Detection of cellular micromotion by advanced signal processing

Abstract

Cellular micromotion-a tiny movement of cell membranes on the nm-µm scale-has been proposed as a pathway for inter-cellular signal transduction and as a label-free proxy signal to neural activity. Here we harness several recent approaches of signal processing to detect such micromotion in video recordings of unlabeled cells. Our survey includes spectral filtering of the video signal, matched filtering, as well as 1D and 3D convolutional neural networks acting on pixel-wise time-domain data and a whole recording respectively.

Figure 1

Experimental setup and data. (a) Experimental setup. A correlative video microscope records a sample of cells in two channels: light transmitted under brightfield illumination and fluorescence of a Ca-active staining. LP long pass, SP short pass, pol. polarizer. (b) Resulting data. A region of several cells is visible in the transmission channel (scale bar: 20 µm). The same region displays spikes of Ca activity in the fluorescence channel. The fluorescence intensity of the whole region is summed to a time trace, which is employed as ground truth for supervised learning.

Figure 2

Signal processing schemes to detect cellular micromotion (a) concept: signal processing is employed to predict fluorescence from micromotion cues in transmission data. (b–d) present schemes processing time-domain data from a single pixel, (e) presents 3D neural networks processing a whole recording in the temporal and spatial domains.

Figure 3

Performance of time-domain signal processing. (a) Definition of correlation score. Predicted fluorescence on the single-pixel level is correlated with observed fluorescence summed over the full region. The maximum correlation is used as a score to assess accuracy of the prediction. (b–e) Pixel-wise maps of correlation score for (b) band-pass filtering, (c) matched-filtering, (d) processing by a 1D CNN as defined in Fig. 2 and (e) fluorescence activity (ground truth) (f) still frame from transmission channel. Labels denote regions of interest displaying strong motion, weak motion and no visible motion that will serve as test cases in the following analysis (Fig. 4). Scale bar: 20 µm.

Figure 4

Performance of all considered schemes. Signals of 1D predictions (upper three lines) have been summed over the regions of interest marked in Fig. 3. The output of filtering approaches (upper two lines) has been squared to produce unipolar data comparable to fluorescence. All approaches manage to correctly predict fluorescence in the strong beating region. Performance varies in the weak beating region, where neural networks yield a clear gain in accuracy. No approach is able to reveal a meaningful signal in the silent region. Length of the recording is 20 s.

Figure 5

3D neural networks. Weights of the fully connected layer (“2D Dense” in Fig. 2e), connecting three activity maps to the final output neuron. Weights are encoded in color and overlayed onto a still frame of the transmission video. The color scale is adjusted for each region; max: maximum weight occurring in all three layers of one region. (a) Strong beating region. Weights are placed on a confined region, presumably a single cell. (b) Weak beating region. Weights are predominantly placed on the border of one cell or nucleus, where intensity is most heavily affected by membrane motion. (c) Silent region. While no meaningful prediction is obtained, the network does place weights preferentially on the border of one cell or nucleus, hinting towards micromotion.