Researchers have used the well-established Potts model-an algorithm mathematicians and biologists use to measure how individual entities behave in relation to their neighbors-to help map the spatial organization of neurons in the brains of Alzheimer's patients and normal volunteers. Due to appear in this week's PNAS early online edition, the results mark a new approach to automate recognition of neurons in microscopic brain images, which could help measure anatomical changes in conditions where neurons die.

First author S. Peng, working with Boston University's Eugene Stanley and Brad Hyman from Massachusetts General Hospital, applied what they call a "parallel Potts segmentation approach" to help identify individual neurons in confocal microscope images of the brain. In this approach the simple Potts algorithm is run several times but with slightly different parameters. Varying the Potts parameters enabled the authors to correct for slight variations in contrast or focus that are inherent across the confocal images. This improved resolution and helped solve a major problem in brain mapping, namely the difficulty in identifying individual neurons when they overlap with others.

Peng and colleagues applied this mapping system to five images taken from a healthy subject and five from a patient suffering from Alzheimer's disease. A simple Potts analysis identified 86 percent of neurons in the former and 77 percent in the latter, but the parallel segmentation approach increased these numbers to 98 and 93 percent, respectively. Given that the method appears to work even in AD images, which often are of poorer quality due to morphological changes in the tissue, the authors suggest it will be useful when applied to other degenerative disorders and tissue types.—Tom Fagan


Make a Comment

To make a comment you must login or register.

Comments on this content

  1. A general problem in neuroanatomy arises from the difficulty to acquire sufficient positional neuronal data to accurately quantify neuronal disruptions in the brain. This problem is particularly serious for assessing disruptions caused by neuropathological diseases, such as Alzheimer's. The task of manually collecting neuronal positions is extremely time-consuming.
    Moreover, the number of neuronal positions needed in, e.g., a study of microcolumnar structures in the human cortical lining of the superior temporal sulcus is immense, ranging up to tens of thousands of neurons.

    Peng et al. present a systematic study that addresses this need. The authors developed a fully automated method, which takes as input digitized pictures of tissue and produces numerical output corresponding to the spatial coordinates of the identified neurons in the picture. The method, called a parallel Potts segmentation method, is based on concepts previously developed in condensed matter physics for understanding magnetic materials. In this method, a computer replaces each pixel by a "virtual magnet" that points in a direction assigned by the relative darkness of the pixel. Then, the method groups together magnets pointing in the same direction and produces a patchwork of "islands" that directly correspond to different islands in the original picture. From each of the thousands of resulting islands, the method then decides whether the island corresponds to a neuron or not, by selecting islands based on size, shape and relative darkness.

    To validate their new method, Peng et al. apply it to digitized pictures of tissue from healthy humans and patients with AD. They find that their method can identify up to 98 percent of all neurons in healthy subjects and up to 93 percent of all neurons in AD cases. The number of false positives do not account for more than 3 percent of all identified neurons.


No Available References

Further Reading

No Available Further Reading

Primary Papers

  1. . Neuron recognition by parallel Potts segmentation. Proc Natl Acad Sci U S A. 2003 Apr 1;100(7):3847-52. PubMed.