Super-Resolution Processing Of Unpaired CBCT Images

With the development of medical imaging technology,Cone Beam Computed Tomography(CBCT)has been widely used in oral clinical diagnosis and treatment.However,with the advancement of medical technology,the resolution and clarity of CBCT images have become difficult to meet clinical demands.In recent years,deep learningbased image super-resolution technology has made significant progress.However,existing CBCT image super-resolution technology still faces challenges such as lack of datasets,difficulty in training unpaired high and low-resolution CT images,and difficulty in aggregating information between different CT sections.This thesis conducts the following research:(1)To address the problem of a lack of datasets,this thesis collaborated with West China Hospital of Stomatology,Sichuan University to collect 5930 low-resolution CBCT images and 56568 high-resolution Micro-CT(Micro Computed Tomography)images of48 extracted teeth and completed pre-processing operations such as grayscale value normalization,resolution adjustment,and semantic segmentation labeling.We established an unpaired CT image super-resolution dataset.At the same time,the collected data is registered based on the iterative closest point algorithm to obtain a paired CT image superresolution dataset.The correctness of the constructed dataset is verified by calculating the mean intersection-over-union.This provides data support for subsequent research on CT image super-resolution.(2)To address the difficulty of feature extraction from a single CT image,this thesis designed a CT image super-resolution network based on self-attention mechanism.The self-attention mechanism helps the model better process global and local information in the image,thereby improving the model’s ability to understand image content.At the same time,to solve the problem that unpaired datasets are difficult to use for image superresolution training,this thesis combined the Cycle-GAN with the above super-resolution network to design three networks: Upsampling before Cycle-GAN,Upsampling within Cycle-GAN and Upsampling after Cycle-GAN.All of them constrain the content of the generated images with cycle consistency loss,and achieve unpaired training.(3)To address the difficulty of aggregating information between different CT sections,this thesis designed a CT image super-resolution network based on inter-layer correlation.The network uses bidirectional second-order propagation and mainly uses deformable convolution for feature alignment during the propagation process,avoiding the accumulation of errors during propagation and achieving more accurate feature extraction.The super-resolution network composed of shallow feature extraction modules and super-resolution modules efficiently and accurately achieves super-resolution processing of CBCT images.The experimental results show that the Cycle-GAN-based unpaired image superresolution network can achieve similar effects as paired image training,and the peak signal-to-noise ratio(PSNR)of Cycle-GAN mid-upsampling is 0.63 d B higher than that of the SRGAN algorithm based on paired image training.The designed CT image superresolution network based on inter-layer correlation has a PSNR that is 0.11 d B higher than the Basic VSR++ algorithm.Its super-resolution results can achieve more continuous 3D modeling effects,helping doctors to identify and analyze important and delicate dental anatomical structures such as root canals.