Figure 1.
Microscopic image examples of different TMA-cores.
(a) Representation of all 3 channel of a fluorescently stained core in RGB colorspace. Glyphs originated due to TMA's preparation. Red representing the stromal marker (Vimentin), green the tumor marker (CK19) and blue the DAPI channel highlighting the cell nuclei; (b) the DAPI channel of (a) as an intensity image: in general tumor cells are darker and tighter connected than stromal cells; (c) another DAPI image of a core with a high density of cells; (d) an example of a core with a lower density of cells shows the high heterogeneity among the cores.
Figure 2.
A Flowchart showing the single steps of our methodology.
After obtaining the images, pre-processing steps enhance the image quality and watershed segmentation for the subsequent segmentation is applied. Accordingly the cell graphs are generated and features are computerized. The last step uses a SVM to classify the graphs as either tumor or stroma.
Figure 3.
The different image processing steps and the graph generation steps.
(a) original image of the DAPI-channel; (b) image after shading correction and noise removal; (c) result of the watershed segmentation, the segmented cells are highlighted by green contour; (d) the image after removal of single cells; (e) showing the cells which were connected via the graph generation step in the same color (cells marked with the same color belong to the same sub-graph); (f) cell graph representation of the cells. The red dots are the nodes which represent the cells, the black lines are the edges between them.
Figure 4.
Conceptional representation of cell graphs.
(a) Artificial sketch of 3 different 3 cell type: tumor cells in blue, lymphocytes in white and in purple fibroblast. (b) Cell graph representation of (a). Cells are depicted as nodes and links between them represent biological relations.
Table 1.
Graph metrics used to train the classifier and their description.
Figure 5.
The results of the classification.
(a–d) showing the original RGB core images; (e–h) showing the corresponding DAPI channel as an intensity image of the cores (a–d); (i–l) results of the classification step, green = cells classified as tumor cells, blue = cells classified as stroma cells.
Table 2.
Accuracy of the watershed cell segmentation.
Table 3.
The F-scores of each feature in descending order.
Table 4.
The average classification accuracies.
Table 5.
The accuracies of the slides of the test set with the additional single node classification.