Precise segmentation of the heart is crucial for reliable calculation of clinical indices such as chambers’ volumes and ejection fraction (EF). Currently, in echocardiography, cardiac segmentation is manually performed through a tedious and time-consuming process. More importantly, this process is prone to inter- and intra-user variability that usually leads to remarkable variability in calculation of EF, particularly in patients with borderline cardiac dysfunction. For many years, studies on automatic segmentation of heart chambers were focused on left ventricle (LV), due to its key role in heart dynamic. It is not a long ago, that a few methods for segmentation of right ventricle (RV) have been investigated as well. Up to our best of knowledge, no fully automatic method for 4-chamber segmentation of echocardiograms currently exists. Here, we describe our approach for fully automatic segmentation of whole heart in echocardiograms using deep-learning algorithms.
Motivated by outstanding performance of fully convolutional networks (FCN) in natural image semantic segmentation tasks, we have implemented and validated a FCN for whole heart segmentation in Echocardiograms. A typical FCN consists of convolution layers, nonlinear activation functions, pooling layers, and up-sampling layers as basic building blocks.Recent studies proved the advantage of using pretrained networks for applications with limited number of training dataset, such as medical image segmentation. Thus, we used the technique of fine-tuning a pre-trained network instead of random initialization. We also took advantage of real-time data augmentation for this study, which is one of the greatest tools for artificially increasing the number of training dataset. So, our network never sees same input twice in the training process, by applying random cropping, rotation, shear transformation and adding Gaussian noise.
The whole network was trained in an end-to-end (image-to-image) scheme, which took an image as an input and directly outputs the probability map. Here, we utilized and modified the FCN-VGG16 model (trained on the PASCAL VOC dataset). We implemented our method using Caffe library on a workstation with four GTX 980 Ti GPUs. It took about 5h to train and 1.89 s to process one test image with size 256×256 pixels.
The functionality of our approach is evaluated using a dataset of 1000 annotated images from 100 normal subjects’ echocardiography record
s available from Loma Linda University. The dataset includes 4-chamber long axis views. After training for 90 epochs, Dice coefficient as a validation metric for the errors between manual and automatic segmentations were calculated as 89.15% (SD=4.0) for LV, 81.13% (SD=10.7) for RV, 86.41% (SD=10.2) for the left atrium (LA), and 88.00% (SD=6.1) for the right atrium (RA), respectively
This study verifies that deep-learning algorithms can be successfully used to solve the challenging problem of automatic cardiac segmentation in echocardiography. To the best of our knowledge, this study is unique with respect to the whole heart segmentation.