Research On Photoacoustic Image Reconstruction Of Tumor Based On Deep Learning

With the development of global economy,as one of the key factors threatening human life and health in the 21 st century,early diagnosis and accurate treatment are particularly important.Due to the restriction of principle,ultrasound,CT and MRI cannot provide accurate tumorous structure and function information.Photoacoustic imaging combines the advantages of ultrasound imaging and optical imaging,such as cross-scale,deep penetration and no radiation,providing a new scheme for the cancer prevention and treatment.However,under the limitation of equipment cost and sampling time,the images reconstructed by traditional algorithms based on sparse sampling contain many under-sampling artifacts,which are difficult to be used for rigorous medical diagnosis.Deep learning,especially the convolutional neural network,has emerged in recent years in the visual field of medical image processing,by virtue of data-driven,rapidity and intelligence.Therefore,the combination of deep learning and tumor photoacoustic image reconstruction has important technical innovation and clinical application value for tumor prevention and treatment.In this paper,the reconstruction method of tumor photoacoustic image based on deep learning is studied,including the simulation datasets establishment,network structure design and reconstruction experiments.The details are as follows:First,the basic principle of photoacoustic imaging was analyzed and the datasets were set up by designing the simulation experiments.Taking the tumor photoacoustic tomography with ring array based on sparse sampling as the research object,the principle of photoacoustic imaging was analyzed and the general mathematical formula of iterative reconstruction algorithms was summarized.On this basis,two tumor phantoms were designed by image processing algorithms and k-Wave,after studying the morphological and structural characteristics of tumor tissue.And two datasets were generated by simulation experiments,which were respectively used for removing artifacts and reconstruction based on the initial pressure signal graph.Then,the SEU-Net,evaluation system and network training strategies were researched and designed.By analyzing the principle of convolution,the convolutional neural network was connected with photoacoustic imaging.By analyzing the characteristics of various visual fields,the semantic segmentation was connected with photoacoustic image reconstruction.An end-to-end reconstruction network was designed based on U-Net,called SEU-Net.Correspondingly,the principle and experimental details were elaborated from the aspects of loss function,evaluation indexes,comprehensive evaluation system,optimizer,training strategies and environmental parameters.Finally,the reconstruction experiments were carried out and the SEU-Net was improved.The loss function was determined based on experiment results and network fitting degree.The effectiveness of GELU activation function was further verified.SEU-Net was improved from three aspects: network trunk,down sampling module and structural parameters.Based on the final network structure,two kinds of simulation datasets were used to train the network progressively to complete two reconstruction tasks.Experiments shows that SEU-Net is superior to similar reconstruction networks and TR algorithm in terms of the quality and efficiency of reconstruction on our datasets.In this paper,a deep learning-based algorithm for tumor photoacoustic image reconstruction is proposed,which can realize rapid and accurate image reconstruction based on sparse sampling.It effectively reduces the equipment cost,and provides the possibility for real-time photoacoustic imaging and clinical transformation.