Towards Improving Single-Cell Segmentation in Heterogeneous Configurations of Cardiomyocyte Networks

To explore the formation and deterioration of cellular networks, we develop systems powered by Artificial Intelligence (AI) that accurately distinguish and quantify the differential configuration of cells in those networks (i.e. single cells, multicellular aggregates) as an initial proof-of-concept approach. We use image data acquired from self-organised cardiac cell networks formed in vitro which are difficult to segment using conventional methods. We used two data pre-processing approaches prior to the application of four segmentation algorithms (including two newly generated configurations of the Cellpose algorithm) for a total of eight segmentation pipelines. We demonstrate the effectiveness of a transfer learning capability in improving the accuracy of Cellpose in identifying discrete cells within complex (heterogeneous) cardiac cell network configurations. Our \(\texttt {Cellpose}^{p_1}_{\texttt {3}}\) segmentation pipeline displays an F1-Score of \(82.34\%\) , a precision of \(88.52\%\) and an accuracy of \(87.84\%\) . Furthermore, in addition to our new method performing best in its ability to detect discrete cells in each network, it also avoided the problem of erroneously identifying cell boundaries in large multicellular aggregates. This preliminary work shows the feasibility of describing the physical and functional properties of cellular networks using accurate indices of cellular arrangement and heterogeneity.