Cross-Domain Exemplars for Cell Counting

Cell counting plays a major role in biological studies and medical diagnosis. To achieve automatic counting of cells in an image, there is a need to create annotated datasets which is not practical. Further, the quality of the exemplars plays an important role in automatic cell counting. In this paper, we introduce a novel approach of weakly supervised learning by automatically selecting good quality exemplars obtained from cross-domain training. We exploit the cell structure similarity to achieve cross-domain knowledge transfer. Firstly, two domain adaptive Faster R-CNN networks are trained on RGB microscopy images and background subtracted images respectively. The candidate exemplars obtained from each trained network are merged based on their confidence scores, and then through K-Means clustering approach good quality exemplars are selected based on their appearance variability. A density map prediction network is trained using the chosen exemplars to count the number of cells. In our experiments on the MBM, DCC and VGG datasets, we achieved close to state-of-the-art results using far less training samples.