Brain MRI reconstruction with cascade of CNNs and a learnable regularization

Tryfonopoulos, Dimitris; Τρυφωνόπουλος, Δημήτρης

dc.contributor.author	Tryfonopoulos, Dimitris	en
dc.contributor.author	Τρυφωνόπουλος, Δημήτρης	el
dc.date.accessioned	2023-04-06T07:53:01Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/57496
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.25193
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	MRI ανακατασκευή εικόνας	el
dc.subject	Μηχανική Μάθηση	el
dc.subject	Επιστήμη Δεδομένων	el
dc.subject	Απεικονιστική Ιατρική	el
dc.subject	Βιοιατρική Απεικόνιση	el
dc.subject	Biomedical Engineering	en
dc.subject	MRI Reconstruction	en
dc.subject	Deep Learning	en
dc.subject	CNN Networks	en
dc.subject	Compressed Sensing	en
dc.subject	Biomedical Imaging	en
dc.title	Brain MRI reconstruction with cascade of CNNs and a learnable regularization	en
heal.type	masterThesis
heal.classification	Machine Learning	el
heal.classification	MRI Reconstruction	el
heal.classification	Compressed Sensing	en
heal.classification	Regularization Learning	en
heal.dateAvailable	2024-04-05T21:00:00Z
heal.language	en
heal.access	embargo
heal.recordProvider	ntua	el
heal.publicationDate	2023-02-01
heal.abstract	In the field of Medical Imaging, the recent years special attention has been given to Magnetic Resonance Imaging (MRI). MR imaging is a non-invasive imaging modality capable of producing cross-sectional images with high spatial resolution. Unlike Computed tomography (CT) the nature of the acquisition does not employ ionising radiation. However, MRI acquisitions are characterized by long scan time (acquisition time) which is directly related to the number of samples acquired in the k-space. Being able to reduce the acquisition time will help increase patient satisfaction, reduce several artifacts (mainly motion), and finally reduce in overall the medical costs. The main initiative on reducing the scan time from a software perspective, is Compressed Sensing (CS) optimization [1]-[3] according to which given some constraints that are fulfilled we can reconstruct highly undersampled images. Following the success of deep learning (DL) in a wide range of applications, neural networks based on CS optimization have received significant interest for accelerating MR acquisitions and reconstruction strategies. In CS signal reconstruction, the iterative algorithm unrolled over a deep neural network. The basic strategy is to train the network to learn the weights (CNN kernels) for dealising undersampled MR images from a large dataset containing pairs of aliased and dealiased images. In this work, we are motivated from the 2D Deep Cascaded CNN-Network (DCNN) [4]. DCNN mainly consists of two blocks the CNN and the Data Consistency block, operating on the image and k-space (sampling domain) respectively, linked by a regularization term which adjusts the data fidelity based on the noise level of the acquired measurements. In this work, we introduce an improved regularization for the data consistency term in two novel settings, a learnable regularization parameter per k-space slice and a spatially learnable regularization parameter. We show that the introduced regularization is independent of the CNN block architecture used and can be incorporated in any DL-CS based optimization network setting. We show that the employment of the proposed regularization in any DL-CS based network highly improves the reconstruction performance without adding any computational burden to the network.	el
heal.advisorName	Karantzalos, Konstantinos	en
heal.committeeMemberName	Vakalopoulou, Maria	en
heal.committeeMemberName	Voulodimos, Athanasios	en
heal.committeeMemberName	Douskos, Makis	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	65 σ.	el
heal.fullTextAvailability	false