TY - CHAP
T1 - Semi-supervised assessment of incomplete LV coverage in cardiac MRI using generative adversarial nets
AU - Zhang, L.
AU - Gooya, A.
AU - Frangi, A.F.
PY - 2017/9/30
Y1 - 2017/9/30
N2 - Cardiac magnetic resonance (CMR) images play a growing role in diagnostic imaging of cardiovascular diseases. Ensuring full coverage of the Left Ventricle (LV) is a basic criteria of CMR image quality. Complete LV coverage, from base to apex, precedes accurate cardiac volume and functional assessment. Incomplete coverage of the LV is identified through visual inspection, which is time-consuming and usually done retrospectively in large imaging cohorts. In this paper, we propose a novel semi-supervised method to check the coverage of LV from CMR images by using generative adversarial networks (GAN), we call it Semi-Coupled-GANs (SCGANs). To identify missing basal and apical slices in a CMR volume, a two-stage framework is proposed. First, the SCGANs generate adversarial examples and extract high-level features from the CMR images; then these image attributes are used to detect missing basal and apical slices. We constructed extensive experiments to validate the proposed method on UK Biobank with more than 6000 independent volumetric MR scans, which achieved high accuracy and robust results for missing slice detection, comparable with those of state of the art deep learning methods. The proposed method, in principle, can be adapted to other CMR image data for LV coverage assessment.
AB - Cardiac magnetic resonance (CMR) images play a growing role in diagnostic imaging of cardiovascular diseases. Ensuring full coverage of the Left Ventricle (LV) is a basic criteria of CMR image quality. Complete LV coverage, from base to apex, precedes accurate cardiac volume and functional assessment. Incomplete coverage of the LV is identified through visual inspection, which is time-consuming and usually done retrospectively in large imaging cohorts. In this paper, we propose a novel semi-supervised method to check the coverage of LV from CMR images by using generative adversarial networks (GAN), we call it Semi-Coupled-GANs (SCGANs). To identify missing basal and apical slices in a CMR volume, a two-stage framework is proposed. First, the SCGANs generate adversarial examples and extract high-level features from the CMR images; then these image attributes are used to detect missing basal and apical slices. We constructed extensive experiments to validate the proposed method on UK Biobank with more than 6000 independent volumetric MR scans, which achieved high accuracy and robust results for missing slice detection, comparable with those of state of the art deep learning methods. The proposed method, in principle, can be adapted to other CMR image data for LV coverage assessment.
UR - http://www.scopus.com/inward/record.url?eid=2-s2.0-85031426647&partnerID=MN8TOARS
U2 - 10.1007/978-3-319-68127-6_7
DO - 10.1007/978-3-319-68127-6_7
M3 - Chapter
SN - 9783319681269
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 61
EP - 68
BT - Simulation and Synthesis in Medical Imaging
A2 - Tsaftaris, Sotirios A.
A2 - Gooya, Ali
A2 - Frangi, Alejandro F.
A2 - Prince, Jerry L.
PB - Springer Cham
CY - Cham
ER -