Abstract Funding Acknowledgements Type of funding sources: Private company. Main funding source(s): Ultromics Ltd Background Myocardial wall motion analysis from echocardiography allows precise assessment of cardiac contractile function. Strain, which assesses myocardial deformation, has been shown to enable earlier detection of myocardial disease [1]. Current analysis software packages [2] use semi-automated methods to compute strain, which frequently require manual endocardial delineation and iterative contour adjustment based on tracking results, respectively, causing significant variability. Purpose We present a fully automated pipeline for tracking left ventricular (LV) wall motion to quantify global and segmental longitudinal strain from 2D echocardiograms, and go on to validate the pipeline with an openly available myocardial infarction (MI) dataset. Methods We applied our existing deep learning-based automated contouring method [3] to delineate the endocardial border in the A4C, A2C and A3C views and combined this with spline-based elastic image registration to track LV motion through time. We sampled points from a region of interest initiated from the endocardial border at the end-diastolic (ED) frame, and tracked subsequent motion by recomputing updated positions of all sample points based on each frame‘s displacement field, enabling us to both track the myocardium throughout the cardiac cycle and calculate longitudinal strain relative to the ED frame. The automated endocardial contour was used to regularise the process. The pipeline was independently tested on the HMC-QU dataset [4] which was downloaded from Kaggle and consists of a single cardiac cycle from the A4C view from 160 patients who were diagnosed with an acute MI and underwent echocardiography either prior to percutaneous coronary intervention or within 24 hours of undergoing the procedure; the dataset includes the labels of ED and end-systolic (ES) frames, as well as the presence of an MI in 6 segments excluding the apical cap (Fig 1a), as determined by the consensus of cardiologists from HMC Hospital in Qatar. The Wilcoxon signed-rank test was used to compare peak strain between the MI and non-MI segments; ROC curves were computed to compare the performance of the automatically derived peak longitudinal strain against the MI labels. Results Fig 1b shows ROC curves of peak segmental longitudinal strain for detecting MI, with the best performance in the mid-anterolateral segment (AUC 0.84), and a lower performance for basal segments than mid and apical segments, consistent with known variation in clinical practice [5]. Fig 2 shows that peak longitudinal strain computed from our pipeline was statistically significantly more positive in segments with an MI. Conclusions We present a fully automated pipeline for calculating segmental strain across a cardiac cycle to identify infarcted segments without any observer variability. Clinical application of this method has the potential to identify and monitor regional myocardial function and benefit patient management. Abstract Figure. Fig1. ROC of peak longitudinal strains Abstract Figure. Fig2.Boxplot of peak longitudinal strain

Type of funding sources: None. Segmentation of cardiac structures in echocardiography is a pre-requisite for accurately assessing cardiac morphology and function. Manual or semi-automated segmentation are both routinely used in clinical practice, although these can be time-consuming, and can introduce high inter- and intra- operator variability resulting in decreased reproducibility. Effective contouring with no manual input has proven to be challenging due to variations in image quality, image noise, motion during the acquisition and the lack of a well-defined geometry. This work proposes a coordinate regression method for automated left ventricle (LV) segmentation, presented in Figure 1 (a). The proposed method is based on a modified U-net architecture that outputs the likelihood of coordinates of landmark points. The obtained likelihood heatmaps are converted to 2D coordinates using a differentiable spatial to numerical transform. The model was trained and validated on UK multisite data (1383 subjects) comprising apical 2 and 4 chamber views for both contrast and non-contrast echocardiographic images. The Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (CAMUS) echocardiographic image segmentation database was used to assess the performance of the proposed method acting as data from a new clinical site. The CAMUS dataset comprises apical 2 and 4 chamber views acquired from 500 patients with manually annotated cardiac structures for end-diastole and end-systole frames. The original CAMUS dataset was split into training (450 patients) and testing (50 patients), with manual contours being available only for the training dataset. Therefore, we used the CAMUS training dataset to both test and improve our model, by using a random sample of 100 studies as an independent testing dataset and the remaining 350 studies were used for retraining the initial model to improve performance for this dataset. The results obtained on the testing images are presented in Figure 1 (b). When the model was trained using no CAMUS data for the LV segmentation, a mean Dice coefficient of 0.890 and a median of 0.911 was obtained. Including 350 studies with the original 1383 UK dataset and retraining the same model improved the average Dice coefficient to 0.930 and the median to 0.939. The CAMUS dataset authors reported the best average Dice coefficient of 0.924 on the 50 CAMUS testing images, therefore the proposed points regression method introduces a promising alternative to mask-based segmentation models. In conclusion, the auto-contouring framework has proven to be effective in terms of its performance and ability to generalise to new data. Furthermore, this work highlights the importance of both evaluating model performance on data from new clinical sites and also enhancing model performance. Abstract Figure.

Abstract Background Assessment of LVEF and myocardial deformation with GLS has shown promise in predicting CAD, which may add prognostic information for patients undergoing SE. However, selection bias precludes an accurate assessment of routine clinical SE workflow due to the exclusion of poor image quality and contrast enhanced studies. We hypothesise that an artificial intelligence (AI) pipeline capable of fully automated contouring of the left ventricle and GLS analysis of both non-contrast and contrast SE images is feasible and can predict CAD. Purpose The aim of this study was to evaluate the prediction of obstructive coronary artery disease (CAD) from fully automated left ventricular ejection fraction (LVEF) and global longitudinal strain (GLS) measures in a large multicentre population of patients undergoing stress echocardiography (SE). Methods 500 patients from five medical centres undergoing SE for the clinical evaluation of ischaemic heart disease were included in this study. LVEF and GLS was automatically calculated using AI in non-contrast and contrast images at rest and peak stress. The primary endpoint was CAD assessed using invasive coronary angiography. Results Patients with significant CAD demonstrated significantly reduced LVEF and GLS at rest and peak stress (all p<0.001) compared to those without CAD. Of the 130 patients who exhibited myocardial ischaemia at peak stress, patients without significant CAD (37%) had significantly reduced LVEF and GLS when compared to those who did. Multivariate analysis demonstrated that a peak LVEF (0.93; 95% CI 0.9–0.96) and peak GLS (1.15; 95% CI 1.07–1.24) were significant independent predictors of CAD. The addition of automated LVEF and GLS to basic models significantly improved the C statistic from 0.78 to 0.83 and 0.85 (both p<0.001), respectively. Conclusions Fully automated LVEF and GLS in non-contrast and contrast SE images is feasible and independently augment the prediction of obstructive CAD above and beyond traditional SE indexes. Funding Acknowledgement Type of funding sources: None.

Abstract Background Stress echocardiography has become established as the most widely applied non-invasive imaging test for diagnosis of coronary artery disease within the UK. However, stress echocardiography has been substantially qualitative, rather than quantitative, based on visual wall motion assessment. For the first time, we have identified and validated quantitative descriptors of cardiac geometry and motion, extracted from ultrasound images acquired using contrast agents in an automated way. Purpose To establish whether these novel imaging features can be generated in an automated, quantifiable and reproducible way from images acquired with perfluoropropane contrast, as well as investigating how these extracted measures compare to those extracted from sulphur hexafluoride contrast and non-contrast studies. Methods 100 patients who received perfluoropropane contrast during their stress echocardiogram were recruited. Their stress echocardiography images were processed through a deep learning algorithm. Novel feature values were recorded and a subset of 10 studies were repeated. The automated measures of global longitudinal strain (GLS) and ejection fraction (EF) extracted from these images were compared to values previously extracted from sulphur hexafluoride contrast and non-contrast images using the same software. Results A full set of 31 novel imaging features were successfully extracted from 79 studies acquired using the perfluoropropane contrast agent with a dropout rate of 14% (n=92, 8 incomplete image sets). Repeated analysis in a subset of 10 perfluoropropane cases demonstrated excellent reproducibility of the extracted feature values (R2=1). Automated values of GLS and EF, at both rest (GLS = −16.4±4.8%, EF = 63±13%) and stress stages (GLS = −17.7±5.8%, EF = 68±11%), were extracted from 83 perfluoropropane studies, with a dropout rate of 16% (n=99, fewer incomplete sets as short axis view not required). The ranges of GLS and EF measures extracted from the perfluoropropane images were comparable to the other contrast studies (n=222) (Rest GLS = −16.8±5.8%, Rest EF = 63±10%; Stress GLS = −19.1±6.7%, Stress EF = 71±9%) and non-contrast studies (n=86) (Rest GLS = −15.7±5.3%, Rest EF = 57±10%; Stress GLS = −17.3±6.4%, Stress EF = 61±14%). Conclusions Novel features and clinically relevant measures were extracted from images acquired using perfluoropropane contrast for the first time in a fully automated and reproducible way using a deep learning algorithm. The analysis failure rate and generated measures are comparable to those extracted from images using other commonly used sulphur hexafluoride contrast agents and non-contrast stress echocardiography studies. These findings demonstrate that deep learning algorithms can be used for automated quantitative analysis of stress echocardiograms acquired using various contrast agents and in non-contrast studies to improve stress echocardiography practice. Funding Acknowledgement Type of funding source: Private company. Main funding source(s): Lantheus Medical Imaging, Inc.