Cell Detection in Domain Shift Problem Using Pseudo-Cell-Position Heatmap

The domain shift problem is an important issue in automatic cell detection. A detection network trained with training data under a specific condition (source domain) may not work well other conditions (target domain). We propose unsupervised adaptation method for using the pseudo-cell-position heatmap, where centroid becomes peak Gaussian distribution map. In prediction result target domain, even if location correct, signal around often has non-Gaussian shape. heatmap re-generated positions predicted to have clear Our selects confident heatmaps Bayesian and adds them next iteration. can incrementally extend from source semi-supervised manner. experiments 8 combinations of domains, proposed outperformed existing methods.