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
Results
Evaluation on synthetic motion artifacts: During classification,
a threshold of 0.5 is selected for the separation of motion-free and motion-perturbed coronary cross-sections.
The achieved mean classification accuracy of 93.26 % +- 1.82 is split into the ratio 46.90 % : 46.36 % : 3.64 % : 3.10 % for the rates TN : TP : FN : FP,
where positive refers to the class artifact.
The regression networks reach a mean absolute error of 0.112 +- 0.007; the corresponding confusion matrix exhibits a clear diagonal structure with few scattering (see Figure 2).
Evaluation on real motion artifacts: Figure 3 provides a qualitative evaluation of the networks performances in clinical practice.
In contrast to the handcrafted measurements,
high network activations are correctly located at areas of motion.
Image patches with consistent observer rankings are correctly assessed by the deep-learning-based measures.
Vessel segments with tiny lumen (see Figure 3g) and noise (see Figure 3e) are identified as potential sources of errors.
The networks show great generalization capabilities as they are trained on clinical data sets with step-and-shoot protocol but do also deliver sensible results for data with helical acquisition mode.
Furthermore,
transferability from synthetic to real motion artifacts is demonstrated.