Keywords:
Artificial Intelligence, Cardiac, Cardiovascular system, CT, CT-Angiography, Image manipulation / Reconstruction, Physics, Computer Applications-Detection, diagnosis, Diagnostic procedure, Image verification, Quality assurance
Authors:
T. Lossau1, M. Vembar2, H. Nickisch1, T. Wissel1, R.-D. Bippus1, M. Morlock1, M. Grass1; 1Hamburg/DE, 2Cleveland/US
DOI:
10.26044/ecr2019/C-0850
Methods and materials
Two convolutional neural networks are trained to recognize typical coronary motion artifact pattern like arc-shaped blurring and intensity undershoots in coronary cross-sectional image patches.
The data required for the supervised learning is generated by introducing simulated cardiac motion to 17 prospectively ECG-triggered CT cases with excellent image quality.
The data sets are acquired with a 256-slice CT scanner and verified as artifact-free by visual inspection.
Data generation: A Coronary Motion Forward Artifact model for CT data (CoMoFACT) is developed for retrospective simulation of cardiac motion during CT acquisition [3].
Synthetic motion vector fields (MVFs) which are spatially limited to the previously segmented coronary artery trees are integrated in a second reconstruction of the projection data [4].
The introduced motion level is controlled by the lengths of the displacement vectors in the MVFs (see Figure 1).
By this procedure,
over 40k cross-sectional image patches with varying motion level are collected for subsequent supervised learning.
The data base is case-wise separated with a ratio of 9 : 4 : 4 into the subsets training,
validation and testing.
Network training: Two convolutional neural networks (CNNs) are trained based on the synthetically motion perturbed clinical data.
A detailed description of the network architecture and the optimization setup is provided in [3].
The networks take one cross sectional image patch of size 60x60x7 with a resolution of 0.4 mm per voxel as input and deliver a value between zero and one as output each.
The first network is a classifier which distinguishes between presence and absence of motion artifacts and outputs the artifact probability.
The second network is a regressor which predicts the artifact level.
Network testing: A five-fold bagging approach is applied for quantitative evaluation of the networks performances on test data with synthetic artifacts.
Furthermore,
the networks are tested on eight additional CCTA data sets from different patients which exhibit real motion artifacts at the coronary arteries.
Given the approximate position of the coronary centerlines,
the neural networks deliver motion artifact measurements for the entire vessel.
The handcrafted motion artifact measures entropy and positivity from Rohkohl et al.
[5] are considered for comparison.
Furthermore,
observer rankings of four research scientists in a five point Likert scale are provided.