Modern microscopy allows the biologist or chemist to produce thousands of cell images in one single experiment. A common analysis of these images usually encompasses detection of the cells (segmentation) and subsequent classification, e.g., to determine the proportion of a specific cell type or their reactivity.
If analysis were performed completeley manually, the procedure would need days. Hence, automatic methods are required. However the drawback is that existing fully-automatic cell image segmentation methods are very problemspecific and currently require time-intensive parametrization or even custom scripting to cover the big variety of cell images.
A promising solution for this dilemma is the incorporation of human knowledge. However, in order to make best use of the human expert, it is crucial to limit the involvement of the expert to a few, carefully selected examples.
Consequently an active segmentation system is firstly initialized by a few user-provided cell samples to create a rough segmentation model. Secondly, the intermediate segmentation results are used to detect regions of high confusion, that is, areas where the segmentation model is "unsure" about the output to query the human expert for a refinement/validation until a satisfying result is reached.
This active segmentation can then be followed by a similar active classification process to build a suitable classifier.