Neonatal Neuroimage Analysis

description

The neonatal studies in our lab focus on understanding, quantifying, and visualizing the effects of complex congenital heart disease on early brain development. This is a challenging task because of  the  low  contrast-to-noise ratio and large variance in both tissue intensities and brain structures, as well as imaging artifacts and partial-volume effects in clinical neonatal scanning. The tools being developed in our lab will be applicable to  studies involving vulnerable infants who are at risk for neurodevelopmental problems.

Brain Tissue Segmentation

Accurate and efficient voxel based segmentation of  white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) is essential for quantitative neonatal brain image analysis. In a Bayesian framework, we integrate complementary information derived from the image as well as priors. The maximum-a-posteriori (MAP) segmentation is obtained using an efficient graph-cut method. We employ a spatially adaptive likelihood model using a data-driven nonparametric statistical technique. The method incorporate probabilistic atlas priors and intensity-based prior to impose additional regularity on segmentation.  In an iterative scheme, the models adapt to spatial variations of image intensities via nonparametric density estimation.

Intensity-Based Tissue Priors

We learn an intensity-based tissue prior for brain tissue segmentation, which relies on the empirical Markov statistics from training data. The intensity-based tissue prior is complementary to the widely-used tissue-probability-map (TPM) prior which provides tissue probabilities based on voxel locations. The intensity-based tissue prior avoids the registration-related problems confronted by TPM-based methods and is robust against manual-segmentation errors and image noise. The construction of such prior relies on nonparametric statistical models, by learning a classifier to distinguish between the Markov intensity statistics of brain tissues using a fuzzy nonlinear support vector machine (SVM).  The decision value of the SVM for each input data is calibrated by a learned sigmoid function into a probability value, which is regarded as the intensity-based tissue prior. Intensity-based tissue priors have proved to be effective in neonatal brain-MR tissue segmentation.