Paper 10574-41 Session 8: Motion, 8:00 AM - 9:40 AM, Salon B Deep-Learning-Based CT Motion Artifact Recognition in Coronary Arteries Tanja Elss, Hannes Nickisch, Tobias Wissel, Holger Schmitt, Mani Vembar, Michael Morlock, Michael Grass Philips Research Hamburg, Digital Imaging Hamburg University of Technology February 13, 2018
Motivation CCTA: = Coronary computed tomography angiography • Used for detection of coronary artery disease • Cardiac motion artifacts may limit evaluation Problem: • Potentially lead to misinterpretations • Motion artifact recognition at the coronary Goal: arteries by a deep-learning-based measure • Assess diagnostic reliability of CCTA images Purpose: • Steering and assessment of algorithms for motion compensation (MC) Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 2
Method Reference data collection Main idea: generate required data for supervised learning by introducing Forward artificial motion to high quality CT cases model Supervised learning Forward model Network Evaluation Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 3
Method Input of the forward model: Reference data • Cardiac CT data sets with collection excellent image quality (no motion reference) Forward • Coronary artery tree including model centerline and lumen contour • Corresponding ECG-triggered Supervised learning projection data Network 9 step-and-shoot Evaluation cases included Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 4
Method Application of the MC-FBP 1 algorithm Reference data • blurred image + true MVF = sharp image collection Forward model • takes estimated motion 𝑛 𝑢 Ԧ 𝑦 of each voxel 𝑦 into account during reconstruction Ԧ Supervised learning Network Evaluation Usual back projection Motion compensated back projection 1 U. van Stevendaal et al., “A motion -compensated scheme for helical cone- beam reconstruction in cardiac CT angiography”, 2008. Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 5
Method Application of the MC-FBP 1 algorithm Reference data • sharp image + artificial MVF = blurred image collection Forward model • takes estimated motion 𝑛 𝑢 Ԧ 𝑦 of each voxel 𝑦 into account during reconstruction Ԧ Supervised learning Network Evaluation Usual back projection Motion compensated back projection 1 U. van Stevendaal et al., “A motion -compensated scheme for helical cone- beam reconstruction in cardiac CT angiography”, 2008. Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 6
Method Generation of the MVF Reference data 𝜍𝑗 collection 𝑛 𝑢 𝑗 Ԧ 𝑦 = w Ԧ 𝑦 ∙ ∙ 𝑡 , 𝑗 ∈ {1,…,5} 𝜍 norm Forward displacement sample vectors: define 5 motion states model weight mask w Ԧ 𝑦 𝜗 0,1 : limits motion area Supervised forces smoothness learning displacement direction: 𝜍 𝑗 ∈ 𝑉 −1,1 3 ) random ( Ԧ Network target motion strength 𝑡 ∈ ℝ + : Evaluation scales displacement lengths Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 7
Method Task: Separate cross-sectional image patches Reference data into classes no artifact and artifact collection Forward model Supervised learning 𝑡 = 8 𝑡 = 10 𝑡 = 0 𝑡 = 2 𝑡 = 4 𝑡 = 6 • Database: ca. 18k samples of size 96x96 pixels balanced classes, case-wise separation Network augmentation: rotation, mirroring, cropping (60x60) Evaluation • Setup: 20-layer ResNet 1 , Adam optimizer 2 1 K. He et al., “Deep residual learning for image recognition”, 2016. 2 D. Kingma et al., “Adam: A method for stochastic optimization”, 2014. Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 8
Results – 2D 4-fold cross-validation (60x60) Reference data mean classification accuracy: 94.4% ± 2.9% collection no motion 1.0 Forward 0.5 0.0 model local motion 1.0 Supervised 0.5 0.0 learning local motion 1.0 Network 0.5 0.0 Evaluation motion level predicted artifact probability normalized entropy 1 normalized positivity 1 1 C. Rohkohl et al., “Improving best -phase image quality in cardiac CT by motio correction with MAM optimization”, 2013. Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 9
Results – 3D 4-fold cross-validation (60x60x11) Reference data mean classification accuracy: 95.6% ± 2.7% collection no motion 1.0 Forward 0.5 0.0 model local motion 1.0 Supervised 0.5 0.0 learning local motion 1.0 Network 0.5 0.0 Evaluation motion level predicted artifact probability normalized entropy 1 normalized positivity 1 1 C. Rohkohl et al., “Improving best -phase image quality in cardiac CT by motio correction with MAM optimization”, 2013. Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 10
Summary Step 1: Preprocessing Step 2: Forward model Step 3: Supervised Learning Reference data collection Artificial motion introduction 3D cardiac CT volume coronary artery tree sampled cross-sectional patches + projection data CNN ECG data artificial MVF MC-FBP no artifact or artifact Conclusions • Demonstrated feasibility of accurate motion artifact recognition in CCTA images using deep learning • Future work: – Increase robustness – Artifact level regression – Testing on real artifacts Tanja Elss 13 th February 2018 Deep-learning-based motion artifact recognition in CCTA images 11
Recommend
More recommend
