Research On High-resolution Photoacoustic Imaging Based On Deep Learnin

Photoacoustic imaging has the advantages of both high optical imaging contrast and high ultrasonic imaging resolution,and has become one of the popular biomedical imaging methods.It has been widely used in the early diagnosis of breast cancer,cardiovascular vulnerable plaque detection and other medical research fields.Photoacoustic imaging technology can be divided into three categories: photoacoustic computed tomography(PACT),photoacoustic microscopy(PAM),and photoacoustic endoscopy(PAE).PAM can be subdivided into acoustic resolution PAM(AR-PAM)and optical resolution PAM(OR-PAM).For any photoacoustic imaging technology,high-speed,large depth,and high resolution are always the goals pursued by imaging systems,but it is difficult to achieve a balance among the three in actual system development.PACT has the ability to image in large depth and large field of view,but under multiplexing technology,data acquisition is slow and requires reconstruction.Due to economic and time cost considerations,sparse sampling PACT(SS-PACT)is commonly used in PACT.However,the original collected data from SS-PACT is sparse,resulting in poor image quality and low contrast resolution after inversion.PAM cannot balance resolution and depth: OR-PAM has a high imaging resolution,but the imaging depth is only about one millimeter;AR-PAM imaging depth can reach centimeter level,but high-resolution imaging range is limited to the focus area.Aiming at the problem that SS-PACT and AR-PAM cannot achieve large depth and high resolution imaging,this article explores reasonable solutions based on deep learning(DL),promoting the in-depth application of photoacoustic imaging technology in various biomedical fields.The specific research content is as follows:(1)High resolution imaging of SS-PACT is realized based on DL method.Based on the improved U-Net method,the SS-PACT in vivo imaging of human blood vessels and rat dorsal vessels was carried out,and the high-resolution reconstruction of SS-PACT low-resolution images was realized.The reconstruction results were compared with the advanced compressed sensing model.It is found that the DL method can significantly improve the contrast resolution and reconstruction speed of the image,but the compressed sensing method shows higher reconstruction accuracy under higher sparse sampling rate.This work can provide different reconstruction schemes for large depth and high resolution photoacoustic imaging according to the actual mission background,which is helpful to expand the application breadth of PACT technology.(2)A two-stage residual network model with attention gate is proposed to expand the high-resolution imaging depth of AR-PAM.The depth of AR-PAM imaging can reach centimeter level,but when the image tissue deviates from the focus area,the imaging resolution will decrease sharply.In order to solve the above problems,a two-stage DL model is proposed creatively.According to the defocus degree of the imaging tissue,this method can reconstruct the high-resolution image in stages.Imaging experiments were carried out on bionic and living rats,and the reconstruction results proved the excellent performance of the proposed method.In addition,in order to improve the ability of the network to reconstruct multi-size blood vessels,this thesis continue to improve the network model and add a multi-scale feature extraction module to the above model.The experimental results show that the improved model achieves more accurate reconstruction effect and improves the deep imaging ability of AR-PAM more effectively.(3)The hybrid convolution network model is designed to realize the high-resolution direct reconstruction of SS-PACT based on the original collected data.At present,most of the high-resolution reconstruction for SS-PACT is the post-processing of the inversion image,and the quality of reconstruction is limited by the performance of traditional inversion algorithms.Therefore,a hybrid convolution network model with residual and inception blocks is proposed in this thesis,which realizes the high-resolution direct reconstruction of SS-PACT based on the original data,and the in vivo vascular reconstruction experiments are carried out in human hands and rats.The reconstruction results prove the superiority of the hybrid convolution network model.Compared with the traditional post-processing method,this method can improve the imaging quality of SS-PACT more effectively.This thesis explores the large-depth and high-resolution imaging methods of PACT and AR-PAM based on DL,which makes photoacoustic imaging suitable for more clinical and application requirements,and provides high-performance imaging methods for a variety of biomedical research.