Berlin, 
rlin, 
located in 
jon of the 
cells used 
eir nuclei. 
terisation, 
mportance 
nal spatial 
ire, from a 
ct to the 
molecules 
that image 
| research, 
long term 
ty cellular 
arance in 
lures were 
r a neural 
iL. (1999) 
cal image 
e methods 
be protein 
a distance 
> from the 
| chosen as 
nventional 
al. (2000) 
e protein 
recognize 
nicroscope 
a primary 
ins within 
'dge-based 
nals. They 
lorescence 
; MURPHY 
ns to train 
0 correctly 
orescence 
determine 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
  
subcellular location, e.g. VELLISTE & MURPIHY (2002) have 
previously described automated systems recognizing all major 
subcellular structures in 2D fluorescence microscopic images. 
They have shown that pattern recognition accuracy is dependent 
on the choice of the vertical position of the 2D slice through the 
cell and that classification of protein localization patterns in 3D 
images results in higher accuracy than in 2D. Automated 
analysis of 3D images provides excellent distinction between 
two golgi proteins whose patterns are indistinguishable by 
visual examination. ROQUES & MURPHY (2002) describe the 
application of pattern analysis methods to the comparison of 
sets of fluorescence microscope images. MURPHY et al. (2002) 
report improved numeric features for pattern descriptions which 
are fairly robust to image intensity changes and different spatial 
resolutions. They validate their conclusions using neural 
networks. DANCKAERT et al. (2002) describe development and 
test of a classification system based on a modular neural 
network trained with sets of confocal focus series. The system 
performed well in spite of the variability of patterns between 
individual cells. 
3. FEATURE EXTRACTION 
In this work, to recognize proteins being active in a cell means 
to visually differentiate between their appearances in images. 
The latter depends on whether there are features which allow 
making a difference between them. This is valid for visual 
judgment by a human observer as well as for a pattern 
recognition algorithm. The criteria used by a human usually are 
directly related to known cell structure. For a pattern 
recognition algorithm, among the multitude of features which 
are present or can be defined in imagery, those have to be 
identified which help to separate different phenotypes of cells 
from each other in feature space. l.e., the pattern characterizing 
a protein has to be parameterized. In general, the parameters to 
be used have to describe the spatial distribution of the protein 
inside of the cell. 
Prior to feature extraction from imagery, a laboratory procedure 
including chemical treatments of probes and microscopic image 
acquisition had to be established. First, antibodies had to be 
found allowing to stain the proteins making them — or the 
organelles as the locations of their activity — visible in the 
imagery. In addition to the protein investigated some organelles 
had to be stained to allow recognition of the most characteristic 
parts cell and, thereby, reference to the cell as such. Lamin was 
chosen as the marker of the membrane of the nucleus of the cell 
allowing separation of the nucleus form the cytoplasm, and 
Golgin97 was used to stain the golgi apparatus. 
On the basis of the membrane of a cell’s nucleus and golgi 
apparatus a reference system allowing translation and rotation 
invariant definition of features describing the proteins was 
defined. The centre of the nucleus is used as central reference 
point in the sense of the origin of a coordinate system. The cell 
is subdivided into sectors inside and outside the nucleus (Fig. 
1). In the system of sectors the direction to the centre of the 
golgi apparatus is used as reference. 
  
Fig. 1: Subdivision of the cell into sectors inside and 
outside the cell nucleus. 
Ten proteins and corresponding antibodies were selected for the 
investigation. It was taken care to choose visually very different 
as well as rather similar proteins. Figs. 2 and 3 show Huntingtin 
and GIT as examples of visually similar proteins statistically 
varying. 
  
Fig. 2: Huntingtin