doctors to create detailed images of the organs and tissue in human body. However, other than for clinical purposes, magnetic resonance images are also widely used in research. There, all the acquired images need to go through a quality control step before the good ones can be used for further analyses. The main goal of this thesis is to automate the quality control process with the help of the state of the art machine learning models for supervised learning. Given two datasets, we tried two different approaches in order to train a binary classifier. Firstly, we used a pre-trained neural network such as Inception or ResNet, to extract feature vectors for the MRI images, then used those to train different classifiers: XGBoost, Random Forest, Logistic Regression and KNN. Once trained, we performed within (we split the same dataset into training and test subsets) and between (one of our datasets was used as the training set and the other as the test set) dataset evaluation and managed to achieve promising results. We got 0.96 accuracy in within dataset evaluation and an accuracy of 0.81 for the between dataset evaluation. We have trained and tuned parameters of multiple classifiers, in the end XGBoost combined with the ResNet pre-trained network gave the best results.

• Language: English
• Keywords: machine learning, magnetic resonance imaging, intelligent systems
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2021
• PID: 20.500.12556/RUL-128686
• COBISS.SI-ID: 72634883
• Publication date in RUL: 23.07.2021

Secondary language

• Language: Slovenian
• Title: Preverjanje kakovosti slik magnetne resonance s pomočjo umetne inteligence

Abstract

Slikanje z magnetno resonanco je priljubljen neinvaziven način za klinično preiskovanje organov in tkiv v človeškem telesu. Poleg klinične uporabe, pa je magnetna resonanca zelo pogosto orodje v različnih raziskavah. Preden se pridobljene slike lahko analizirajo, za potrebe raziskava, je potrebno preveriti njihovo kvaliteto. Glavni cilj te naloge je avtomatizirati postopek preverjanja kvalitete slik magnetne resonance, s pomočjo najsodobnejših modelov za nadzorovano učenje. S pomočjo dveh naborov podatkov smo razvili dva različna pristopa za treniranje binarnih klasifikatorjev. V naših pristopih, smo najprej s pomočjo pred-trenirane nevronske mreže (Inception oziroma ResNet) pridobili značilke iz bitnih slik. Nato smo pridobljene značilke uporabili za treniranje različnih klasifikatorjev: XGBoost, Random Forest, Logistic Regression in KNN. Uspešnost razvitih modelov smo nato ovrednotili na dva načina. V prvem smo model trenirali in evalvirali na istem naboru podatkov, dosegli smo natančnost 0.96. Pri drugi evalvaciji smo model trenirali na enem naboru podatkov ter evalvirali na drugem, tukaj smo dosegli natančnost 0.81. V diplomski nalogi smo preizkusili več načinov za avtomatsko preverjanje kvalitete slik magnetne resonance. V našem primeru smo najboljše rezultate dosegli z algoritmom XGBoost v kombinaciji z pred-trenirano nevronsko mrežo ResNet.

• Keywords: mašinsko učenje, magnetna resonanca, inteligentni sistemi