Segmentation of phase or DIC images can be difficult because methods based on intensity thresholding are not appropriate. In order to overcome this problem, we used an approach based on machine learning, implemented in ImageJ as a plugin called Advanced Weka Segmentation. In its simplest form, one must train the system by marking areas which are assigned to a class (in our case, cell and background), and selecting classifiers (eg, variance, various projections, etc). This training set can then be used to classify other images.
As is often the case, we had available, in addition to the DIC images, a high quality DAPI staining of the nuclei, as well as immunofluorescent staining of one other protein of interest. This provided addition information which was used to select training areas automatically. Since all cells are assumed to contain nuclei, and all nuclei are assumed to be contained within cells, the DAPI images were used to provide the initial training regions for cells, which the segmentation plugin then used to classify the rest of the image. However, this still required that the background training regions be manually selected.
This problem was solved by recognizing that cells have a much larger spatial variance than background. Although variance alone could not be used to obtain an accurate segmentation, it could, in conjunction with suitable morphological operators, provide a good estimate of regions which are, with high probability part, of the background. Thus, our segmentation strategy could automatically select a training set, and then use machine learning to refine the segmentation.
Finally, we recognized that the approach taken to automatically select the background train- ing set could be inverted to select the foreground (cells) training set. This was done in the latest version of an ImageJ/FIJI macro, which also processes an entire directory (in this case, 64 images) automatically.
In this presentation, we will describe the use of machine learning in the segmentation of DIC images of lymphocytes. In particular, we will present the methods which enabled us to automatically select the foreground and background training sets (thus enabling fully automatic segmentation. Then we will show the output which was generated by the machine learning plugin. We will also discuss the segmentation of time lapse phase images of PC3 cells (a prostate cancer line), which are highly textured, and for whom a variance based approach was more appropriate.
segmentation, DIC , phase , microscopy
Presenting author: Aryeh Weiss
Organisation: Bar Ilan University
co-authors: Keren Ailon, Hermona Soreq, Naomi Melamed-Book, Tzufit Lifshitz, Hana Panet