Claudia Dach

Suche


A4 Automated and interactive learning image-analysis for fluorescence-microscopy by the example of microbial effectors in host cells

Principal investigator
Thomas Wittenberg

Mentor
Ralf Palmisano

Interactive cell image-segmentation for fluorescence-microscopy based on the GraphCuts algorithm

In order to obtain further knowledge about the impact of microbial effectors on host cells, automated image processing tools robustly evaluating the image content are strongly required. Hence, as a first step, a detection and segmentation of cells depicted in the fluorescent micrograph is needed. In some cases this segmentation can be performed in a fully automatic way. In more complex cases where e.g. cells are seeded more densely on the slide and where only a subset of the depicted cells can be used for evaluation (see Figure 1) a manual segmentation is needed. Nevertheless, the interactive segmentation task is very time consuming, tedious as well as error-prone. Thus, an interactive segmentation approach combining both, manual and automated cell image analysis would be reasonable, as the accuracy and expert knowledge of the manual approach can be combined with the speed of automatic segmentation methods.

Figure 1: Macrophages seeded densely on the slide (left side) and the corresponding hand segmentation (right side). Note, that even an expert is not able to correctly determine the boundary of all cells. Therefore, only the best visible cells are selected and marked.

In order to obtain an appropriate interactive method for cell image segmentation, the so-called GraphCuts approach proposed by Boykov & Kolmogorov will be used as a guideline in this research project. The fundamental idea of this interactive segmentation approach is that the biologist roughly marks the objects of interest (connected foreground pixels) as well as pixels representing the image background. Based on this exemplary input data, the GraphCuts algorithm tries to segment the objects of interest (see Figure 2). If results are unsatisfactory, a correction by marking regions in which the algorithm failed can be conducted. After incorporating the new input data, the GraphCuts algorithm will be employed again.
A short description of how the GraphCuts algorithm works will be explained in the following: A graph consists of a set of nodes as well as a set of edges interconnecting the nodes. Within an image the pixels are regarded as the nodes of the graph, while the proximate neighbourhoods of adjacent pixels are regarded as directed and weighted edges that connect them. The weights of the edges between adjacent pixels are obtained by a cost function. Additionally, the graph contains two special nodes referred to as terminal nodes. One is called the source node, while the other is denoted as the sink node. In order to visualize the concept of the so-called max-flow min-cut theorem, one can interpret graph edges as directed pipes with capacities equal to the edge weights. The maximum amount of water sent from the source node to the sink node can be seen as the maximum flow. The saturated path found by the maximum flow which corresponds to the cost of a minimum cut separates the pixels belonging to the foreground (the source) and the pixels belonging to the background (sink).
In this current research project, the GraphCuts algorithm will be employed and extended as needed with the target that a variety of fluorescent stained specimens can be detected and segmented reliably. The feasibility of results will be determined with the help of ground truth data provided by an expert.

Figure 2: Illustration of the GraphCuts principle: The left image depicts the lines drawn by the user: non-transparent blue lines indicate typical background regions while the red lines represent the cells. The (simulated) result is depicted in the right image. Both images have been created with Photoshop for illustration purposes.

Publications

Dach, C., Held, C., Wenzel, J., Gerlach, G., Lang, R. and Palmisano, R. (2011). Evaluation of an interactive cell segmentation for fluorescence.. PMIAAB. pdf

Dach, C., Held, C., Palmisano, R., Friedl, S. and Wittenberg, T. (2011). Evaluation of input modalities for the interactive image segmentation of fluorescent micrographs. In: Proc's Annual Scientic Workshop Biomed. Technol., Suppl. 1. doi:10.1515/BMT.2011.571.

Dach, C., Held, C., Wenzel, J., Lang, R., Friedl, S., Palmisano, R. and Wittenberg, T. (2011) Evaluierung der Interaktiven Zellsegmentierung mit Hilfe des Graph Cut Algorithmus in Fluoreszenz-Mikroskop-Bildern. In: Proc's 56. GMDS-Jahrestagung & 6. DGEpi-Jahrestagung, 510- 511.

 

Presentations

September 2010 2nd Annual Retreat, Erlangen School of Molecular Communication, Kloster Banz, Bad Staffelstein, Germany
Interactive cell image-segmentation for fluorescence-microscopy based on the GraphCuts algorithm”
Talk