Image Processing

The images output from the microscope (see Figure 3.1) were processed by first applying nearest neighbor deconvolution [26] to each three image stack in order to remove out of focus fluorescence from the central image plane. See Section 2.2.2 (p. ) for methodological details and Figure 3.2 for an example of a deconvolved image. The background fluorescence, defined as the most common pixel value in the region, was subtracted from all pixels. Next, a threshold value was selected for each image using the method described by Ridler and Calvard [46]. This iterative method starts by creating an N element histogram of intensity values and setting an initial threshold, T₀, at the minimum non-zero value in that histogram. At each step the threshold, T_k, is updated using [47]

$$T_{k+1} = \frac{\sum_{i=0}^{T_k}in(i)}{2\sum_{i=0}^{T_k}n(i)} + \frac{\sum_{i=T_{k+1}}^Nin(i)}{2\sum_{i=T_{k+1}}^Nn(i)}$$ (3.1)

where i is an intensity value, and n(i) is the number of pixels with that intensity. The iteration continues until there is no change from T_k to T_k+1.

Mask images were constructed for each background-subtracted image such that pixels that were at or above the threshold in the background-subtracted image were equal to one in the mask, and those below the threshold were set to 0. A majority filter was then applied to the mask image to remove noise (the majority filter sets a given pixel to 1 if a majority of its 8 neighbors are also 1). See Figure 3.3 for an example of a mask image after thresholding and filtering. Finally, a polygonal region containing a single cell was manually defined for each deconvolved image (see Figure 3.4). Pixels in the background-subtracted, thresholded, and filtered fluorescence image that were outside this crop region were set to 0. Pixels inside this region retained their processed values. See Figure 3.5 for an example of a fully processed image. All processing steps were applied to both the protein localization and corresponding DNA images. The source code implementing these image processing steps is included in Section 5.1 (p. ).

  

Figure 3.1: An example of an unprocessed image as output from the microscope and collected using the cooled CCD camera. This particular image depicts the localization pattern of a mitochondrial protein.

  

Figure 3.2: An example of a deconvolved image corresponding to the image in Figure 3.1.

  

Figure 3.3: The mask image generated for the deconvolved image from Figure 3.2.

  

Figure 3.4: The manually defined crop region generated for the image in Figure 3.2.

  

Figure 3.5: An example of a fully processed (background subtracted, thresholded, majority filtered, and cropped) image.

After all processing was complete, those pairs of images that had no signal left in either the protein or DNA localization image were excluded from further analysis. Table 3.2 summarizes the number of images from each class that were available for the feature extraction methods described below.

 

 

Table 3.2: The number of images in each of the ten classes of the HeLa data set, after processing. For each image of a protein localization pattern, there is a corresponding DNA image (in the case of the DNA localization class, each image is used twice).

Localization Pattern	Number of images
An ER Protein	86
Giantin	87
GPP130	85
LAMP2	84
A Mitochondrial Protein	73
Nucleolin	80
f-Actin	98
Transferrin Receptor	91
Tubulin	91
DNA	87
Total	862