Dissertação de Laura Amaya Torres

Aluna: Laura Amaya Torres

Orientador: Prof. Dr. Anderson Maciel

Coorientadora: Profa. Dra. Mariana Recamonde Mendoza Guerreiro

Título: Automatic Generation of Patient-Specific 3D Models of Organs Using an Unified Deep Learning Approach.

Linha de Pesquisa: Computação Gráfica e Visualização de Dados

Data: 16/12/2019

Hora: 13h40min.

Local: Prédio 43412 – Sala 218 (sala de videoconferência) – Instituto de Informática UFRGS

Banca Examinadora:

Prof. Dr. Bruno Castro da Silva (UFRGS)

Prof. Dr. Leandro Augusto Frata Fernandes (UFF – por videoconferência)

Prof. Dr. Marcelo Walter (UFRGS)

Presidente da Banca: Prof. Dr. Anderson Maciel

Abstract: Reconstruction of 3D shapes from images using convolutional neural networks (CNN) has become a very studied field in recent years and has demonstrated great performance. Rigid and non-rigid objects have been reconstructed using several types of 3D representations and approaches with single or multiple images. However, the reconstructed objects are part of the outside world (they are visible and can be photographed). The most used imaging techniques to obtain visual information from organs are computerized tomographies (CT) and magnetic resonance imaging (MRI), which do not generate regular 2D images. Our main objective is to evaluate the feasibility of using deep learning approaches to directly reconstruct 3D models of organs from medical images using free-form deformations (FFD). To do this, we combined existing algorithms used for segmentation of medical images in 3D space with 3D object reconstruction techniques in a fully automatic convolutional network model. We tested our proposed method by training models for two different organs with higher and lower shape complexity (liver and prostate, respectively). The reconstructed models generated by our network are coherent with the overall shape of the organs demonstrating that it can be learned. However, more work needs to be done to obtain organ models that truly represent their actual structure.

Keywords: Mesh generation. organ reconstruction. free-form deformation. deep learning.