Detection and tracking of overlapping cell nuclei for large scale mitosis analyses
 Yingbo Li†^{1, 2},
 France Rose†^{1},
 Florencia di Pietro^{2},
 Xavier Morin^{2} and
 Auguste Genovesio^{1}Email author
https://doi.org/10.1186/s1285901610309
© Li et al. 2016
Received: 10 March 2016
Accepted: 9 April 2016
Published: 26 April 2016
Abstract
Background
Cell culture on printed micropatterns slides combined with automated fluorescent microscopy allows for extraction of tens of thousands of videos of small isolated growing cell clusters. The analysis of such large dataset in space and time is of great interest to the community in order to identify factors involved in cell growth, cell division or tissue formation by testing multiples conditions. However, cells growing on a micropattern tend to be tightly packed and to overlap with each other. Consequently, image analysis of those large dynamic datasets with no possible human intervention has proven impossible using state of the art automated cell detection methods.
Results
Here, we propose a fully automated image analysis approach to estimate the number, the location and the shape of each cell nucleus, in clusters at high throughput. The method is based on a robust fit of Gaussian mixture models with two and three components on each frame followed by an analysis over time of the fitting residual and two other relevant features. We use it to identify with high precision the very first frame containing three cells. This allows in our case to measure a cell division angle on each video and to construct division angle distributions for each tested condition. We demonstrate the accuracy of our method by validating it against manual annotation on about 4000 videos of cell clusters.
Conclusions
The proposed approach enables the high throughput analysis of video sequences of isolated cell clusters obtained using micropatterns. It relies only on two parameters that can be set robustly as they reduce to the average cell size and intensity.
Keywords
Background
Mitosis, the eukaryotes division, is a complex cellular process involving multiple proteins. In multicellular organisms, the precise orientation of cell divisions relative to their environment plays a crucial role in the development, growth, and homeostasis of many tissues [1]. For example, divisions within the plane of epithelial structures contribute to the expansion of the tissue surface and to the maintenance of the epithelial monolayer organization [2], while divisions perpendicular to the epithelial plane contribute to tissue stratification, binary fate decisions and regulation of stem cell pools [3, 4]. Defective control of spindle orientation may be a step in the transformation process leading to cancer [5, 6]. In vertebrate cells, multiple molecular pathways contribute to spindle orientation in response to a variety of stimuli that include intrinsic cell polarity, adhesion to the extracellular matrix, and contacts with their neighbors [1]. Remarkably, these mechanisms are shared by cells grown in a culture dish, and in vitro studies in adherent cells have contributed a lot to our current understanding of spindle orientation.
The aim of the biological study, for which the following development was set, is to identify new regulators involved in the orientation of cell division through a midthroughput RNAi screen in vitro. To this end, we have developed a specific model of oriented cell division between pairs of cells grown on adhesive micropatterned disks. The precise molecular design of this spindle orientation assay is beyond the scope of the current study and will be described elsewhere, in combination with the results of the RNAi screen (di Pietro et al. in preparation). Here, we present the image analysis approach that we designed with the aim to automatically 1) identify events of cell divisions and 2) measure their orientation relative to their neighbors. Cell culture on micropatterned surfaces is increasingly used in cell and developmental biology studies using single [7–9], pairs [10], or larger groups of cells [11, 12], owing to the possibility that micropatterning offers to control numerous parameters of the cells environment and therefore reduce intercellular variability. Hence the proposed method for the first step can be generally useful to the parallel study of any event of interest arising in a growing cluster of cells.
For low throughput microscopy image analysis, a variety of semiautomated methods were proposed and are currently largely used to detect cells [14]. By semiautomated we mean an imaging throughput that is low enough (a few images or videos) for manual intervention to help or correct the detection. An exhaustive description of those available semiautomatic methods is out of the scope of this paper. However, as soon as full automation is required because of the throughput, the number of concretely working methods shrink to a few and require the data to meet with some strong hypotheses. One of those hypotheses is that cells must contain a single nucleus [15]. Another important hypothesis that is often made is that nuclei can touch each other but should not overlap [16]. Eventually, the accurate monitoring of topological changes, that is tracking splitting objects over time, highly relies on the accuracy of the cell identification process at each time frame.
Despite fruitful debates about the capabilities of levelset methods to uncover the topological changes in a group of objects to detect at low throughput [17], methods currently used at high throughput for cells detection are rarely based on those approaches because of their lack of robustness in a fully automated process. Instead, the cell detection relies most often on two steps: seed identification followed by segmentation [18–20]. The identification step consists in defining a seed for each object and the segmentation step consists in applying a region growing algorithm initialized by those seeds to uncover objects boundaries. An example of naive approach to seed detection is the local maxima detection after smoothing, which is heavily used at high throughput because of its simplicity, its speed and its robustness for many cell based applications. Regarding the detection step, seeded watershed and coupled explicit or implicit active contours can be used [21–24]. The former methods are currently common practice and proved to be very efficient in detecting millions of regular cells in monolayer where nuclei do not overlap [25] while the later are more rarely seen in practice because of their inherent instability. However the whole process depends primarily on the identification step. That is, the results tend to significantly degrade when nuclei overlap with one another and that seed cannot be correctly identified (see Fig. 3). This is precisely the problem we ran into while using micropatterns.
In the literature, those small fibronectin patterns have mostly been used for experiments with a single cell per pattern (a few exceptions with two cells or more do exist but the pattern makes the position of cells obvious and non overlapping [7, 10]). Moreover, most of the studies were not dynamic and focused on getting reproducible cell shape in order to quantify cytoskeleton organization [26]. Therefore, with a few exceptions, tracking cells on single micropatterns has not yet been an issue using this technology.
In our experiment, the chosen pattern is a disk and the number of cells growing onto them is variable and unknown. Furthermore, the pattern introduces physical constraints that tend to pack cells together as they are dividing, making their individual detection or even a simple counting often difficult (see Fig. 3). Indeed when more than two cells are present on a pattern, their shape differ from cells duplicating freely on an unbounded fibronectin slide. Consequently, nuclei shape and distances between nuclei are impacted. Furthermore, when clusters contain three or more cells, they often overlap with each other, making the detection intractable with previously cited methods. We therefore had to propose a new way to extract information from those packed clusters of cells.
In order to detect in each sequence the first frame showing three cells, our approach consisted in modeling the cell cluster by Gaussian Mixture Models. Hence, a selection process based on the sequence would allow us to determine the number of cells and their positions at each frame. Since the event we were looking for in our study was the second mitosis (that is when one of the two cells divides in a cluster of two cells only), we proposed to fit two hypothesis models to the cell cluster at each time frame: a 2component and a 3component 2D Gaussian mixture models (GMM). Fitting a GMM to count and detect biological objects in microscopy images was proposed in the past mostly to model small fluorescent spots or on static images. Thomann et al. [27] used a 3D Gaussian model to approach the point spread function and detect the number of spots reaching superresolution. A χ ^{2} test was then used to choose the right number of Gaussians in the Gaussian mixture. However, the number of degrees of freedom of the χ ^{2} test was defined as the number of pixels lying on the object (a few in the case of spots) which would be unrealistic in our case. Other methods are based on mutual information [28] or are dedicated to mitosis detection in histopathology images [29] but they gave poor results on our data because the cells are more densely packed on micropatterns. However, a close approach was proposed in [30] where numerous cells are tracked in 3D using GMM. The difference with our approach lies in the fact that because the throughput is much higher in our case, images could not be acquired in 3D. Therefore, unlike in 3D imaging, the view is incomplete and cells can overlap with each other and appear out of focus which are the major issues we had to deal with (see Fig. 3).
Method
Extraction of individual sequences from a video
Figure 1 shows the pattern image obtained at the beginning of the sequence acquisition. Each bright area in the pattern image is a micropattern possibly containing an individual and isolated cell cluster. We name a cell cluster a set of cells close to each other that mostly originate from a single cell. Figure 1 also shows a random frame of the video sequence of the H2BmCherry signal corresponding to the same field of view. Each condition of a screen will be made of two such acquisitions. Since each cell cluster is independent from the other, we set up a system to automatically crop a window around each micropattern over time, thus producing one video sequence per micropattern with a possible cluster on it (see Fig. 1).
In order to take into account the illumination bias (on Fig. 1, intensity at the center of the image is brighter than around the borders) we applied an adaptive equalization of the histogram [31]. Once corrected, the pattern image is fairly easy to segment and a smoothing followed by a cropping around local maxima was sufficient to obtain hundreds of cropped movies, each containing one micropattern location as shown by Fig. 1. From this point, those movies could be analysed independently with the following proposed method.
Characterization of cell nuclei by Gaussian mixture model
GMM as a cell cluster model
In theory, whatever the signal, more components in a GMM leads to a better reconstruction. It is therefore not possible to directly compare the fitting residuals obtained by the two models as the 3component model would always show a lower error. This model selection issue was discussed in general in the litterature and universal criteria for model selection were proposed in the past as the Akaike Information Criterium (AIC) [32] or the Bayesian Information Criterium (BIC) [33]. Our experience using those criteria independently at each time frame of the sequence led to a totally erroneous identification of the correct cell division frame. We therefore took a different approach as we describe further. However, prior to discussion on model selection, we describe how an accurate fit of the two GMM with two and three components could be achieved at high throughput: that is, on each of the 400 frames of each of the thousand individual movies of cluster we extracted.
Fitting the model to the data
So each component is fully characterized by a set of 6 parameters P _{ k }={w,μ(x),μ(y),σ _{1},σ _{2},σ _{12}} and the concatenated set of parameters Θ _{ K }={P _{1},…,P _{ K }} fully characterizes a K components mixture. Following the two hypotheses model with two and three components we are interested in testing, we build GMMs with 12 or 18 parameters respectively. We use the Powell algorithm [34] to minimize the leastsquare residual between a frame image I and the K component GMM image model M _{ K } that reads: \(f_{\text {err}} = \sum _{x, y} \left [I(x,y)  M_{K}(x,y)\right ]^{2}\).
Parameters initialization
One of the main difficulties in minimizing such a residual is that given the large number of parameters (12 or 18), the convergence toward the global minimum is not systematic. In order to ease this convergence, it is therefore crucial to set the initial parameters with values close to the optimal solution.
For the first image of the sequence, we take advantage of the fact that an average nucleus diameter \(\bar {d}_{\text {nuc}}\) and intensity \(\bar {w}_{\text {nuc}}\) can be easily estimated from the data. As \(\bar {d}_{\text {nuc}}\) can be modeled as the Full Width at Half Maximum [35], we first define a 2D Gaussian kernel with \(\bar {\sigma }_{\text {nuc}}=\bar {d}_{\text {nuc}}/(2\sqrt {2ln(2)})\). Local maximas are then detected on an image smoothed by this kernel and limited by a foreground defined with the Otsu method [36]. Those local maximas are then used as initial values for Gaussian component locations. If the number of detected maxima is lower than the number of components of the model (e.g. when cells overlap), then additional random locations on the foreground are added. The intensity \(\bar {w}_{\text {nuc}}\) is directly used to initialize w. The remaining parameters σ _{1}, σ _{2} and σ _{12} are initialized with median values of a set of previously fitted GMM components with random initialization.
For the rest of the frames in the sequence, parameters are initialized with values obtained from the fitting at previous frames and from observations obtained from the current frame. In two consecutive frames with no mitotic event (that is in the large majority of the cases), the position, the intensity and the shape of the cells are not supposed to change much given the time interval between video frames (in video duration of 7 min). Therefore, the parameters μ and w could be initialized on the next frame by the values obtained for the same parameter at the last frame. This would read \(\hat {\boldsymbol {\mu }}_{t}=\boldsymbol {\mu }_{t1}\) and \(\hat {w}_{t}=w_{t1}\). However, in the case where mitosis happens, the location and the intensity of some of the cells suddenly change. To take into account this event, local maxima of the image are also precomputed on each image and the locations (resp. the intensity) of each component are initialized by a value half way between the location (resp. the intensity) obtained at the previous frame and the location (resp. the intensity) of the closest local maxima possibly detected on the current frame. This reads \(\hat {\mu }_{t}=(\mu _{t1} + {\mu ^{D}_{t}})/2\) and \(\hat {w}_{t}=(w_{t1} + I({\mu ^{D}_{t}}))/2\) where \({\mu ^{D}_{t}}\) is the location of the closest detected maxima on frame t. This simple method ensures that the fitting process will be initialized a priori as close as possible from the optimal solution while it is not known if a mitotic event occurs or not.
Furthermore, we observed that while the shape of a nucleus is not changing much between two consecutive frames (except at a mitotic event time), on the contrary its orientation is quite dynamic (cells are often rotating). Therefore, we decided to uncouple the shape and the orientation of each component at each time frame in order to properly initialize the fitting process on the next frame for each of those parameters. In the formulation we use, shape and rotation are mixed into the covariance matrix. By diagonalizing the covariance matrix S _{ k,t−1} of each component k resulting from the previous frame we obtain λ _{1} and λ _{2} the eigenvalues corresponding respectively to the length of the major and the minor axes of the ellipse and the corresponding eigenvectors v _{ 1 } and v _{ 2 } from which the angle of the ellipse’s major axis can be computed: \(\theta _{k,t1}=\arctan (v_{1}(y)/v_{1}(x))\).
Constraints to ensure convergence
As our model includes 12 parameters in the case of 2 components and 18 parameters in the case of 3 components, even with a precise initialization the fitting process may diverge (e.g. one component may easily collapse or move outside the frame). We enforced the convergence by adding penalty terms to our error function.
Note that each of those additional constraints prevents the optimization process to move toward absurd values by artificially increasing the total error outside an acceptable range. Therefore, they drastically modify the objective function outside an acceptable range of parameter values while they preserve the function within this range. The consequence is that they do not modify significantly the minimum of the function.
Time features computed from the GMMs
At this stage, large sets of data can be fully automatically processed by extracting all single pattern videos and automatically fitting a 2component GMM and a 3component GMM on each of their time frames. Two parameters only need to be set: the approximated nuclear diameter \(\bar {d}_{\text {nuc}}\) and intensity \(\bar {w}_{\text {nuc}}\). Those values can be easily recovered.
In order to identify the first frame onto which three cells can be observed (that is right at the second division) on each of those videos, we propose to compute the derivative over time of three features. Those features are the fitting error ratio between both models, the minimum distance between the three component centers and the variance of intensity between the closest component centers. None of those require any parameter and they are described below.
F _{1}: fitting error ratio
F _{2}: distance between the closest components
As shown in Fig. 6 c, the distance between the two closest centers in the 3component model \(F_{2}(t)=\min \{\\mu _{i}(t)\mu _{j}(t)\_{2}, \forall (i,j) \in \{1,2,3\}^{2}, i\neq j\}\) becomes much larger when the mitotic event of interest happens. This is because when a 3component GMM is used to model two cells, one of the cells ends up being modeled by two components and therefore shows two very close centers. However, when one of the cells splits into two, the 3component GMM correctly models the cluster, and each component matches a single cell. Consequently, the minimum distance between any two centers suddenly increases.
F _{3}: variance of intensity between the closest components
Along with the distance between the closest centers, the intensity variation of the pixels between those two closest centers also provides information. Indeed, if the variance is high, it denotes that both foreground and background pixels were considered in the calculation, while if the variance is low, it means that only foreground pixels were used. Therefore, this feature tends to measure whether or not the two closest components of a 3component model are separated by some background or not and therefore if they model or not the same cell. The feature F _{3} over time for an example cluster can be seen in Fig. 6 d.
Identification of the division time of interest
In order to detect sudden changes over time using the features described above, we compute their derivatives. Hence we search for a sudden peak in those features’ derivatives (see Fig. 7). In practice, there is a large variability of events we have to deal with when processing hundreds of videos of that kind. Using those three features simultaneously increases the ability of the approach to detect the division time of interest. We show on an example (see Fig. 7) and on a larger study (data not shown) that using the product of those features’ derivatives over time allows to extract this event with a better accuracy than using only one or two of them.
Computation of the division angle
The method described above enables detection of the time of the first anaphase image on a movie with two cells. Detecting the right time is essential in order to measure the correct angle, because cells move and rotate from one time point to the next, especially when there are more than two cells on a pattern. Moreover cells can die or image acquisition can have started when three or more cells were already on the pattern. In those last cases, the error model would not fit. This allows us to exclude sequences where a division angle cannot be measured.
Once the right image is selected, the parameters of the fitting give the positions μ _{ k } and sizes S _{ k } of the corresponding underlying nuclei (see Fig. 5). From those measures, nuclei issued from the last division are chosen to be the two smallest Gaussian objects. Using those, the extraction of the angle described by the Fig. 2 is straightforward.
Results
To our knowledge, no available software could provide a full solution dedicated to the type of assay we propose (i.e. an automated tracking of overlapping cells on thousands of individual movies). Therefore, it was not possible to strictly compare our approach to another possibly existing method. However, a freely available software program that could have matched our need was Cellprofiler [25] because in principle, it enables the tracking of cells over time in a large set of image sequences, using the Hungarian algorithm. However, cell detection in Cellprofiler is based on a maxima detection followed by a seeded Watershed segmentation so we expected it not to perform well in detecting overlapping and dividing cells. In accordance, the results we obtained were dramatically poor. A quantitative comparison here would be meaningless, as almost no mitotic event could be identified this way. However, it was possible to compare our automated approach to a large set of data (4000 sequences) that has been exhaustively analyzed by a human tester, and considered thereafter as the “ground truth” for our method.
Experimental data
The dataset we created to validate the method is made of several videos of hundreds of cell divisions under three biological conditions. Those conditions are as follows: as a negative control, we used an siRNA targeting Cyclophilin, which is proposed as one of several standard negative controls by GEDharmacon in their ONtarget+ human siRNA libraries. LGN (LeucineGlycineAsparagine repeat protein) was used as a positive control: LGN is an adaptor molecule involved in the localized recruitment of dynein motor complexes at the cell membrane, which direct forces exerted on astral microtubules. LGN is a central regulator of spindle orientation in many animal cell types (reviewed in [1]). Our pairedcell assay (di Pietro et al, in preparation) is designed to specifically depend on the “LGNcomplex” molecular cascade. siRNA against LGN therefore significantly alters spindle orientation in this assay. The third siRNA targets p62, which is part of the dynactin molecular complex and as such a candidate for the regulation of dynein activity and spindle orientation. It is therefore expected to differ significantly from the negative control, and to yield results similar (but not necessarily identical) to the positive control. As we aim at using this method on a large set of conditions for which we will have a variation in the number of patterns we will obtain per condition, each condition for our test was respectively made of 4, 5 and 6 videos covering each field of view. Each field of view was made of about 250 frames of size 2048×2048 pixels. Figure 1 shows a frame of such a video captured by a wide field fluorescence microscope and containing about 280 patterns (excluding those touching the borders).
The fitting process is the most timeconsuming step of the analysis. It takes 2 seconds for each image on a PC with Intel Core i74800MQ 2.7 GHz with 16 GB RAM. As analyzing one sequence requires to test two models on 250 frames, the overall process for one cluster containing cells takes 25 min. However, we used a computing cluster to process hundreds of cell clusters simultaneously.
Precision of the event detection
Pattern count on raw data (A and B), after a rough preprocessing step (C, D and E) to discard empty patterns or patterns containing obviously more than 2 cells on the first frame of the sequence. Eventually, the number of pattern where a transition from two to three cell was detected automatically (F) or manually (G). While a lower number of event is selected automatically, more than 80 % of the events selected automatically were also part of the manual selection whatever the experimental condition
Cyclophilin  LGN  p62  

siRNA  siRNA  siRNA  
A) Total number of fields of view (=large videos)  4  5  6 
B) Total number of micropatterns (=single cluster videos)  1116  1393  1668 
C) Micropatterns with no cells (excluded)  400  607  719 
D) Micropatterns with too many cells at time 0 (excluded)  51  64  120 
E) Micropatterns with a low number of cells at time 0  665  722  829 
F) Events selected automatically from (E)  122  135  97 
G) Events selected manually  184  197  227 
H) Proportion of (F) also in (G)  85.9 %  82.5 %  81.8 % 
Accuracy of the angle distributions
The angle samples obtained from a manual selection or an automated analysis are similar: the null hypothesis of a KS test (“both samples come from the same angle distribution”) cannot be rejected at a 10 % significance level
Manual (M)  Automated (A)  (M) vs (A)  

median  stdev  median  stdev  KStest pvalue  
Cyclophilin siRNA  23  30  31  31  0.229 
LGN siRNA  71  26  72  25  0.620 
p62 siRNA  62  28  54  30  0.246 
The angle distributions obtained from a manual selection and an automated analysis reach similar conclusions: the null hypothesis of a KS test (“both samples come from the same angle distribution”) is rejected for any two couple of conditions at a 10 % significance level
KStest pvalue  Cyclophilin siRNA  Cyclophilin siRNA  LGN siRNA 

vs  vs  vs  
LGN siRNA  p62 siRNA  p62 siRNA  
Manual  3.22e15  9.47e11  5.64e02 
Automated  2.09e07  1.65e02  2.72e04 
Discussion
In order to factor out some possible issues that may have occured we performed additional tests.
Possible bias induced by the statistical test
Possible bias produced by the pattern
Conclusion
In this paper we proposed a high throughput method to automatically detect the transition of a cell cluster from two to three cells in thousands of videos. The proposed algorithm performs a robust implicit tracking of cells even when they are packed, overlap or are not clearly distinguishable. The approach is based on a robust fitting of twodimensional Gaussian mixture models with two and three components on each frame of the video. We showed that the derivatives of the residual ratio between the two models, the distance between the two closest centers and the variation of intensity between them was sufficient to detect the exact time of an event of interest. We showed, using three independent conditions, that the distributions of angles obtained automatically were very similar to those obtained through a very tedious manual annotation that took several days and would be impossible to concretely extend to hundreds of conditions. While the focus of our study was to monitor the division orientation, the same principle can easily be extended to many other questions through the calculation of other features obtained using the proposed approach.
Availability of data and materials

Project name: livespin

Project home page: https://github.com/biocompibens/livespin

Archived version: https://github.com/biocompibens/livespin.git

Operating system(s): Platform independent

Programming language: Python

Other requirements: Python 2.7

License: GNU GPL 3.0
Declarations
Acknowledgements
We thank ANR10LABX54 MEMO LIFE, ANR11IDEX000102 PSL* Research University, ANR12LIVESPIN and ARC for funding, Felipe Delestro Matos for helping with the figure formating and Mary Ann Letellier for helpful comments on the manuscript. We acknowledge Léo Valon, for helpful discussions and his initial work at testing various methods to perform an automated analysis.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
Authors’ Affiliations
References
 Morin X, Bellaïche Y. Mitotic spindle orientation in asymmetric and symmetric cell divisions during animal development. Dev Cell. 2011; 21(1):102–19. doi:10.1016/j.devcel.2011.06.012.View ArticlePubMedGoogle Scholar
 Fleming ES, Zajac M, Moschenross DM, Montrose DC, Rosenberg DW, Cowan AE, Tirnauer JS. Planar spindle orientation and asymmetric cytokinesis in the mouse small intestine. J Histochem Cytochem. 2007; 55(11):1173–1180.View ArticlePubMedPubMed CentralGoogle Scholar
 Williams SE, Beronja S, Pasolli HA, Fuchs E. Asymmetric cell divisions promote Notchdependent epidermal differentiation. Nature. 2011; 470(7334):353–8.View ArticlePubMedPubMed CentralGoogle Scholar
 Quyn AJ, Appleton PL, Carey FA, Steele RJ, Barker N, Clevers H, Ridgway RA, Sansom OJ, Nathke IS. Spindle orientation bias in gut epithelial stem cell compartments is lost in precancerous tissue. Cell Stem Cell. 2010; 6(2):175–81.View ArticlePubMedGoogle Scholar
 Noatynska A, Gotta M, Meraldi P. Mitotic spindle (dis) orientation and disease: cause or consequence?J Cell Biol. 2012; 199(7):1025–1035.View ArticlePubMedPubMed CentralGoogle Scholar
 Pease JC, Tirnauer JS. Mitotic spindle misorientation in cancer–out of alignment and into the fire. J Cell Sci. 2011; 124(Pt 7):1007–1016.View ArticlePubMedGoogle Scholar
 Thery M, Racine V, Pepin A, Piel M, Chen Y, Sibarita JB, Bornens M. The extracellular matrix guides the orientation of the cell division axis. Nat Cell Biol. 2005; 7(10):947–53.View ArticlePubMedGoogle Scholar
 Saadaoui M, Machicoane M, di Pietro F, Etoc F, Echard A, Morin X. Dlg1 controls planar spindle orientation in the neuroepithelium through direct interaction with LGN. J Cell Biol. 2014; 206(6):707–17.View ArticlePubMedPubMed CentralGoogle Scholar
 Thery M, JimenezDalmaroni A, Racine V, Bornens M, Julicher F. Experimental and theoretical study of mitotic spindle orientation. Nature. 2007; 447(7143):493–6.View ArticlePubMedGoogle Scholar
 Tseng Q, DucheminPelletier E, Deshiere A, Balland M, Guillou H, Filhol O, Théry M. Spatial organization of the extracellular matrix regulates cell–cell junction positioning. Proc Natl Acad Sci. 2012; 109(5):1506–1511.View ArticlePubMedPubMed CentralGoogle Scholar
 Dupin I, Sakamoto Y, EtienneManneville S. Cytoplasmic intermediate filaments mediate actindriven positioning of the nucleus. J Cell Sci. 2011; 124(Pt 6):865–72.View ArticlePubMedGoogle Scholar
 Warmflash A, Sorre B, Etoc F, Siggia ED, Brivanlou AH. A method to recapitulate early embryonic spatial patterning in human embryonic stem cells. Nat Methods. 2014; 11(8):847–54.View ArticlePubMedPubMed CentralGoogle Scholar
 Degot S, Auzan M, Chapuis V, Béghin A, Chadeyras A, Nelep C, CalvoMuñoz ML, Young J, Chatelain F, Fuchs A. Improved visualization and quantitative analysis of drug effects using micropatterned cells. J Vis Exp.2010;46. doi:10.3791/2514.
 Meijering E. Cell segmentation: 50 years down the road [life sciences]. IEEE Signal Proc Mag. 2012; 29(5):140–5.View ArticleGoogle Scholar
 Wählby C, Lindblad J, Vondrus M, Bengtsson E, Björkesten L. Algorithms for cytoplasm segmentation of fluorescence labelled cells. Anal Cell Pathol. 2002; 24(23):101–11.View ArticlePubMedPubMed CentralGoogle Scholar
 Mathew B, Schmitz A, MuñozDescalzo S, Ansari N, Pampaloni F, Stelzer EH, Fischer SC. Robust and automated threedimensional segmentation of densely packed cell nuclei in different biological specimens with linesofsight decomposition. BMC Bioinforma. 2015; 16(1):187.View ArticleGoogle Scholar
 McInemey T, Terzopoulos D. Topology adaptive deformable surfaces for medical image volume segmentation. IEEE Trans Med Imaging. 1999; 18(10):840–50.View ArticleGoogle Scholar
 Gauch JM. Image segmentation and analysis via multiscale gradient watershed hierarchies. IEEE Trans Image Process. 1999; 8(1):69–79.View ArticlePubMedGoogle Scholar
 Meyer F, Beucher S. Morphological segmentation. J Vis Commun Image Represent. 1990; 1(1):21–46.View ArticleGoogle Scholar
 Yang F, Mackey MA, Ianzini F, Gallardo G, Sonka M. Cell segmentation, tracking, and mitosis detection using temporal context. Med Image Comput Comput Assist Interv. 2005; 8(Pt 1):302–9.PubMedGoogle Scholar
 Lindblad J, Wählby C, Bengtsson E, Zaltsman A. Image analysis for automatic segmentation of cytoplasms and classification of rac1 activation. Cytometry Part A. 2004; 57(1):22–33.View ArticleGoogle Scholar
 Jones TR, Carpenter A, Golland P. Voronoibased segmentation of cells on image manifolds. In: Proceeding CVBIA’05 Proceedings of the First international conference on Computer Vision for Biomedical Image Applications: 2005. p. 535–43.Google Scholar
 Zimmer C, OlivoMarin JC. Coupled parametric active contours. IEEE Trans Pattern Anal Mach Intell. 2005; 27(11):1838–1842. doi:10.1109/TPAMI.2005.214.View ArticlePubMedGoogle Scholar
 Zhang B, Zimmer C, OlivoMarin JC. Tracking fluorescent cells with coupled geometric active contours. In: Biomedical Imaging: Nano to Macro, 2004. IEEE International Symposium on: 2004. Vol. 1. p. 476–9, doi:10.1109/ISBI.2004.1398578.
 Kamentsky L, Jones TR, Fraser A, Bray MA, Logan DJ, Madden KL, Ljosa V, Rueden C, Eliceiri KW, Carpenter AE. Improved structure, function and compatibility for cellprofiler: modular highthroughput image analysis software. Bioinformatics. 2011; 27(8):1179–1180.View ArticlePubMedPubMed CentralGoogle Scholar
 Duong T, Goud B, Schauer K. Closedform densitybased framework for automatic detection of cellular morphology changes. Proc Natl Acad Sci. 2012; 109(22):8382–387.View ArticlePubMedPubMed CentralGoogle Scholar
 Thomann D, Rines DR, Sorger PK, Danuser G. Automatic fluorescent tag detection in 3D with superresolution: application to the analysis of chromosome movement. J Microsc. 2002; 208(Pt 1):49–64.View ArticlePubMedGoogle Scholar
 Yang ZR, Zwolinski MSM. Mutual Information Theory for Adaptive Mixture Models. IEEE Trans Pattern Anal Mach Intell. 2001; 23(4):396–403.View ArticleGoogle Scholar
 Khan AM, ElDaly H, Rajpoot N. A gammagaussian mixture model for detection of mitotic cells in breast cancer histopathology images. J Pathol Inform. 2013; 4:11. doi:10.4103/21533539.112696.PubMedPubMed CentralGoogle Scholar
 Amat F, Lemon W, Mossing DP, McDole K, Wan Y, Branson K, Myers EW, Keller PJ. Fast, accurate reconstruction of cell lineages from largescale fluorescence microscopy data. Nat Methods. 2014; 11(9):951–8.View ArticlePubMedGoogle Scholar
 Stark JA. Adaptive image contrast enhancement using generalizations of histogram equalization. IEEE Trans Image Process. 2000; 9(5):889–96.View ArticlePubMedGoogle Scholar
 Akaike H. A new look at the statistical model identification. IEEE Trans Autom Control. 1974; 19(6):716–23. doi:10.1109/TAC.1974.1100705.View ArticleGoogle Scholar
 Schwarz G. Estimating the dimension of a model. Ann Statist. 1978; 6(2):461–4. doi:10.1214/aos/1176344136.View ArticleGoogle Scholar
 Powell MJ. An efficient method for finding the minimum of a function of several variables without calculating derivatives. Comput J. 1964; 7(2):155–162.View ArticleGoogle Scholar
 Weisstein EW. Full width at half maximum. MathWorld–A Wolfram Web Resource. http://mathworld.wolfram.com/FullWidthatHalfMaximum.html.
 Otsu N. A threshold selection method from Graylevel histograms. IEEE Trans Syst Man Cybern. 1979; 9(1):62–6. doi:10.1109/tsmc.1979.4310076.View ArticleGoogle Scholar