Cardiac ventricle segmentation from cine MR images of pigs using 3D convolutional neural networks

Authors: Maialen Stephens Txurio Karen López-Linares Román Arnoldo Santos Ángel Gaitán Jesús Ruiz Iván Macía Oliver

Date: 08.06.2020


Abstract

Cardiac Magnetic Resonance (MR) Imaging is widely applied for the diagnosis and follow up of cardiaovascular diseases. Particularly, in patients with Pulmonary Hypertension (PH) MR images aid detecting right ventricle (RV) hypertrophy, which is a specific sign that characterizes the disease. PH related to left heart disease is the form that accounts for most of the cases. Hence, a previous segmentation of the cardiac ventricles is essential to extract imaging biomarkers that help better characterizing PH. Lately, Convolutional Neural Networks (CNNs) based on the U-Net architecture have shown to improve the results of previous approaches for accurate cardiac ventricle segmentation, yet, the performance of automatic RV segmentation techniques is still poor. Thus, in this study we aim at comparing different approaches to segment both cardiac ventricles using 3D CNNs together with the active contour-based loss function. We propose two strategies: (1) train one model for the segmentation of each ventricle separately, and (2) train a model to segment both ventricles at once. Results suggest that specific models for each ventricle have a higher accuracy than the joint one. Moreover, the proposed architecture together with the active contour-based loss function seems to outperform previous RV segmentation approaches with a dice score of 0.89.

BIB_text

@Article {
title = {Cardiac ventricle segmentation from cine MR images of pigs using 3D convolutional neural networks},
pages = {150-152},
keywds = {
Convolutional Neural Networks, Segmentation, Cardiac ventricles, Pulmonary hypertension
}
abstract = {

Cardiac Magnetic Resonance (MR) Imaging is widely applied for the diagnosis and follow up of cardiaovascular diseases. Particularly, in patients with Pulmonary Hypertension (PH) MR images aid detecting right ventricle (RV) hypertrophy, which is a specific sign that characterizes the disease. PH related to left heart disease is the form that accounts for most of the cases. Hence, a previous segmentation of the cardiac ventricles is essential to extract imaging biomarkers that help better characterizing PH. Lately, Convolutional Neural Networks (CNNs) based on the U-Net architecture have shown to improve the results of previous approaches for accurate cardiac ventricle segmentation, yet, the performance of automatic RV segmentation techniques is still poor. Thus, in this study we aim at comparing different approaches to segment both cardiac ventricles using 3D CNNs together with the active contour-based loss function. We propose two strategies: (1) train one model for the segmentation of each ventricle separately, and (2) train a model to segment both ventricles at once. Results suggest that specific models for each ventricle have a higher accuracy than the joint one. Moreover, the proposed architecture together with the active contour-based loss function seems to outperform previous RV segmentation approaches with a dice score of 0.89.


}
date = {2020-06-08},
}
Vicomtech

Parque Científico y Tecnológico de Gipuzkoa,
Paseo Mikeletegi 57,
20009 Donostia / San Sebastián (Spain)

+(34) 943 309 230

close overlay