AMES: Automated evaluation of sarcomere structures in cardiomyocytes

Abstract

BackgroundArrhythmias are severe cardiac diseases and lethal if untreated. To serve as an in vitro drug testing option for anti-arrhythmic agents, cardiomyocytes are being generated in vitro from induced pluripotent stem cells (iPSCs). Unfortunately, these generated cardiomyocytes resemble fetal cardiac tissue rather than adult cardiomyocytes. An automated tool for an unbiased evaluation of cardiomyocytes would highly facilitate the establishment of new differentiation protocols to increase cellular maturity. ResultsIn this work, a novel deep learning-based approach for this task is presented and evaluated. Different convolutional neural networks (CNNs) including 2D and 3D models were trained on fluorescence images of human iPSC-derived cardiomyocytes, which were rated based on their sarcomere content (sarcomerisation) and the orientation of sarcomere filaments (directionality) beforehand by a domain expert. The CNNs were trained to perform classifications on sarcomerisation, directionality ratings, and cell source, including primary adult and differentiated cardiomyocytes. The best accuracies are reached by a 3D model with a classification accuracy of about 90 % for sarcomerisation classification, 63 % for directionality classification, and 80 % for cell source classification. The trained models were additionally evaluated using two explanatory algorithms, IGrad and Grad-CAM. The heatmaps computed by those explainability algorithms show that the important regions in the image occur inside the cell and at the cellular borders for the classifier, and, therefore, validate the calculated regions. ConclusionIn summary, we showed that cellular fluorescence images can be analyzed with CNNs and subsequently used to predict different states of sarcomere maturation. Our developed prediction tool AMES (https://github.com/maxhillemanns/AMES) can be used to make trustworthy predictions on the quality of a cardiomyocyte, which ultimately facilitates the optimized generation of cardiomyocytes from iPSCs and improves the quality control in an automated, unbiased manner. The applied workflow of testing different CNN models, adjusting parameters, and using a variety of explanatory algorithms can be easily transferred to further image based quality control, stratification, or analysis setups.