Research On Methods Of Endoscopic Photoacoustic Imaging Based On Deep Learning

Endoscopic photoacoustic tomography(EPAT)is a newly developed imaging technique for detecting lesioned tissues in biological cavities such as arterial vessel and digestive tract.The scanning geometry of the detector is enclosed in a cavity or vascular lumen,so that the acquisition of photoacoustic signals is subject to several limitations including the mechanical structure of the catheter,spatial location and imaging time.Consequently,the detector usually collects incomplete signals within a limited viewangle,which subsequently reduce the quality of the reconstructed images.The purpose of this thesis is to reconstruct high-quality EPAT images based on deep learning from photoacoustic signals acquired by sparse measurement in a limited-view.The main work includes two aspects.First,a method for recovering the images representing the spatial distribution of the optical absorption on the cavity cross-sections from limited-view sparse measurements is proposed by combining the traditional variational iteration algorithm with the convolutional neural network(CNN).The forward operator of imaging and its adjoint operator are separated from the network training and are embedded into each layer of a network unit.The influence of sparse measurement on the reconstruction quality is reduced and the reconstruction time is shortened by learning the gradient information.Second,a CNN model is constructed and trained by computer-simulated sample data,which is used to recover the maps representing the spatial distribution of the optical absorption coefficient on the cavity cross-sections by taking the optical absorption images as the input.The proposed methods have been demonstrated by computer simulation.Results indicate that our method is superior to the postprocessing method by CNN in fast reconstructing the optical absorption images with high quality from the limited-view measurements with low sampling rate acquired by limited detectors.Besides,the quantitative imaging method based on CNN is able to improve the quality of the quantitative reconstruction effectively by recovering the missing information caused by limited-view scanning.