Circularity Features for Nuclei Segmentation

Computational Aspects of Biological Information, 2016

We introduce a model for circumference and circle likelihoods that proves effective as nuclei boundary and interior descriptors in a multi-layer random forest for nuclei segmentation. We thus add to recent work on learning 'context cues' for semantic segmentation by adding 'geometric cues' through such circularity features. Our predictive model consists of three stacked random forests, where the forests at layers 2 and 3 see the output of the previous trained layer -- in addition to the features computed from the original image. It is on these layers that we compute circle boundary and interior likelihoods (in addition to context features) from the nuclei probability maps produced as output by the previous layers. The model, trained on less than a dozen labeled images, was successfully deployed to automatically segment thousands of nuclei in images of mouse visual cortex.

Application: We developed an in vivo paradigm to study sensory-dependent gene expression in primary visual cortex. Adult male mice (6-7 weeks) were housed in complete darkness for one week followed by one hour of light exposure; control mice remained dark-housed (0 hr condition). Fluorescence in situ hybridization (FISH) for the early response transcription factors (ERTFs) cFos and Nr4a1 from the visual cortex of mice after the control and one hour light stimulus showed robust gene-induction between the two conditions. ERTFs regulate the diversity of subsequent transcriptional responses to sensory activity in the brain. While many ERTFs have been identified, it remains unknown how their expression levels are coordinated at the level of single cells, and among different cell-types. Do individual cells express different subsets of ERTFs? Are there compensatory transcriptional mechanisms that regulate the total number of ERTFs? Testing these hypotheses requires single-cell resolution for unambiguous cell-type identification and mRNA puncta quantification. We took advantage of our automated pipeline to address these questions in multiple cell types across several early response transcription factors. We find that in several cell types, including excitatory neurons, the expression levels of early response transcription factors are highly correlated. Correlations between the mRNA transcripts of ERTFs Nr4a1, cFos, and Egr1 are shown for thousands of individual Vglut1+ cells.This evidence suggests that sensory stimulation results in a tightly regulated co-induction of a core set of ERTFs across cell-types which act in concert to stimulate downstream transcription.


BibTeX: @misc{CABICFNS2016, author = {M. Cicconet and D. Hochbaum and D. Richmond and H. Elliott}, title = {Circularity Features for Nuclei Segmentation}, howpublished = {Computational Aspects of Biological Information [Poster]}, year = {2016}, note = {Boston} }