Research On The Key Issues Of Machine Learning Based Low-dose CT Imaging

Due to high spatial and temporal resolution,X-Ray Computed Tomography(X-CT)has increasing application in clinical practice,CT can not only image the ultrastructure of the human body non-invasively but also provide four-dimensional imaging for motor organs.However,when the cumulative dose of X-Ray s inside the body exceeds the tolerable dose for human organs,the internal structure of the human body will change probabilistically,even gene mutation.Therefore,to reduce the dose of X-Ray radiation during the CT scan for patient is urgent and informative.Theoretically,more X-Ray dose means higher quality CT image,and balancing image quality and X-Ray dose level has become a well-known trade-off problem in CT.According to the working mechanism of CT,various strategies can reduce the dose of X-Ray radiation during the CT scan for patient: lowering the single exposure dose of the X-Ray source,improving the utilization of single exposure dose of X-Ray source,reducing the projection acquisition rate,increasing the physical size of a single detector pixel and improving the detector detection efficiency,etc.Different strategies have different effects on CT image quality.In this work,we reduced the dose of X-Ray radiation during the CT scan for patient from three aspects,the first strategy is to reduce the projection acquisition rate,according to Fourier section theorem,low projection acquisition rate means incomplete data in projection space and it will result in streak artifacts in the reconstructed CT image.Through analysis,the shape and position of streak artifacts remained relatively stable under a certain scanning protocol.In this work,a deep-learning based method has been proposed to extract the common features of streak artifacts and then these common features can be utilized to help improve image quality.The proposed method employed the CT image reconstructed by filtered backprojection(FBP)as the input and the proposed neural network can be trained in an end-to-end manner.In addition,the number of parameters of the proposed network is 1/3 of those neural networks which have the same depth and width due to the feature reuse,and fewer parameters means that the proposed neural network can avoid overfitting to some extent in the case of small number of medical images.The second strategy is to increase the physical size of a single detector pixel.When the output of X-Ray source is less than the normal dose,increasing the physical size of the detector pixel will improve the X-Ray photon receiving area of a single X-Ray detector pixel and the signalto-noise ratio(SNR)of projection.However,this operation will lower the resolution of projection and induce a blurred CT image.In this work,we first introduce the relationship among the four images: high resolution sinogram,low resolution sinogram,clear CT image and blurred CT image.Further,a hybrid model by integrating the super-resolution model,deblur model and CT reconstruction model was proposed and converted into a convolutional neural network to make the hybrid model more flexible,and then the proposed neural network can be trained well from a training dataset of 210 samples to estimate the blurred kernels and penalty function.Several experiments have been carried out to validate the generalization and robustness of the proposed model for CT super-resolution reconstruction and the proposed model not only greatly improve the spatial resolution of CT system but also lower the cost of CT hardware indirectly.The third strategy is to improve the utilization of single exposure dose of X-Ray source.For conventional X-Ray source,the cathode needs to be heated to 2000 K to generate X-Ray,and the X-Ray source cannot be started and stopped instantaneously due to hyperthermy,thereby,we cannot precisely control single exposure dose of X-Ray source.Fortunately,carbon nanotube based on X-Ray source can generate and stop X-Ray at nanosecond level due to field emission and the X-Ray dose can be controlled by adjusting the grid voltage.Since the cold cathode does not generate much heat,the size of X-Ray source can be greatly reduced.We can design a novel CT system with several CNT X-Ray sources and X-Ray detectors and employ electronic impulse to activate CNT X-Ray sources installed at different positions to replace the rotation of traditional CT mechanical structures,and then it will greatly speed up CT scanning and lower the dose of X-Ray radiation during the CT scan for patient.However,the CT system with several CNT X-Ray sources and X-Ray detectors will decrease the number of projections,increase hardware costs and sometimes limit the size of the field of view of the CT system in some mechanical designs,different defects will induce different artifacts and features of these artifacts is relatively fixed.To obtain CT images with the same quality as traditional CT systems,we will adopt the above two data-driven models to restore high-quality CT images from low-resolution and sparse projections and improve the quality of CT images caused by changes in the system architecture.Conclusively,we comprehensively analyzed the influence of X-Ray dose on CT images during CT scanning,different approaches have been proposed to reduce the dose of X-Ray radiation during the CT scan for patient during CT scanning.Depending on the method to lower the dose of X-Ray radiation,the corresponding algorithm has been proposed to improve the quality of the degraded CT image.For different method,abundant experiments have been conducted to verify the effectiveness of the corresponding algorithm.