Ultrasound Beamforming using MobileNetV2

Ultrasound Beamforming using MobileNetV2Sobhan Goudarzi1, Amir Asif 1, Hassan Rivaz1, 1Department of Electrical and Computer Engineering, Concordia University, Montreal, QC, Canada Background, Motivation and Objective In the past few years, the success of deep learning has led to a transformation in several high-level tasks in computer vision and medical image analysis such as classification and segmentation. Deep learning has also caused positive disruptions in several low-level tasks such as CT and MR image reconstruction. In this work, we are proposing a novel deep learning based approach for the low-level task of ultrasound image reconstruction from the pre-beamformed channel data. More specifically, we adapt MobileNetV2 to train a model that mimics Minimum Variance Beamforming (MVB). Statement of Contribution/Methods Herein, we consider the fact that all mathematical transformations can only represent the underlying information existing in input domain, and none of them, including deep learning approach, can generate new information. Therefore, all necessary preprocessing steps are applied to raw RF channel data before feeding to the network, and the network input contains all required information for estimating the result of MVB. More specifically, first, IQ demodulation is applied on the RF channel data since MVB requires complex signals to compute complex weights allowing for beampatterns that are asymmetrical around the center of the beam. Second, time delays are compensated to reduce the load on the network. Finally, the F-number is fixed for all image depths in order to make the image quality uniform. It has to be mentioned that the input to the network is supposed to be in range otherwise RF channel data has to be scaled proportionally. Each pixel of the image is reconstructed separately as the case for MVB. Network’s input is a matrix in which first two channels are real and imaginary parts of IQ data, is the number of channels and is the length of the window considered for temporal averaging to preserve the speckle statistics. The network output is a two dimensional vector containing real and imaginary parts of the beamformed data. As mentioned before, MobileNetV2 is used as the network structure since it is a leading architecture for networks with low computational complexity and memory requirement. This is of critical importance for commercial success of deep learning beamforming given the ultrahigh very large ultrasound frame-rate and very limited computational resources, especially in mobile ultrasound devices. As for training details, the network’s output for each images is reconstructed using MVB code provided by UltraSound ToolBox (USTB). The model is implemented using PyTorch library, and AdamW with L1-loss is used for training. The network is trained with a variety of imaging settings such as the acquisition center frequency, the sampling frequency, the transducer shape, and the number of transducer elements. Results/Discussion The results will be investigated on the test data of CUBDL.