Left ventricle segmentation in cardiac magnetic resonance imaging with deep learning and deformable models containing shape restrictions

Author(s)

Matheus Alberto de Oliveira Ribeiro

• Document type: Master's Dissertation
• Press: São Paulo.
• Institution: Universidade de São Paulo (USP). Escola de Artes, Ciências e Humanidades (EACH)
• Defense date: 2021-10-20

Examining board members

Fátima de Lourdes dos Santos Nunes Marques; Esther Luna Colombini; Marco Antonio Gutierrez

• Advisor: Fátima de Lourdes dos Santos Nunes Marques

Abstract

Images from cardiac magnetic resonance exams are recognized as the gold standard for diagnosing various heart diseases. Biomarkers estimated from the segmentation and analysis of the left ventricle in these images can be used in diagnosis. However, manual segmentation of the left ventricle in various images of magnetic resonance exams demands time and repetitive effort from the expert, which can increase quality variability of diagnosis. In recent years, several automatic and semiautomatic approaches for left ventricle segmentation in these images have been proposed in the literature. The main strategies involve the use of methods based on atlas, graphs, deformable models, artificial intelligence, classical image processing techniques and hybrid combinations of these, together with anatomical restrictions regarding the shape of the ventricle. Despite showing good results, no method has yet reached the excellence of the expert due to wide variations in structures represented in magnetic resonance images. From a systematic mapping, it was found that the use of hybrid methods integrating artificial intelligence and shape restrictions have obtained promising results and offer a possible solution to the segmentation problem, but still without reaching the desired excellence. The main objective of the present work is to develop a hybrid method that combines deep learning and deformable models with shape restrictions to automatically segment the left ventricle in cardiac magnetic resonance images. Deformable models favor the production of precise segmentations and are able to impose anatomical shape restrictions, reducing the production of segmentations with anatomical errors, one of the most common problems in recent methods. The results indicate that the method produces results comparable to the literature and is anatomically more consistent, while demonstrating generalization ability between different image databases. However, the method still performs poorly in apical slices and specific cases of heart disease. In addition to offering a contribution to the Medical Image Processing area, the proposed method contributes to the area of diagnosis in Cardiology (AU)

• FAPESP's process: 19/22116-7 - Automatic segmentation of the left ventricle in cardiac magnetic ressonance exams
• Grantee: Matheus Alberto de Oliveira Ribeiro
• Support Opportunities: Scholarships in Brazil - Master