Medical Image Processing - Signal Processing Systems - page 368

to sum the resampled colored and shaded intensities weighted by the opacity values.
Common optimizations such as early ray termination (when cumulative opacity
reaches 1) have been implemented. Another significant acceleration technique is
shear-warp factorization [ 17 ] that allows compositing along data axis as opposed to
camera axis. An example of volume rendering is shown in Fig. 15 .

9 Medical Image Processing Platforms

Medical image processing is potentially a time demanding task. Time constraints
are involved in a variety of clinical scenarios. The time from acquisition to when
physicians must have the information contained in those images may be a matter
of minutes in the case of shock trauma injury in which diagnosis and treatment
plans must be made immediately, or less than a second in the case of image-guided
surgery in which decisions about a patient must be made in real time. Even when
these time demands are relaxed in the case of planning a procedure days after the
image is acquired, practical issues of volume load imply that any image processing
to be done must still complete in a reasonable amount of time.
Many automatic, robust and effective medical image processing algorithms
are iterative in nature. Examples discussed here include iterative reconstruction
from noisy projections, deformable model segmentation, and intensity-based image
registration. These tend to be computationally intensive, typically taking minutes to
hours to solve on high-end processors. Such complexity has inhibited the adoption
of these technologies in the clinical workflow. While specific requirements vary
from application to application, the computation time must be on the order of
seconds to be viable in most image-guided interventions scenarios.

9.1 Computing Requirements

Computational and memory requirements are driven by the size of the data that must
be processed and stored. These can vary widely by the particular medical image
processing scenario, so we will attempt to quantify the gamut here. Voxel intensity
(sometimes referred to as gray-scale value) is typically stored in 8 bit or 16 bit
fixed point data with signed and endianness varying across equipment. Scanning
equipment precision often does not produce byte aligned values, but because of
computing considerations, all images are stored in this format. The number of voxels
in a volume depends on the modality, the equipment used, but these images can vary
in size from 256

256 8-bit voxels (e.g., a moderate viewing size of a low resolution
2D ultrasound) to 512

×

512 16-bit voxels (e.g., a full-body scan with a high
resolution CT). Therefore, the memory requirements for image processing platforms
may be light, requiring 32 KB or they may be significant, requiring 26 MB or more.

× 512 ×