While most neurons are
generated during the embryonic and postnatal periods, it is
now well accepted that some regions of the brain keep
producing new neurons throughout adulthood. Understanding the
processes that govern the birth and development of neurons
could lead to future treatments to a variety of pathological
and neurodegenerative conditions using adult neural stem
cells. In order to understand these processes,
neuroscientists need to study the movement and morphology of
neurons as they migrate and develop. We have developed
state-of-the-art methods for detecting the nuclei of
developing neurons and tracking them as they migrate within
the brain.
To detect neurons in microscopy
data, we developed a detector which learns the appearance of neurons,
even though it may vary greatly from neuron to neuron. An Adaboost
learning framework is used with a new class of image features called
Rays. Rays are capable of characterizing deformable or irregular
shapes such as neurons with consistent responses, whereas standard
features including Haar and HOG are less reliable. They can also
provide contextual information outside of the detector window.
Our neuron detector is applied to
2-photon time-lapse microscopy data, containing hundreds of migrating
neurons. Each neuron is tracked by linking the detections using a
batch MCMC data association approach that estimates the MAP of a
Bayesian filtering distribution. Two general constraints help us
limit the search space. The first constraint relaxes the standard
assumption that target motion and appearance are independent, and
learns a correlation between the appearance of a neuron and its motion
- the neuron nucleus elongates in the direction of motion (increasing
elongation with speed). The second constraint forces neurons to enter
and exit the scene at image boundaries.
We have also
developed methods to track neurons and reduce the amount of labeling
necessary. In the video on the right, migrating neurons are tracked
and automatically annotated. Users have clicked on the neuron
locations in 1/32 of the frames, denoted by cyan bounding boxes. In
all other frames, the neuron is detected and tracked
automatically. Results are shown by the blue bounding boxes. Gray
bounding boxes indicate the true location of the neurons.
