Summary
Objectives: The cultivation of adherently growing cell populations is a major task in the field
of adult stem cell production used for drug discovery and in the field of regenerative
medicine. To assess the quality of a cell population, a crucial event is the mitotic
cell division: the precise knowledge of these events enables the reconstruction of
lineages and accurate proliferation curves as well as a detailed analysis of cell
cycles. To serve in an autonomous cell farming framework, such a detector requires
to work reliably and unsupervised.
Methods: We introduce a mitosis detector that is using a maximum likelihood (ML) estimator
based on morphological cell features (cell area, brightness, length, compactness).
It adapts to the 3 phases of cell growth (lag, log and stationary phase). As a concurrent
model, we compared ML with kernel SVMs using linear, quadratic and Gaussian kernel
functions. All approaches are evaluated for their ability to distinguish between mitotic
and nonmitotic events. The large, publicly available benchmark data CeTReS (reference
data set A with > 240,000 segmented cells, > 2,000 mitotic events) is used for this
evaluation.
Results: The adaptive (unsupervised) ML approach clearly outperforms previously published
non-adaptive approaches and the linear SVM. Furthermore, it robustly reaches a performance
comparable to quadratic and Gaussian SVM.
Conclusions: The proposed simple and label free adaptive variant might be the method of choice
when it comes to autonomous cell farming. Hereby, it is essential to have reliable
and unsupervised mitosis detection that covers all phases of cell growth.
Keywords
Time-lapse imaging - cell tracking - mitosis detection - image processing - phase
contrast microscopy
