Abstract
Supervised learning for 3D left ventricle (LV) ultrasound segmentation is difficult due to the challenges of acquiring large amounts annotated data. In this work, pre-training on a weakly labeled dataset, combined with augmentations and fine-tuning on a limited dataset using a straightforward 3D convolutional U-net type neural network was investigated. The results indicate that an accuracy close to both state-of-the-art and inter-observer can be achieved with such an approach. The resulting neural network was highly efficient (17 ms on laptop GPU) and was used to create a real-time application for fully automatic LV volume and ejection fraction measurements over multiple heartbeats to enhance practical use in the echo lab.