Abstract 173: Development of an Auto-Segmentation Technique Using a Convolutional Neural Network for the Segmentation of the Ventricular Cavity in Zebrafish

Abstract

Light sheet fluorescence microscopy (LSFM) with a synchronization algorithm is a unique imaging technique that can be used to image the beating zebrafish heart in 4-dimensions (4-D). However, natural arrhythmic diastolic and systolic contraction and relaxation of the heart can cause aberrations in 4-D image reconstruction. These aberrations are observed as blurred lines, spikes, or holes in the myocardium of the heart, which can lead to complications when analyzing biomechanics by computational fluid dynamics (CFD). In this study, we imaged a beating zebrafish heart at 4 days post fertilization (dpf) using a 10x objective lens with our single-sided LSFM system scanning with a 2 μm step size. The obtained images were analyzed in 3-D to ensure that there were limited aberrations in the wall of the myocardium. Then, the ventricular cavity of these 3-D images was hand segmented at various time points that were spaced within two cardiac cycles. These hand segmented images were used to train a convolutional neural network (CNN) which would have the capability of performing segmentation of the ventricular chamber automatically. This auto-segmentation tool was then used to segment the ventricular cavity of ten samples contained within one cardiac cycle that had not previously been used for training purposes. The ventricular volume of these ten images calculated after auto-segmentation was found to be comparable to the images when hand segmented. Then, a Gaussian wavelet filter was used after auto-segmentation to remove aberrations and minimize errors in the CFD simulation. Both the auto-segmented images and the hand-segmented images were analyzed using a CFD solver to simulate the hemodynamic shear stress through the cardiac cycle. The results from these simulations showed no significant difference between shear stress when the images were hand-segmented or auto-segmented, and there was no leakage observed between the two sets. The results using the auto-segmentation code are promising for future applications in cardiac research where low-quality images obtained under fast acquisition times can be enhanced without the traditional costly deconvolution techniques. This can lead to faster and more accurate analysis of the ventricular cavity for related diseases.