Ultrasonic Tomography Image Reconstruction Method Based On Machine Learning

Ultrasound Tomography（UT）has the advantages of non-invasiveness and low cost,and has broad application prospects in the fields of industrial measurement and medical monitoring.Among them,the image reconstruction algorithm process can reconstruct the spatial distribution of the medium in the measured field,which is of great significance for process monitoring and control.However,in the UT system,due to the fixed number of probes and limited emission angles in ultrasonic measurement,there are measurement blind areas,resulting in low accuracy and poor robustness in the application of ultrasonic tomographic image reconstruction algorithms.Therefore,it is necessary to improve the accuracy of image reconstruction.Important scientific issues in the development of UT technology.This topic takes ultrasound tomography method as the research object.Aiming at the problem of low accuracy of traditional reconstruction algorithm,convolutional neural network（CNN）is used to complete image reconstruction.In the training process,the small-step convolution and shrinkage are repeatedly used to extract the data characteristics of the boundary measurement values,and finally,the inclusion distribution of the image is reconstructed through the pooling and connection operations at the same time.The specific research work is:（1）Establish an ultrasound tomography database.Aiming at the problem that deep learning network model training requires a large number of samples,COMSOL Multiphysics^?and MATLAB co-simulation are used to establish a database of ultrasonic measurement models with different inclusion distributions.Each piece of data contains inclusion distribution information and the corresponding measured sound pressure value.The distribution information of the inclusions is binarized,and the corresponding measured sound pressure is normalized to reduce model training errors.The database contains 23355 samples without noise,and Gaussian white noise is used to expand the database samples,which solves the problem of insufficient data in the network training process.（2）Ultrasonic tomography method based on U-NET convolutional neural network.In order to improve the accuracy of image reconstruction,a pre-imaging convolutional neural network based on U-NET is proposed,which takes pre-imaging pictures as the input of the network.U-NET fully extracts multi-dimensional image features through multiple downsampling,and the network output layer merges A variety of feature maps effectively improve the reconstruction accuracy,and at the same time have good noise resistance,which solves the problem of serious artifacts in the reconstruction of traditional algorithms.（3）Ultrasonic tomography method based on multi-channel convolutional neural network.Ultrasonic detection systems are usually limited by the transmitter angle of the sensor,leading to insufficient prior information and serious artifacts in reconstructed images.In order to improve the reconstruction accuracy,a multi-channel convolutional neural network（CNN）method is proposed,which uses the pixel matrix of the pre-imaging picture with 48effective measurement values as channel one,and all 256 measurement values as channel two.The use of channel two data to guide the feature extraction during the training process solves the problem that the ultrasound measurement data is limited by the emission angle of the transducer and effectively improves the accuracy of image reconstruction.