| Photoacoustic tomography(PAT)is a hybrid imaging mode based on photoacoustic effect.It uses short laser pulses to stimulate the biological tissues,creating a photoacoustic wave with high contrast optics and high-resolution acoustic intelligibility.However,a photoacoustic tomography image represents the photoacoustic information of the object corrupted by distortions inherent from the imaging system,which includes the impulse response of the ultrasound transducer,the propagation media properties,and errors in the reconstruction algorithm.Such systematic distortion can be modeled as the point spread functions.Image degradation caused by PSF can be eliminated in PAT image deconvolution,thus,resolution can be improved.Herein,we developed the deconvolution method based on spatially variant PSFs for PAT image.At present,the deconvolution restoration methods of PAT image are mainly model-based approaches,and the premise of achieving a good recovery result is to accurately measure the PSFs of the PAT system.Therefore,we firstly design an experimental method to realize the accurate measurement of the spatially variant PSFs by directly obtaining the response of the system to a point source.Secondly,we conduct simulation,phantom and in vivo animal PAT imaging restoration experiments using the model-based image deconvolution method.The results show that adding measured PSFs into the image deconvolution model can effectively improve the PAT image resolution.Moreover,in recent years,deep-learning-based image deconvolution approaches have shown great promise for image recovery and are expected to be applied to PAT image degradation restoration.Given the spatially variant resolution and the lack of ground truth data in PAT imaging,a novel learning strategy tailored for PAT imaging is highly desired.Herein,we use the deep neural networks as the prior to account for the unique characteristics of PAT imaging.Based on deep image prior,a new unsupervised deep learning framework including image prior and PSF prior is constructed.The overall framework is named Deep Hybrid Image-PSF Prior(DIPP).DIPP is an unsupervised method for obtaining high-quality restoration results only from the original degraded image.We can also incorporate the experimentally measured PSFs of the specific PAT system as a reference to further boost performance during the PSF estimation.To show the algorithm performance in dealing with multiple degradations in PAT,we conduct extensive experiments on simulation images,publicly available datasets,phantom images as well as in vivo small animal imaging data.Compared with both classical analytical methods and state-of-the-art deep learning models,DIPP obtains substantially enhanced restoration results in terms of image details and contrast. |
