Hongjun An1, Hyeong-Geol Shin1, Sooyeon Ji1, Woojin Jung1, Sehong Oh2, Dongmyung Shin1, Juhyung Park1, and Jongho Lee1
1Department of Electrical and computer Engineering, Seoul National University, Seoul, Korea, Republic of, 2Division of Biomedical Engineering, Hankuk University of Foreign Studies, Gyeonggi-do, Korea, Republic of
1Department of Electrical and computer Engineering, Seoul National University, Seoul, Korea, Republic of, 2Division of Biomedical Engineering, Hankuk University of Foreign Studies, Gyeonggi-do, Korea, Republic of
A new deep-learning method correcting for
respiration-induced artifacts is designed. This method extracts B0
fluctuation from a GRE image, providing reliability in the correction. The
results show a successful correction of the artifacts.
Figure
1. Overview
of DeepResp. (a) Overview of DeepResp. DeepResp is designed to extract the
phase error from a GRE image. (b) Architecture of a two-stage neural network in
DeepResp. The first stage extracts the differential values of the phase errors.
The second stage accumulates the differential values, generating the phase
errors.
Figure
2. In-vivo
results of DeepResp in deep breathing. (a) Two slices (first and third rows)
and their zoomed-in images (second and fourth rows) are shown. Artifacts (red
and yellow arrows) are clearly reduced in DeepResp- and navigator-corrected
images. The phase errors show very high correlations between the results of
DeepResp (red line) and the navigator (black line). (b) Quantitative metrics of
NRMSE report improvements in all subjects (red: DeepResp-corrected; blue: uncorrected
images)