Deep Learning-base Adaptive Visual Enhancement And Tissue Segmentation Methods Of CT Images

Computed tomography is an essential component of modern medical assessment.CT has gradually become a common examination method due to its non-invasive,fast imaging,and high spatial resolution.Nevertheless,during the CT image reconstruction process,radiology experts should pick the appropriate filter kernel based on the diagnostic task in order to get image quality that fulfills the diagnostic demands.Upon the completion of the image reconstruction,the doctor should pick the proper window width and window level setting based on the CT value of the target tissue to improve the target tissue’s visual effect.The above situation can be divided into two challenges:(1)There are many filter kernel choices,and one filter kernel could typically only fulfill the diagnostic tasks of several specific tissues,the sharp filter kernel,for example,is appropriate for chest examination,whereas the smooth filter kernel is appropriate for mediastinum and other soft tissue examination;(2)There are various window width/level options,and CT images are high dynamic range image.As a result,during clinical diagnosis,the window width and level must be modified,but this will cause the visual information loss of non-target tissues.To solve the first problem,some researchers attempt to combine deep learning(DL)technology with CT image filter kernel style to achieve filter kernel conversion technology in the spatial domain.For the second problem,some scholars have proposed the "All-in-one" visualization concept,which attempts to merge image representations of multiple window widths and levels into a single image.Unfortunately,prior studies only targeted one of the problems while ignoring the other,so there are certain shortcomings in terms of visual effects.As a result,this work focuses on the shortcomings in CT image visualization research,explores how to achieve high-quality visual enhancement of CT images using deep learning technology,and examines the superior performance of visually enhanced images in multi-tissue segmentation tasks.The following is an overview of the specific research work:(1)Aiming at the characteristics of multiple filter kernel style,multiple window width and window level options of CT images,this thesis proposes a tissue-aware grayscale adaptive CT image representation conversion net(TACNet)based on deep learning.The network uses a style conversion loss function to learn the style differences and grayscale variation rules between different filter kernel reconstruction images and common window width/level settings,so that the CT image can adaptively select the most suitable reconstruction filter kernel and grayscale mapping curve according to the tissues.Following that,an improved self-supervised loss function based on the mean filtering method is proposed to improve tissue boundary visualization.The suggested TACNet could merge multiple filter kernels,window widths,and window level contrasts into a single image to improve visual effects,and this method outperforms the comparison methods in both qualitative and quantitative terms.(2)Due to the fact that convolutional neural network simulating human perception system for feature extraction,this thesis proposes a visual enhancement and multi-organ segmentation model for CT images based on generative adversarial network(VC-SegAN).This method incorporates TACNet with a discriminator that can extract multi-scale feature information into a conventional segmentation task,so the"Enhancement-Segmentation-Adversarial" work flow is formed.The segmentation sub-network then combines the enhanced image and the original image to predict masks for abdominal multiple tissues.The results of the experiments indicate that the VC-SegAN can produce high-precision abdominal tissue masks.