Image-Based Cell Quality Assessment: Modeling of Cell Morphology and Quality for Clinical Cell Therapy - Computational Modeling in Tissue Engineering - page 211

Fig. 7 Schematic illustration of the cell contamination evaluation model and its construction.
a Schematic illustration of decision tree model for cell morphology modeling. Combination of
morphological parameters (left) could be determined using decision tree (right). b Practical
scheme for constructing cell contamination evaluation model. Model can be constructed by single
cell type culture images if all cell objects could be effectively extracted from the image data.
From the unknown image of co-cultured cells, fibroblasts mixed in keratinocytes in this case, the
constructed model can predict the contamination rate

b

analysis. Considering the existence of ''sub-populations'' in both types of cells,
a multiple-grouping method, decision tree modeling, was applied. The concept of the
cell contamination evaluation model using cellular images is illustrated in Fig. 7 .
To obtain input features from cell culture images, we collected a total of
270 phase contrast microscopic images (10 9) of cultured normal human dermal
fibroblasts (NHDF; KURABO, Osaka, Japan) and normal human epithelial
keratinocytes (NHEK; KURABO) grown at 37C in the presence of 5% CO 2 .
NHDFs were maintained in modified Eagle's medium (MEM, Life technologies)
with 10% FBS, and NHEKs were maintained in EpiLife-KG2 (KURABO). For
image acquisition, NHDFs were mixed with NHEK (0%/1%-30%/100%) in an
NHDF contamination model of NEHKs. A total of 1912 phase contrast micro-
scopic images (49) were acquired from five view fields per well of a six well plate
over 8 h periods for 5 days using BioStation CT (Nikon Corporation). All images
(.bmp files) were processed using MetaMorph software (Molecular Devices) and
our original program with original filter sets, as described in the prior section.
Using integrated morphometry analysis, the number of objects was measured
together with 19 individual morphological features in MetaMorph, such as total
area, hole area, relative hole area, perimeter, width, height, length, breadth, fiber
length, fiber breadth, shape factor, elliptical form factor, inner radius, outer radius,
mean radius, equivalent radius, pixel area, area, orientation. From both cell types,
totally 2,792,527 objects were measured with these morphological parameters.
In the modeling process, input features were selected by recursive partitioning and
regression tree (rpart) modeling using R statistics.
For the modeling of teacher signals in rpart, we carefully classified the cellular
morphology types of both cells, and identified nine types of objects in the image
data: (1) F_n: objects with the most typical fibroblast morphology (2) F_s: objects
with small fibroblast morphology in the process of expansion (3) F_o: objects that
represent a fusion of several overlapping fibroblasts (4) K_n: objects with the most
typical keratinocyte morphology (5) K_s: objects with small keratinocyte mor-
phology in the process of expansion (6) K_o: objects with holes after binarization,
indicating the halo in the middle of cell (commonly caused by hill-top structure of
cells) (7) K_c: objects with a ''C'' shape, indicating that the halo is relatively large
in cell area (8) n_s; tiny non-cellular objects (9) n_l: long, but tiny non-cellular
objects (Fig. 8 ). Groups 8 and 9 represented the common noise found in both cell
types. From the modeling, we chose 100 objects from the total cells which fit to
each cluster from 100% fibroblast and 100% keratinocyte images, constructed a
decision tree model by tenfold cross validation, and examined the accuracy of the