Encrypted login | home

Program Information

Statistical Iterative CBCT Reconstruction Using Convolutional Neural Network

S Tan

Z Gong1 , Q Shi1 , L Liu1 , J Wang2 , S Tan1*, (1) Huazhong University of Science and Technology, Wuhan, Hubei, China (2) UT Southwestern Medical Center, Dallas, TX


SU-I-GPD-I-7 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall

Purpose: To propose a novel method for CBCT statistical iterative reconstruction (SIR) using Convolutional Neural Network (CNN).

Methods: A novel objective function was designed to account for the blur process in CBCT SIR using the half quadratic splitting method. The reconstruction was implemented using an alternative optimization algorithm consisting of two parts. The first part corresponded to SIR using Hessian, while the second part corresponded to a de-blur problem. A CNN was adopted for the de-blue problem in an iterative way. We designed an end-to-end, pixels-to-pixels network to directly capture the characteristics of blurry objects via learning the map between the low-dose and normal-dose CT image pairs. The blurring kernel was estimated using the image of point sources in the CatPhan 600 phantom. The reconstruction of the CatPhan 600 phantom and the Shepp-Logan phantom was tested. Visual inspection and the full width at half maximum (FWHM) were used for evaluation.

Results: SIR using Hessian, SIR using total variation (TV), and the proposed method all had a good ability in reducing noise. SIR using TV preserved edges well but led to the staircase effect in regions with smooth intensity transition, while SIR using Hessian suppressed the staircase effect but slightly blurs object edges. The proposed method had a high performance: it preserved the smooth transition regions better than TV and yielded more pleasurable image resolution than Hessian. Particularly, the FWHMs of SIR using TV, SIR using Hessian and the proposed method were 1.81, 2.45 and 1.94, respectively.

Conclusion: Our experiments indicated that the proposed method had an excellent ability in reducing noise, suppressing staircase effect and preserving edges.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by National Natural Science Foundation of China (NNSFC), under Grant Nos. 61375018 and 61672253.J. Wang was supported in part by grants from the Cancer Prevention and Research Institute of Texas (RP130109 and RP110562-P2), the National Institute of Biomedical Imaging and Bioengineering (R01 EB020366).

Contact Email: