Low Dose Cone Beam Breast CT Based On Deep Learning

Cone beam breast computed tomography(CBBCT),a dedicated imaging technique for breast,is of great significance for breast cancer screening and diagnosis.Compared with the two dominant x-ray breast imaging modality mammography and tomosynthesis,CBBCT can avoid breast extrusion,provide 3D breast images with high spatial resolution.However the increased radiation dose hinders clinical application of CBBCT.Women breasts are highly sensitive to radiation dose that can increase the risk of suffering cancer.Therefore there are strong needs in developing low dose CBBCT.Sparse-view CT and low-mA CT are the two most common way to lower the radiation dose.Sparse-view CT acquires fewer projections;low-mA CT reduces the x-ay flux by decreasing the current or shortening the exposure time of the X-ray tube.Both of them lead severe artifacts in CT images reconstructed by filtered back projection(FBP)which could compromise the diagnostic performance.Recently deep learning appeals great attention in CT imaging.The key point of applying deep learning to low dose CT is reducing artifacts while maintaining contrast and preserving details.The fine glandular structures in breast and low fat-glandular contrast,pose deep learning in low dose breast CT great challenges.This paper aims at improving low dose breast CT image quality via avoiding over-smoothing and detail lost problem in sparse-view and low-mA CBBCT respectively.This paper contains two parts as below.For sparse-view CBBCT this paper proposes a image post-processing algorithm which is a conditional generative adversarial network(GAN)constrained by image edges,called ECGAN.The generator of ECGAN is the improved U-net;the discriminator is the combine of patchGAN and LSGAN.The improvement of U-net and the addition of the adversarial loss are for better train convergence and preservation of high frequency information.To further preserve subtle structures and micro calcifications,edge images of CBBCT are extracted and added in conditional GAN to guide the learning.Evaluated on clinical datasets of CBBCT,the results shows ECGAN presents better results in terms of artifacts suppression and avoiding over-smoothing compared with other two state-of-the-art methods.These results indicate that ECGAN successfully reduces radiation dose of CBBCT by a factor of 4 with only small image degradations.For low-mA CBBCT reconstruction,a projection domain and image domain cascaded network PI-Net is proposed.The PI-Net contains three parts:projection denoising network PD-Net,FBP and image enhancement network IE-Net.The PI-Net divides the low dose CBBCT reconstruction into two parts,each parts concentrates on one mission.Specifically PD-Net focuses on reducing artifacts via denoising the two dimensional projection in each view in CBBCT.The IE-Net focuses on solving the over-smoothing problem and improving image quality further.The design makes networks training easier.The results on clinical simulation datasets show the PI-Net achieves great improvement in reducing artifacts,contrast retention compared with other state-of-the-art methods.