RPS 2405a - Artificial intelligence (AI) in cardiovascular imaging

Author Block: A. Chami, C. v. Schack, R. Adolf, N. Nano, E. Hendrich, A. Will, S. Martinoff, M. Hadamitzky; Munich/DE
Purpose or Learning Objective: Coronary Computed Tomography Angiography (CCTA) is an established modality for assessing coronary artery disease (CAD). Its role for prognosis assessment is still limited. Deep convolutional neural networks (CNNs) might improve this process by using plaque characteristics that are currently not used.
Methods or Background: The Consecutive CCTAs from patients with suspected CAD examined between October 2004 and January 2018 were analyzed. The primary endpoint was a composite of all-cause mortality, myocardial infarction and late revascularization. The training endpoint additionally included early revascularization. The clinical risk was assessed by Morise score; for conventional CCTA assessment, extent of CAD (eoCAD) and segment involvement score (SIS) were used. Semiautomatic post-processing was performed for vessel delineation and annotation of calcified and non-calcified plaque areas. Two-step training of a densenet-121 CNN was done: The full network was trained using the training endpoint, then the feature layer was trained using the primary endpoint. Five times cross-validation was performed to ensure that each CNN was evaluated on an unseen set of data.
Results or Findings: The study population comprised 5468 patients. During the follow-up of 7.2 years, 334 patients reached the primary endpoint; in addition, 405 early revascularizations occurred. The outcome correlation of CNN showed an AUC of 0.720±0.010 and 0.631±0.015 for training endpoint and primary endpoint resp. Combining CNN with conventional CT parameters showed an improvement of AUC from 0.791 to 0.821 (p<0.0001) and from 0.766 to 0.773 (p<0.0001) for eoCAD and SIS resp. In a stepwise model including clinical risk, conventional CT parameters and CNN, the latter improved the prediction from 0.813 to 0.819 (p=0.0022) and from 0.772 to 0.775 (p<0.0013) for eoCAD and SIS resp.
Conclusion: CNNs are a promising tool to further improve the prediction of major cardiovascular events after CCTA.
Limitations: No limitations identified.
Ethics committee approval: Ethics committe approval was received.
Funding for this study: No funding was received for this study.

Author Block: y. Wang; Zhengzhou/CN
Purpose or Learning Objective: To compare coronary CT angiography (CCTA) contrast opacification between a fixed trigger delay and patient-specific trigger delay bolus tracking in a third-generation dual-source scanner.
Methods or Background: 100 consecutive patients were randomly divided into two groups to perform CCTA scans in the bolus tracking method; group A with a fixed trigger delay time of 5 seconds and group B with an automatic patient-specific trigger delay time estimated from monitored CT values. All CT scanning and contrast media injection protocol parameters were kept identical. CT value of aorta root (AO), coronary segments and superior vena cava (SVC) were measured for the objective image quality evaluation. Subjective evaluation of the image quality was performed by two independent blinded reviewers using a 5-point scale (5 = excellent, 1 = poor). Independent sample t-test and the Wilcox-Mann-Whitney test were used to compare quantitative and qualitative data, respectively.
Results or Findings: The trigger delay time in group B ranged from 4-8 seconds (mean, 6.6±1.4 seconds). Group B had higher mean enhancement in AO and coronary (407±61Hu vs 360±48Hu for pRCA, 423±58Hu vs 367±57Hu for pLAD, and 408±55Hu vs 351±56Hu for pLCX, all p＜0.05) than group A. The opacification of the SVC was significantly lower in group B than in group A (147Hu vs 261Hu; p<0.05). Subjective image quality was higher in group B than in group A (4.5 vs 4.1; p<0.05).
Conclusion: Compared with a fixed delay time, a patient-specific trigger delay bolus tracking estimation provided significantly higher attenuation and improved the image quality for coronary CT angiography.
Limitations: The diagnostic accuracy of our study needs to be further validated against invasive angiography.
Ethics committee approval: This study protocol was approved by the local ethics committee.
Funding for this study: No funding was received for this study.

Author Block: F. Huang, J. LIAN, K. S. NG, V. Vardhanabhuti; Hong Kong/HK
Purpose or Learning Objective: The purpose of this study is to utilize a graph convolutional neural network (GCN) to predict the coronary artery disease reporting and data system (CAD-RADS) based on coronary CT angiography (CCTA) using the quantitative vascular biomarkers derived from fundus images of the same subjects.
Methods or Background: This prospective single-centre study included 145 subjects who had received both CCTA and fundoscopy examinations on the same day in our local imaging centre in 2019. The CCTA scans were stratified by CAD-RADS scores by expert readers, which were then binarized into two classes (i.e. 0 (normal), 1 (minimal) and 2 (mild) were in class 0, and 3 (moderate), 4 (severe) and 5 (occluded) were in class 1). The vascular biomarkers were extracted from their eye images using a retinal health information and notification system. A graph was constructed, where each graph nodes represented a fundus image, and the node features were the vascular biomarkers relevant to blood vessel width and curvature. The graph edges were determined by the similarity of age and gender of paired subjects. A GCN model was employed on the graph to predict the binarized CAD-RADS score for each node. The image data of 115 subjects (80%) were used for training and 30 subjects (20%) for testing. We also trained multiple traditional machine learning models for comparison.
Results or Findings: The GCN model showed a sensitivity, specificity, accuracy and area under the curve of 75%, 81.03%, 79.27% and 0.864, respectively. The performance outperforms the same evaluation metrics obtained from the traditional machine learning models (p<0.01).
Conclusion: The changes in fundus vasculature had potential predictive value for CAD-RADS scores and significant coronary artery diseases.
Limitations: Small sample size. Proof of concept study.
Ethics committee approval: Approved by the local institution.
Funding for this study: No funding was received for this study.

Author Block: M. J. Pamminger, D. Obmann, C. Kremser, P. Poskaite, F. Troger, S. Reinstadler, B. Metzler, M. Haltmeier, A. Mayr; Innsbruck/AT
Purpose or Learning Objective: To compare fully automated segmentation of left ventricular late gadolinium enhancement (LGE) as evaluated by a convolutional neuronal network (CNN) with manual segmentation in chronic myocardial infarction.
Methods or Background: Cardiac magnetic resonance imaging, including two-dimensional LGE imaging, was performed in 191 patients on a 1.5 T clinical scanner 12 months after ST-elevation myocardial infarction. LGE images were presented to a trained CNN for automated determination of left ventricular myocardium and consequently LGE volume. Manual LGE segmentation according to the +5-SD method was used as the reference standard. Image quality was assessed according to a 3-point Likert scale (2 = perfect image quality, 1 = some artefacts without impaired LGE delineation, 0 = strong artefacts with impaired LGE delineation). Regression and Bland-Altman analyses were performed.
Results or Findings: In 191 included patients (182 male, mean age 57 years), the LGE volume was 9.7 [IQR 3.6 to 16.2] cm3 according to manual segmentation and 8.3 [3.2 to 17.6] cm3 according to CNN segmentation. The Bland-Altman analysis showed little average difference (-0.5 cm3, p=0.257), however, the limits of agreement ranged from -18.4 cm3 to 17.5 cm3. The linear correlation was fair (0.57, p<0.001). The subgroup analysis according to the image quality showed comparable performance of CNN segmentation in all three groups.
Conclusion: Our fully automated LGE segmentation based on a CNN in two-dimensional data sets provides measurements with little average difference compared to very time-consuming manual segmentations. However, dispersion is substantially and limits the current application of this approach on a per-patient basis. Image quality does not affect CNN performance.
Limitations: Manual segmentation according to the +5-SD method is dependent on investigator experience and is limited in circumferential myocardial LGE.
Ethics committee approval: Local ethics committee approval was provided.
Funding for this study: No funding was received for this study.

Author Block: M. A. Fink¹, C. Seibold², H-U. Kauczor¹, R. Stiefelhagen², J. Kleesiek³; ¹Heidelberg/DE, ²Karlsruhe/DE, ³Essen/DE
Purpose or Learning Objective: We aimed to develop a deep learning (DL) model based on detector-based spectral dual-energy angiography CT (DE-CTPA) data, yielding predictions of low-keV acquisitions to improve automatic pulmonary embolism (PE) detection in conventional single-energy CT scans.
Methods or Background: We used two data sets: our institutional DE-CTPA data set D1 comprising standard arterial series and the corresponding virtual monoenergetic images (VMI) at low-energy levels (40 keV) with 7,892 image pairs, and a 10% subset of the RSNA Pulmonary Embolism Detection Challenge (2020) data set D2, which consists of 161,253 polyenergetic images with dichotomous slice-wise annotations (PE/no PE). We trained a fully convolutional 9-block ResNet encoder-decoder network to generate VMI predictions from D1, which are then fed into a ResNet50 network for the PE classification task on single-energy CT scans from D2. We evaluated our VMI reconstruction results in terms of Peak-Signal-to-Noise-Ratio (PSNR) and Structural Similarity Index Measure (SSIM). For PE identification, we performed a binary classification on slice level and reported the area under the curve (AUC).
Results or Findings: The quantitative results on the reconstruction ability of the DL model revealed high-quality visual VMI predictions with reconstruction results of 0.984 ± 0.002 (SSIM) and 41.706 ± 0.547 (PSNR). The PE classification yielded an AUC of 0.84 for our framework, which improves PE classification compared to other naïve PE classification approaches with AUCs up to 0.81.
Conclusion: Our results demonstrate that the prediction of synthetic VMI from polyenergetic CT scans can improve the automated detection of PE. This could help rescue CTPA studies with suboptimal opacification of the pulmonary arteries from single-energy CT scanners.
Limitations: Class imbalance per subset could bias the results.
Ethics committee approval: The study was approved by our IRB (S-236/2020).
Funding for this study: No funding was received for this study.

Author Block: D. Sobotka, A. Herold, M. Perkonigg, L. Beer, N. Bastati-Huber, A. Sablatnig, A. Ba-Ssalamah, G. Langs; Vienna/AT
Purpose or Learning Objective: Liver vessel segmentation in MR imaging is crucial for the computational analysis of vascular remodelling. Existing techniques rely on contrast-enhanced MR (MRce), which are not uniformly acquired. Non-contrast images are acquired more frequently, but vessels are hard to distinguish from other structures due to lack of opacification. Here, we propose a convolutional neural network to segment liver vessels on non-contrast images with the help of auxiliary contrast-enhanced data available only during training. The approach improves segmentation accuracy on non-contrast images and reduces the need for annotated examples.
Methods or Background: A multi-task learning approach trains a Y-net style convolutional neural network for liver vessel segmentation on MR imaging data. During training, MR with and without contrast together with vessel annotations on a sub-set of the data are available. The auxiliary MRce provides variability to the encoder training of the model. This improves the vessel segmentation accuracy, even if no MRce is available during application. We investigate overall- and vessel thickness specific segmentation accuracy.
Results or Findings: Using auxiliary contrast-enhanced sequences improves the Dice score for vessel segmentation on non-contrast MR by 0.10 from 0.45 to 0.55. For small vessels (0-10 mm), the score increases by 0.03, for bigger vessels (>10mm) by 0.13.
Conclusion: Highly-informative contrast-enhanced sequences improve vessel segmentation models for non-contrast imaging data. It allows for a reduction of the number of annotated examples necessary for vessel segmentation model training.
Limitations: Number of vessel annotations used in evaluating the proposed framework.
Ethics committee approval: The local ethics committee approved this study protocol (EK 2027/2017), which was performed in accordance with the Helsinki Declaration.
Funding for this study: This study was partially funded by Austrian Science Fund (FWF): P 35189, Vienna Science and Technology Fund (WWTF): LS20-065, Novartis Pharmaceuticals Corporation.