Detecting abnormal cell behaviors from dry mass time series

Abstract

The prediction of pathological changes on single cell behaviour is a challenging task for deep learning models. Indeed, in self-supervised learning methods, no prior labels are used for the training and all of the information for event predictions are extracted from the data themselves. We present here a novel self-supervised learning model for the detection of anomalies in a given cell population, StArDusTS. Cells are monitored over time, and analysed to extract time-series of dry mass values. We assessed its performances on different cell lines, showing a precision of 96% in the automatic detection of anomalies. Additionally, anomaly detection was also associated with cell measurement errors inherent to the acquisition or analysis pipelines, leading to an improvement of the upstream methods for feature extraction. Our results pave the way to novel architectures for the continuous monitoring of cell cultures in applied research or bioproduction applications, and for the prediction of pathological cellular changes.

Figure 1

(a) Dry mass measurements (in pg) for the cell of interest as a function of time. (b) 72-h timelapse acquisition of a HeLa cell. Cell tracking and cell segmentation are computed together to obtain this time-lapse series. Each cell successfully tracked is depicted with a different color. Every cropped image is $100\times 100\mu$ m$_{}^2$. The cell of interest is centered in the cropped image (light red). Four green asterisks point cell divisions. For readability purposes, the time between two images is 20 min, even if the acquisition rate is one image every 10 min. The four cell divisions are associated with a decrease in cell dry mass by a factor of two.

Figure 2

Full scheme of the StArDusTS model. (a) Shows the window-wise anomaly detection based on the representation extracted with a 1D-CNN. (b) Shows the whole StArDusTS model for a cell, including the 1D-CNN representation learning block, the window-wise anomaly detection and the aggregation of window-wise results with the anomaly score.

Figure 3

Time-lapse acquisition of adherent cells (Hela) that were detected as abnormal by StArDusTS. Each subplot depict an example of a subclass of anomalies. Every cropped image is $100\times 100\mu$  m$_{}^2$. The time between two images is 20 min. Cell tracking and cell segmentation are computed together to obtain time-lapse series. Each cell successfully tracked is depicted with a different color. The cell of interest is centered in the cropped image. Under each cell, the dry mass (in pg) time series from which the anomalies have been detected is printed. The time steps corresponding to the pictures are framed in red.

Figure 4

ROC curve of the anomaly detection on datasets $\mathcal{C}$ and $\mathcal{D}$. Each point is a threshold value of ${\tau }_w$ for the consideration of a window as abnormal. The red cross is the ${\tau }_w$ value computed from the training dataset $\mathcal{C}$ such that the time series outside the 95% interval of confidence are abnormal.