Banca de DEFESA: Gabriel Luis de Araujo e Freitas

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : Gabriel Luis de Araujo e Freitas
DATE: 15/07/2022
TIME: 10:00
LOCAL: Sala de Seminários (Espaço Sérgio Barroso)
TITLE:

Novel Approach for Direct Methods in Medical Imaging using Compressive Sensing

KEY WORDS:

Keywords: Medical Imaging, Compressive Sensing, Non-Convex Minimization, Direct Method.

PAGES: 104
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Telecomunicações
SPECIALTY: Sistemas deTelecomunicações
SUMMARY:

ABSTRACT Title: Novel Approach for Direct Methods in Medical Imaging using Compressive Sensing Author: Gabriel Luis de Araújo e Freitas Supervisor: Vinícius Pereira Gonçalves, PhD. Co-Supervisor: Cristiano Jacques Miosso, PhD. Graduate Program in Electronic and Automation Systems Engineering Brasília, XXXXX XXXX, 2022 With the support of medical imaging technologies, healthcare workers are provided with relevant information about a patient’s condition when planning and following up on treatment. X-ray Computed Tomography (CT) and Magnetic Resonance (MR) are two of the most consolidated technologies in the field. These techniques yield anatomical images of specific planes or volumes. Although CT and MR exploit different physical principles, both collect measurements that can be modeled as the Fourier Transform coefficients of the image to be reconstructed. The reconstruction procedure refers to the stage of computing the desired image from the measurements acquired by the medical equipment. The acquisition usually requires the patient to stay in the same position for long periods, and, in the case of CT, there is the emission of ionizing radiation. Thus, such procedures should take place as safely and quickly as possible. A possible approach to address this issue is the development of reconstruction algorithms that can generate meaningful images for clinical practice from a reduced amount of measurements. Concepts of Compressive Sensing (CS) have been adopted in the devising of new algorithms for medical imaging to achieve a more efficient acquisition. This area of knowledge studies the reconstruction of signals from incomplete measurements by solving underdetermined linear systems. The signal of interest is the solution whose most of the coefficients are null. That is, the reconstructed signal is assumed to have a sparse representation in a known domain. Minimizing `p (0 < p ≤ 1) is a strategy often exploited by CS algorithms. Adopting `p metrics with smaller values of p, even leading to non-convex problems, opens up the possibility of further reductions in the number of measurements. Images are large signals. For this reason, CS-based reconstruction techniques rely on indirect methods to perform matrix operations because the storage of the matrices that model the problem is impractical during the execution of the algorithms. The stability and convergence of indirect methods are affected by reducing the value of p so that this strategy cannot be well exploited when performing matrix operations indirectly.

In this background, the present research devises the Direct Reconstruction Structure (DRS) for medical image formation through the composition of lower-dimensional signals, which are obtained through `p minimization. First, we present the mathematical formalism for generic implementations of this structure, which makes no assumptions about the operation for composition. Following, we derive the mathematical model and the minimization problem for a formulation that composes the image from onedimensional signals, which contain the information of a row of measurements in the frequency plane. We implemented that specific DRS formulation using the Iteratively Reweighted Least Squares (IRLS) as the minimization algorithm and prefiltering for sparse representation. We conducted four numerical experiments to investigate the behavior of the CS algorithms when reducing the value of p and evaluate the performance of DRS compared to techniques using an indirect method. In our tests, we used both artificial signals and actual image data. The results suggest that DRS can satisfactorily reconstruct medical images in good sparsity conditions. Prefiltering did not achieve the same effect in sparsifying the signals reconstructed by DRS compared to the case of algorithms using the indirect method. 

BANKING MEMBERS:
Externo à Instituição - HOMERO SCHIABEL - USP
Interna - 1609346 - MYLENE CHRISTINE QUEIROZ DE FARIAS
Externo ao Programa - 1769665 - VINICIUS DE CARVALHO RISPOLI
Presidente - 1415757 - VINICIUS PEREIRA GONCALVES