Detection and characterization of apoptotic and necrotic cell death by time-lapse quantitative phase image analysis

Cell viability and cytotoxicity assays are highly important for drug screening and cytotoxicity tests of antineoplastic or other therapeutic drugs. Even though biochemical-based tests are very helpful to obtain preliminary preview, their results should be confirmed by methods based on direct cell death assessment. In this study, time-dependent changes in quantitative phase-based parameters during cell death were determined and methodology useable for rapid and label-free assessment of direct cell death was introduced. Our method utilizes Quantitative Phase Imaging (QPI) which enables the time-lapse observation of subtle changes in cell mass distribution. According to our results, morphological and dynamical features extracted from QPI micrographs are suitable for cell death detection (76% accuracy in comparison with manual annotation). Furthermore, based on QPI data alone and machine learning, we were able to classify typical dynamical changes of cell morphology during both caspase 3,7-dependent and independent cell death subroutines. The main parameters used for label-free detection of these cell death modalities were cell density (pg/pixel) and average intensity change of cell pixels further designated as Cell Dynamic Score (CDS). To the best of our knowledge, this is the first study introducing CDS and cell density as a parameter typical for individual cell death subroutines with prediction accuracy 75.4 % for caspase 3,7-dependent and -independent cell death.


Introduction
Analysis of cell viability and the distinction of specific cell death subtype represents a key aspect in many areas of cell biology. This kind of information is also highly important for drug screening and cytotoxicity tests of antineoplastic or other therapeutic drugs. Cell death is considered reversible until a first 'point-of-no-return' is overstepped. While no exactly defined biochemical event can be taken as an undisputable proof of this point-of-no-return, cell should be taken as dead when any of these situations occur: (1) the cell has lost the integrity of its plasma membrane; (2) the cell and its nucleus has undergone complete fragmentation into discrete bodies (apoptotic bodies); (3) the cellular corpse has been engulfed and digested by a neighbouring cell 1 . Cells that are arrested in the cell cycle should be counted as viable 2 .
Methods for cell death analysis are usually based on various basic cell functions such as enzyme activity, semi-permeability of the mitochondrial or cellular membrane, cell adherence, ATP production, the presence of specific markers, or changes of functionality due to specific inhibitor (genetical or pharmacological) 3 . Methods for cell death detection form two main groups: a) methods that directly measure cell death; and b) methods that analyse biochemical processes or features characteristic for viable cells 4 . Even though indirect tests are very helpful to obtain preliminary preview, their results should be confirmed by methods based on direct cell death assessment 3 . According to NCCD (Nomenclature Committee on Cell Death), the currently accepted definition of cell death and its subroutines is based particularly on genetical, biochemical, pharmacological, and functional parameters, rather than morphological aspects 1 . Specific methods for apoptosis and necrosis detection are focused on typical biochemical parameters such as visualization of phosphatidylserine exposure, executioner caspases activation, or DNA fragmentation in the case of apoptotic cell death; and loss of barrier function and subsequent permeabilization of the plasma membrane with the release of specific death associated molecular patterns (DAMP) during necrotic cellular demise 1 . Nevertheless, almost all these methods are based on fluorometric or colourimetric endpoint visualization of the analyzed parameter and belong among indirect assays. Such indirect endpoint analyses are prone to the misleading results as we have shown in our previous work 5 . Although morphological aspects of cell death are not generally recommended to determine cell death subroutines 1 , it would be a mistake to completely ignore them. Recent progress in Quantitative Phase Imaging techniques (QPI) has enabled the observation of time-dependent subtle changes unrecognizable to the naked eye (such as cell mass distribution) on micrographs. These changes in cell mass distribution, cell density, micro-blebbing of the cell membrane, nuclear shape, homogeneity of cell content, and many other parameters, which can be typical for individual cell death subroutines, can be observed without fixation, labelling or cell harvesting.
In this article, we demonstrate a methodology useable for rapid assessment of direct cell death that is based on NCCD recommendations. Using advanced quantitative label-free phase imaging (QPI), we were also able to observe typical dynamical changes of cell morphology during both caspase 3, 7-dependent and -independent cell death subroutines and to determine the moment of cell death. Dynamical, time-dependent changes in cell mass distribution maps were used for the label-free distinction between typical morphological patterns appropriate for caspase-dependent and caspase-independent cell death subroutines.
As model cells, prostate cancer cell lines DU145, LNCaP, and benign cell line PNT1A established by immortalization of normal adult prostatic epithelial cells were used. These cell lines were not selected because of the clinical relevance, but rather because of their radically different size and accessibility for the automatic segmentation algorithm. We expected cellular size to be the factor influencing the rate of morphological changes during cell death.
Cell death was triggered by doxorubicin and staurosporine, two commonly known inducers of apoptosis with a distinct induction of caspase cleavage and consequently a distinct morphological manifestation 6,7 . Cell death induced by black phosphorus (BP) was detected as an example of difficult detection conditions 8 .
Cell tracking is an essential step in cell image analysis. Even though many methods exist 9 , there is no universal and sufficiently robust method applicable to QPI, especially for touching cells and if correct tracking through whole image sequence is needed. Simple available tools like TrackMate 10 have shown to be insufficient, thus we have developed a new tracking method tailored for our dataset.
The number of existing methods for the detection of cell death in label-free time-lapse images is very limited. In 11 authors detect cell death event in Phase Contrast Microscopy using features time-series (size, roundness, speed etc.) and classify each time point as alive or death with transductive support vector machine. In comparison, our technique based on quantitative microscopy data uses more features and more advanced Long Short-Term Memory (LSTM) neural network.
There are several techniques for static cell image classification including extraction of features and application of classifier 12,13 or application of convolutional neural network 14 .
These techniques can be also possibly applied for a distinction of different types of cell death, however, we decided for using only two features for this distinction in order to make our results easily interpretable, while also incorporating new features describing cell dynamics.  17 . DU-145 cell line is derived from the metastatic site in the brain and contains P223L and V274F mutations in p53. This cell line is PSA and ARnegative and androgen independent 18 . All cell lines used in this study were purchased from HPA Culture Collections (Salisbury, UK) and were cultured in RPMI-1640 medium with 10 % FBS. The medium was supplemented with antibiotics (penicillin 100 U/ml and streptomycin 0.1 mg/ml). Cells were maintained at 37°C in a humidified (60%) incubator with 5% CO2 (Sanyo, Japan). Cell lines were not tested on mycoplasma contamination.  (1) Integrated phase shift through a cell is proportional to its dry mass, which enables studying changes in cell mass distribution 20 .

Cell dry mass tracking
The custom method for automatic cell tracking and measuring of selected features was developed and implemented in MATLAB. Although tracking is not the main contribution of this article, it is a necessary step for our analysis and there is no available method suitable for QPI. The proposed tracking method consists of two main steps -nuclei tracking followed by expansion of each nucleus region to the whole cell. initialized as an identity matrix. The whole algorithm is summarized in Fig. 1 and for more details see 23 . This method does not recognise division of cells. For the next analysis, we use only cells, which occur in the whole image sequence without division or joining with other cell (which appears mainly due to tracking and segmentation errors).
The previous step provides tracked nuclei, but for extraction of most of the cell features, we need cell segmentation. However, nuclei can be used as seeds for segmentation in each frame.
We can use simple thresholding for segmentation of foreground in QPI cell image, where the threshold value can be the same in all frames, which provides sufficient results for a distinction between foreground and background, but a separation of single cells is very problematic. We use tracked nuclei as seeds for proper division of the foreground binary mask However, the simple application of the watershed algorithm leads to a fast mass exchange between cells due to segmentation errors, which is very undesirable for precise feature extraction. For this reason, we introduced a simple modification named Movement Regularized Watershed (MRW), in order not to allow dramatic contour changes between frames. This can be achieved by incorporating the mask from the previous frame to the actual frame watershed calculation. This can be done by modifying  Besides the classical cell features, we introduce tailored feature Cell Dynamic Score (CDS).
CDS is a mean Euclidian distance between cell pixels in the actual and the following frame computed as ‫ܯ‬ is a set of poisitons defined by the cell segmentation mask in the n-th frame and ‫ܫ‬ is the n-th frame of QPI. CDS provides information about the speed of change of the cell pixel values due to both movement and morphological changes, but it is not much dependent on the segmentation quality, because the same mask is used in both frames. All these features were evaluated in all frames, where the result is a set of signals describing the cell behaviour in time.

Label-free algorithm for cell death detection
Besides the significant mass decrease of dead cells, various feature evolvement types were observed during detected cell deaths, which complicate the expert specification of the detected phenomenon, thus machine learning approach was chosen for this task. Bidirectional Long Short-term Memory (BiLSTM) networks 25  According to 27 , the network was set to two BiLSTM layers (with 100 units) and 3 fully connected layers (with 100, 50 and 100 neurons) with ReLu and dropout with probability 0.5.
The whole network was optimized using Mean Square Error loss and ADAM optimizer 28 with learning rate 10 -3 , β 1 = 0.9, β 2 = 0.999, gradient clipping to norm 1 29 , weight decay 10 -3 , batch size 256 and 40 epochs. We augment our training dataset with random clipping (shortening each signal by at most 1/3 of length). One network was trained for all 9 cell

Cell death type identification
We observed two distinct dynamical patterns of cell mass manipulation during cellular death, where each cell was manually labelled as one of these types. To quantify these morphological

Automatic cell death detection
After induction of cell death by using 0.5 µM staurosporine, 0.1 µM doxorubicin, or 400µg/mL of BP, respectively, 11 morphological parameters detectable by QPI (see

Cell density and Cell Dynamic Score reflect subroutine of cell death
Based on QPI data, we were able to distinguish two different subroutines of cell death: a) cells having high cell density and intensive blebbing of the plasma membrane (high CDS) followed by plasma membrane rupture and b) cells having low cell density and low CDS before plasma membrane rupture (see Fig. 7 and Supplementary video 1-2). These variants occupied extreme positions on the CDS versus cell density plot (see Fig.6b and Supplementary fig. 1). Fig 6a and

Discussion
The basic requirement for a good method of cell death detection is the ability to detect a robust parameter in a highly reproducible and if possible inexpensive manner under changing conditions caused by the cell heterogeneity. In this study, we suggest that morphological and dynamical features extracted from the QPI cell image are suitable for cell death detection. QPI enables the time-lapse observation of subtle changes in the cell mass distribution unrecognizable by the naked eye, and therefore provides detailed information on cell morphology and cell mass topography during cell death. Cell dry mass can be calculated directly from phase values detected in each pixel 31,32,33 . Cell dry mass is the direct result of biosynthetic and degradative processes within a cell and is, therefore, a promising indicator of cell fate including cell death 34,35 . The distribution of cell mass during the processes of cell death is significantly changing in time with a typical steep decrease of the cell mass due to the rupture of the plasma membrane or cell fragmentation. This phenomenon is bona fide universal for all cell types and was also observed in target cells after contact with a cytotoxic T-cells 34 . Plasma membrane rupture can be otherwise measured by quantification of the release of intracellular enzymes from the cell into the cell culture medium. Nevertheless, the enzymatic activity of these enzymes can be seriously affected by differences in the pH between intracellular environment and culture medium and by time spent in the extracellular space 3 .
Based on QPI data, we were also able to distinguish two specific subroutines of cell death. While the cell death subroutine a) (bona fide apoptosis) showed a high cell density and strong fluctuation in CDS before cell death; during the subroutine b) (bona fide necrosis) the cell density was low and CDS was relatively stable after the initial decline. These variants occupied extreme positions on the CDS versus cell density plot. Cells near the dividing line (see Fig.6b.) probably succumb to another subroutine of cell death such as pyroptosis or necroptosis, as neither staurosporine nor doxorubicin is a specific apoptosis inductor 36,37 . The a) type of cell death was characteristic by a decrease in the cell area, a gradual cell rounding and a loss of surface contact (this phenomenon was not observed in the case of staurosporine treatment) followed by membrane blebbing (known as "dance of death"); see Fig. 6 and Supplementary fig. 2 to 12, 38 . Since membrane blebbing during apoptosis results from caspase-mediated activation of ROCK I, it can be assumed that apoptosis was involved in these cases 39 . At the final stages of apoptosis, the actin cytoskeleton is degraded 40,41 . Due to the hydrophobic nature of the apoptotic bodies, they undergo the plasma membrane fusion in the culture medium. The membrane of this post-apoptotic body subsequently cracks and culminates in secondary necrosis (in our case depicted by a steep decrease of cell mass) in the absence of phagocytes 42,43 . The b) type of observed cell death presented persisting large cytoplasmic membrane blebs or multiple small blebs and cell swelling leading to the final cell membrane rupture. This type of dying cells was adherent during the whole process. It can be assumed that necrosis was involved in these cases 3 . Although several studies based on QPI detection of cell death have been published earlier, they do not include the distinguishing of specific subroutines of cell death and have failed to capture the entire process of cell death, including the early stages due to short intervals of QPI capturing 44,45,46 . From an image analysis point of view, we introduce a new powerful technique for cell tracking and segmentation, capable of robustly track cell thought the whole image sequence. We also introduced a completely new method for the detection of cell death in time-series of measured cell features using LSTM neural network. To the best of our knowledge, this is also the first study introducing CDS and cell density as a parameter typical for individual cell death subroutines.
Annotated quantitative phase image dataset (with cell tracking masks and labels of cell death timepoints and types of death) used in the manuscript is available at Zenodo repository (www.zenodo.org), https://doi.org/10.5281/zenodo.2601562