Deep Learning Based Microwave-Induced Thermoacoustic Tomography Reconstruction Algorithm

Microwave thermoacoustic tomography is a rapidly developing non-invasive and non-ionizing biomedical imaging technology in recent years.It combines the advantages of microwave imaging and ultrasound imaging technologies to provide high-contrast,high-resolution and high-penetration imaging capabilities in the areas of breast cancer,angiography and brain imaging.In microwave thermoacoustic tomography technology,image reconstruction algorithm is an important part,which directly affects the efficiency of imaging,as well as the quality and reliability of results.Existing classical reconstruction algorithms often require a trade-off between computational efficiency and reconstruction quality,and are affected by hardware equipment and data integrity that can produce severe artifact interference,yielding some unsatisfactory results.With the development of deep learning techniques,more and more fields are being profoundly affected,including the field of biomedical imaging.In recent years,various deep learning-based image reconstruction methods have been applied to computational tomography,MRI and photoacoustic imaging,etc.Inspired by the successful applications in these fields,this paper presents the first proposal to introduce deep learning techniques into microwave thermoacoustic tomography image reconstruction.However,deep learning-based methods often require a large amount of data support,which is often difficult to produce and acquire in large quantities in microwave thermoacoustic tomography.Besides,the models used in existing deep learning-based direct reconstruction algorithms are still immature and suffer from difficulties in training,weak network model expression,low reconstruction quality,and overfitting.In this paper,we propose a set of technical solutions applicable to microwave thermoacoustic tomography mainly for the deep learning direct reconstruction algorithm,and realize the direct reconstruction from the sinogram to the initial energy loss density distribution.The proposed methods are validated on simulation and microwave thermoacoustic phantom experiments.This paper mainly includes the following three aspects:1.For scenarios where realistic training data are scarce or unavailable,a data synthesis method based on finite element simulation is proposed in this paper.The deep learning model is trained through the simulation synthesis training set,so that the model has the ability to reconstruct microwave thermoacoustic tomography images.2.Due to the inconsistent distribution between the simulation data and the actual experimental data,there is a problem of domain gap.In this paper,a set of simulation and experimental data preprocessing scheme is proposed to solve the problem of domain gap,when the model trained by simulation data is transferred to the actual thermoacoustic experimental data.3.The existing deep learning-based direct reconstruction algorithm is analyzed and discussed,a series of improvements are proposed,and a new TAT-Net model architecture is designed.The model is validated in simulations and microwave thermoacoustic phantom experiments,and achieves excellent performance in terms of reconstruction quality and robustness.In addition,the quantitative reconstruction capability of TAT-Net is investigated in this paper and demonstrated on simulations and thermoacoustic phantom experiments.Compared with other state-of-the-art reconstruction methods,the TAT-Net method can reduce the root mean square error to 0.0143,and increase the structure similarity and Peak Signal-to-Noise Ratio to 0.988 and 38.64,respectively.The results obtained indicate that the TAT-Net has great potential for improving image reconstruction quality and fast quantitative reconstruction.