Integrating Input from Human Experts into Prototype-Based Classifier Learning

An expert is able to tell the system developer in many image-related tasks what a prototypical image should look like. Usually he will choose several prototypes for one class, but he cannot provide a good and large enough sample set for the class to train a classifier. Therefore, we mapped his technical procedure into a technical system based on proper theoretical methods that assist him in acquiring the knowledge about his application and furthermore in developing a classifier for his task. This system helps him to learn about the clusters and the borderlines of the clusters even when the data are very noisy as is the case for microscopic cell images in drug discovery, where it is unclear if the drug will produce the expected result on the cell parts. We describe in this paper the necessary functions that a prototype-based classifier should have. We also use the expert’s estimated similarity as a new knowledge piece and based on that we optimize the similarity. The test of the system was carried out on a new application on microscopic cell image analysis the study of the internal mitochondrial movement of cells. The aim was to discover the different dynamic signatures of mitochondrial movement. Three results of this movement were expected: tubular, round, and dead cells. Based on our results we can show the success of the developed method.