Detection and classification of Ca²⁺ release events in cardiomyocytes using 3D-UNet neural network.

Abstract

Global Ca²⁺ increase in the cytosol of cardiomyocytes is crucial for the contraction of the heart. The malfunctioning of proteins involved in this process can trigger local events (e.g., sparks and puffs) and global events (e.g., waves). These are thought to be involved in the development of pathological conditions, such as arrhythmias. To understand the underlying mechanisms, it is important to detect and identify arrhythmogenic Ca²⁺ release events. We present a novel approach, based on a 3D U-Net architecture, to perform these tasks automatically. We employed data obtained with fast xyt confocal imaging of cardiomyocytes and provide a dataset of xyt image series where the subcellular Ca²⁺ events are manually segmented and labelled. We trained the neural network to infer comparable segmentation as outputs and processed them to obtain single event instances that are available for further analysis. We obtained promising results despite the relatively small amount of available annotated data and the challenges that it exhibits.