Identificação automatizada da composição corporal em adultos a partir da tomografia computadorizada

• Main Author: Cittolin, Pedro Francescon
• Publication Date: 2023
• Other Authors: Souza, Viviane Lima Bonfim Moroni de
• Format: Bachelor thesis
• Language: por
• Source: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Download full: http://repositorio.utfpr.edu.br/jspui/handle/1/35536
• topic: Composição corporal; Tomografia; Interfaces de usuário (Sistemas de computação); Coluna lombar; Algorítmos - Desenvolvimento; Body composition; Tomography; User interfaces (Computer systems); Lumbar vertebrae; Algorithms - Development; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA

The computed tomography (CT) technique emerges as one of the current reference methods for assessing a patient’s body composition (BC). It can provide important information for several clinical applications. The proper use of the information obtained by this method depends on local reference values. In Brazil, however, there is a large gap regarding these values for its population, motivating a retrospective population study to establish such normative reference values. In order to enable and facilitate the extensive analysis of a large number of CT exams, several authors in the computational area have worked to automate the processes of segmentation and diagnostic assistance using neural networks. Therefore, the purpose of this work was to develop a system to automate the analysis of computed tomography scans to enable the study of body composition in adults. It was necessary to automate the selection of the ideal slice (3rd lumbar vertebra, L3) on the CT for the analysis of the BC and, from the selected slice, automate the process of mapping the different tissues (bone, muscle and fat) that compose it. As a result, a cross-platform system in Python was developed for processing CT files in the DICOM format. The system performs the chaining of convolutional neural network (CNN) algorithms, such as U-Net, SpatialConfiguration-Net and EfficientNet-B6, in order to automate the processes of L3 selection and tissue segmentation. The system has a simple and friendly interface, intended for use by health professionals, optimizing time and human and computational efforts to perform the study.