Discriminative Segmentation of Microscopic Cellular Images

Microscopic cellular images segmentation has become an important routine procedure in modern biological research, due to the rapid advancement of fluorescence probes and robotic microscopes in recent years. In this paper we advocate a discriminative learning approach for cellular image segmentation. In particular, three new features are proposed to capture the appearance, shape and context information, respectively. Experiments are conducted on three different cellular image datasets. Despite the significant disparity among these datasets, the proposed approach is demonstrated to perform reasonably well. As expected, for a particular dataset, some features turn out to be more suitable than others. Interestingly, we observe that a further gain can often be obtained on top of using the "good" features, by also retaining those features that perform poorly. This might be due to the complementary nature of these features, as well as the capacity of our approach to better integrate and exploit different sources of information.