Hypertrophic cardiomyopathy (HCM) is a major contributor to sudden cardiac death (SCD) in young adults and athletes. Effective management and treatment of HCM require accurate assessment of the disease severity and risk of SCD. Several studies showed that scar volume measured from late gadolinium enhancement (LGE) can be used for predicting SCD risk in HCM. The clinical utility of scar volume as a prognostic marker requires a standardized analysis method that conveniently integrates into the clinical workflow. However, current methods rely on manual analysis and thus are tedious and lack standardization. We recently addressed this challenging limitation by showing the strong potential of deep convolutional neural networks (CNN) to segment the scar in LGE images. We also identified two limitations of the method hindering the translation of this novel scar quantification method to clinical practice. That is, the developed CNN showed: i) low performance at regions with low definition of LV boundaries; and ii) underestimation of scar volume due to the dominance of cases with small scars in the training image set. In this research, our goal is to substantially improve the performance of automatic HCM scar quantification by introducing two novel modifications to the architecture and training scheme of our CNN network. First, we propose to resolve the low conspicuity of LV boundaries in LGE images by modifying the CNN architecture to allow simultaneous learning of cine images, in addition to LGE images. The rationale is that, in practice, conventional cine images are often needed by human readers to accurately identify ambiguous scars and myocardium boundaries. Secondly, we propose to employ an ensemble of CNN models, rather than a single network, to improve the segmentation accuracy of cases underrepresented in the training set. The premise is that ensembles of network models have been successfully used and showed significant potential to overcome imbalanced training sets in many image analysis applications. Finally, we will dedicate a large portion of project time to evaluate the reproducibility and validity of the developed method using a large (n=400) independent validation dataset for HCM patients. The outcomes of this project will pave the way for more effective treatment and management of HCM through facilitating accurate and objective evaluation of scar burden.
|Program type||Institute - AI and ML|
|Effective start/end date||07/01/2019 → 06/30/2021|