Modern network science has introduced exciting new opportunities for understanding the brain as a complex system of interacting units in both health and disease and across the human lifespan. Despite the rapidly growing interdisciplinary science of complex networks, which spans the range from genetic and metabolic networks all the way up to social and economic systems, it remains a formidable challenge to identify the most representative and shared brain alterations caused by a specific disorder (e.g., Alzheimer’s disease), namely ‘disorder signature’, in a population of brain networks, let alone multi-modal brain networks where each brain network is derived from a particular neuroimaging modality (e.g., functional or diffusion magnetic resonance imaging (fMRI or dMRI)). Such integral signature can be revealed by what I name as a multimodal connectional brain template (CBT), which would constitute an unprecedented contribution to network neuroscience, rooted in firstly learning brain connectivity normalization and secondly foreseeing its evolution. During this fellowship, I will develop NormNets, a novel technique leveraging the power of geometric deep-learning to meet this challenge by normalizing a population of multimodal brain networks. Particularly, NormNets tools will substantially advance the field of network neuroscience by estimating not only an integral but also a predictive mapping of neurological disorders. In addition to the multi-disciplinary high-quality training I will receive at both host and secondment institutions, this fellowship will remarkably consolidate and accelerate my career on the international landscape scene in the new cross-disciplinary area of “geometric deep learning integration connectomics” I will pioneer during this fellowship. With the endorsement of international multi-sectoral stakeholders, open NormNets resources will impact on and contribute towards the development of connectomics-rooted predictive precision medicine.

This project will investigate, implement and compare different robust machine learning approaches for analysing and interpreting medical images. While medical image data presents a wide spectrum of modalities (macroscopic images, X-rays, MRI, CT), the processing pipeline for treating any modality encompass two main stages: feature extraction and classification [1]. With the advent of deep learning frameworks, the end-to-end methods emerged as competitive alternative. In this paradigm, both features and classification are learnt in a unified learning module. Despite this advances, several problems remains unsolved, on the top of them come, the image representation, biomarkers identification, class imbalance, the scarcity of data samples to name just few. This thesis will address some aspects of these problems considering specific diagnosis cases, while testing and validating several varieties of deep learning architecture.

With yearly costs of about 800 billion euros and an estimated 179 million people afflicted in 2010 in Europe, neurological disorders (ND) are an unquestionable emergency and a grand challenge for neuroscientists. Specifically, in Turkey, dementia is usually diagnosed at the end of the first stage or at the beginning of the middle stage. An accurate diagnosis for neurological disorders in a very early stage using non-invasive magnetic resonance imaging (MRI), including Alzheimer’s disease (AD), is one of medicine’s holy grails. However, meeting this formidable challenge is faced with several hurdles cluttering the following ideal wishful ambitions: (1) one would ideally use high-resolution data for diagnosis, however, high-resolution MRI data acquired using 7T scanners scarce and generally are not integrated in the clinical diagnosis routine in comparison with low-resolution 3T MRI scanner, (2) one would ideally diagnose fast and accurately using a single timepoint however, progressive neurodegenerative disorder might show very subtle changes in their early stage. Hence acquiring a few-month spread-out brain MRIs can improve diagnosis however spoil early ND diagnosis from a single observation upon 1st visit to hospital, (3) one would ideally diagnose using multiple available MR imaging modalities (i.e., multiple medical sources) including functional and diffusion magnetic resonance imaging (fMRI or dMRI), which have demonstrated great potentials in diagnosing brain dementia compared to solely using structural T1-weighted MRI. However, these are not conventionally acquired in the clinical routine in Turkey and worldwide, and (4) suppose that all challenges (1-3) can be met with groundbreaking medical data analysis technologies, reproducing disorder-specific brain biomarkers to target for preventive treatment in a very early stage within a brain dataset of interest, let alone across different datasets with same disorder, remains unsolved. Such formidable challenges can be solved by the development of what I name as “REproducible PRedictive Intelligence in MEdicine (RepPRIME)” to diagnose in the earliest stage using minimal clinically non-invasive data: a low-resolution, single timepoint, and conventional T1-weighted MRI modality. This would constitute a breakthrough in early ND diagnosis as it would allow accurate early diagnosis using high-resolution, multimodal MRI data, and follow-up observations all predicted from only T1-weighted MRI acquired at observation timepoint.

Finding brain biomarkers of Alzheimer’s disease (AD) in its very early stage (i.e., mild cognitive impairment (MCI)) can facilitate efficient treatment development. Specifically, developing accurate predictive neuroimaging-based methods to identify the late outcome of an MCI patient from baseline exam might aid in individualizing treatment for MCI patients who will remain stable, convert to AD or reverse to normal state over a certain time period. However, despite the large literature on AD and major advances in neuroimaging technologies and brain image analysis and learning methods, MCI prognosis research has not progressed as desired since: (i) a few methods targeted MCI outcome prediction from baseline as opposed to a plethora of advanced research methods focusing on classification tasks (e.g., stable MCI vs. converted MCI), (ii) absence of public neuroimaging-based MCI outcome prediction software, which limits comparability of existing methods and reproducibility of results, and (iii) absence of methods that identify early MCI cortical brain morphology alterations on a connectional (brain network) and shape (cortical surface) levels for late outcome prediction. To address these issues, we propose to develop the first freely-available software, ‘Dementia-free Ware’ (DFW.1.0), which aims to accurately detect subtle neurodegeneration very early in the course of AD by devising two novel brain analysis methods: one first learns connectional changes in brain morphology, structure and function for MCI outcome prediction, and the second predicts the evolution trajectory of the cortical surface from baseline MCI and its status (i.e., stable, converted or reversed).