Image-Based Cell Quality Assessment: Modeling of Cell Morphology and Quality for Clinical Cell Therapy - Computational Modeling in Tissue Engineering - page 216

points. We also used a fully auto-scheduled and auto-focusing image acquisition
system, BioStation CT (Nikon corporation) to minimize the image acquisition bias.
(ii) Population bias: we have shown that population bias can be classified using decision
tree models, too. We also remove the ''common population'' before analysis by using
our original noise-reduction algorithm. (iii) Cell and culture bias: we routinely used
three or more lots or passages of cells in every experiment. We also set image acqui-
sition periods of less than 8 h to obtain more information during the average doubling
time. (iv) Processing bias: we optimized the universal threshold by examining the
threshold ranges of randomly picked images from 5 to 10 cell lines at different time-
points for the best threshold that resulted in minimum error in the cell count. We also
strongly recommend using pattern matching algorithms to recognize objects in the
image, such as the software ''CL-Quant (Nikon Corporation).
To assist clinical cell therapy with such machine learning models combined with
image analysis, ''what to predict'' should be carefully researched. There are strong
expectations and requirements in clinical cell therapy toward non-invasive cell quality
assessments. A large number of ''qualities'' and ''events'' are expected to be predicted
in the clinical setting. However, to build a prediction model, the teacher signal should
be carefully selected and defined. For example, beta galactosidase activity is a
biological marker of cellular senescence [ 12 ]. However, the efficiency of beta galac-
tosidase detection in a single stained image is actually very low. It may correlate with
senescence when the data are averaged in ''1 well,'' but not in ''images that do not cover
all the wells.'' In such cases, differences in detection power exist between biological
assays and image detection. Without knowing the detection power of the target event,
inconsiderate modeling and data acquisition will result in no harvest. Moreover, there
are some cases where a ''1 marker measurement'' is not sufficient for definition. For
example, in the case of defining stemness of cells, the single marker staining result
would not produce the expected stemness prediction model.
Additionally, the ''acceptable accruracy'' of model should be also carefully dis-
cussed by the users of these machine-learning models for cell quality evaluations.
The accuracy should also be carefully checked with the aspect of consistency and
reproducibility. In many cases, the non-labeling and real-time estimation of cell
quality are difficult to compare its performance with conventional methods, because
the image-based cell quality assessment has uncomparable advantage features.
However, such comparison difficulty never means that conventional assay can be
eliminated. The acceptable accuracy of models should be defined by each medical
facility with their specific verification data, and should be used only to reduce the
excessiveness and unstableness in the conventional cell production process.
7 Conclusion
Machine learning modeling algorithms have great potential in image-based cell
quality assessments. Based on the selection of appropriate models with sufficient
quantities of images for predicting target events, our results suggest a potential of