Automated neuroanatomical microvessel segmentation, nuclei labeling, and geometric computations with 3D two photon laser scanning microscopy

John Kaufhold
Science Applications International Corporation (SAIC)

J. P. Kaufhold, P. S. Tsai, P. Binder, B. Friedman, P. D. Lyden, D. Kleinfeld and A.I. Ifarraguerri

We describe the development a suite of computer algorithms that operate on volumetric stacks of neuro-anatomical images obtained with All Optical Histology, a technique that combines two photon laser scanning microscopy (TPLSM) with ablation by ultra-short laser pulses to process large blocks of tissue. The computer algorithms extract the locations of all labeled cellular nuclei and blood vessels, separate cell subtypes based on a fluorescent indicator, correct for imaging anisotropy, and calculate three-dimensional (3-D) statistics, e.g., the separation of each neuron to the nearest vessel (figure).

The automated isolation of features from large data sets of neuro-anatomical images with low signal-to-noise ratio and noinuniform backgrounds present challenges. Techniques based on simple thresholds can miss structures or erroneously label background features. Our techniques respect the 3-D statistics of the vessels and cells in the underlying tissue. For microvasculature, a mask of all vessels is generated based on matched filtering of the data to 11 µm x 1 µm x 1 µm rods that span 82 directions in 3-D space. Anisotropy inherent in the TPLSM images leads to masks that are artificially elongated along the axial dimension with increasing severity as a function of imaging depth. To correct this anisotropy, we statistically model the depth dependence of vessel elongations and locally reshape the mask to effectively perform a binary deconvolution. We distinguish capillaries from larger vessels based on a vessel mask whose diameter is less than 6 µm. With regard to cells, a mask of all nuclei is generated from an imaging channel that contains DAPI-stained nuclei. A preliminary cell mask is generated by local mean subtraction and 3-D Gaussian matched filtering. Separation of nuclei with connected masks is achieved by computing a local distance transform for each contiguous mask component, the peaks of which correspond to nuclei centers. A watershed transform on the negative distance transform then labels each nuclear footprint in 3-D. Finally, nuclei are classified as neuronal or non-neuronal based on data from a channel that contains all neuronal nuclei labeled with the neuron-specific antibody NeuN. An initial neuronal score is assigned to each nuclei mask based on the signal in the corresponding position in the neuronal channel. A mean-field algorithm iteratively classifies each cell based on the ratio of its score to that of neighboring cells classified as neurons. The vessel and cell masks provide the basis for volume and geometrical analysis.

Sponsors: NIBIB (EB003832), NIMH (MH72570), NCRR (RR021907), and NINDS (NS043300).

Presentation (PowerPoint File)

Back to Image Analysis Challenges in Molecular Microscopy