Research On Transcranial Metamaterial Parameters Determined By Deep Neural Networks

Cranial ultrasound is an important method for the diagnosis of cranial diseases.However,the skull has a very strong attenuation and distortion effect on ultrasound.It is difficult for ultrasound to effectively penetrate the skull to achieve intracranial tissue and blood flow imaging in a conventional way.In recent years,Harvard University,Massachusetts Institute of Technology and Langevin Institute in France and other world-renowned research institutions have successively carried out cutting-edge exploration of transcranial ultrasound imaging,but most of them are innovative studies,which require the removal or thinning of the skull.The development of metamaterials has improved the ability to explore cranial ultrasound.It is found that by adjusting the parameters of acoustic metamaterials,the influence of skull distortion on the acoustic wave can be eliminated.However,no material that can be used to eliminate skull distortion layer has been prepared.Moreover,the interaction mechanism between transcranial imaging and metamaterials is unknown so that the mapping relationship between transcranial image quality and parameters of metamaterials cannot be established.In this project,a method for predicting the parameters of transcranial metamaterials based on deep neural network is proposed to solve the above problems.Specifically,the mapping relationship between transcranial image quality and metamaterial parameters is explored by using deep neural network and ultrasonic transcranial images collected.The specific research contents of this thesis include:(1)Ultrasound transcranial images based on metamaterials are collected,and the role of metamaterials in transcranial imaging was verified.The cranial imaging experimental system built in our laboratory was used to image the resolution mold placed inside the skull without and with metamaterials respectively,and the transcranial images at different skull thickness(1.53 mm,2.86 mm and 4.11mm)were collected.The analysis of ultrasonic transcranial images at different skull thickness was as follows: without metamaterials materials,the contrast ratios(CR)of images at 1.53 mm,2.86 mm and 4.11 mm skull thickness were 3.568 d B,0.154 d B and 0.049 d B,respectively;The CR values are 10.677 d B,6.998 d B and 0.286 d B in the presence of metamaterials,respectively.The experimental results show that the metamaterials can improve the quality of transcranial image.In order to ensure the validity of the follow-up experimental data,the skull thickness of 1.53 mm was determined as the sampling point for the follow-up experimental data,and 1456 groups of intracranial image data were collected(including 112 groups of ultrasonic images of metamaterials at 13 imaging depths).(2)The intracranial image data were analyzed and processed,and the neural network data set was constructed.By offline processing of the image data,the horizontal resolution and vertical resolution of each group of data were calculated respectively(the full width at half maximum at the energy-8d B was taken as the measurement standard of resolution).By cleaning 1456 sets of data,915 sets of effective data were obtained,and the corresponding data measurement method was changed,which provided a good data sources for the subsequent neural network model.(3)Single-output and multi-output transcranial metamaterials parameter prediction models based on deep back propagation neural network are built,and the mapping relationship between transcranial image quality and metamaterial parameters was established by using the network model.For the mean square error(MSE)of the error evaluation index,the three single-output networks with average particle size,filling ratio and thickness output are 0.098,0.059 and 0.106,respectively,while the MSE of the multi-output networks is 0.025.The result shows that the multi-output network model has higher prediction accuracy and is more suitable for the data set of this thesis.Finally,using the multi-output prediction model with high prediction accuracy,the actual prediction of the transcranial metamaterial parameters with different image quality is carried out.