made PC-AT-based image workstation with a scaleable- 
performance architecture. 
An overview of the image processing system in use is 
given in Figure 1. A scaleable image processor (SIP), 
consisting of a network of processing elements (PE), is 
surrounded by a host system providing local high resolution 
display and graphics capabilities as well as basic I/O- 
services. Selected PE's are connected to a dedicated Ethernet- 
PE which links the image processing system to a network of 
RS/6000 3D-graphics workstations. 
A detailed description of the SIP hardware can be found 
in [2]. In summary, our SIP consists of 8 custom transputer 
modules (PE, T805, 25MHz, 8MB RAM [3]) arranged in a 
cube-shaped network which is controlled by the 'Root- 
module (RPE, T805, 25MHz, 12MB RAM). A fast DMA- 
link (4.5MB/s peak data throughput) allows the RPE to work 
directly into the host's memory system, thus providing fast 
image and graphics output to the high resolution display 
hardware. The SIP-network is closely coupled to a dedicated 
Ethernet-module (EPE, T805, 25MHz, 8MB RAM) which 
establishes a fast connection to the powerful RS/6000 
workstation network of the IBTZ. This data link becomes 
crucial for the ever-increasing storage and processing 
requirements in the field of 3D-image analysis. The raw 
computing power of the SIP-network equals 120 MIPS and 
19 MFLOPS. 
  
     
   
    
  
   
   
   
   
  
   
  
  
   
  
  
   
  
  
   
  
  
   
   
  
  
  
   
  
4. IMAGE PROCESSING SOFTWARE 
The stack of up to 100 confocal laser microscope image 
slices with a resolution of 512 by 512 pixels is mapped onto 
the cube-shaped SIP in a way which minimizes 
communication pathlength. In order to be able to quantify the 
spatial density distribution of the photo-sensitive agents in 
the cell population, we use a multi-level, parallel and true 
3D-processing scheme. First, a spatial nonlinear filter process 
usually referred to as Inisotropic Diffusion (ID) [5], is 
applied in order to eliminate noise artefacts and to simplify 
edge detection within the image volume in preparation of the 
subsequent step. ID is a numerical heat equation solver where 
temperature is substituted by voxel intensity. This diffusion 
process minimizes intensity fluctuations whereas sharp edges 
are preserved by reducing the heat conductivity along them. 
Subsequent binarisation splits the volume into 3D-objects and 
background. Parametrisation converts the objects into data 
structures which describe their geometry, topology and 
texture. The main challenge of this step is to generate a 
connectivity graph that leads to a closed surface or 'skin' of 
the object. The resulting surface-descriptors can be fed into 
todays standard solid modellers. Our own modeller features 
an additional smoothing step prior to raster conversion. In 
either case any arbitrary three-dimensional view of the cell 
population obtained with the CLSM becomes available for 
detailed analysis and quantification of the structures of 
interest. 
  
  
  
  
Display System Harddisk 
1280*1024*10 Á AT-386/387 |* A4 ^ 200 MB 
a 
3 i = 
Ma d S| | Optical Disk 
3 9 > 800 MB 
É Z 
  
Graphics Engine 
HD-63484 [|< 
  
i 
  
  
Figure 1: 
Local 
Area 
LAN-Processor 
Overview of the image processing hardware consisting of the host system and the 
transputer based image processor. PE = Processing Element. RPE = Root Processing 
Element with dedicated high speed DMA-channel to host system. EPE = Ethernet 
Processing Element connected to LAN-module.