Research On Ultrafast Ultrasound Imaging Based On Deep Learning

Ultrasound imaging has become one of the most widely used medical imaging techniques,owing to its non-invasiveness,non-ionizing radiation,cost-effectiveness,and portable property.In recent years,new imaging modalities based on ultrasound imaging have been introduced,which grant new insight into medical diagnosis and raise higher requirements in terms of imaging frame rate.Conventional ultrasound schemes transmit focused beams in a line-by-line scanning manner,which limits the imaging frame rate.Ultrafast ultrasound imaging allows an ultra-high frame rate with single emission of unfocused transmission beams,such as plane waves(PW)and diverging waves(DW),while producing images of poor quality.High-quality PW and DW imaging reconstruction generally rely on the coherent compounding of multiple successive steered emissions while decreasing the frame rate in turn,which leads to a tradeoff between the image quality and the frame rate in practice.In the past few years,deep learning-based approaches have been introduced in various general image processing tasks,as well as medical imaging problems.In this thesis,we applied deep learning to ultrasound imaging and developed a novel high-quality ultrafast ultrasound imaging approach.The contributions of this thesis are the following.A high-quality reconstruction approach based on deep learning has been introduced to tackle the trade-off problem between frame rate and imaging quality in compoundingbased ultrafast ultrasound imaging.The proposed method consists of solving an optimal compounding operator from data using supervised learning,i.e.,by training a convolutional neural network(CNN)to reconstruct high-quality images using a small number of transmissions.A large number of in vitro and in vivo samples,composed of paired lowand high-quality PW images,were acquired for supervised training.Experimental results demonstrated that the proposed method could produce high-quality reconstructed images using a small number of PWs,yielding an image quality equivalent to those obtained with compounding of a large number of PWs,i.e.,providing high-quality PW imaging at an ultra-high frame rate.A novel CNN architecture composed of the concatenation of multi-scale convolution kernels has been introduced to tackle the inconsistency between the quality of different local regions in the reconstructed DW images by conventional CNN architecture.In order to deal with the specific property of acoustic pressure field and data sampling of DW imaging,multi-scale convolution kernels were used to exploit corresponding image features and maxout units were used to activate depth-dependent features.A large number of in vitro and in vivo samples,composed of paired low-and high-quality DW images,were acquired for supervised training.Experimental results demonstrated that the proposed method produced high-quality reconstructed images using only 3 DWs,yielding a globally-consistent image quality equivalent to those obtained with the compounding of 31 DWs,i.e.,providing high-quality DW imaging at an ultra-high frame rate.A complex-valued convolutional neural network(CCNN)has been introduced for the high-quality reconstruction of ultrafast ultrasound imaging from in-phase/quadrature signals from modern ultrasound systems,which cannot be processed by real-valued CNN.The proposed model consisted of complex-valued weights and was optimized using complexvalued gradient descent.A large number of in vitro and in vivo data composed of paired low-and high-quality I/Q samples were used for supervised training.Experimental results demonstrated that the proposed method provided equivalent image quality to that obtained from RF-trained CNNs while using a reduced sampling rate and model complexity,which may be better incorporated in modern ultrasound systems.Moreover,the proposed method has been experimentally evaluated in realistic applications of dynamic imaging.Realistic motion velocities were analyzed in a controlled in vitro spinning disk phantom with anechoic cysts.Experimental results demonstrated that the proposed method yields robust image quality under a large range of rotational velocities,resulting in consistent contrast-to-noise ratio of the cysts and motion field of speckle tracking.In vivo experiments of cardiac imaging demonstrated that the proposed method yields superior image quality using a small number of transmissions compared with the standard compounding approach,in terms of static frame quality and speckle tracking of consecutive frames.