Neuron Image Reconstruction
J. Pinezich (Applied Math, SUNY at Stony Brook)
M. Vazquez (Brookhaven National Lab)
Future missions in space may involve long-term travel beyond the magnetic field
of the Earth, subjecting astronauts to radiation hazards posed by solar flares
and galactic cosmic rays. Thus, it is critical to determine if there will be
any reversible or irreversible detrimental neurological effects from this
prolonged exposure to space. A question of particular importance focuses on
the long-term effects of space radiation on the central nervous system (CNS)
neuroplasticity. The major sources of such radiation are solar flares and
galactic cosmic rays, the components of which are energetic charged particles,
protons as well as fully ionized nuclei of all elements. Of particular concern
are the high-Z and -energy (HZE) particles such as Fe, due to their high
charge and resulting high energy deposition.
A series of experiments, utilizing the Alternating Gradient Synchrotron at
Brookhaven National Lab, has been conducted to measure the effects of such
radiation exposure on the CNS plasticity. An in vitro model, in which
neuronal regeneration occurs under controlled conditions was employed. Retinal
explants taken from chick embryos were exposed to incremental doses of Fe ions,
then cultured and incubated for three days. At the end of incubation
two-dimensional 256-gray-scale images (see figure 1) were taken for the purpose
of morphometric analysis of neurite outgrowth. The analysis is human-labor
intensive, requiring neurite tracing and reconstruction by hand, followed by
an automated count of the number of pixel locations occupied by neurite. Due
to the labor involved, only a small number of the more than 1000 images obtained
have been analyzed.
We have developed a new algorithm based on pattern recognition, with
the aim of fully automated reconstruction of the neurite outgrowth
images. Complications in processing arise due to the complexity
of the neurite growth patterns (neurite branching, overlapping,
finite focal depth) and the relatively poor quality of the images (dead
tissue debris, bubbles in culture medium). Our approach is to first apply a
standard edge-sharpening algorithm, and then a ridge tracking algorithm, which
detects and reconstructs neurites (see figure 2). The basic idea of the ridge
tracking algorithm is to compute, at each point-under-test, a polar function
representing neurite density as a function of angle about the point. To
facilitate this computation and provide a certain level of smoothing, the
image data, originally specified on a regular grid (approx. 640x470) is extended
via bilinear interpolation. The image data representation is such that the dark
regions (indicating the presence of neural matter) correspond to larger numbers,
the white regions to small numbers. Thus the algorithm detects maxima in the
computed polar functions. Moreover, neurite signature qualities (user-specified
parameters), such as maximum neurite turning angle, neurite direction with
respect to explant, noise thresholds, are applied to the polar function
filtering out the most likely directions for neurite presence at the
point-under-test. Once a direction has been determined to coincide with a
neurite, a new polar function is computed at a point one distance unit in that
direction. An analysis is again performed for the presence of neurite,
augmented by an angular analysis for agreement to the previously determined
direction. If the determination is positive, the point is accepted, and the
tracking procedure iterated. If it is negative, the point is bypassed and a
new polar function is computed at a point two distance units from the last
accepted point-under-test. The analysis is conducted once again, and if
positive, the previously bypassed point is filled in (accepted). In this manner
a weak neurite (e.g. drifting in and out of focal plane) may be reconstructed;
the maximum allowable number of sequential bypass points is user-specified.
As currently implemented, all user-specified parameters are applied with equal
probability. We propose to upgrade this algorithm by providing a valid
statistical basis and measure of uncertainty. This will be done through
the introduction of a probability based model in the polar function
filtering and reconstruction steps. We believe that this reconstruction
algorithm will be of broader interest in the image processing community.
J. Pinezich funded by the Swartz Institute.
 RAW IMAGE
|
RECONSTRUCTED IMAGE
|
| TEXT postscript file (gzipped)
|
| FIG 1 PDF file
|
| FIG 2 PDF file
|
| FIG 1 and 2 PDF file
|