Learning to Count and Detect Cells with Partial Annotation
In high through-put screening there is a need to detect and count cells of a particular type, for example detecting cancerous cells in histopathology images whilst ignoring other non-cancerous cells. This can be achieved automatically using computer vision and machine learning. However, in order to train the system a user must first provide ground truth annotation of the cells, e.g. by `dotting' the cells of interest in multiple samples. This can be quite onerous.
The purpose of this project is to relieve the burden of complete annotation, and investigate machine learning methods where a user provides only partial annotation, for example by annotating some examples of cells of interest and some examples of those that are not.
Data from existing studies will be used to evaluate the new methods which will be compared against machine learning method(s) using complete annotations.