International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
  
  
Fig. 3: GIT 
[Immune fluorescence imaging the probes in a confocal 
microscope, the red channel was used to image the protein, and 
the green channel to image the reference organelles, i.e. the cell 
nucleus membrane and the golgi apparatus. Fig. 4 shows an 
example of the colocalizations becoming visible. 
  
Fig. 4: Images of a cell prepared to show Huntingtin: the 
protein (or its marker HDI) in red, the reference 
organelles (markers Lamin and Golgin97) in green 
and the color image showing colocalizations (from 
left to right). 
For each protein to be investigated ten images of different cells 
were acquired. In order to explore the statistic behaviour of the 
visual appearance, ie. the feature vector of a protein, two 
visually similar proteins, Huntingtin and GIT (Figs. 2 and 3) 
were imaged 100 times. Half of the images were used as 
training data, the other half as test data. 
The features to be used for classification are for instance 
statistical measures describing protein localisation inside of the 
nucleus, e.g. variance and entropy, edge segments appearing 
inside and outside of the nucleus, and the visibility of the golgi 
apparatus being attached to the nucleus. If these “features” are 
no numerical values directly, they have to be transformed into 
numeric measures such as edge length or strength. As the 
occurrence of the protein inside of a cell is a natural event more 
or less varying statistically, the statistic behaviour of the 
extracted features is of major importance for the performance of 
the method and, therefore, has to be taken into account by the 
algorithm, e.g., by using the probability density distributions of 
the features for classification. 
The feature vector actually used includes the following features; 
c.f. (STECKLING & KLÓTZER, 2003; STECKLING et al., 2003): 
I. White pixels: number of pixels whose grey value is 
greater than the average of all grey values of the 
image. 
2. White segments: number of image segments fulfilling 
the same condition. À segment is defined as a four- 
connected neighbourhood (BOLAND & MURPHY, 
2001). 
J 
four-connected pixels with grey values lower or equal 
than the average of all grey values of the image 
(BOLAND & MuRPHY, 2001). 
4. Expectation value: 
] K 
m=E(x)=— > x, (1) 
= 2x 
k=l 
(BOLAND & MURPHY, 2001). 
5. Energy: second angular moment 
N, 
> Ps (f (2) 
i=0 
where x and y are the coordinates (row and 
column) of an entry in the co-occurrence matrix, 
and p,(i) is the probability of co-occurrence 
matrix coordinates summing to x+y 
(BOLAND & MURPHY, 2001; HARALICK ct al., 1973). 
6. Difference entropy: 
S a 
- VY p, G)log[p, .()] G3) 
i=0 
(BOIAND & MURPHY, 2001; HARALICK ct al., 1973). 
7 Lines: number of segments extracted with a line 
extraction method. 
4. CLASSIFICATION 
A maximum likelihood classification was used. To illustrate the 
separability of the clusters of two visually similar proteins 
based on the feature vector defined in the previous section, Fig. 
5 shows the sub-space of the three most informative features. 
3. Black segments: number of segments consisting of 
    
    
   
   
   
   
   
   
   
   
   
   
  
   
  
   
   
   
  
  
  
   
   
  
   
   
  
  
   
  
  
  
  
  
  
  
  
   
   
   
  
  
  
  
  
  
  
  
  
  
  
    
Inter. 
87 V 
STEC 
who 
class 
to 10 
class 
hand 
throu 
possi 
aae vaivraices 
Fig. 
The 
stati: 
dete; 
diffe 
but ¢ 
the 
acqu 
prop