| X-ray imaging quality is the basis for providing spatial correspondence for medical CT image analysis and processing.Due to the complexity of image appearance modeling and the computational complexity of deformable registration models,how to image high-quality CT images under low-dose X-ray energy irradiation is a difficult problem that needs to be solved urgently.X-ray thermoacoustic imaging technology,as an emerging imaging technology that has developed rapidly in recent years,can combine the high contrast of CT images with the high resolution of acoustics,and has become a promising imaging method in medical imaging.Traditional imaging techniques(including CT,B-ultrasound,and MRI)have low contrast,specific tissue types,and other risks associated with high-dose ionizing radiation.Significant work has been carried out to improve the cutting-edge method of CT imaging quality and the cutting-edge method of reducing X-ray dose.However,due to various practical reasons,it is still challenging to establish a low-dose and high-resolution CT imaging system and apply it to a wide range of clinical applications.First of all,the iterative reconstruction method used in the frontier field of X-ray thermoacoustic imaging requires a heavy computational resource.And most of the previous work only analyzed the X-ray thermoacoustic effect,and did not integrate the advantages of traditional CT imaging.Moreover,designing a high-resolution reconstruction method for CT images is also a challenge.Some methods based on deep learning have been proposed,but a large number of high-resolution CT samples are required to train the model,and the cost of labeling training samples is high.This thesis is dedicated to solving the above-mentioned problems of X-ray thermoacoustic imaging of the lung and its application in improving the quality of lung CT images.In response to the demand for low-dose imaging and further improving the quality of CT imaging,this thesis focuses on the key issues of rapid X-ray thermoacoustic imaging,multi-modal image registration and CT image super-resolution reconstruction.The main work of this thesis are as follows:(1)This thesis uses human lung tissue as the research site,creates a digital lung phantom,and builds an X-ray thermoacoustic imaging simulation platform.Then,this thesis simulates the X-ray source energy,the initial sound pressure of the thermoacoustic effect in the lungs and the ultrasonic detector in the circle distribution on the platform.In this paper,the thermoacoustic signal acquired by the detector is reconstructed by filtered back projection method to meet the needs of low-dose CT imaging.During the reconstruction process,this thesis analyzed the minimum dose required for the application of this method in lung imaging.(2)Aiming at the feature segmentation requirements of the fusion registration of thermoacoustic images and CT images,this thesis proposes a new semi-supervised 3D deep neural network.The model trains the network by using the strong labels in the DSB data set and the weak labels in the LUNA data set to output all suspicious tumor nodules in lung CT images.The model can extract two kinds of labels synchronously in sequence,and combine the two objects through a weighted transfer function during model training.Experimental results prove that the proposed method can significantly improve the accuracy of lung nodule detection.(3)This thesis first simulated digital phantoms of lungs of different patients,and performed X-ray thermoacoustic imaging and CT imaging on these phantoms at the same time,and finally got registration training data sets.Then,this thesis segmented and extracted the tumor features of the image by using these data sets based on the segmentation model proposed above.And this thesis uses the large-deformation micro-morphometric metric mapping model as a benchmark for the fluid registration characteristics of medical images,and parameterizes the model through a convolutional neural network,and optimizes the parameters of the neural network on a set of images.Finally,this paper gives a pair of new registration training methods.That method can quickly calculate the deformation field by directly using voxel registration through network parameters.(4)Based on the GNN,this thesis establishes a super-resolution reconstruction model that can further improve the quality of CT images,and can accurately restore high-resolution CT images from low-resolution counterparts.In particular,this thesis uses a deep unsupervised network composed of 16 residual blocks to design the generator.Then this paper constructs a discriminator based on the supervised network to reduce the dimensionality of the output of each hidden layer.This thesis also proposes four types of loss functions to construct a new loss function to strengthen the mapping between the generator and the discriminator.Finally,this thesis deletes the batch normalization layer in the common residual network to construct a new type of residual network.The innovations of this thesis mainly include:(1)A new semi-supervised convolutional transfer deep neural network is proposed to perform initial tumor feature segmentation on 3D lung CT images,as the basic coding and decoding framework for registering X-ray thermoacoustic images to original transmission CT images.Compared with the mainstream segmentation methods,the Dice coefficient of this method can be improved by about 6% in the absence of labeled data.(2)A fluid model-based registration method for lung CT and thermoacoustic images is proposed.This method first performs pathological image segmentation to make the registration more accurate,which is expected to speed up medical image analysis and registration.At the same time,it provides a new direction for deep learning-based registration and its applications.(3)A method of GNN is proposed to perform super-resolution reconstruction of CT images to further improve image quality.In terms of experiments,advanced methods to conduct an objective and subjective comprehensive evaluation is used.Compared with the mainstream deep learning methods,the peak signal-to-noise ratio of our proposed network structure in super-resolution image reconstruction is improved by about 5%,and the reconstruction speed is accelerated by about 30%. |
PDF Full Text Request