![](https://csdnimg.cn/release/download_crawler_static/10960690/bg4.jpg)
detection (Bao et al., 2006; Keller et al., 2008, 2010)is
computationally efficient but subject to shrinking bias,
which technically serves the purpose of detection rather
than segmentation. Segmentation by gradient flow
tracking is very sensitive to object texture (Li et al., 2007).
The Watershed method is also fast yet produces loose
boundaries that cover the background (Fernandez et al.,
2010; Olivier et al., 2010). Discrete Markov random field
optimization allows for incorporation of prior information
such as shape (X.L. et al., 2012, IEEE, conference), but the
underlying pipeline is overly complicated and, thus,
impractical. It is important to note the growing trend of
developing generic, trainable software frameworks that
are based on machine learning methods that can interact
with biologists and solve a variety of problems (Carpenter
et al., 2006; A. Sommer et al., 2011, IEEE, symposium).
However, the tradeoff for such generality is specificity.
These tools usually do not capture the very essential char-
acteristics of nuclear imaging and so do not provide a pre-
cise analysis and quantitation.
To meet this current need, our goal was to develop a tool
that would make automated cell segmentation feasible and
efficient in analyzing data from higher organisms, as has
been applied to less complex data from lower organisms,
such as in bacteria (Locke and Elowitz, 2009). The objective
was to assemble a segmentation framework that is accurate
enough to allow high-fidelity analysis over a variety of
images while being robust enough to make it practical for
routine use across laboratories. A major goal was that the
software had to be a simple and intuitive application that
could run on a desktop computer having routine process-
ing power. Usability is a particularly important feature
that we sought to incorporate in modular interactive
nuclear segmentation (MINS), as this has been raised as
an important issue for bioimaging software development
(Carpenter et al., 2012). We wanted to allow biological re-
searchers to analyze large 3D imaging data with only a
few mouse clicks and minimum parameter tuning.
Here, we report an efficient and user-friendly nuclear seg-
mentation and quantitation framework (i.e., MINS) for the
analysis of cohorts of cells in both stem cell cultures and
in preimplantation stage mouse embry os. Our method
consists of three core cascaded modules: detection, seg-
mentation, and classification. Detection provides accurate
localization of cell nuclei only. Segmentation expands the
detection output to cover the full nuclear body. Finally,
classification serves multiple purposes, including the sepa-
ration of multiple embryos and removal of outliers, as well
as the classification of inner and outer cells (ICM versus TE
cells) within blastocyst-stage embryos. MINS is hosted at
http://katlab-tools.org.
RESULTS
Core Algorithmic Components of MINS
We chose specimens of increasing complexity for analysis
using MINS software (Figure 1C). Mouse XEN stem cells
representing the PrE lineage of the blastocyst grow as
adherent monolayer cultures (Artus et al., 2012; Kunath
et al., 2005; Niakan et al., 2013). By contrast, mouse ES cells
grow as adherent multilayered dome-shaped cultures
(Nichols and Smith, 2011). Like ES cells, preimplantation
mouse embryos comprise spatially complex cohorts of cells
(Rossant and Tam, 2009). A direct segmentation of every
nucleus is computationally difficult and also error prone;
for example, the active contour method is prone to under-
segmentation (Roeder et al., 2012). We therefore opted to
break down the problem into two steps: a detection module
that identifies each nucleus followed by a segmentation
module that propagates this ID information to the entire
body of the respective nucleus. We also added a classifica-
tion module for TE versus ICM cell lineage ID, which was
based on inner versus outer cell position, respectively,
within preimplantation mouse embryos (Figure 1D).
In brief, the core of MINS consists of three algorithmic
components: detection, segmentation, and classification.
Each component was devised and tailored according to
the specific characteristics of cell culture and mouse
embryo imaging experiments. The underlying algorithms
are chosen so that the overall pipeline satisfies our goals:
high performance and high usability.
Step 1: Detection
For the detection of nuclei, we applied the multiscale blob
detection technique developed previously (X.L. et al.,
2012, IEEE, conference). It uses the very robust eigenvalue
of the image Hessian matrix to identify nuclei and also uses
scale-space analysis to suppress noise because noise does
Figure 1. Image Analysis of Cells and Mouse Embryos and a Schematic of Preimplantation Embryo Development
(A) Schematic showing the experimental setup used for static and live imaging of stem cell and mouse embryo specimens. Notably, samples
are maintained in liquid culture, and images are acquired on inverted microscope systems.
(B) Examples of imaging acquisition of 3D static immunostaining (left) or 3D live imaging of fluorescent reporter (right).
(C) Schematic diagram showing 2D, 3D, and 4D image data acquisition and analysis.
(D) Differential interference contrast (DIC) images of CAG:H2B-GFP transgenic fluorescent reporter expressing embryos at two-cell,
compact morula, early, and late blastocyst stages merged with 2D and 3D renderings of GFP channel showing nuclei labels and a schematic
diagram of lineage specification during preimplantation development (Schrode et al., 2013). Scale bar, 20 mm.
Stem Cell Reports j Vol. 2 j 382–397 j March 11, 2014 j ª2014 The Authors 385
Stem Cell Reports
MINS: Modular Interactive Nuclear Segmentation