Public defence in Engineering Physics, Biomedical Engineering, M.Sc. Antonios Thanellas

Public defence from the School of Science, Department of Neuroscience and Biomedical Engineering
A nearly profiled angle captures a semi-transparent human head merged with a brain MRI slice, featuring various labelling techni

Title of the thesis: Advancing Segmentation of Intracranial Structures in Brain Imaging

Doctoral student: Antonios Thanellas
Opponent: Prof. Jussi Tohka, University of Eastern Finland
Custos: Prof. Matti Hämäläinen, Aalto University School of Science, Department of Neuroscience and Biomedical Engineering

The growing demand for radiological imaging and the shortage of radiologists, highlights the need for robust software solutions that would ease clinical workload pressures by aiding diagnosis and simplifying image interpretation. 

In response, this thesis seeks to address several key challenges in the field of anatomical neuroimaging: 
1) Dealing with Limited Datasets: We found ways to overcome the difficulties posed by small datasets in brain volumetric analyses. 
2) Enhancing Algorithm Efficiency: We explored methods to optimize existing algorithms for more efficient labelling of brain structures in radiological images. 
3) Improving Algorithm Generalizability: We developed algorithms capable of accurately labelling brain findings across various hardware and acquisition settings. 

To address these challenges, we specifically: 
1) Examined Dataset Limitations: Investigated how limited datasets affect brain volumetric analyses from Magnetic Resonance Images (MRIs). 
2) Introduced Innovative Brain Structure Localization Methods: Developed novel approaches for locating intracranial structures and pathologies from MR and Computed Tomography (CT) images. 
3) Demonstrated Effective Training Material Creation: Showcased ways for creating effective training data for machine learning (ML) methods aimed at delineating brain blood from CT images across various hardware and acquisition setups. 

Key findings: 
1) Effective Metric Usage: We highlight the effectiveness of specific metrics in accurately distinguishing between healthy and non-healthy subjects, amidst biases and confounders in brain volumetric analyses with limited datasets. 
2) A Novel Segmentation Method: We combined segmentation fusion with a marker-controlled watershed transform resulting in superior segmentation performance for delineating the brain from MR images compared to conventional techniques. 
3) Machine Learning Algorithm Generalizability: We demonstrated promising results in delineating intracranial blood from CT images acquired with various hardware and acquisition settings. The network, trained using recommended methods, maintains effectiveness across diverse datasets. 

In conclusion, the challenges posed by limited datasets require careful consideration as they impact the development of ML and classical image processing methods for delineating structures within the head. While standalone classical image processing methods may decline in use, their integration with ML approaches is likely to grow significantly.

Key words: magnetic resonance imaging, computed tomography, segmentation, deep learning, intracranial hemorrhage, anatomical neuroimaging

Thesis available for public display 10 days prior to the defence at: 

Doctoral theses of the School of Science:

  • Published:
  • Updated:
URL copied!