Abstract
We introduce software assistants – bots – for the task of analyzing image-based transcriptomic data. The key steps in this process are detecting nuclei, and counting associated puncta corresponding to labeled RNA. Our main release offers two algorithms for nuclei segmentation, and two for spot detection, to handle data of different complexities. For challenging nuclei segmentation cases, we enable the user to train a stacked Random Forest, which includes novel circularity features that leverage prior knowledge regarding nuclei shape for better instance segmentation. This machine learning model can be trained on a modern CPU-only computer, yet performs comparably with respect to a more hardware-demanding state-of-the-art deep learning approach, as demonstrated through experiments. While the primary motivation for the bots was image-based transcriptomics, we also demonstrate their applicability to the more general problem of scoring “spots” in nuclei.
Footnotes
↵* HMS Image and Data Analysis Core. Corresponding author for software and image analysis matters, such as the machine learning model. cicconet{at}gmail.com
↵† HMS Department of Neurobiology, Harvard University Society of Fellows, and Howard Hughes Medical Institute. Corresponding author for all experimental considerations, such as microscopy and in situ techniques. drhochbaum{at}gmail.com
↵‡ HMS Image and Data Analysis Core.
↵§ HMS Department of Neurobiology, and Howard Hughes Medical Institute.
