Embargo until
Journal Title
Journal ISSN
Volume Title
Johns Hopkins University
Machine learning and, in particular, deep learning have been sweeping many disciplines in recent years. Advancements in neural networks, the primary tool of deep learning, have made them the go-to approach in a variety of fields, ranging from computer vision to natural language processing. While democratizing deep learning, high-level software packages offering \qts{black-box} neural network models can also potentially exacerbate the divide between machine learning and classical statistical learning approaches. In this manuscript, we aim to combine machine learning and statistical techniques synergistically in a way that uses the remarkable feature extraction capabilities of deep learning, without compromising statistical rigor. We apply the techniques developed while tackling two important problems in the field of Computational Cardiology: \begin{enumerate}[label=(\alph*),leftmargin=0cm,itemindent=0cm,listparindent=0cm] \item We propose a novel deep learning solution for contrast-enhanced cardiac magnetic resonance (CMR) image analysis, which produces anatomically-accurate myocardium and scar/fibrosis segmentations and uses these to calculate clinical features. Visualizing scarring and fibrosis in the heart on CMR imaging with contrast enhancement (LGE) is paramount in characterizing disease progression and quantifying pathophysiological substrates of arrhythmias. However, segmentation and scar/fibrosis quantification from LGE-CMR is an intensive manual process prone to large inter-observer variability. We trained and tested deep neural networks which apply a 3-stage approach that identified the left ventricle (LV) region of interest, segmented the ROI into viable myocardium and regions of enhancement, and, lastly, used a post-processing network to ensure that segmentations results conformed to anatomical constraints. The segmentations were used to directly compute clinical features, such as LV volume and scar burden. LV and scar segmentations predicted by our model achieved $96\%$ and $75\%$ balanced accuracy, respectively, when compared to trained expert segmentations. The difference in mean scar burden between ground truth and predicted segmentations was $2\%$. All resulting segmentations passed morphological checks to ensure LV anatomical accuracy. We developed and validated a 3-stage deep neural network for automatic, anatomically accurate expert-level LGE-CMR segmentation, which was used for direct extraction of important clinical measures. As our model has been trained on scans from heterogeneous cohorts, it has the potential to be extended to multiple imaging modalities and patient pathologies. \item We develop a new deep learning technology, that directly uses LGE-CMR images to construct a parametric model in survival analysis. Sudden cardiac death from ventricular arrhythmia continues to be a major cause of mortality worldwide and a vast public health and economic burden. Current approaches to arrhythmic death risk prediction represent broad guidelines and fail to incorporate personalized, complex, large-scale clinical data and individualized phenotyping. Deep learning (DL) approaches are ideal for such data, however, most of the DL work related to arrhythmia has focused on disease classification and detection from ECG time series data. Furthermore, although mechanistically arrhythmia results from the heterogeneous scar distribution in the heart, DL on raw imaging scans, which visualize this distribution, has not been explored for risk analysis. In this work we develop a novel DL approach which blends neural networks and survival analysis to predict patient-specific survival curves from raw contrast-enhanced cardiac magnetic resonance images and clinical covariates for patients with ischemic cardiomyopathy. The DL-predicted survival curves offer accurate arrhythmic sudden death predictions at all times up to 10 years and allow for estimation of patient-specific uncertainty in predictions. The performance of this learning architecture was evaluated on multi-center internal validation data, and tested on an external, independent test set. It achieves both high risk discrimination (concordance index of 0.83 and 0.74, respectively) and high calibration (10-year integrated Brier score of 0.12 and 0.14, respectively). We additionally demonstrate that our DL approach learning from only raw, unsegmented cardiac images outperforms standard survival models constructed using both non-imaging and imaging clinical covariates. Brought to clinical practice, this technology has the potential to transform clinical decision-making by offering accurate, generalizable, and interpretable predictions of patient-specific survival probabilities of arrhythmic death over time.
Deep Learning, Sudden Arrhythmic Death