 Methodology article
 Open access
 Published:
Cell segmentation methods for labelfree contrast microscopy: review and comprehensive comparison
BMC Bioinformatics volume 20, Article number: 360 (2019)
Abstract
Background
Because of its nondestructive nature, labelfree imaging is an important strategy for studying biological processes. However, routine microscopic techniques like phase contrast or DIC suffer from shadowcast artifacts making automatic segmentation challenging. The aim of this study was to compare the segmentation efficacy of published steps of segmentation workflow (image reconstruction, foreground segmentation, cell detection (seedpoint extraction) and cell (instance) segmentation) on a dataset of the same cells from multiple contrast microscopic modalities.
Results
We built a collection of routines aimed at image segmentation of viable adherent cells grown on the culture dish acquired by phase contrast, differential interference contrast, Hoffman modulation contrast and quantitative phase imaging, and we performed a comprehensive comparison of available segmentation methods applicable for labelfree data. We demonstrated that it is crucial to perform the image reconstruction step, enabling the use of segmentation methods originally not applicable on labelfree images. Further we compared foreground segmentation methods (thresholding, featureextraction, levelset, graphcut, learningbased), seedpoint extraction methods (Laplacian of Gaussians, radial symmetry and distance transform, iterative radial voting, maximally stable extremal region and learningbased) and single cell segmentation methods. We validated suitable set of methods for each microscopy modality and published them online.
Conclusions
We demonstrate that image reconstruction step allows the use of segmentation methods not originally intended for labelfree imaging. In addition to the comprehensive comparison of methods, raw and reconstructed annotated data and Matlab codes are provided.
Background
Microscopy has been an important technique for studying biology for decades. Accordingly, fluorescence microscopy has an irreplaceable role in analyzing cellular processes because of the possibility to study the functional processes and morphological aspects of living cells. However, fluorescence labeling also brings a number of disadvantages. These include photobleaching, difficult signal reproducibility, and inevitable phototoxicity (which results not only from staining techniques but also from transfection) [1]. Labelfree microscopy techniques are the most common techniques for live cell imaging thanks to its nondestructive nature, however, due to the transparent nature of cells, methods of contrast enhancement based on phase information are required.
The downside of contrast enhancement is an introduction of artifacts; Phase contrast (PC) images contain halo and shadeoff, differential image contrast (DIC) and Hoffman Modulation Contrast (HMC) introduce nonuniform shadowcast artifacts (3Dlike topographical appearance). Although various segmentation procedures have been developed to suppress these artifacts, a segmentation is still challenging.
On the other hand, quantitative phase imaging (QPI), provides artifactfree images of sufficient contrast. Although there are no standardized methods for the segmentation of QPIbased images, fundamental methods for segmentation of artifactfree images (e.g. from fluorescence microscopy) will be utilized.
In this review, we describe and compare relevant methods of the image processing pipeline in order to find the most appropriate combination of particular methods for most common labelfree microscopic techniques (PC, DIC, HMC and QPI). Our aim is to evaluate and discuss the influence of the commonly used methods for microscopic image reconstruction, foregroundbackground segmentation, seedpoint extraction and cell segmentation. We used real samples  viable, nonstained adherent prostatic cell lines and captured identical fields of view and cells manually segmented by a biologist. Compared to microscopic organisms like yeast or bacteria, adherent cells are morphologically distinctly heterogeneous and in labelfree microscopy, the segmentation is therefore still a challenge. We will use the most common imaging modalities used by biologist and we will provide a recommendation of image processing pipeline steps for particular microscopic technique.
The segmentation strategies tested herein are selected to provide the most heterogeneous overview of recent state of the art excluding the simplest and outdated methods (e.g. simple connected component detection, ultimate erosion, distance transform without hmaxima etc.). Deeplearning strategies are intentionally not included due to their distinct differences, high demands on training data and the range of possible settings (training hyperparameters, network architecture, etc.).
Results
In the paragraphs below we provide a detailed summary of each image processing step from the pipeline (see Fig. 1), followed by short description of achieved results. We start with description of “allinone” tools and continue with image reconstruction, foregroundbackground segmentation, cell detection and final single cell segmentation (i.e. instance segmentation).
Due to the large number of tested methods and approaches, we have decided to introduce a specific designation of the methods. We used prefix in order to refer to image reconstruction (‘r’), foregroundbackground segmentation (‘s’) and cell detection (‘d’) and finally to allinone tools (‘aio’). The list of these designations, number of parameters to be adjusted in these methods and computational demands are provided in Table 1.
“Allinone” tools
First, we performed an analysis with the available commercial and freeware “allinone” tools including FARSIGHT [2], CellX [3], Fogbank [4], FastER [5], CellTracer [6], SuperSegger [7], CellSerpent [8], CellStar [9], CellProfiler [10] and QPHASE’ Dry mass guided watershed (DMGW) [11]. As shown in Table 2 the only algorithm providing usable segmentation results for raw images is Fogbank, which is designed to be an universal and easy to set segmentation tool. Very similar results were provided by CellProfiler, which is easy to use tool allowing to crate complete cell analysis pipelines, however, it works sufficiently only for reconstructed images. The QPI’ dedicated DMGW provided exceptional results, but for this microscopic technique only. The remaining methods did not provide satisfactory results on label free data; FastER, although userfriendly, failed because of the nature of its maximally stable extremal region (MSER) detector. FARSIGHT failed with the automatic threshold during foreground segmentation. CellX failed in both the cell detection with gradientbased Hough transform and in the membrane pattern detection because of indistinct cell borders. The remaining segmentation algorithms  CellStar, SuperSegger, CellSerpent  were completely unsuitable for labelfree nonround adherent cells with Dice coefficient <0.1 and thus are not listed in Table 2 and Fig. 4.
Because of the low segmentation performance of the examined “allinone” methods, we decided to divide the segmentation procedure into four steps  (1) image reconstruction (2) background segmentation, (3) cell detection (seed expansion) and (4) segmentation tailored to the specific properties of individual microscopic techniques (see Fig. 1).
Image reconstruction
As shown, the performance of most “allinone” methods is limited for labelfree data, in particular due to the presence of contrastenhancing artifacts in microscopic images. Image reconstruction was therefore employed to reduce such artifacts. Methods by Koos [12] and Yin [13] (further abbreviated rDICKoos and rDICYin, respectively) were used for DIC and HMC images. Images of PC microscopy were reconstructed by TopHat filter involving algorithm by the Dewan [16] (rPCTopHat), or Yin method (rPCYin) [14].
Generally, following conclusions apply for image reconstructions:

No distinctive differences in image reconstruction efficacy were observed between the microscopic methods apart from QPI, as shown in Fig. 2 (described by area under curve, AUC, see Methods for details).

The AUC of QPI was distinctly higher with values near 0.99

Computationally moredemanding methods (rDICKoos and rPCYin) perform better except for relatively simple rPCTopHat, which provides similar results

Probability maps generated by sWeka or sIllastik can be used like reconstructions in later segmentation steps. The advantage of this approach is the absence of the need to optimize parameters.
DIC and HMC reconstructions
With regard to the morphology of reconstructed images, rDICKoos provides a detailed structure of the cells with distinctive borders from the background. For rDICYin [13], details of the reconstructed cells are more blurred and uneven background with the dark halos around the cells (see Fig. 2) complicating the following segmentation. As a result, AUC of rDICYin was distinctly lower as compared with the others.
Both rDICKoos [12] and rDICYin [13] methods work on the principle of minimizing their defined energy function. The main difference is that betterperforming Koos [12] uses l1norm (instead of l2) for sparse regularization term. Yin’s l2norm, on the other hand, enables derivation of closed form solution, which is much simpler and thus faster to compute. Time needed for the reconstruction is dramatically different  2.1 s, 36.6 min, 13.1 min and 0.17 s for rDICKoos, rDICYin, rPCKoos and rPCTopHat, respectively. rDICKoos also introduces a parameter for the number of iterations, which is however insensitive within the tested range.
Although these methods were not designed for use on HMC images, the same conclusions also apply for the reconstruction of those images, which showed only slightly worse results. The results of reconstruction accuracy can be seen in Fig. 2. Combinations of the bestperforming parameters are listed in the Additional file 1.
Phase contrast reconstruction
From the perspective of cellular morphology of reconstructed images, rPCTopHat creates artifacts between closely located cells with the borders precisely distinguishable. Reconstruction based on rPCYin [14] causes an even background without observable artifacts around the cells, however cell borders are missing and mitotic cells are not properly reconstructed (see Fig. 2).
The optimization of the PSF parameters of rPCYin reconstruction is problematic. The PSF parameters of a particular microscope are not always listed or known. Searching for these parameters with optimization proved to be complicated. Because the optimizing function is not smooth and contains many local local extrema, the result changes significantly and chaotically even with a small change of parameters or, at the same time, combinations of parameter settings give very similar (near optimal) results.
Regarding the computational times, the rPCYin reconstruction works very similarly as the rDICKoos approach for DIC, with similar computational difficulties. The result of a simple tophat filter unexpectedly turned out to be comparable to the complex and computationally difficult rPCYin method. For the reconstruction performance see Fig. 2, for optimal parameter setting see the Additional file 1.
Foregroundbackground segmentation
In the next step of the workflow, the image foreground (cells) was segmented from the image background. Both unprocessed and reconstructed images were used. Following strategies were used for the foregroundbackground segmentation: (a) Thresholdingbased methods: simple threshold (sST), automatic threshold based on Otsu et al. [17] (sOtsu), and Poisson distributionbased treshhold (sPT) [2], (b) featureextracting strategies: empirical gradient threshold (sEGT) [18] and approaches specific for PC microscopy by Juneau et al. (sPCJuneau) [19], Jaccard et al. (sPCPhantast) [21], and Topman (sPCTopman) [20]), (c) LevelSetbased methods: Castelles et al. [22] (sLSCaselles), and ChanVese et al. [23] (sLSChanVese), (d) Graphcut [24], and (e) Learningbased Ilastik [25], and Trainable Weka Segmentation [26].
Based on the obtained results, this step can be considered the least problematic in segmentation, with the following general findings:

Wellperforming methods (e.g. sWeka, sIllastik, sLSCaselles,sEGT, sPCJuneau) are robust enough to work even on unreconstructed data.

Image reconstruction improves foregroundbackground segmentation efficacy and once reconstructed, there are no distinct differences in segmentation efficacy between microscopic techniques

QPI performs dramatically better even unreconstructed

Learningbased methods (sWeka and sIlastik) perform better by a few units of percents. Its performance can further be improved with GraphCut.

More timeconsuming methods (sLSCaselles, sLSChanVese, sGraphCut, sWeka, sIlastik) does not necesarily provide better results. For detailed results, see chapters below and Fig. 3.
Thresholdbased approaches
The Simple threshold (sST) provides better results than automatic thresholding techniques assuming Poisson distribution (sPT) or Otsu method (sOtsu). The potential of these automatic techniques lies in the segmentation of images, where optimal threshold value varies between the images. However, this is not necessary for QPI images (constant background value increases success of sST) and for reconstructed images with background removal (background values are close to zero, so the histogram cannot be properly fitted with Gaussian or Poison distribution, see Table 2). There are not any parameters to optimize for sOtsu and sPT methods, which is the main advantage. The results of thresholding could be potentially improved by morphological adjustments. Regarding the computational times, these are the simplest and thus the fastest possible methods, which are listed mainly to provide basic idea about the segmentability of our data.
Featureextractionbased approaches
The featurebased approaches  sEGT, sPCTopman, sPCPhantast and sPCJuneau are all mainly based on the extraction of some feature image, which is then thresholded and morphologically modified. Because of feature thresholding strategies, the segmentation is possible without the image reconstruction. Thus these methods are among the most straightforward approaches to extract and threshold some local features (e.g. absolute value of gradient or local standard deviation).
All these methods can be easily adjusted, have the same number of parameters and the segmentation performance is very similar (see Table 1) with slightly betterperforming sEGT. Compared to the other featureextractionbased methods, sEGT includes elimination of small holes.
The performance of featureextraction methods is techniquedependent with the highest scores for DIC and QPI and the lowest (but still high) for PC. This is mostly due to halos in PC; although sPCTopman and sPCPhantast are extended by the elimination of PC artifact regions, sPCTopman have even worse results than sEGT or sPCJuneau and sPCPhantast leads to a slight improvement only for a cost of more parameters to be set.
From feature thresholding methods, sEGT was shown to be the best with only a small number of parameters and great versatility. Because of its percentile based threshold, it can be used even with a default setting, which achieves e.g. 0.84 Dice coefficient value for QPI. Compared to thresholdbased methods, featureextraction strategies perform approximately 10% better. Considering the computational demands, these methods are very simple and fast  comparable to simple thresholding.
Levelsetbased approaches
Both sLSCaselles [22] and sLSChanVese [23] active contours tended to shrink too much, which was compensated by setting additional force to negative sign, which leads to a tendency of the contour to grow. The increase of the additional force leads to a better Dice coefficient value until a breaking point, after which it leads to the total divergence of the contour. Still, the value of additional force had a much greater influence than the smoothness parameter.
Compared to the abovementioned foregroundbackground segmentation strategies, the levelset based methods are relatively complicated and computationally difficult (tens of seconds vs. less than 1 s per FOV, Table 1). In their basic forms, two parameters are needed to be set. Another great disadvantage is that proper initialization is required, mainly the sLSCaselles method is very sensitive to initialization. Based on segmentation results, sLSChanVese is applicable on reconstructed images only, and does not even reach the segmentation efficacy of simple threshold results. On the other hand, sLSCaselles is applicable on raw images, but only for PC images it surpasses the otherwise much faster sEGT.
Graphcut
There is a large number of methods and modifications based on GraphCut. Herein, we tested the basic model only. When Graphcut was employed on the reconstructed images (sGraphCut), the highest Dice coefficient was obtained among nontrainable approaches except for rPCTophat, being surpassed by sLSChanVese. Nevertheless, GraphCut does not outperform simple threshold dramatically, providing roughly 2% increase in Dice coefficient and is only suitable for reconstructed data.
Regarding differences between microscopic methods, the Graphcut approach was most suitable for reconstructed DIC images, followed by PC and HMC. Regarding the computational times, this method performs similarly as the levelsetbased strategies (tens of seconds per FOV  Tables 1 and 2). Optimized values are shown in Additional file 1.
Trainable approaches
Trainable Weka segmentation (sWeka) and Ilastik (sIlastik) were employed in this step. Similarly to the featureextracting approaches, these are applicable on raw, unreconstructed data. Both sIlastik and sWeka outperformed all tested foregroundbackground segmentation methods with Dice coefficient up to 0.94 for QPI and up to 0.85 for DIC, HMC and PC.
Regardless of the imaging modality used, there was an identifiable “breakpoint” in the dependency between the area size used for learning and the segmentation efficacy after which no dramatic increase in Dice coefficient was observed, see Fig. 3. For DIC, PC, and HMC it was approx. at the size 70×70 px., for QPI, distinctly smaller area was necessary, approx. 25×25 px. These areas roughly correspond to the cell size. However, to demonstrate the theoretical maximum of this method and to compare it with Ilastik, learning from one whole FOV for DIC, HMC, and PC and from 3 FOVs for QPI was deployed (see Table 2.
Next, an effect of learning from one continuous area in one FOV, or smaller patches of same sizes from multiple FOVs was tested. On DIC data it was demonstrated that learning from multiple areas causes significant, but slight 2% increase increase in Dice coefficients.
No increase of Dice coefficient was observed when different filters were enabled apart from the set of default ones (“default” vs “all”) as well as changing of minimum/maximum sigma. This was tested with a random search approach and with the Dice coefficient varying ±0.01. Both Weka and Ilastik provide almost the same segmentation results and are identically timedemanding.
There are two parameters to be optimized: terminal weights and edge weight. Edge weight (designated as “smoothness” in the GUI, range 010) reflects a penalty for label changes in the segmentation (higher values cause smoother result).
Furthermore, probability maps generated by sWeka and Ilastik under optimal settings were exported and these maps were further segmented by GraphCut (sWekaGraphCut/sIlastikGraphCut) and optimized in a same manner as sGraphCut on reconstructed data. A slight increase of the segmentation efficacy caused the sWekaCraphCut/sIlastikCraphCut combination to be the most efficient foregroundbackground segmentation method for QPI, HMC, and PC, only being surpassed by EGT on raw DIC image data. More importantly, this was achieved without the need of the image reconstruction.
Cell detection (seedpoint extraction)
Once the foreground (cells) is separated from the background, the next step is to identify individual cells (seed points). The following strategies were used: (a) Cell shapebased, Laplacian of Gaussian (LoG) variants Peng et al. [27] (dLoGmPeng), Kong et al.[28] (dLoGmKong), Hessian Zhang et al.[30] (dLoGhZhang), generalized Kong et al. [28] (dLoGgKong), generalized Xu et al. [29] (dLoGgXu), (b) Cell shapebased, generalized radial symmetry transform [32] (dGRST), fast radial symmetry transform [31] (dFRST), (c) Qi et al.[33] radial voting (dRVQi), (d) distance transform [15] (dDTThreshold, dDTWeka), (e) Maximally Stable Extremal Region [34] (dMSER), and (f) dCellDetect [35]. Following general conclusions are applicable for this segmentation step:

Seedpoint extraction is crucial step of cell segmentation

The requirement of reconstructed images is a significant bottleneck of the seedpoint extraction

multiscale and generalized LoG are among the most robust and to some extent work also on unreconstructed data

Radial symmetry transformbased strategies perform well

Seedpoint extraction is exceptional on QPI data

Learningbased approach dCellDetect provide exceptional results on reconstructed data.
Laplacian of Gaussianbased strategies
Multiscale LoG filters (dLoGmPeng and dLoGmKong) perform similarly as generalized versions (dLoGgKong and dLoGgXu), but Hessianbased LoG (dLoGhZhang) were significantly worse in some cases. As for the traditional microscopic methods, LoG approaches enables the highest achievable segmentation efficacy. It was found out that particular combinations of reconstructionLoG filter perform better than others; an optimal reconstructionseedpoint extraction combination is rDICKoos followed by dLoGmPeng for DIC, rDICKoos plus dLoGmKong for HMC, and rPCTophat plus dLoGmPeng for PC. Moreover, there were dramatic differences in cell detection between QPI and the remaining contrastenhancing microscopic methods. On the other hand, there were no differences with Dice coefficient 0.9 for both QPI and DIC with dLOGmKong (Fig. 4).
Hessian variant dLoGhZhang achieved low segmentation efficacy on our samples of adherent cells (of various sizes) due to the use of one estimated optimal kernel size only (see Table 2). dLoGgKong originally completely fails for some modalities due to the wrong cell size estimation caused by subcellular structures, which produce higher signal then cells. This was eliminated by introducing a new σ_{min} parameter, limiting the lower scale.
Regarding the computational times, LoGbased are among faster techniques, being surpassed only by the distance transform.
Radial symmetry transformbased strategies
Compared to the computationallysimple LoGbased techniques, the dFRST [31] and generalized dGRST [32] provide better results for unreconstructed QPI images and, notably, for unreconstructed HMC and PC images. On reconstructed data, a possible application is for PC data with results very close to QPI segmentation. Nevertheless, computational times in the orders of hundreds of seconds need to be taken into account.
Radial voting
Radial voting (dRVQi) approach [33] does not achieve the results of fast LoGbased strategies for all microscopic modalities, either raw or reconstructed, while being computationally comparable to radial symmetry transformbased approaches. Thus, it is considered not suitable for such data.
Distance transform
The strong advantage of the distance transform [15] is its speed, which is the highest among other seedpoint extraction strategies. Segmentation efficacy of the tested version with optimal thresholding (dDTThreshold) is the highest among all microscopies except for PC, but image reconstruction is needed. An alternative approach is to use WEKA for binary image generation (dDTWeka), where cells are less separated than in a case of optimal threshold.
Maximally stable extremal region
Compared to the relatively consistent performance of LoG between microscopic techniques, the dMSER approach [34] is distinctly more suitable for HMC reconstructed by rDICKoos and PC reconstructed by rPCYin, where the segmentation performance as well as computational requirements are identical or similar to LoG.
CellDetect
The CellDetect approach uses [35] maximally stable extremal region for segmentation. Adherent cells in unreconstructed DIC/HMC/PC images are, however, dramatically heterogeneous structures. Thus, there are no elements registered for learning and thus the performance of CellDetect was similar to aioFastER methods. On the reconstructed data, it performs similarly as LoG or distance transformbased methods. Nevertheless, because the trainable nature of this technique, enormous computational time demands must be taken into account (up to 100fold higher than DT). Segmentation of microscopic elements of low shape heterogeneity (e.g. yeast) would profit from CellDetect significantly.
Single cell (instance) segmentation
The data which underwent reconstruction, foreground segmentation and seedpoint extraction were finally segmented by Markercontrolled watershed (MCWS) applied on distance transform or on images directly. Compared to previous steps, errors generated by this step have only minimal impact on overall segmentation quality, providing fewpixelshifts to one or other adjacent cells. The distance transform approach is more universal but, in case the cells are wellseparated, MCWSonly approach can provide better results. When compared to “allinone” segmentation strategies, the approach proposed by us provides dramatically better results except of proprietary software for quantitative phase imaging (see Table 2). With this in regard, the development of a new method which is strictly based on the nature of massdistributionQPI images could provide even better results.
Finally, it was assessed how the segmentation accuracy’s individual steps are affected by morphological aspects of cells. Following aspects were studied (Fig. 5): cellular circularity and level of contact of cells with other cells (isolated cells vs cells growing together in densely populated areas, expressed as a percentage of cellular perimeter in contact with other cells). The circularity ranged 38.2 to 63.5%, median 51.2%, (percentage of cells with a circularity 100%: 2.1%), the percentage of perimeter ranged 4.1–41.9%, median 22.0% (percentage of cells with no contact with others 21.7%). Cells with circularity ranges 0–40% and 70–100% were considered low and highcircularity cells. Regarding the degree of contact with other cells, cells whose 0–15% and 50–100% of perimeter was in contact with other cells were designated “isolated” and “growing together”, respectively.
It was found out that the reconstruction method does not affect a difference in segmentation accuracy between highly and lowcircular cells (the segmentation accuracy in highly circular cells is in average 15% better for all reconstruction methods) without significant variations for individual methods. Seedpoint extraction, however, is much more cellshapedependent (Fig. 5c). Because these methods are blob detectors by nature, the result is better for more circular cells with most methods. However, the dDTTreshold and dCellDetect are not affected by circularity and are among the most efficient segmenting tools at the same time.
Regarding the effect of a degree of contact with other cells, method of image reconstruction does not affect a difference in segmentation between densely and sparsely populated areas (20% better segmentation results for isolated cells). Seedpoint extraction accuracy is however even more profoundly affected by a level of contacts with other cells (in average 25% better segmentation for isolated cells).
Discussion
During the last two decades, the amount of approaches to segment microscopic images increased dramatically. The precise segmentation of labelfree livecell microscopic images remains challenging and not completely solved task. Furthermore, different microscopic techniques make this task more difficult due to different image properties provided.
Accordingly, the aim of this study was to compare the most heterogeneous spectrum of segmentation methods to real data of the same cells from multiple contrast microscopic modalities. The properties of each processing step has been evaluated and segmentation accuracy has been compared.
We used human adherent cells, which are much more heterogeneous in shape and thus much bigger challenge for segmentation than the segmentation of spherical bacteria or yeast. Based on the described results, we can now summarize, discuss and suggest several findings directed to both biologists and bioinformaticians from different points of view.
Segmentability of microscopic techniques
When considering a microscopy technique for labelfree segmentation, there were no dramatic differences in the segmentation efficacy between DIC, HMC or PC. However, the highest segmentation efficacy was obtained when QPI microscopy was used due to the higher image quality (without significant artefacts and high image contrast). In principle, approaches originally intended for fluorescence segmentation are applicable for these images. QPI technique should be also the choice, when a fast, high throughput segmentation is desirable, because no image reconstruction is needed and simple thresholding with MSER  Seeded watershed provides satisfactory results.
Performance of segmentation steps
Regarding individual processing steps, the most crucial are image reconstruction and seedpoint extraction methods. Foregroundbackground segmentation, on the other hand, can be considered the least problematic part, where no dramatic differences between methods were observed, except that learningbased approaches scored better. Regarding the seedpoint extraction, however, a reconstructed image is needed for almost all approaches (except dDTWeka), making seedpoint extraction dependent on precise reconstruction. Not all foregroundbackground segmentation methods need reconstructed images, because some are compatible with raw DIC or PC images (e.g. sWeka, sLSCaselles, sEGT) and generally perform well. Omitting the reconstruction step will need the seed point extraction methods applicable to raw data (eg. dDTWeka or DT with different foregroundbackground segmentation), which can slightly reduce the quality of cell segmentation. It was also evident, that not all reconstruction algorithms are suitable for the seedpoint extraction (high Dice coefficient in the foreground segmentation step does not guarantee suitability for the seedpoint extraction). It also cannot be stated that the timeconsuming methods are dramatically betterperforming in the seedpoint extraction. Here, the learningbased approach provided better results, too. Below we provide short workflow summary for each microscopic technique:

QPI – this technique usually provides images with the best image properties with respect to automated image processing. We observed that Weka probability map segmented with Graphcut, followed by seedpoint extraction with dGRST and finally segmented by Markercontrolled watershed gives superior results. In general, any segmentation approach used, QPI gained the highest segmentation efficacy.

PC – for this modality we suggest simple and fast reconstruction with TopHat filter, and dGRST or dFRST for seedpoint extraction. Graphcut applied to probability Weka probability maps produce best foregroundbackground segmentation. Final segmentation is slightly improved if Markercontrolled watershed is applied to distance transform image (instead of intensity image).

DIC/HMC – the images from these modalities are similar, which leads us to suggestion that the same pipeline can be applied to both. We suggest to use rDICKoos method for reconstruction and Graphcut applied to probability Weka probability maps for foregroundbackground segmentation. Thresholding with distance transform (dDTThreshold) is best for seedpoint extraction, finally segmented by Markercontrolled watershed. Although DIC and HMC have a lot of similar features, DIC produce generally better results.
Allinone packages are extremely popular in biologist community and more or less provide the complex solution for single cell segmentation task. However, these packages implement common image processing methods (some of them described here) and together with graphical user interface and interactions, provide rich possibilities for segmenting the images. We can conclude that FogBank and CellProfiler tools achieve the highest segmentation efficacy among these approaches (without need of programming skills) and it is also universal for various imaging modalities. Both FogBank and CellProfiler use a similar generalizable approach based on the combination of watershed and distance transform, however, CellProfiler also includes a possibility to build complete cell analysis pipelines and as such should be default choice without programming.
Deeplearning remarks
Intentionally, our focus was set on a spectrum of traditional strategies while the rapidlydeveloping spectrum of deeplearningbased segmentation was omitted. The main practical limitation of application of deployment of pretrained Unet or other deep learning method (transfer learning) is the need for sufficiently large training dataset (covering different modalities and cell types/shapes). However, the image databases for segmentation tasks are not as large and complex as ImageNet [36], which became a standard for pretraining of classificationbased networks. For this reason, available models use only pretrained encoder [37], which is pretrained for classification on ImageNet. As such, we leave this investigation for future work, where deepercomparison is highly needed, especially for different amounts of training data and from view of computational requirements.
Despite the tremendous success of deep learning approaches applied in many computer vision tasks including live cell imaging, there is no straightforward way how to use these methods for cell segmentation of touching dense populated cells. One of the approaches to achieve separated mask for each cell is to predict simple binary foreground mask, but giving higher weight to correct prediction on the boundary as in [38]. Another simple solution is to predict three pixel classes – background, foreground and cell boundary as in [39], which provides better separation of cells. On the other hand, deep learning can be also used for cell detection by regression of 2D Gaussians on the position of centroids as in [40]. In [41], authors combined detection and segmentation into simultaneous prediction with one UNet network, where one prediction map predicts distance to cell boundary (after thresholding we obtain foregroundbackground segmentation) and second map predicts distance to cell centroid (detections are obtained with local maxima detection). These techniques are very promising, however, their testing is out of scope of this paper because our dataset is not sufficiently large for training of these algorithms and there is no standard way how to use deep learning for cell segmentation, leading to enormous number of possible setups to test in order to achieve fair comparison to classical methods.
Remarks and limitations
Because of Matlab platform was used, the information regarding computational time is approximate with a large softwaredependent space for its reduction. All segmentation steps were performed in a sequential way. Thus parallel processing may provide a distinct improvement for most of the methods, but this was beyond the scope of this study. Based on a distinctive difference in a segmentation accuracy between “allinone” methods and individual methods reviewed herein, wellperforming methods usually have more than three parameters to be set (usually not even corresponding with morphological features of the cells). Thus it is still difficult, if not impossible, to automatize the whole segmentation process. In a spite of this, deeplearning approaches provide some alternative – instead of setting optimized parameters, user needs just to provide a training dataset.
Although there are several excellent reviews on such segmentation, a study practically comparing the todate bestperforming approaches on real data from various microscopic techniques is still missing. In [42] the authors review a broad spectrum of segmentation methods to segment histological images. In [43] the authors focus on available tools with GUI. The author of [44] summarizes historical progress of cell segmentation methods. There are also works on comparing QPI, DIC and PC, but in [45], the authors compare modalities without segmentation and in [46] authors test algorithms only on QPI data, without considering image reconstructions. In Ulman et al. [47] the authors compared segmentation and tracking on various microscopic methods, including 2D fluorescent, DIC and PC. Many detection errors can be eliminated with tracking. Thus the comparisons with our review might be rather relative. Similarly to our results, one of the bestperforming algorithms “KTHSE” used a relatively simple thresholding together with a precise seedpoint extraction (tracking in their case). This underlines the fact that a precise seedpoint extraction is the most crucial segmentation aspect. Also a segmentation performance was significantly lower in the “FluoC2DLMSC” dataset characteristic by low circularity of cells.
Our study has several limitations. These include the focus on the segmentation of adherent cells, not those cultivated in the 3D matrix or suspensioncultured counterparts. Also the ground truth manual segmentation was performed by a human, although experienced biologist. The problem of overlapping cells was present, although relatively rare. Using the learningbased approaches it was demonstrated that those surpass the transitional strategies. This predicts a future success for deeplearning methods and probably also their future superiority. Also, in accordance with cell timelapse trend in microscopy, cell segmentation is just the first part of the story with cell tracking being another one.
Conclusion
In this study, we performed a comprehensive testing of image processing steps for single cell segmentation applicable for labelfree images. We searched for published methods, which are used by biologists and bioinformaticians, we assessed the suitability of used data and we carefully tested particular segmentation steps (image reconstruction, foregroundbackground segmentation, seedpoint extraction and cell segmentation) and compared them with available “allinone” approaches. As expected, learningbased methods score among the bestperforming methods, but welloptimized traditional methods can even surpass these approaches in a fraction of the time. We demonstrated that the image reconstruction step makes it possible to use segmentation methods not directly applicable on the raw microscopic image.
Herein we collected a unique set of similar fieldofview image of the same cells acquired by multiple microscopic techniques and annotated by experienced biologist. The raw and reconstructed data is provided, together with the annotated ground truth and Matlab codes of all approaches.
Methods
Dataset
Cell culture and culture condition
PNT1A human cell line was used in the experiment. This cell line was derived from normal adult prostatic epithelial cells immortalized by transfection with a plasmid containing SV40 genome with defective replication origin. The cell line was purchased from HPA Culture Collections (Salisbury, UK). PNT1A cells were cultured in RPMI1640 medium supplemented with antibiotics (penicillin 100 U/ml and streptomycin 0.1 mg/ml) with 10% fetal bovine serum (FBS). Prior microscopy acquisition, cells were maintained at 37^{∘}C in a humidified (60%) incubator with 5%CO_{2} (Sanyo, Japan). Intentionally, high passage number of cells was used (>30) in order to describe distinct morphological heterogeneity of cells (rounded and spindleshaped, relatively small to large polyploid cells). For acquisition purposes, cells were cultivated in Flow chambers μSlide I Luer Family (Ibidi, Martinsried, Germany).
Microscopic image acquisition and dataset characteristics
QPI microscopy was performed on Tescan QPHASE (Tescan, Brno, Czech Republic), with objective Nikon CFI Plan Fluor 10×/0.30 captured by Ximea MR4021MC (Ximea, Münster, Germany). Imaging is based on the original concept of coherencecontrolled holographic microscope [48, 49], images are shown as a 32bit file with values corresponding to pg/ μm^{2} recalculated from radians according to Barer and Davies [50, 51].
DIC microscopy was performed on Nikon A1R microscope (Nikon, Tokyo, Japan) with a Nikon CFI Plan Apo VC 20×/0.75 objective captured by a Jenoptik ProgRes MF CCD camera (Jenoptik, Jena, Germany).
HMC microscopy was performed on Olympus IX71 microscope (Olympus, Tokyo, Japan) with Olympus CplanFL N 10×/0.3 RC1 objective captured by Hamamatsu Photonics ORCAR2 CCD camera (Hamamatsu Photonics K.K., Hamamatsu, Japan).
PC microscopy was performed on a Nikon Eclipse TS100F microscope, with a Nikon CFI Achro ADL 10×/0.25 objective captured by Jenoptik ProgRes MF CCD camera.
The captured dataset characteristics are summarized in Table 3. All data were manually segmented by an expert in cell biology as ground truth for segmentation and detection. Same areas of sample were captured using these microscopes, but due to the cell movement and different FOV size the overlap is not absolute.
Allinone segmentation tools
Here are described “allinone” approaches (designated with “aio” prefix).
aioFARSIGHT
FARSIGHT toolkit 0.4.5 module Nucleus editor [2] consists of an automatic Poisson threshold binarization refined with graphcut (applied on a binary foreground image) and produces initial segmentation containing cell clusters. Next, Multiscale LaplacianofGaussian is used to produce feature map (image where blobs are enhanced  see “LoG filters” section for more details), which is segmented by local clustering algorithm. This clustering algorithm then produces rough cell cluster separation. Finally αExpansions (multilabel graph cut) is used to refine segmentation, with novel method of Graph colouring for more efficient computation (see [2] for more details).
The first set of parameters was cellshapederived: “min scale” and “max scale” (the minimum and maximum scale of the multiscale Laplacian of Gaussian filter) were set based on a measured radius of equivalent circle of cells, “xy clustering res” (resolution of the local maximum clustering) was set similarly as “min scale”, and “min object size” was set as the area of the smallest cell. The second set of parameters was optimized: “high sensitivity” (enable/disable high sensitivity binarization), “finalize segmentation” (enable/disable the segmentation refinement step), “use distance map” (enable/disable the use of the distance constraint to select the LoG scales), and “refinement range” (parameter sets the maximum distance that an initial contour can be shifted).
aioCellX
Dimopoulos et al. [3] approach consists of seed generation with gradientbased Hough transform, construction of membrane patterns images for each seeed (crosscorrelation with estimated membrane profile) and segmentation of each such image with graphcut. After that, statistical morphological outliers are removed and individual regions are combined (almost identical regions are merged and overlaps are resolved).
CellX includes a GUI, where user can interactively set cell size range, maximal cell length and estimated membrane profiles.
aioFogbank
In Chalfoun et al. [4] Fogbank, foreground is segmented with EGT. Seeds are detected as connected regions after percentile thresholding (with some distance and size constraints). Pixels above a defined percentile level are then connected to the nearest seedpoint. Either intensity or gradient image and either Euclidean or geodesic distance are used for computation.
aioFastER
Hilsenbeck et al. [5] FastER applies MSER to construct component tree and SVM for classification of regions into groups of cells or false detections. Finally nonoverlapping regions with the highest score are selected. It shares CellDetect similarities (see “CellDetect” section), but this algorithm uses 9 features for SVM classification only and does not compute globally optimal solution, thus being computationally faster. To achieve complete segmentation (not only detection as CellDetect), authors modified their approach on the algorithmic level. Size constraints of cells (measured min/max cell size) were set and “denoise” parameter setting were optimized (off/on/strong). Larger number of FOVs used for training were tested without improvement.
Dry massguided watershed
The dry massguided watershed method (designated as aioDMGW) is a thresholdingbased approach, implemented as a part of Analyzer module of QPHASE software 6.803 (Tescan, Brno, Czech Republic). First the phase image is slightly smoothed and foreground is separated from background using thresholding. Then watershed starting from the local maxima is performed. The decision of merging of touching segments, or leaving them separated, is based upon the sums of pixel values (i.e. dry mass) in each touching segment. The optimized parameters are: threshold; min segment sum (the minimum accepted sum of pixel values in each segment used to filter out noise and cell debris); max merge sum (the threshold of sum of pixel values of touching segments used to decide if the segments should be merged or left separated).
aioCellProfiler
The CellProfiler [10] is a strong segmentation tool, however, we perceive it more as a platform where a substantial part of the segmentation strategies used here can be reproduced. Nevertheless, we evaluated output of “IndetifyPrimaryObject” module, which combines thresholding and watershed. Watershed is used twice, for seedpoint extraction and final single cell segmentation, and it is applied to either intensity or DT image. Additionally, module uses some smoothing and it remove seedpoints bellow some allowed distance. Measured range of cell radiuses and optimal threshold (see Additional file 1) were used and we optimized betwen application to intensity or DT image for both steps.
Other allinone tools
Following algorithms were reviewed but not used in comparison with reasons stated below:
CellTracer [6] consists of 3 steps – foreground segmentation, border segmentation and cell segmentation by model fitting. This approach is more suitable for yeast or bacteriashaped objects (coccus or bacillusshaped with distinctive borders). Similar issues were observed in SuperSegger [7], CellSerpent [8] or CellStar [9].
Image reconstruction techniques
DIC, HMC and PC image formation process can be described as convolution between the original image of the scene and 2D PSF. For PC images PSF is [13]
where δ(·) is Dirac delta function and airy(·) is Airy pattern. This leads to halo and shadeoff artifacts (see Fig. 2). For DIC image PSF is difference of two Gaussians [52]:
where σ is Gaussian standard deviation and u=[u v]^{T} is unit vector specifying shear direction. It means that DIC image is derivation under shear direction visible as 3Dlike topographical appearance (see Fig. 2). The inverse PSF then can be used for image reconstruction. The goal of these reconstruction algorithms is to produce image of bloblike cells qualitatively corresponding to cell mass (similar to QPI). The methods described bellow are designated with prefix “r” (reconstruction), original modality and author, where possible.
DIC reconstruction methods were well reviewed in [12]. Based on the results of this study, two methods were chosen: (a) fast, computationallyefficient Yin et al. approach [13] (in following parts designated as “rDICYin”) and (b) more computationallydemanding Koos et al. [12] (designated as “rDICKoos”). HMC images have the similar properties as DIC and therefore the same reconstruction algorithms were tested.
For PC reconstruction [14], two methods were chosen (a) more complex computationallydemanding method based on PSF model (designed as “rPCYin”) (b) simple Tophat filtering (designated as “rPCTophat”).
rDICKoos
Method proposed by Koos [12] (rDICKoos) uses an energy minimization with data term and total variation regularization term
where · denotes dot product, ▽ denotes gradient, u=[u v]^{T} is unit vector specifying shear direction, Ω is image domain and K is kernel which approximate PSF without derivative (Gaussian function), where ▽K=PSF_{DIC}(x,y). EulerLagrange equation of data term for symmetric kernel K leads to
where ∂_{x} and ∂_{y} denotes partial derivatives and W is a local window (with size of kernel). Finally, this can be solved with gradient descent iterative method as
where \(\hat {f}^{(t+1)}\) is reconstructed image in next iteration, div denotes divergence. Last term is proposed by total variation regularization.
Besides of shear angle, which is assumed to be known (or recognizable from image  typically multiple of 45), rDICKoos method has three parameters  weight of smoothness (total variation) regularization w_{s}, step size of gradient descent w_{a} and number of iteration it. Smooth regularization sets compromise between noise elimination and details preservation. Too large step size leads to method divergence and too small step size leads to slow convergence. Number of iterations has a small influence on the result; default value 20000 was used. For setting of other parameters see Additional file 1.
rDICYin
Yin et al. [13] presented a reconstruction method for DIC images (rDICYin) working with multiple shear directions, but with some simplification in equations it also works on images with one shear angle direction. Authors assumed that distortion of the microscope can be modeled by convolution with PSF
where d is PSF (in general a directional firstderivativeofGaussian kernel, but simple difference without Gaussian is used for simplification), g is acquired image and f is original image. Simple inverse filtering leads to highly noisy images, which can be reduced by regularization. This can be achieved with optimization of energy function which must be minimized over whole image domain
This equation is composed of data term, smooth term and sparse term (all with l_{2} penalization, where w_{s} and w_{r} are weights for the smooth and sparse regularizations, respectively). \(\hat {\mathbf {f}}\) is reconstructed image (approximation of f). Smoothness is achieved by setting a restored pixel value to be close to the average of its neighbors (where a=[1,1,1;1,−8,1;1,1,1]/8). Sparse regularization causes the value of background pixels to be close to zero. Optimization of function has closeform solution in Fourier space (\(\hat {\mathbf {F}}=\mathcal {F}\{\hat {\mathbf {f}}\}\) etc.)
where “ ⊘” and “ ⊙” denotes elementwise division and multiplication, respectively.
Besides shear angle, rDICYin has two parameters only, w_{s} and w_{r}, which set smoothness and sparse regularizations, respectively.
rPCYin
In [14] Yin et al. used a deconvolution with sparse constraint regularization to reconstruct PC images. This method was further expanded with dictionary of diffraction patterns [53], which deals with problematic mitotic cells. This method is in fact a segmentation method as presented in the Su at al. paper [53] and it therefore cannot be used as preprocessing (i.e. reconstruction) step. rPCYin [14] is very similar to rDICYin [13] with modified equation 7 to linear equation system with l_{1} penalization for the sparse term.
where \(\overline {\hat {\mathbf {f}}}\) and \(\overline {\mathbf {g}}\) are vectorized restored and acquired images, H is the transfer matrix of the image formation model and L is Laplacian matrix (corresponding to different expression of operators d and a in the equation 7). Λ is positive diagonal matrix defining sparseness, w_{s} and w_{r} are weights for the smooth and sparse regularizations. Because of l_{1} penalization of sparseness (known to be better than l_{2}) there in no closedform solution. It can be solved with an iterative algorithm which is based on nonnegative multiplicative updating (for more implementation details see [14]). PSF (which leads to H) is then modeled by the equation 1, where airy pattern is
where R and W are PSFdependent parameters  outer radius (R) and ring width (W) of phase ring and J_{1}(·) is the first order Bessel function of the first kind. rPCYin has also optimization parameters w_{s} and w_{r} which define weights of components of optimized energy function. Other parameters not discussed in [14] were set to default value (radius=2,epsilon=100,gamma=3,scale=1,maxiter=100,tol=10^{−16}). Because of large computational time, optimization of PSF and optimization parameters was done separately  first proper PSF was found (other parameters set to default value w_{s}=1 and w_{r}=0.5) and then optimal W and R values were used in optimization of w_{s} and w_{r}.
rPCTophat
Tophat filtering (referred here as rPCTophat) was used by Thirusittampalam et al. [15] and Dewan et al. [16] for halo artifacts elimination. This simple heuristic approach shows very promising results and it is considered as the next PC reconstruction technique in this paper.
Reconstruction based on tophat filtering with diskshaped structuring element has only one adjustable parameter  radius of structuring element, which is roughly equal to the radius of the cell, with optimal value r=16.
Foregroundbackground segmentation
Thirteen methods has been tested and to make it more clear, the methods are designated with prefix “s” (segmentation), original modality and the author, where possible.
Thresholding
Three thresholdbased techniques were used for the foregroundbackground segmentation. Simple threshold (named as sST) and two automatic threshold algorithms, Otsu [17] (sOtsu) and Poisson distribution [2] (sPT).
Automatically determined thresholds varies between FOVs, so a better result can be expected. sOtsu assumes that grayscale values are mixture of two Gaussian distributions. Nevertheless, for the adherent cell images the mixture of two Poisson distributions is sometimes more suitable [2], thus sPT was tested. For ST, threshold value was optimized with 100 steps between minimal and maximal value.
Empirical gradient threshold
Chalfoun et al. [18] described an empirical gradient threshold method (referred here as sEGT), which uses empirically derived model for threshold estimation. sEGT was described to work with different microscopic modalities (PC, DIC, brightfield and fluorescence) and is applicable also on the others, including raw (unreconstructed) images. sEGT utilizes a Sobel operator to compute absolute value of gradient, then the percentilebased threshold is found, followed by the binary morphological operations. Three parameters must be set beforehand: minimal cell size (removing small objects), minimal hole (removing small holes) and manual finetune (decreasing or increasing the estimated threshold). For all these methods minimal object size was determined from a ground true mask of the training images.
sPCJuneau
Juneau et al. [19] described simple segmentation method (referred here as sPCJuneau) designed for PC images. It computes a range map (difference between minimum and maximum in local window), which is then thresholded. Consequently, all holes and small objects in the binary image are removed. Thus these parameters are optimized: window size, threshold and minimal object size. Although originally designed for PC images, it is applicable for other modalities as well.
sPCPhantast
Jaccard et al. [21] developed a software toolbox PHANTAST consisting of foreground segmentation techniques specialized for PC microscopy images. It computes local contrast
where G is a Gaussian kernel with standard deviation σ. The resulting localcontrast image is then globally thresholded and halos are corrected. For halos correction, the gradient direction is computed by eight Kirsch filters (8 directions). Halo pixels are initialized with boundary pixels of binary image, then iteratively each halo pixel points to its gradient direction and two adjacent directions, where each of these three pixels is marked as halo if it is considered foreground (for bright halos gradient points in and for dark cells gradient points out). Maximum cell area fraction removed as halo is restricted and after elimination of halos, small objects and holes are removed. This leads to 5 parameters  Gaussian σ, threshold, halo area fraction, minimal hole size and minimal object size.
sPCTopman
Topman et al. [20] described another method for foreground segmentation originally intended for PC images. This approach applies two filters, one with a small and one with a large local window computing the standard deviation, where both are thresholded. The result is an intersection of these two binary images, where binary image from large window is morphologically eroded (with morphological element of half the size of the large window) and final image is morphologically opened and closed. This leads to 4 parameters  two window sizes, threshold, and morphological element size.
LevelSets
Matlab implementation of levelset method with function activecontour was used. This implementation includes an edgebased method [22] (referred as sLSCaselles) and regionbased method [23] (referred as sLSChanVese). Both methods use a SparseField implementation [54] for contour evolution and both have two adjustable parameters  smoothness of the result contour and additional force, which leads to a tendency of the contour to grow or shrink. While sLSChanVese segments the image into two regions based on the mean region intensities, sLSCaselles segments the image based on the edges. The levelset methods were initialized with morphologicallydilated binary results of Weka segmentation, because it provides similar initial contours for all modalities. Number of iterations of the evolution was set to 1000, which was shown to be enough for all types of images and all parameter settings.
Trainable Weka Segmentation
Next, a machine learning tool for microscopy pixel classification Trainable Weka Segmentation v.3.2.13 was used [26] (designated as sWeka). Compared to previous foregroundbackground segmentation strategies, this approach was primarily used directly on the raw data. Weka was trained using the following default training features (Gaussian blur, Sobel filter, Hessian eigenvalues, difference of Gaussians filter, membrane projections) as well all remaining available filters (variance filter, minimum filter, maximum filter, median filter, anisotropic diffusion, bilateral filter, lipschitz filter, kuwahara filter, gabor filters, Sobel filter, laplacian filter, structure, entropy filter). For these filters it is also possible to set a σ range, which specifies the filter size. Other parameters were set to default values, random forest classifier was set to 200 trees (WEKA FastRandomForest). Because of learning nature of this approach, the effect of following factors on segmentation efficacy was optimized: (a) number of fields of view used for learning (b) training features used for learning (“all” and “default” training features), (c) effect of various fields of view used for training (one continuous area in one FOV, or smaller patches of same sizes from multiple FOVs), (d) size of FOV used for learning (increasing the area from 6×6 px to 1360×1024 px). Moreover, probability maps were exported and used for further analyses.
Ilastik
Another tested machine learning tool for pixel classification was Ilastik v.1.3.0 [25]. Ilastik uses a random forest classifier [55] with 100 trees and is very similar to WEKA. Ilastik was launched using the following settings: enabled all training features: raw intensity, gradient magnitude, difference of Gaussians, Laplacian of Gaussian, structure tensor eigenvalues and the Hessian matrix eigenvalues  all with 7 Gaussian smoothings with σ=0.3−10px.
Ilastik was optimized accordingly as Weka. It allows a computationally expensive automatic selection of suitable features. Based on a first optimization step, there was no significant difference between “optimal” and “all” features. Thus, in a spite of this and a fact that Ilastik has less available features then WEKA, “all” features were used in further steps.
Graphcut approach
An ImageJ plugin for GraphCut (v. 1.0.2) based on the reimplementation of Kolmogorov’s maxflow v3.01 library [24] was used. The following data were used as an input for GraphCut: (a) Probability maps generated by Weka (referred as sWekaGraphCut), (b) images reconstructed with approaches described in “Image reconstruction approaches” and (c) raw image data (both referred as sGraphCut). There are two parameters to be optimized: terminal weight and edge weight. Edge weight (designated as “smoothness” in the GUI, range 010) reflects a penalty for label changes in the segmentation (higher values cause smoother result). Terminal weights (designated as “foreground bias”, range 01) correspond to a cost of assigning background pixels to the foreground.
Terminal weights (foreground bias in GUI) affect the segmentation efficacy distinctly, thus its optimization is crucial. On the other hand, edge weight (smoothness) corresponds to the size of individual cells and has been roughly estimated from 0.4 to 0.8 for used cell sizes (between 1000 and 4000 pixels, respectively).
Cell detection (seedpoint extraction)
The cell detection (seedpoint extraction) plays a key role in the segmentation of the overlapping objects. For densely clustered and touching cells a precise cell detection has the most significant influence to the final segmentation accuracy. The primary goal in the cell detection is to recognize the presence of the individual objects in the image. Finally, combination of successful foregroundbackground separation followed by identification of individual cells enable to segment individual cells. There is a considerable amount of methods for cell detection and the mostly used and cited methods are described and evaluated in this paper. Because most of the described methods require bloblike cells, image reconstruction is necessary in most cases (except LoG and generalized LoG filters by Kong et al. [28]).
The tested seedpoint extraction methods usually include parameters related to the cell radius (minimal and maximal). For this reason we estimated these values from the ground truth masks. Background segmentation from the previous step was used to eliminate falsely detected seeds in the background. Some of the tested methods already include this step (e.g. dLoGgXu [29]). The binary background masks produced by trainable Weka segmentation were used for this purpose. For clarity, the methods described bellow are designated with prefix “d” (detection), image processing approach and author, where possible.
LoG filters
Because of distinctive popularity of the LoG filter for the blob object detection, many modifications of this detector exist, e.g. multiscale LoG, Hessian LoG, generalized LoG. LoG filter at a scale σ is defined by equation
where G is 2D Gaussian function, x=(x,y) and · is Euclidean norm [27]. In principle, this filter works as a matched filter for blobs.
Multiscale LoG filtering uses a bag of LoG filters with m different sigma values, which leads to m−D feature space. As proved by Lindeberg [56], LoG responses must be normalized LoG(x,σ)_{norm}=σ^{γ}LoG(x,σ) for scale invariance, where γ=2 for scale invariance, but it can be refined for a preference of larger or smaller objects.
Peng et al. [27] used Maximum Intensity Projection (MIP) of the series of LoGfiltered images \(MIP(\mathbf {x})=\max \limits _{\sigma }({LoG}_{norm}(\mathbf {x},\sigma))\), with threshold applied to resulting 2D image, where binary objects correspond to the detected cells. This method (further designated as dLoGmPeng) has the following parameters: minimal sigma σ_{min}, maximal sigma σ_{max}, sigma step Δσ, γ and threshold.
Kong et al. [28] searched for local maxima higher than defined threshold in whole m−D LoG scale space with elimination of overlapping regions by a pruning process. In the pruning process, the overlapping blobs are eliminated, where only blob with larger value in m−D scale space is preserved. This method has these parameters: σ_{max}, sigma step Δσ, γ, threshold and maximal overlap ratio. Here for σ the logarithmic step is used. This method is referred as dLoGmKong.
Hessian analysis of LoG (referred as dLoGhZhang) described by Zhang et al. [30] uses the same bag of LoGfiltered images, but optimal scale identification and cell center detection is different. It is known, that local Hessian matrix is positive definite for bloblike structures. The Hessian H (computed from LoGfiltered image) at position (x,y) can be approximated with differences in 2×2 neighborhood. Each connected region with a positive definite Hessian is considered as cell, where H is a positive definite matrix when H_{11} is positive and det(H) is positive.
Optimal is considered a such scale where the mean intensity of the LoGfiltered image is maximal on the positive definite locations, and these positive definite regions are the detected cells. Method is insensitive to choice of range and steps of σ, which leaves only γ parameter to be optimized. Zhang [30] also uses unsupervised classification to identified true cells only, but in our case this leads to deterioration of the results only and thus was not included in the testing.
Intuitively rotationallysymmetric LoG kernels are very sensitive to irregular cell shape. For this reason Kong et al. [28] proposed a generalized LoG filter (referred as dLoGgKong) for the detection of the elliptical shapes. They derived an equation for elliptical kernel with two standard deviations σ_{x}, σ_{y} and orientation Θ. Method also includes a specific scale normalization with a parameter α and automatic choice of sigma range based on the initial analysis with circular LoG filters. For every pixel position, a feature image is created as a sum of all filter responses and detected cells are local maxima in this image (see [28] for more details). Thanks to the automatic σ estimation, there is one parameter only  α. Method uses integer kernel sizes smaller than estimated σ_{max}. Small kernels produce false peaks on a subcellular structures in our data. These artefacts are eliminated by adding a σ_{min} parameter, which corresponds to a minimal cell radius.
Xu et al. (referred as dLoGgXu) [29] sped up this technique by summation of the filters with the same kernel orientation Θ, which is possible thanks to the distributive property of convolution. Instead of automatic estimation of σ range, they estimate it from cell radii. Moreover this technique includes a different normalization (without parameter) and meanshift clustering for elimination of multipletime detected seeds. Parameters of this method are: σ range and mean shift window size.
A similar approach was described also in Peng et al. [27] method. Parameter range of σ is estimated based on cell radius as \(\sigma =r/\sqrt {2}\). For dLoGmPeng we used estimated σ_{max} and σ_{min}. Step of σ (Δσ) is insensitive parameter, therefore we set it to 1. For setting of other parameters see Additional file 1. Authors [27] used γ=2, which is proven to lead to the theoretical scale invariance. When γ<2 the smaller objects are preferred, for γ>2 the larger objects are preferred. Appropriate setting of γ leads to mean Dice coefficient improvement +0.089 for dLoGmPeng method and for this reason we add γ to optimized parameters for both dLoGmPeng and dLoGmKong methods. Similarly for dLoGmKong we used estimated σ_{max} and σ_{min} with 13 logarithmic steps like the authors[28] (for other parameter settings see Additional file 1). Extension by γ parameter leads to 3 parameters (besides of cell radii), which are sensitive and must be properly set. Both generalized LoG methods try to avoid parameters setting, where dLoGgXu has cell sizerelated parameters only (we set it based on estimated radius) and dLoGgKong has one adjustable parameter  scale normalization factor, but cell size estimation is automatic. Both generalized LoG methods are computationally expensive (see Table 1), but dLoGgXu reduces the computational time by a reduction of number of convolutions.
Distance transform
Distance transform (DT) of foreground image is defined as a distance to the nearest background pixel (Euclidean distance is chosen as metric). Local maxima of the generated distance map are considered as cells. This method often detects many false cells. For this purpose hmaxima transform is used [15], which uses a grayscale morphology for elimination of small local maxima, where parameter h sets the depth of local maxima to be eliminated. We used two modifications of this method; dDTThreshold, where binary foreground is computed with optimized threshold and dDTWeka, where foreground from Weka segmentation is used. Other parameter of this method is maximal size of objects and holes, which are eliminated before applying of the DT.
Fast radialsymmetry transform
Fast radialsymmetry transform [31] (referred as dFRST) is a general method for the detection of circular points of interest applicable to approximately circular objects. Pixels with absolute value of gradient greater than threshold β vote in its gradient direction at the distance of radius r∈R, where R is set of radii, determined based on object/cell size. If bright blobs are only considered detection, positions of affected pixel is given by an equation
where g(x) represents the gradient and round operator rounds each vector element to its nearest integer. On position P(x), an orientation projection image O_{r} is increased by 1 and magnitude projection image M_{r} by ∥g(x)∥. Transformation is defined as mean over all radii
where
where G_{r} is a Gaussian kernel, α is the radial strictness parameter and k is a scaling factor normalizing different radii (where typically k≈10). Inspired by Ram et al. [57], we use a grayscale dilatation to small local maxima suppression in S. Local maxima are then considered as cells. As R we use all integer values between estimated minimal and maximal cell radius. The parameters for this method include: radial strictness α, scaling factor k, size of morphology structuring element δ, and gradient threshold β.
Generalized Radialsymmetry transform
The generalized radialsymmetry transform as described by Bahlman et al. [32] (referred as dGRST) is able to deal with elliptical shapes because affine transform is employed.. Similarly to generalized LoG filters, we can compute response for different axis scalings and rotations. The dGRST principle is similar to dFRST method, but the gradient g(x) is transformed to
where
and G is affine transformation matrix  for ellipse it is rotation and scaling with parameters θ, a and b. We can set r=1 and used a and b to adjust the size of the desired ellipse axis. All integer values between estimated minimal and maximal cell radius with a>b and 6 steps for θ were used for a and b. Bahlmann at al. [32] mentioned also a Gaussian kernel specified by affine transformation parameters θ, a and b. For consistency with dFRST, we use Gaussian kernel with σ=0.5 distorted with G transformation. Remaining parameters are identical to dFRST.
Radial voting
Qi et al. [33] presented a modification of radial voting for cells in histopathology specimens (reffered here as dRVQi). It is based on an iterative radial voting described previously [58], but works as a singlepath voting followed by a meanshift clustering. Every pixel with position x=[x,y] vote in Gaussian smoothed gradient direction α(x), with cone shapedkernel function (voting area).
where θ is an angle of vector α(x), {r_{min},r_{max}} is kernel radial range and Δ is the kernel angular range. In addition, voting sector is weighted by Gaussian function with center located at kernel center. Every pixel (with gradient above certain threshold) update voting image V by adding voting pixel gradient magnitude g(x) to all pixels under kernel. Voting image is then thresholded with several thresholds and results are summed and clustered with meanshift algorithm. For more details see [33]. We used estimated r_{min} and r_{max} from the ground truth images, thresholds were set to 0.2, 0.3,...0.9times the maximum of image, and we optimized sigma of gradient Gaussian smoothing, sigma of Gaussian for kernel and mean shift bandwidth. We also vote with all pixels, not only with pixels with high gradient magnitude, because computational time of our implementation is not dependent on number of voting pixels. Besides [33] we also tested original [58] and newer [59] methods, but both were less suitable for adherent cells.
Maximally stable extremal region
Extremal regions of grayvalue image are defined as connected components of thresholded image I_{t}=I>t for some t in this method (designated as dMSER). As described in [34], dMSER produces stable extremal regions of image which are stable in sense of area variation w.r.t. changing threshold t. Minimal stability of extracted region can be set with two parameters  threshold step δ defining the percentage of intensity range and maximal relative area change with this step. This method generates many regions which can overlap. Finally, the smallest regions generated with the highest threshold are picked. This is achieved by finding of the local maxima in the sum of binary images of all regions. Another option is the usage of most stable region from the overlapped ones, but this was shown to be noneffective in our case.
CellDetect
Arteta et al. [35] implementation of CellDetect uses MSER to identify the candidate regions, followed by a classification of true regions. Method extracts 92dimensional feature vector with object histograms and shape descriptors. Training proceeds in two phases. In the first phase, training of binary SVM and its evaluation is done, which produces score for each region. Region with one seedpoint and highest score (one for each seedpoint) is used as ground truth for the next phase. In the second phase, structured SVM is used for classification of the regions within each tree created from the overlapping regions, but nonoverlap constrains are included. For more details see [35]. Method requires few training images with simple dot annotation and proper setting of MSER detector to achieve high recall.
Single cell (instance) segmentation
After image reconstruction, foregroundbackground segmentation and seedpoint extraction, individual cells were segmented using Markercontrolled (or seeded) watershed [60]. This step showed to be less crucial, because inaccuracy in placing border between cells has a small influence to segmentation efficacy only. Thus, for combining of foreground and seeds into the final segmentation, we test only this simple but very robust technique. Note that more advanced methods exist  e.g. graphcut [61], or levelset [62] based.
Makercontrolled watershed is similar to classic watershed approach, with restriction of local minima positions into detected seeds location, which can be simply done with mathematical morphology operations. Besides of straightforward application on our images, we proposed a second approach applied on DT image, which does not require an intensity valley between separated cells. For DT image we use geodesic distance transform [63] with distances from seeds (the distance within the foreground pixels only, ignoring the background).
General parameter optimization strategy
Grid search with 10 steps was used for the optimization of parameters of all methods, where suitable range was selected experimentally by a few manual tests. Parameters with large searched range (relatively large difference between lower and upper bound) were searched with logarithmic scale. The same parameters ranges were used for all modalities. All parameters were properly set for training images and then these values were used for all testing images. For background segmentation and detection methods Dice coefficient was used as an objective function (used e.g. in [18]). For image reconstruction techniques the area under ROC curve (AUC) generated by thresholding was used (as well as in [14] or [12]). Because of large computational difficulty of some methods, we attempted to eliminate such parameters from optimization, which does not influence the objective function. If threshold is optimized parameter, its value was optimized between a minimal and maximal intensity of image pixels, with 100 steps for simplicity. Before application of each method, images were normalized into interval [0,1], where minimal and maximal values of the first image of each sequence were used as a reference for the normalization.
Evaluation of results
The F_{1} score (Dice coefficient) was used as a measure of segmentation accuracy for (1) foregroundbackground segmentation, (2) seedpoint extraction, and (3) single cell segmentation, with following modifications:
Foregroundbackground segmentation evaluation
For the evaluation of cell segmentation, Dice coefficient was used as follows:
where · is number of pixels of region, X and Y are ground truth and result segmentation, respectively. Dice coefficient is equal to F_{1}score, but this term is used for pixelwise evaluation. Another metric used for segmentation evaluation is Jaccard index, which is related to Dice coefficient as:
which is monotonically increasing function on interval <0,1> (the range of Dice values). This means that order of quality of segmentation algorithms w.r.t. Jaccard is same as w.r.t. Dice coefficient and for this reason we evaluated only Dice coefficient.
Dice coefficient was computed for evaluation of the foreground segmentation results using all pixels in the image.
Seedpoint extraction evaluation
Single dot labels (seeds) are considered as cell detection results. If some method produces presegmented regions, then centroids are used as labels. Because our ground truth corresponds to the binary segmented cells, we consider as TP (true positive) such cells having one seed only. As FP (false positive) are considered cells with additional seeds in one cell and with seeds outside cells. FN (false negative) are cells without any seed. To evaluate the performance of the cell detection, Dice coefficient (F_{1} score) was used
In some papers the accuracy of the centroid positions is also evaluated. Nevertheless, these positions are not very significant for cell segmentation. Therefore, we didn’t evaluate this accuracy.
Single cell segmentation evaluation
For single cell segmentation evaluation the F_{1} score (Dice coefficient) is used in a similar manner as for foregroundbackground segmentation evaluation with following modifications: We dealt with correspondence of objects. We used same evaluation of correspondence as [64] in their SEG evaluation algorithm – cell are considered as matching if:
which ensures unambiguous pairing. The final measure of Dice was calculated as the mean of the Dice coefficient of all the reference objects. The cells which are on the image boundary were labeled and they are not included in the evaluation.
A computer with following specifications was used to estimate computational times: Intel Core i56500 CPU, 8 GB RAM.
Abbreviations
 AUC:

Area under curve
 DIC:

Differential image contrast
 DMGW:

Dry massguided watershed
 DT:

Distance transform
 EGT:

Empirical gradient treshold
 FOV:

Field of view
 FRST:

First radial symmetry transform
 GRST:

Generalized Radial symmetry transform
 HMC:

Hoffman modulation contrast LoG: Laplacian of Gaussian
 MCWS:

Markercontrolled watershed
 MIP:

Maximum intensity projection MSER: maximally stable extremal region
 PC:

Phase contrast
 PD:

Poisson distribution
 PSF:

Point spread function
 PT:

Poisson treshold
 ROC:

Receiveroperator curve
 RV:

Radial voting
 ST:

Simple treshold
References
Wang Z, Millet L, Chan V, Ding H, Gillette MU, Bashir R, Popescu G. Labelfree intracellular transport measured by spatial light interference microscopy. J Biomed Opt. 2011; 16(2):026019–0260199. https://doi.org/10.1117/1.3549204.
AlKofahi Y, Lassoued W, Lee W, Roysam B. Improved automatic detection and segmentation of cell nuclei in histopathology images. IEEE Trans Biomed Eng. 2010; 57(4):841–52. https://doi.org/10.1109/TBME.2009.2035102.
Dimopoulos S, Mayer CE, Rudolf F, Stelling J. Accurate cell segmentation in microscopy images using membrane patterns,. Bioinformatics (Oxford, England). 2014; 30(18):2644–51. https://doi.org/10.1093/bioinformatics/btu302.
Chalfoun J, Majurski M, Dima A, Stuelten C, Peskin A, Brady M. FogBank: a single cell segmentation across multiple cell lines and image modalities,. BMC Bioinformatics. 2014; 15(1):6598. https://doi.org/10.1186/s128590140431x.
Hilsenbeck O, Schwarzfischer M, Loeffler D, Dimopoulos S, Hastreiter S, Marr C, Theis FJ, Schroeder T. Faster: a userfriendly tool for ultrafast and robust cell segmentation in largescale microscopy. Bioinformatics. 2017; 33(13):2020–28. https://doi.org/10.1093/bioinformatics/btx107.
Wang Q, You L, West M. Celltracer: Software for automated image segmentation and lineage mapping for singlecell studies.Duke University; 2018. http://www.stat.duke.edu/~mw/MWextrapubs/Wang2008.pdf.
Stylianidou S, Brennan C, Nissen SB, Kuwada NJ, Wiggins PA. Supersegger: robust image segmentation, analysis and lineage tracking of bacterial cells. Mol Microbiol. 2016; 102(4):690–700.
Bredies K, Wolinski H. An activecontour based algorithm for the automated segmentation of dense yeast populations on transmission microscopy images. Comput Vis Sci. 2011; 14(7):341–52.
Versari C, Stoma S, Batmanov K, Llamosi A, Mroz F, Kaczmarek A, Deyell M, Lhoussaine C, Hersen P, Batt G. Longterm tracking of budding yeast cells in brightfield microscopy: Cellstar and the evaluation platform. J R Soc Interface. 2017; 14(127):20160705.
Carpenter AE, Jones TR, Lamprecht MR, Clarke C, Kang IH, Friman O, Guertin D. a., Chang JH, Lindquist R. a., Moffat J, Golland P, Sabatini DM. CellProfiler: image analysis software for identifying and quantifying cell phenotypes,. Genome Biol. 2006; 7(10):100. https://doi.org/10.1186/gb2006710r100. http://arxiv.org/abs/arXiv:1201.3109v1.
Tescan AS. QPHASE: Quantitative, Labelfree Imaging Cytometry. https://www.tescan.com/enus/technology/lightmicroscopy/qphase. Accessed 30 Apr 2018.
Koos K, Molnár J, Kelemen L, Tamás G, Horvath P. DIC image reconstruction using an energy minimization framework to visualize optical path length distribution. Sci Rep. 2016; 6:30420. https://doi.org/10.1038/srep30420.
Yin Z, Ker DFE, Kanade T. Restoring DIC microscopy images from multiple shear directions. In: Lecture Notes in Computer Science (including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). vol. 6801 LNCS: 2011. p. 384–97. https://doi.org/10.1007/9783642220920.
Yin Z, Kanade T, Chen M. Understanding the phase contrast optics to restore artifactfree microscopy images for segmentation. Med Image Anal. 2012; 16(5):1047–62. https://doi.org/10.1016/j.media.2011.12.006. NIHMS150003.
Thirusittampalam K, Hossain MJ, Ghita O, Whelan PF. A novel framework for cellular tracking and mitosis detection in dense phase contrast microscopy images. IEEE J Biomed Health Inf. 2013; 17(3):642–53. https://doi.org/10.1109/TITB.2012.2228663.
Dewan MAA, Ahmad MO, Swamy MNS. A method for automatic segmentation of nuclei in phasecontrast images based on intensity, convexity and texture. IEEE Trans Biomed Circ Syst. 2014; 8(5):716–28. https://doi.org/10.1109/TBCAS.2013.2294184.
Otsu N. A threshold selection method from graylevel histograms. IEEE Trans Syst Man Cybern. 1979; 9(1):62–6. https://doi.org/10.1109/TSMC.1979.4310076.
Chalfoun J, Majurski M, Peskin A, Breen C, Bajcsy P, Brady M. Empirical gradient threshold technique for automated segmentation across image modalities and cell lines. J Microsc. 2015; 260(1):86–99. https://doi.org/10.1111/jmi.12269.
Juneau PM, Garnier A, Duchesne C. Selection and tuning of a fast and simple phasecontrast microscopy image segmentation algorithm for measuring myoblast growth kinetics in an automated manner. Microsc Microanal. 2013; 19(4):855–66.
Topman G, SharabaniYosef O, Gefen A. A method for quick, lowcost automated confluency measurements. Microsc Microanal. 2011; 17(6):915–22.
Jaccard N, Griffin LD, Keser A, Macown RJ, Super A, Veraitch FS, Szita N. Automated method for the rapid and precise estimation of adherent cell culture characteristics from phase contrast microscopy images. Biotech Bioeng. 2014; 111(3):504–17.
Caselles V, Kimmel R, Sapiro G. Geodesic active contours. Int J Comput Vis. 1997; 22(1):61–79.
Chan TF, Vese LA. Active contours without edges. IEEE Trans Image Process. 2001; 10(2):266–77. https://doi.org/10.1109/83.902291.
Boykov Y, Kolmogorov V. An experimental comparison of mincut/max flow algorithms for energy minimization in vision. IEEE Trans Pattern Anal Mach Intell. 2004; 26(9):1124–37. https://doi.org/10.1109/TPAMI.2004.60.
Sommer C, Straehle C, Köthe U, Hamprecht FA. Ilastik: Interactive learning and segmentation toolkit. In: 2011 IEEE International Symposium on Biomedical Imaging: From Nano to Macro: 2011. p. 230–3. https://doi.org/10.1109/ISBI.2011.5872394.
ArgandaCarreras I, Kaynig V, Rueden C, Eliceiri KW, Schindelin J, Cardona A, Seung HS. Trainable weka segmentation: a machine learning tool for microscopy pixel classification. Bioinformatics. 2017. https://doi.org/10.1093/bioinformatics/btx180.
Peng H, Zhou X, Li F, Xia X, Wong STC. Integrating multiscale blob/curvilinear detector techniques and multilevel sets for automated segmentation of stem cell images. In: 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. IEEE: 2009. p. 1362–5. https://doi.org/10.1109/ISBI.2009.5193318. http://ieeexplore.ieee.org/document/5193318/.
Kong H, Akakin HC, Sarma SE. A generalized laplacian of gaussian filter for blob detection and its applications. IEEE Trans Cybern. 2013; 43(6):1719–33. https://doi.org/10.1109/TSMCB.2012.2228639.
Xu H, Lu C, Berendt R, Jha N, Mandal M. Automatic nuclei detection based on generalized laplacian of gaussian filters. IEEE J Biomed Health Inform. 2017; 21(3):826–37.
Zhang M, Wu T, Bennett KM. Small blob identification in medical images using regional features from optimum scale. IEEE Transactions on Biomedical Engineering. 2015; 62(4):1051–62. https://doi.org/10.1109/TBME.2014.2360154.
Loy G, Zelinsky A. Fast radial symmetry for detecting points of interest. IEEE Trans Pattern Anal Mach Intell. 2003; 25(8):959–73. https://doi.org/10.1109/TPAMI.2003.1217601.
Bahlmann C. Fast radial symmetry detection under affine transformations. In: Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). CVPR ’12. Washington, DC: IEEE Computer Society: 2012. p. 932–9. http://dl.acm.org/citation.cfm?id=2354409.2354741.
Qi X, Xing F, Foran DJ, Yang L. Robust segmentation of overlapping cells in histopathology specimens using parallel seed detection and repulsive level set. IEEE Trans Biomed Eng. 2012; 59(3):754–65.
Matas J, Chum O, Urban M, Pajdla T. Robust widebaseline stereo from maximally stable extremal regions. Image Vis Comput. 2004; 22(10):761–7. https://doi.org/10.1016/j.imavis.2004.02.006.
Arteta C, Lempitsky V, Noble JA, Zisserman A. Learning to Detect Cells Using Nonoverlapping Extremal Regions. Miccai. 2012; Figure 1:1–8. https://doi.org/10.1007/9783642334153.
Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M, et al. Imagenet large scale visual recognition challenge. Int J Comput Vis. 2015; 115(3):211–52.
Iglovikov V, Shvets A. Ternausnet: Unet with vgg11 encoder pretrained on imagenet for image segmentation. 2018. arXiv preprint arXiv:1801.05746.
Ronneberger O, Fischer P, Brox T. UNet: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and ComputerAssisted Intervention. 2015. https://doi.org/10.1007/9783319245744_28.
Sadanandan SK, Ranefall P, Wählby C.Feature Augmented Deep Neural Networks for Segmentation of Cells. Computer Vision – ECCV 2016 Workshops. In: Computer Vision – ECCV 2016 Workshops. Amsterdam: Springer International Publishing: 2016. https://doi.org/10.1007/9783319466040_17.
Van Valen DA, Kudo T, Lane KM, Macklin DN, Quach NT, DeFelice MM, Maayan I, Tanouchi Y, Ashley EA, Covert MW. Deep learning automates the quantitative analysis of individual cells in livecell imaging experiments. PLoS Comput Biol. 2016; 12(11):1005177.
Schmidt U, Weigert M, Broaddus C, Myers G. Cell detection with starconvex polygons. In: Medical Image Computing and Computer Assisted Intervention  MICCAI 2018  21st International Conference, Granada, Spain, September 1620, 2018, Proceedings, Part II: 2018. p. 265–73. https://doi.org/10.1007/9783030009342_30.
Xing F, Yang L. Robust nucleus/cell detection and segmentation in digital pathology and microscopy images. IEEE Rev Biomed Eng. 2016; 9:234–63. https://doi.org/10.1109/RBME.2016.2515127.
Wiesmann V, Franz D, Held C, Münzenmayer C, Palmisano R, Wittenberg T. Review of free software tools for image analysis of fluorescence cell micrographs. J Microsc. 2015; 257(1):39–53.
Meijering E. Cell segmentation: 50 years down the road [life sciences]. IEEE Signal Process Mag. 2012; 29(5):140–5.
Kasprowicz R, Suman R, O’Toole P. Characterising live cell behaviour: Traditional labelfree and quantitative phase imaging approaches. Int J Biochem Cell Biol. 2017; 84:89–95.
Loewke NO, Pai S, Cordeiro C, Black D, King BL, Contag CH, Chen B, Baer TM, Solgaard O. Automated cell segmentation for quantitative phase microscopy. IEEE Trans Med Imaging. 2017.
Ulman V, Maška M, Magnusson KEG, Ronneberger O, Haubold C, Harder N, Matula P, Matula P, Svoboda D, Radojevic M, Smal I, Rohr K, Jaldén J, Blau HM, Dzyubachyk O, Lelieveldt B, Xiao P, Li Y, Cho SY, Dufour A, OlivoMarin JC, ReyesAldasoro CC, SolisLemus JA, Bensch R, Brox T, Stegmaier J, Mikut R, Wolf S, Hamprecht FA, Esteves T, Quelhas P, Demirel Ö, Malström L, Jug F, Tomančák P, Meijering E, MuñozBarrutia A, Kozubek M, OrtizdeSolorzano C. An objective comparison of cell tracking algorithms. Nat Methods. 2017. https://doi.org/10.1038/NMETH.447.
Kolman P, Chmelík R. Coherencecontrolled holographic microscope. Opt Express. 2010; 18(21):21990–2004. https://doi.org/10.1364/OE.18.021990.
Slabý T, Kolman P, Dostál Z., Antoš M., Lošťák M, Chmelík R. Offaxis setup taking full advantage of incoherent illumination in coherencecontrolled holographic microscope. Opt Express. 2013; 21(12):14747–62. https://doi.org/10.1364/OE.21.014747.
Davies H, Wilkins M. Interference microscopy and mass determination. Nature. 1952; 169(4300):541. https://doi.org/10.1038/169541a.
Barer R. Refractometry and interferometry of living cells. J Opt Soc Am. 1957; 47(6):545–56. https://doi.org/10.1364/JOSA.47.00054.
Koos K, Molnar J, Horvath P. DIC Microscopy Image Reconstruction Using a Novel Variational Framework. 2015. International Conference on Digital Image Computing: Techniques and Applications (DICTA). 2015. https://doi.org/10.1109/DICTA.2015.7371252.
Su H, Yin Z, Kanade T, Huh S. Phase contrast image restoration via dictionary representation of diffraction patterns. Med Image Comput Comput Assist Interv. 2012; 15(3):615–22.
Whitaker RT. A levelset approach to 3d reconstruction from range data. Int J Comput Vis. 1998; 29(3):203–31.
Breiman L. Random forests. Mach Learn. 2001; 45(1):5–32. https://doi.org/10.1023/A:1010933404324.
Lindeberg T. Feature detection with automatic scale selection. Int J Comput Vis. 1998; 30(2):79–116.
Ram S, Rodríguez JJ. Symmetrybased detection of nuclei in microscopy images. In: 2013 IEEE International Conference on Acoustics, Speech and Signal Processing: 2013. p. 1128–32. https://doi.org/10.1109/ICASSP.2013.6637826.
Parvin B, Yang Q, Han J, Chang H, Rydberg B, BarcellosHoff MH. Iterative voting for inference of structural saliency and characterization of subcellular events. IEEE Trans Image Process. 2007; 16(3):615–23. https://doi.org/10.1109/TIP.2007.891154.
Lu C, Xu H, Xu J, Gilmore H, Mandal M, Madabhushi A. Multipass adaptive voting for nuclei detection in histopathological images. Sci Rep. 2016; 6:33985. https://doi.org/10.1038/srep33985.
Parvati K, Rao P, Mariya Das M. Image segmentation using grayscale morphology and markercontrolled watershed transformation. Discret Dyn Nat Soc. 2008; 2008:384346. https://doi.org/10.1155/2008/384346.
Daněk O, Matula P, OrtizdeSolórzano C, MuñozBarrutia A, Maška M, Kozubek M. Segmentation of touching cell nuclei using a twostage graph cut model. In: SCIA: 2009. Oslo: Springer. p. 410–19. https://doi.org/10.1007/9783642022302_42.
Zhou Y, Kuijper A, Heise B, He L. Cell Segmentation Using Level Set Method. Linz: Johannes Kepler University; 2007. https://pdfs.semanticscholar.org/33b2/7d8696026b5f4f7eac7095ff30e9c920621a.pdf.
Ikonen L, Toivanen P. Shortest routes on varying height surfaces using graylevel distance transforms. Image Vis Comput. 2005; 23(2):133–41.
Maška M., Ulman V, Svoboda D, Matula P, Matula P, Ederra C, Urbiola A, España T, Venkatesan S, Balak DMW, Karas P, Bolcková T., Štreitová M, Carthel C, Coraluppi S, Harder N, Rohr K, Magnusson KEG, Jaldén J, Blau HM, Dzyubachyk O, Křížek P, Hagen GM, PastorEscuredo D, JimenezCarretero D, LedesmaCarbayo MJ, MuñozBarrutia A, Meijering E, Kozubek M, OrtizdeSolorzano C. A benchmark for comparison of cell tracking algorithms. Bioinformatics. 2014; 30(11):1609–17. https://doi.org/10.1093/bioinformatics/btu080.
Acknowledgements
We thank prof. Radim Chmelik from Brno University of Technology for enabling the DIC microscopy in their facility and Tomas Slaby from Tescan a.s., Brno, for their kind help with operating the quantitative phase microscopy and with processing of the data using their software.
Funding
This work was supported by the Czech Science Foundation GACR 1824089S and by funds from the Faculty of Medicine, Masaryk University to Junior researcher (Jaromir Gumulec). Josef Jaros was supported by project of Masaryk University (MUNI/A/1298/2017). We acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research. None of the funding bodies had any role in the design of the study and collection, analysis and interpretation of data, and in writing the manuscript.
Availability of data and materials
Annotated image dataset and image reconstructions used in our study are available for download in the Zenodo repository (https://zenodo.org, Digital Object Identifier: https://doi.org/10.5281/zenodo.1250729). Matlab code is available at GitHub https://github.com/tomasvicar/Cellsegmentationmethodscomparison.
Author information
Authors and Affiliations
Contributions
TV designed the workflow, selected segmentation methods, performed analysis in Matlab and Python and wrote manuscript. JB performed in vitro experiments, designed experiment, JJ performed HMC and helped with in vitro experiments, FJ designed foregroundbackground segmentation and seedpoint extraction structure, helped with trainable approaches and graph cut, RK helped with selection of segmentation strategies and corrected the manuscript, MM provided ideas for segmentation, supported in vitro experiment, JG designed experiment, wrote manuscript and coordinated work. All the authors have read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
not applicable.
Consent for publication
not applicable.
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Additional file
Additional file 1
Optimal values for parameters of individual reconstruction methods (xlsx table). * highest value not reducing sensitivity, ** not learned because of identification of small number of regions. nan, not a number. (XLSX 17 kb)
Rights and permissions
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.
About this article
Cite this article
Vicar, T., Balvan, J., Jaros, J. et al. Cell segmentation methods for labelfree contrast microscopy: review and comprehensive comparison. BMC Bioinformatics 20, 360 (2019). https://doi.org/10.1186/s1285901928808
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s1285901928808