A Pilot Study on Convolutional Neural Networks for Motion Estimation From Ultrasound Images

Abstract

In recent years, deep learning has been successfully applied to the analysis and processing of ultrasound images. To date, most of this research has focused on segmentation and view recognition. This paper benchmarks different convolutional neural network algorithms for motion estimation in ultrasound imaging. We evaluated and compared several networks derived from FlowNet2, one of the most efficient architectures in computer vision. The networks were tested with and without transfer learning and the best configuration was compared against the particle-imaging-velocimetry method, a popular state-of-the-art block-matching algorithm. Rotations are known to be difficult to track from ultrasound images due to a significant speckle decorrelation. We thus focused on images of rotating disks, that could be tracked through speckle features only. Our database consisted of synthetic and in-vitro B-mode images after log-compression, and covered a large range of rotational speeds. One of the FlowNet2 sub-networks, FlowNet2SD, produced competitive results with a motion field error smaller than 1 pixel on real data after transfer learning based on simulated data. These errors remains small for a large velocity range without the need for hyper-parameter tuning, which indicates the high potential and adaptability of deep learning solutions to motion estimation in ultrasound imaging.

Myocardial Function Imaging in Echocardiography Using Deep Learning

“…In a recent pilot study, efforts were also made into benchmarking different networks components of FlowNet 2.0, with fine-tuning on simulated US data. The results were on par with current state of the art for flow estimation [17]. In sum, these studies indicate a potential and adaptability for CNN based ME in US image analysis.…”

LU-Net: A Multistage Attention Network to Improve the Robustness of Segmentation of Left Ventricular Structures in 2-D Echocardiography

“…By definition, these techniques are optimal to the data they are trained on. Such approaches have been applied in the context of left ventricular structures analysis [10], [11], in particular for segmentation. In 2012, Carneiro et al exploited deep belief networks and the decoupling of rigid and nonrigid classifiers to improve robustness in terms of image conditions and shape variability [12].…”

A Pipeline for the Generation of Synthetic Cardiac Color Doppler

“…In this context, ultrasound simulations can play a key role in building such datasets. In this regard, our recent pilot study has shown that DL techniques can learn from synthetic ultrasound sequences to improve a targeted task; in this case, the estimation of rotational movements on in vitro data [12].…”