Image Collection and Processing

To address the aforementioned problems with the 5-class data set, an entirely new set of images was collected using HeLa cells. The details of the image processing methods can be found in Section 3.2.2, above, but are described briefly here. Cells were grown to sub-confluent levels on collagen-coated microscope coverslips, fixed in paraformaldehyde, and permeabilized with saponin. They were then incubated with one of 8 antibodies or rhodamine phalloidin (a label for filamentous actin). When necessary, the cells were subsequently incubated with a fluorescent secondary antibody. All cells were also labeled with the DNA intercalating dye DAPI. The coverslips were then mounted on microscope slides and prepared for imaging. The coverslips were scanned manually using a transmitted light mode of the microscope (differential interference contrast) to identify cells that were well spread and separated from their neighbors. Images of protein and DNA-associated fluorescence were collected using a microscope like the one in Figure 1.5. To facilitate deconvolution, each field of view was collected as a stack of three images separated by 0.237 $\mu$m. The out-of-focus component of each central plane was removed via nearest neighbor deconvolution [26], as described in Section 2.2.2 (p. ). The remaining background fluorescence was then subtracted from the deconvolved image and small, isolated spots of fluorescence were removed with a majority filter. The image intensity was then thresholded and single cells were isolated using a manually defined polygon. After these steps, the images were used in the feature extraction steps described below.

The first improvement these data represent over the CHO data is that there are ten classes of localization patterns, represented by: an endoplasmic reticulum (ER) protein, the Golgi protein giantin, the Golgi protein GPP130, the lysosomal protein LAMP2, a mitochondrial protein, the nucleolar protein nucleolin, filamentous actin, transferrin receptor, tubulin, and a DNA intercalating dye. These data include patterns from all of the major organelles of a eukaryotic cell and will provide better insight into how the pattern recognition techniques perform with many classes. Representative images selected from each class via a systematic method (HTFR in [34]) are shown in Figure 3.7.

The second reason that this data set represents an improvement over the CHO data is that some pairs of patterns were purposely selected to be difficult to distinguish visually. The two Golgi patterns should be very similar, as should the patterns of transferrin receptor and LAMP2. The mitochondria and ER patterns are also somewhat similar in their perinuclear distribution. These labels were chosen so that it would be possible to assess how robust this approach might be in the face of ever larger numbers of localization patterns.

  

Figure 3.7: (Following page). Representative images from the sets used as input to the classification systems described in the text. These images have bad background fluorescence subtracted and have had all pixels below a threshold set to 0. Images are shown for cells labeled with antibodies against an ER protein (A), the Golgi protein giantin (B), the Golgi protein GPP130 (C), the lysosomal protein LAMP2 (D), a mitochondrial protein (E), the nucleolar protein nucleolin (F), transferrin receptor (H), and the cytoskeletal protein tubulin (J). Filamentous actin was labeled with rhodamine-phalloidin (G) and DNA was labeled with DAPI (K). Scale bar = 10 $\mu$m.