Back to search

University of Thessaly (UTH)

Quantification and Machine Learning classification of imaging biomarkers towards the optimization of MRI differential diagnosis

Abstract

dc:description

The role of biomedical imaging has expanded remarkably since the advent of digital technology, demonstrated from a wide range of powerful, versatile, and precise diagnostic imaging protocols that constitute integral part of most clinical procedures. Modern clinical imaging has brought forth new levels of pathophysiology assessment non-invasively, by exploiting various macroscopic quantitative imaging parameters that sensitively correlate with biological endpoints. However, the digital transformation of medical imaging technology poses new challenges in the rational interpretation of the dense quantitative imaging information, creating the need for development and usage of novel inferential approaches, to overcome the limitations of current qualitative evaluation and exploit its full potential. The present doctoral research aims to address challenges in the interpretation of multiparametric Magnetic Resonance Imaging (MRI) in neurologic and oncologic applications, and focuses on the incorporation and extension of the emerging data-driven inferential approaches that are based on advanced image analysis and statistical learning methodologies. The first study investigated the molecular substrates of the brain functional responses to clinically relevant ongoing pain stimuli, as observed through the Arterial Spin Labelling (ASL) perfusion technique. More specifically, the study examined the spatial correlation of hemodynamic responses of two ASL datasets of acute and chronic ongoing pain patients, with neurotransmitter receptor distributions, as informed from Positron Emission Tomography data of independent studies, and from transcriptomic data of the human Allen Brain Atlas. The linear regression models have revealed moderate to strong associations of the hemodynamic responses with the distributions of opioid and dopamine receptor subtypes (R2 values ranging from 0.14 to 0.35), which play a key role in the cerebral processing of pain signals, and especially the endogenous pain modulation mechanism. Differential associations were found for the serotonin receptors examined. Additionally, the results highlight the utility of the adopted methodology as an easily applicable tool for enhancing the molecular specificity of functional MRI signals. The second study aimed to develop a data-driven diagnostic model, based on an advanced multiparametric MRI dataset, for supporting the clinical task of differentiation between low- and high-grade gliomas. The multiparametric MRI data was analysed with a robust radiomic analysis workflow, including semi-supervised image segmentation, texture feature quantification, and machine learning - based feature selection and classification. The most discriminative features identified, were those obtained from the advanced MRI techniques, i.e., MR Spectroscopy (MRS), Diffusion Tensor Imaging (DTI) and Dynamic Susceptibility Contrast (DSC). The diagnostic model was based on a Support Vector Machine (SVM) classifier, evaluated with cross-validation, demonstrating a classification accuracy of 95.5%. Overall, the study demonstrates the value of utilizing multiparametric signal from advanced MRI and quantitative image analysis in the evaluation of brain tumor aggressiveness. The third study presents a radiomic analysis application of multiparametric MRI data for differentiating benign from malignant breast lesions. The specific objective was to evaluate the potential of Ensemble Learning techniques in feature selection and classification steps of the analysis workflow, to increase the relatively moderate diagnostic accuracy of the previously reported models that were based on individual classifiers. The key findings of the study were the superiority of Diffusion Weighted Imaging (DWI) derived features that completely outperformed and substituted the T2-weighted (T2-w) MRI derived features, as well as the superiority of Boosting Ensemble Learning compared to SVM classification models (Area under Receiver Operating Characteristic curve (AUC)= 0.95 vs. 0.88).Overall, the doctoral research demonstrates the enhanced perception regarding the underlying biological processes, gained from the combination of advanced imaging protocols, while further exploitation of advanced computational tools facilitates the development of powerful diagnostic models for supporting personalized interventions in the clinical setting.

Degree

thesis:*
Grantor dc:publisher
University of Thessaly (UTH)
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Vamvakas, Alexandros-Chrysovalantis
  • Βαμβακάς, Αλέξανδρος-Χρυσοβαλάντης

Subjects

dc:subject × 52

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/53355

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Vamvakas, Alexandros-Chrysovalantis; Βαμβακάς, Αλέξανδρος-Χρυσοβαλάντης. Quantification and Machine Learning classification of imaging biomarkers towards the optimization of MRI differential diagnosis. University of Thessaly (UTH), 2023. http://hdl.handle.net/10442/hedi/53355