C. Brandorff, P. Rautek (1), W. Weninger (2), E. Gröller (1)
„ONTOVIS“– Illumination Correction of Histological Cut Imagery for Ontology Data
The “Center for Anatomy and Cell Biology” (2), a department of the Medical University of Vienna, developed a technique to generate series of exactly aligned images of biomedical specimens. This technique is used to study the impact of different genes during morphogenesis. However this technique, also introduces artefacts which make it hard to use automatic or semi automatic segmentation algorithms. The topic of this internship was to reduce the artefacts in the images. In this work we concentrated on the artefact resulting from the uneven illumination.
Figure 1: A specimen embedded in a polymer block.
The specimens are subjected to immunohistochemistry or whole mount in situ hybridization in order to specifically stain proteins or sites of gene expression. After this the tissue is embedded in a medium based on polymer to generate a small hard block for slicing. An example for a block containing a specimen can be seen in Figure 1. Then the block-surface is photographed with a digital camera from the front using two separate filters, one to capture the morphology (morpho image), the other to capture the signal generated by the marker (signal image). After these images are generated a thin slice is cut from the front. The process of taking photographs and slicing is repeated until stopped manually or the object of interest is completely cut. Since the thickness of the slices is in the order of µms, a block of an embryonic chicken generates typically about 2000 slices. The image generation takes several hours. The data generated with this technique is in the order of several gigabytes (two images per slice each 2560 x1920 stored with 8 bit precision results in approximately 2.3GB).
Figure 2: A morpho image of a chicken embryo.
Figure 2 shows a morpho image and Figure 3 shows the corresponding signal image of a chicken embryo. As stated above there are several artefacts in this images. In the morpho image of Figure 2, the uneven illumination across the image can be seen. Figure 4shows the slicing artefacts generated by the blade. Since the images are always taken from the front, the structures below the photographed slices shine through. This effect results in the shadow-artefacts shown in Figure 5. Figure 6 shows some remaining blood that cannot be removed before the tissue is embedded into the block. All these artefacts make the segmentation of the data very hard, and standard automatic segmentation techniques completely fail. Therefore most of the segmentation tasks are currently carried out manually which is very time consuming.
Figure 3: A signal image of a chicken embryo.
|
Figure 4: Slicing artifacts caused by the blade. |
Figure 5: Shadow artifacts caused from objects shining through the top of the block. |
|
Figure 6: Remaining blood which could not be extracted before the embedding. |
|
In this work we concentrated on the illumination correction of the images. An easy way to correct the illumination is to capture a blankfield (an image without the embedded object) and to use the blankfield to correct the illumination. However it is not possible to take a blankfield for every block. Therefore we tried several standard Image Processing techniques like the TopHatFilter. Unfortunately, the results were not very pleasing and an alternative approach had to be implemented. The novel approach is to interpret the image as points in three dimensional space (using the brightness as z-value).
Figure 7: The pixels of the image in Figure 8 plotted as points in 3D-space.
Figure 8: Morpho image of a slice if a chicken heart.
Figure 7 shows the image pixels (plotted red) in 3d of the image in Figure 8. The background pixels approximately form a parabolic surface (plotted green). This led us to the hypothesis that the illumination artefact can be described by a parabolic function. We fit a quadric surface in the background pixels. As fitting algorithm we used a simple least squares technique (without weights or iterations). We tried this algorithm on a blankfield to explore which pixels are the most important for the fitting algorithm. We did masked out several regions from the original image and measured the root mean square error. We found out that the most important pixels are the border pixels. This can be seen in Figure 9 and 10, while we started masking out the center pixels, and let the mask grow towards the border in Figure 9, we started with the border pixels and let the mask grow towards the center in Figure 10, this results in a higher error if we only use the center pixels for the fit. When using this technique on real images (images with objects embedded), we had to spare out as much object pixels as possible while preserving as much from the background pixels (and especially the background border pixels) as possible. Making use of our findings we implemented a tool to apply our algorithm to a large number of images.
Figure 9: A square mask growing form the center was used to delete data points. The plot here shows the RMS error made with our computation compared to the original blankfield image (1 = 100%).
Figure 10: The same computation as in Figure 9 with the inverted mask (1 = 100%).
Illumination Correction Tool
Figure 11: This image shows the graphical user interface of the illumination correction tool
