Spatio-temporal feature learning with reservoir computing for T-cell segmentation in live-cell hbox {Ca}^{2+} fluorescence microscopy

Abstract

Advances in high-resolution live-cell hbox {Ca}^{2+} imaging enabled subcellular localization of early hbox {Ca}^{2+} signaling events in T-cells and paved the way to investigate the interplay between receptors and potential target channels in hbox {Ca}^{2+} release events. The huge amount of acquired data requires efficient, ideally automated image processing pipelines, with cell localization/segmentation as central tasks. Automated segmentation in live-cell cytosolic hbox {Ca}^{2+} imaging data is, however, challenging due to temporal image intensity fluctuations, low signal-to-noise ratio, and photo-bleaching. Here, we propose a reservoir computing (RC) framework for efficient and temporally consistent segmentation. Experiments were conducted with Jurkat T-cells and anti-CD3 coated beads used for T-cell activation. We compared the RC performance with a standard U-Net and a convolutional long short-term memory (LSTM) model. The RC-based models (1) perform on par in terms of segmentation accuracy with the deep learning models for cell-only segmentation, but show improved temporal segmentation consistency compared to the U-Net; (2) outperform the U-Net for two-emission wavelengths image segmentation and differentiation of T-cells and beads; and (3) perform on par with the convolutional LSTM for single-emission wavelength T-cell/bead segmentation and differentiation. In turn, RC models contain only a fraction of the parameters of the baseline models and reduce the training time considerably.