Abstract:
Cardiac diseases, including myocardial infarction (MI), are the leading cause of death worldwide. Dysfunction of the heart results from remodeling processes affecting electrophysiology and electro-mechanical tissue properties. Conceptual and computational models of diseases rely on accurate microstructural parameters, such as passive conductivity tensors, and their alteration during remodeling. However, our knowledge on these parameters is limited. To close this knowledge gap, a comprehensive computational framework for the reconstruction and analysis of normal and infarcted myocardial microstructure was implemented, particularly focusing on the derivation of passive conductivity tensors. 3D image stacks of control and MI ventricular tissue were acquired by confocal microscopy with a sub-micrometer resolution. Myocardial microstructure was assessed with a multi-label staining protocol, simultaneously acquiring signals from five fluorescent labels. Two synergistic segmen- tation approaches were used to predict the mask and boundaries of cardiomyocytes. Intended for scenarios when no previously annotated image stacks are available, a method requiring minor human interaction, based on hand-crafted image features and ensembles of decision trees, was implemented. For the scenario when multiple annotated stacks are available, e.g. due to segmen- tation using latter method based on ensembles of decision trees, a fully automated segmentation pipeline utilizing convolutional neural networks was implemented. The predictions were used to generate comprehensive 3D reconstructions of cardiac microstructure, including cardiomyocytes, extracellular space, fibroblasts, myofibroblasts, and vessels. To overcome bottlenecks in the reconstruction pipeline, a graph-based agglomeration method was used in the reconstruction process. 3D reconstructions of tissue microstructure (control: 5 stacks, MI: 5 stacks) were used to construct computational models of passive intra- and extracellular conductivity tensors. Predictions of the decision tree-based segmentation approach resulted in mean Matthews corre- lation coefficients (MCCs) between 0.90 and 0.96 for increasing amounts of available training data. The deep learning-based segmentation approach yielded mean MCCs of 0.87 and 0.91 for the prediction of the mask and boundary of cardiomyocytes, respectively. The graph-based agglomeration method achieved a mean V-Measure of 0.84. Computational models of passive intra- and extracellular conductivity tensors were verified against analytic solutions and yielded errors in the order of 107 and 108, respectively. The mean intracellular conductivities in control tissue were 0.257, 0.021, and 0.001 S/m in fiber, transverse, and normal direction, respectively. For MI tissue, the mean intracellular conductivities were 0.090, 0.051 and 0.0002 S/m in fiber, transverse, and normal direction, respectively. The mean extracellular conductivities in control tissue were 0.388, 0.174, and 0.117 S/m in fiber, transverse, and normal direction, respectively. For MI tissue, the mean extracellular conductivities were 0.556, 0.433 and 0.224 S/m in fiber, transverse, and normal direction, respectively. Overall, this work implemented a computational framework capable of reconstructing normal and infarcted cardiac microstructure with minimal human interaction. Proposed segmentation algo- rithms showed promising results. Intra- and extracellular conductivity tensors in MI tissue were significantly altered compared to control tissue. The next step is the application to newly acquired and large-scale image data, putting the new framework into practice. The novel quantitative data can serve as a foundation for future modeling studies on, for instance, cardiac conduction in hearts with MI.