Deep Learning-Based Cone-Beam CT Scatter Correction&Image Enhancement

Computed Tomography is an important auxiliary tool in clinical diagnosis.Cone-beam CT(CBCT),which has appeared in recent years,gradually demonstrates its great potential for clinical use because of its higher spatial resolution,shorter scan time,higher radiation utilization,and lower radiation dose.However,due to the influence of beam hardening and other factors,CBCT is often accompanied by more serious defects such as scattering artifacts and low contrast in the imaging process.These defects limit the further use of CBCT images.This paper first proposes a deep residual learning-based CBCT scattering correction method.Instead of using phantom data,real CBCT-CT image pairs collected from Image-guided radiation therapy system are used to construct our dataset with a two-steps registration applied to eliminate the valid information in residual labels.Deep CNN is adopted to implicitly learn the degradation model of CBCT scater.By subtracting the learned residual images from original CBCTs,scatter correction can be implemented.Implied by experiment result,this method is sufficient to suppress the scattering artifacts while retaining the CBCT details,and the potential performance of subsequent applications(such as threshold-based CBCT image segmentation)can be further improved.Secondly,this paper explores the further improvement of CBCT image quality from the perspective of improving spatial resolution.A CBCT chest dataset is adopted to complete the training of the super-resolution model.Using the network's full convolution characteristics and generalization performance,the original resolution of CBCT image is 4 × enhanced.The experimental results show that by using this method,the detail and edges of CBCT images are both enhanced.