2D left ventricle segmentation using deep learning

Abstract

Automatic segmentation of the left ventricle (LV) can become a useful tool in echocardiography. Deep convolutional neural networks (CNNs) have shown promising results for image classification and segmentation on several domains, however CNNs seem to require a lot of training data. In this work, CNNs are investigated for LV ultrasound image segmentation. We study if the need for manual annotation can be reduced by pretraining a CNN using a previously published automatic Kalman filter (KF) based segmentation method. The results show that a CNN is able to achieve similar accuracy to that of the automatic method, by only training with generated data. The dice similarity coefficient was measured to be 0.86 ± 0.06 for the CNN versus 0.87 ± 0.06, while the Hausdorff distance was better at 5.9 ± 2.9 mm for the CNN versus 7.5 ± 5.6 mm for the KF method. In future work, this may enable CNNs to exceed state-of-the-art with a small set of expert annotations for fine-tuning.