FBE- İşletme Mühendisliği Lisansüstü Programı - Doktora

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/55

Gözat

Konu "Bilgisayarlı görüntüleme" ile FBE- İşletme Mühendisliği Lisansüstü Programı - Doktora'a göz atma
  • Öge
    Multi-modal neuroimaging data prediction: Estimation of connectional brain template and multigraph classification with application to gender fingerprinting
    (Lisansüstü Eğitim Enstitüsü, 2021) Chaari, Nada ; Akdağ Camgöz, Hatice ; Rekik, İslem ; 709909 ; İşletme Mühendisliği
    The work developed in this Ph.D. thesis concerns the design of machine learning and geometric deep learning models that estimate a holistic representation of a population of multigraph brain connectivity and use the learnable integration networks for classification tasks with application to gender fingerprinting. Male and female brains are demonstrated to be highly distinguishable. Understanding sex differences in the brain has implications for elucidating variability in the incidence and progression of the disease, psychopathology, and differences in psychological traits and behavior. Decoding the brain construct using diverse neuroimaging techniques seems to be the ultimate pursuit of neuroscientists as well as brain-imaging analysts to extract the difference in genders' brains, thus boosting the neurological disorder diagnosis and prognosis related to sex. Currently, where an increasing number of brain imaging is being collected to investigate both women and man brains at their different modalities, more advanced analytical tools are required to meet new challenges revealed by large, complex, and multi-source sets of brain networks. On one hand, the estimation of a connectional brain template (CBT) integrating a population of brain networks while capturing shared and differential connectional patterns across individuals remains unexplored in gender fingerprinting. On the other hand, multigraphs with heterogeneous views present one of the most challenging obstacles to classification tasks due to their complexity. Several works based on feature selection have been recently proposed to disentangle the problem of multigraph heterogeneity. However, such techniques have major drawbacks. First, the bulk of such works lies in the vectorization and the flattening operations, failing to preserve and exploit the rich topological properties of the multigraph. Second, they learn the classification process in a dichotomized manner where the cascaded learning steps are pieced in together independently. Hence, such architectures are inherently agnostic to the cumulative estimation error from step to step. To overcome these drawbacks, in this thesis, we propose a medical computer-aided diagnosis tool enabling us to address the key challenges related to brain networks collected from multiple sources/modalities. First, we proposed how to estimate representative and centered brain network atlases, which can be leveraged to identify discriminative brain connectivities between male and female populations across heterogeneous datasets. Perhaps one of the greatest scientific challenges is to create a representative map of a brain network population acting as a connectional fingerprint. A very recent concept -connectional brain template (CBT), presents a powerful tool for capturing the most important and discriminative traits of a specific population while preserving its topological patterns. The idea of a CBT is to integrate a population of heterogeneous brain connectivity networks into a unified representation. Specifically, we present the first study to estimate gender-specific CBTs using multi-view cortical morphological networks (CMNs) estimated from conventional T1-weighted magnetic resonance imaging (MRI). Specifically, each CMN view is derived from a specific cortical attribute (e.g. thickness), encoded in a network quantifying the dissimilarity in morphology between pairs of cortical brain regions. To this aim, we propose Multi-View Clustering and Fusion Network (MVCF-Net), a novel multi-view network fusion method, which can jointly identify consistent and differential clusters of multi-view datasets in order to capture simultaneously similar and distinct connectional traits of samples. Our MVCF-Net method estimates a representative and well-centered CBTs for male and female populations, independently, to eventually identify their fingerprinting regions of interest (ROIs) in four main steps. First, we perform multi-view network clustering model based on manifold optimization which groups CMNs into shared and differential clusters while preserving their alignment across views. Second, for each view, we linearly fuse CMNs belonging to each cluster, producing local CBTs. Third, for each cluster, we non-linearly integrate the local CBTs across views, producing a cluster-specific CBT. Finally, by linearly fusing the cluster-specific centers we estimate a final CBT of the input population. MVCF-Net produced the most centered and representative CBTs for male and female populations and identified the most discriminative ROIs marking gender differences. The most two gender-discriminative ROIs involved the lateral occipital cortex and pars opercularis in the left hemisphere and the middle temporal gyrus and lingual gyrus in the right hemisphere. Second, to address the major issues in classifying complex data, we put forward an integration learning which fuses multigraphs brain connectomes with the aim to boost classification performance using the integrated networks. Specifically, we introduce Multigraph Integration and Classifier Network (MICNet), the first end-to-end graph neural network-based model for multigraph classification. First, we learn a single-view graph representation of a heterogeneous multigraph using a GNN based integration model. The integration process in our model helps tease apart the heterogeneity across the different views of the multigraph by generating a subject-specific graph template while preserving its geometrical and topological properties. Second, we classify each integrated template using a geometric deep learning block which enables us to grasp the salient graph features of a specific population. We train, in end-to-end fashion, these two blocks using a single objective function to optimize the classification performance. We evaluate our MICNet in gender classification using brain multigraphs derived from different cortical measures. We demonstrate that our MICNet significantly outperformed its variants thereby showing its great potential in multigraph classification. Finally, we review current graph integration methods that estimate well-centered and representative brain connectional templates (CBTs) for populations of single-view and multigraph brain networks. Then, we conducted a comparison study on these generated CBTs by single-view and multigraph fusion methods to evaluate their performances, separately, based on the following criteria: centeredness, discriminability (biomarker-reproducibility), and topological soundness (node-level similarity, global-lever similarity, and distance-based similarity). We demonstrate that deep graph normalizer (DGN) method significantly outperforms other multi-graph and all single-view integration methods for estimating CBTs using a variety of healthy and disordered datasets in terms of centeredness, discriminability (i.e., graph-derived biomarkers reproducibility that disentangle the typical from the atypical connectivity variability), and preserving the topological traits at both local and global graph-levels.