V-FCNN: Volumetric Fully Convolution Neural Network for Automatic Atrial Segmentation

Abstract

Atrial Fibrillation (AF) is a common electro-physiological cardiac disorder that causes changes in the anatomy of the atria. A better characterization of these changes is desirable for the definition of clinical biomarkers, furthermore, thus there is a need for its fully automatic segmentation from clinical images. In this work, we present an architecture based on 3D-convolution kernels, a Volumetric Fully Convolution Neural Network (V-FCNN), able to segment the entire volume in a one-shot, and consequently integrate the implicit spatial redundancy present in high-resolution images. A loss function based on the mixture of both Mean Square Error (MSE) and Dice Loss (DL) is used, in an attempt to combine the ability to capture the bulk shape as well as the reduction of local errors products by over-segmentation. Results demonstrate a reasonable performance in the middle region of the atria along with the impact of the challenges of capturing the variability of the pulmonary veins or the identification of the valve plane that separates the atria to the ventricle. A final dice of 92.5% in 54 patients (4752 atria test slices in total) is shown.

A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging

“…The 17 out of 27 teams who provided their methodologies and performances either in full STACOM papers (Xia et al 2018, Bian et al 2018, Vesal, Ravikumar and Maier 2018, Puybareau et al 2018, Chen, Bai and Rueckert 2018a, Borra et al 2018, de Vente et al 2018, Preetha et al 2018, Qiao et al 2018, Nuñez-Garcia et al 2018, Savioli, Montana and Lamata 2018 or online (Huang 2018, Xu 2018 were included in this benchmarking study. Summaries of their methodologies are shown in Table 1 sorted by the final challenge rankings.…”

“…This involved the use of additional residual connections (Xia et al 2018), replacing all layers with dilated convolutions (Vesal et al 2018), improved methods of training such as the use of custom loss functions, deep supervision (Yang et al 2018), multi-task learning (Chen et al 2018a), and attention mechanisms throughout the network (Li et al 2018). The three teams which did not use U-Net as a baseline approach implemented enhanced versions of existing architectures such as ResNet (Bian et al 2018, Szegedy et al 2017, VGGNet (Simonyan andZisserman 2014, Puybareau et al 2018), and Fully-CNNs (Long et al 2015, Savioli et al 2018, Xiong et al 2019 which have been widely used on the ImageNet database (Deng et al 2009).…”

CEREBRUM: a fast and fully-volumetric Convolutional Encoder-decodeR for weakly-supervised sEgmentation of BRain strUctures from out-of-the-scanner MRI

“…Different works have considered the potential improvements of removing said volume partitioning [16,13]. Such fully-volumetric approach has already been applied to prostate [17], heart atrium [18], and proximal femur MRI segmentation [19], but not yet in the context of brain MRI segmentationwhere it could prove particularly useful given the complex geometry and the variety of structures characterising the brain anatomy. Here, we discuss how both hardware limitations and the scarcity of hand-labelled ground truth data can be overcome.…”

Segmentation of the left atrium and proximal pulmonary veins based on dimensional decomposition attention

“…Savioli et al 24 proposed a volumetric fully CNN (V‐FCNN) that utilized 3D convolution kernels which employed a loss function that combined mean square error and Dice to capture the overall shape feature of the left atrium and reduce local errors, thereby improving segmentation results. Vesal et al 25 introduced residual links to the 3D U‐Net encoder and incorporated dilated convolutions at the bottom layer, enabling effective fusion of local and global information.…”