An Image Post-Processing Approach Based on Fully Dense U-Net for Microwave Induced Thermo-Acoustic Tomography

Abstract

In microwave induced thermo-acoustic tomography (TAT), circular scanning sensor array is the general signal acquiring mode. For higher sensitivity, large aperture detectors are widely preferred in the real TAT imaging system. However, large aperture of detector usually results in the limited spatial sampling points. Meanwhile, for considering the safety and comfort, the microwave radiation power is also limited. Due to these limitations, the signal to noise ratio (SNR) of TA signals may not be ideal to reconstruct an image with satisfactory quality. In addition, the TA image often suffers the artifacts resulted by sparsely sampling data. Herein, we propose a deep learning-based scheme to process the TA images which are reconstructed with time reversal (TR) method. Since the obvious artifacts removing performance in photoacoustic tomography (PAT), fully dense U-Net (FD U-Net) is employed to be the base deep learning architecture. Due to the amount limitation of real experimental TA data, the network is mainly trained by synthetic data which are generated in TAT simulating scheme. Furthermore, in order to ensuring the trained network could correctly detect the targets in the TA image, we propose a scheme termed data augmentation FD U-Net (DAFD U-Net) in which a few experimental TA images are added to the training procedure and perform data augmentation.The performance of DAFD U-Net is verified by an experimental image. Results show that the proposed scheme could significantly improve the quality of TAT image.