Magnetoacoustic Tomography Image Reconstruction Based On Conditional Generative Adversarial Network

Studies have shown that the electrical properties(electrical impedance,electrical conductivity,and dielectric constant)of biological tissues undergo changes during the initial stage of lesion development and have not yet changed its morphological structure.Therefore,early diagnosis and treatment of some diseases may be achieved by characterizing the electrical properties of the tissue.Biological Electrical Impedance Tomography(BEIT)is a novel functional imaging technology after morphological and structural imaging.Magneto-Acoustic Tomography with Magnetic Induction(MAT-MI)is a novel multi-physics coupled bioelectrical impedance imaging method using ultrasound as a carrier wave.It generates induced eddy current inside the test sampleusing the principle of electromagnetic induction,which effectively eliminates the shielding effect of biological tissue,avoids the safety issues of current injection,and also improves the detection sensitivity.The image reconstruction problem of MAT-MI is a non-linear inverse problem.At present,investigators typically conduct research with the assumption that the electrical conductivity distribution of the test sample consists of regions of constancy with discontinuous change at the regional boundaries and the shape of regions of constancy are highly regular.Investigators have proposed a variety of image reconstruction methods to solve the inverse problem,but there remain many problems that restrict the practical application of MAT-MI.Due to the complexity of the conductivity reconstruction inverse problem in current magnetic-acoustic coupled imaging,the ultrasound signal is mostly concentrated at the boundary between regions of conductivity,and that ultrasound raw data has complex factors such as directionality,the reconstruction of the conductivity has been restricted.Therefore,currently there is no wellaccepted analytical or computational solution for image reconstruction of electrical conductivity.Following the rapid development recently,deep learning has been used in the fields of medical imaging successfully.Conditional generative adversarial network(c GAN)is a widely used network structure in recent years,and have been applied to the field of medical imaging.Based on the current theory and basic model of magnetic-acoustic coupled imaging,a simulation model of the positive problem of MAT-MI on a finite element model(COMSOL)was built;a data set consisting of approximately 20,000 pairs of multi-parameter,multi-modal simulation sample data pairs was generated;a training network to reconstruct the conductivity distribution based on c GAN was constructed;pairs of conductivity distribution and current density is used as a training data set to train a model for reconstructing the conductivity distribution;according to the trained model input vector sound source data,conductivity distribution image is reconstructed;a series of comparative verification experiments were performed.The final result shows that training the c GAN deep learning method with a large,constructed and diversified data set,the conductivity distribution of the samples with various shapes and characteristics can be fully reconstructed.The main innovations and conclusions are as follows: 1)Different changes in the shape,location,number,area,and hardness of the edge of the lesions do not impact the success rate and accuracy of conductivity reconstruction results,and the different samples are almost consistent.2)When the complete displacement vector(from ultrasound data)was replaced by only a directional component,and the network retrained,the success rate and accuracy of the reconstruction resultsis unchanged,which provides new possibilities for the MAT-MI with reduced data requirement.3)Both the internal areas and boundaries of the conductivity regions with irregular shapes are completely reconstructed,the reconstructed distribution has high spatial resolution and the computation speed of the reconstruction is fast;adding noise to the test data does not affect the success rate of reconstruction with only slight and predictable degradation of the image quality.4)It was found to be superior to the traditional reconstruction methods in terms of image quality and the quantitative accuracy of the conductivity value.5)For strongly deviant or pathological samples whose characteristics are drastically different from the training data,the reconstruction of morphological distribution and the distribution of conductivity values was found to be reasonable.There are no serious failure cases,no pathological reconstruction results,and with our best effort,overlearning is not found to be present.