Super-resolution Research Of Cbct Images Based On Weakly Supervised Learning

With the continuous development of medical image technology,medical images of various modes are more and more used in health monitoring,clinical diagnosis,auxiliary surgery and other fields.However,due to the limitation of imaging technology and hardware equipment,there are often noise and artifacts in medical images,and their clarity is relatively low,which is not conducive to the subsequent clinical application of medical images.Therefore,researchers have proposed the work of super resolution of medical images to improve the quality of medical images.Deep learning technology has developed rapidly in recent years,researchers use convolutional neural networks to achieve super resolution,effectively improving the quality of medical images.However,medical images are often difficult to obtain,and large-scale data sets cannot be constructed for supervised network training.Meanwhile,the generality of training results needs to be improved.To solve the above problems,this thesis constructs two CBCT image super resolution network based on prior GAN and CBCT super resolution method based on implicit image function.The main work of this thesis is summarized as follows:1)This thesis designs an image super resolution method of CBCT based on prior GAN.The pre-trained GAN network learns the hidden vector space that conforms to the distribution of reference images,and then the pre-trained GAN model is embedded into the U-shaped network to generate the target image with near reference rich features.The prior GAN method can capture the rich prior knowledge of texture and shape in Micro-CT,and improve the resolution of teeth in CBCT images to close to the effect of Micro-CT.In this thesis,a tooth detection module is used to locate a single part in the original medical image,and the matching problem between Micro-CT and CBCT in the network is solved through the mapping of the prior generator in the hidden vector space.Because of the different scanning equipment used,there are often different noises,artifacts and other information between CBCT images and Micro-CT images.In order to solve the gap in this field,wavelet transform is used to extract the high-frequency noise information of CBCT,and the extracted high-frequency noise is added to the construction of low-resolution to high-resolution image pairs,so that the generator can ignore the distribution information of CBCT noise in the process of super-resolution,so as to generate images more in line with the information distribution of Micro-CT.2)This thesis designs a CBCT super resolution method based on implicit image function.The feature information of each point in the image is extracted by the feature extraction module.The coordinates related to the target pixel in the whole world are obtained by calculating the movable vector,and the above coordinates and the adjacent points are integrated to obtain sufficient feature information.Wavelet transform and inverse wavelet transform are used to distinguish high frequency and low frequency feature acquisition before and after acquiring feature information.Finally,an implicit function is used to generate a super resolution image with arbitrary magnification.In this thesis,sufficient experiments are carried out on the CBCT data set,and it is verified that the super resolution effect of the proposed model is more competitive than the existing super resolution methods.After obtaining high quality medical images,computer aided diagnosis and treatment can be more efficient.