Doctoral defence in Electrical and Computer Engineering - Hans Emil Atlason

The defence will be streamed live

Ph.D. student: Hans Emil Atlason

Dissertation title: Deep learning for segmentation of brain MRI - validation on the ventricular system and white matter lesions

Opponents: Dr. Bennett Landman, Professor and Department Chair of Electrical and Computer Engineering, with secondary appointments in Computer Science, Biomedical Engineering, Radiology and Radiological Sciences, Psychiatry and Behavioral Sciences, Biomedical Informatics, Vanderbilt University, USA; and Director of the Center for Computational Imaging, Vanderbilt University Institute of Imaging Science, Vanderbilt University Medical Center.

Dr. Pierre-Louis Bazin, Senior Scientist at the Integrative Model-based Cognitive Neuroscience research unit, Department of Psychology, University of Amsterdam, Netherlands and Departments of Neurophysics and Neurology, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany

Advisor: Dr. Lotta María Ellingsen, Associate Professor and Vice Chair of the Faculty of Electrical and Computer Engineering, University of Iceland

Doctoral committee: 
Dr. Vilmundur Guðnason, Professor of Cardiovascular Genetics, University of Iceland. Director of the Icelandic Heart Association
Dr. Magnús Örn Úlfarsson, Professor and Department chair of the Faculty of Electrical and Computer Engineering, University of Iceland

Chair of Ceremony: Dr. Jakob Sigurðsson, Associate Professor and substitute of the Head of the Faculty of Electrical and Computer Engineering, University of Iceland

Abstract:
Magnetic resonance images (MRIs) enable neuroradiologists to investigate the human brain to look for possible causes of disease. The clinical interpretation of these images is, however, mostly limited to subjective assessment or a rough measurement of the area or volume of brain structures and presence of lesions. Multiple automatic methods have been developed to label different brain structures from MRIs, from which size, shape, and location of these structures and lesions can be extracted. These types of measurements enable researchers to perform comparisons in large scale studies. Multiple conventional whole-brain segmentation methods are based on finding a geometric transformation from MRIs with manually delineated brain structures to a target MRI that will consequently have the corresponding brain structures automatically labeled. These methods have worked very well for transforming labels between subjects and they have been extensively used in brain MRI studies. However, their main disadvantages are: 1) They are very slow (6+ hours for labelling one image), 2) their results are often inaccurate when the brain is deformed, e.g., due to atrophy, and 3) it is not possible to transform brain lesions from one subject to another, since their placement in the brain is variable. Current state-of-the-art brain segmentation methods are often based on deep neural networks (DNNs). DNNs can learn to approximate any function between an input and output given enough training data. After training, the DNNs can be used to analyze images fast and accurately. However, it can be very expensive and time consuming to generate enough training data for DNNs. A DNN trained on one MRI data set often does not work adequately well on another data set where different MRI parameters or scanners are used. Therefore, it would be beneficial to develop DNN methods that minimize the need for manually delineated training data. We have developed novel, automatic methods to label the ventricular system and WM lesions in the brain. Ventricular enlargement and WM lesions are associated with neurodegenerative diseases, e.g. Alzheimer’s disease, vascular dementia, and adult hydrocephalus. Our method, called SegAE, is the first unsupervised convolutional neural network for simultaneous segmentation of tissues and WM lesions from brain MRIs. SegAE‘s output are images that show the proportion of tissues and WM lesions in each voxel, but labelling WM lesions automatically has thus far been a challenging problem to solve, e.g. due to the variability of lesion load and location, and the inhomogeneous nature of MRI signal intensities within tissues. Furthermore, we use the output from SegAE to make images that have the same contrast irrespective of scanner type and parameters. That is advantageous because one major challenge in automatic medical image analysis is the lack of consistency of results using different data sets. This way we can make use of a DNN trained on manually labelled images from one data set and use it on another where the DNN input is a standardized image of the materials that cause the signal intensities in the MRI sequences. We have validated our methods on various data sets, including the AGES-Reykjavik study, that includes thousands of brain MRIs with a large variability of ventricular volumes and WM lesions. The methods have been compared to state-of-the-art methods and manual delineations by neuroradiologists. Our results indicate that the methods are accurate and robust to different scanners, and variability in brain structure, as well as being significantly faster than conventional methods.

 About the doctoral candidate:

Hans Emil Atlason graduated with a BSc degree in Electrical and Computer Engineering from the University of Iceland in 2014 and two years later with a MSc degree in biomedical engineering from Chalmers University of Technology in Sweden. He initiated the PhD studies at the University of Iceland in 2017. In 2020 he founded Visk ehf, a company that provides computer vision solutions for industrial applications.