Image-Based Cell Quality Assessment: Modeling of Cell Morphology and Quality for Clinical Cell Therapy - Computational Modeling in Tissue Engineering - page 206

cell yield. The correlation ratio of the model using increase rate of cell count was
0.50 (Fig. 4 d), and the model using viable cell density after 1 day was 0.59
(Fig. 4 e). These results strongly suggest that morphological feature selection
combined with not only ''morphologies'' but also ''changes of morphologies'' is
effective in image-based analysis. Since many previous image-based cell analysis
works depend on ''intentionally selected'' morphological features, our work points
out the importance of introduction of machine learning approach to construct
better models for practical usage. Cell count, one of the easiest parameter that
could be raised for cell yield prediction, was selected in the best prediction model.
However, it should be noted that cell count was the ''last parameter'' selected for
constructing the model, indicating that it only works in combination with mor-
phological information. This is also clear from the above-compared prediction
results in the model using a single parameter, ''cell growth rate.'' Together with
such combinational effects of parameters, another important finding using this
model is that the ''exact period of culture'' was quantitatively pointed out. For
example, cell count (nearly equal to cell growth) is important in the first 24 h, and
not very informative in the latter period. Such a timing definition is extremely
important for setting the image acquisition schedule, and also for defining the
prediction date in the early cell culture process.
Fibroblasts change their morphology from the sharp spindle shape to the flat
and polygonal shape when their growth activity decreases. The automatically
determined combination of morphological parameters directly correlated with this
known

morphological change, indicating that the expert's feeling could be

effectively modeled with this technique.
From the data, we arrived at three conclusions: (1) morphological cell infor-
mation is informative for cell growth prediction, (2) objectively selected param-
eters are more effective in cell growth prediction than the ones selected on the
basis of feeling, and (3) a combination of multiple parameters is more effective in
the prediction than a single parameter. It should also be noted that such quanti-
tative cell quality prediction can be further extended to cell differentiation rate
prediction [ 7 ]. Kagami et al. [ 7 ] have shown that osteogenic differentiation status
of human mesenchymal stem cells after 2 weeks of cell culture could be predicted
by the morphological features priorily.

4 Discriminant Function Model for Image-Based
Cell Quality Assessment

Discriminant function analysis is a statistical analysis to predict a categorical
dependent variable using one or more continuous or binary independent variables.
Compared to the regression analysis model, the discriminant function model
satisfies clinical cell therapy requirements for assessing ''binary categorical
events.'' This is because many events in the cell production process cannot be
quantitatively measured or measured data is usually categorized even if they could