| The human brain undergoes drastic development in its anatomy and organization from the last trimester to the first few years of life. A full knowledge of characteristics of this early phase of brain development is essential to understand not only normative brain development but also developmental origins of brain diseases. MR imaging has emerged as a dominant noninvasive technique to visualize this in vivo process in both scientific and clinical studies. Quantitative analyses of neonatal brain MR images can identify regional or global morphological changes in a consistent fashion. Reliable automatic tissue segmentation is the key to success of such quantitative methods. The dynamic nature of brain development and inherent limitations of MR imaging, however, cause low image contrast and high variability of intra-tissue intensities, which therefore pose great challenges to automatic tissue segmentation.;We propose a sound probabilistic framework that adopts a new strategy of probabilistic modeling, model integration, and efficient optimization. Discriminative models are learned based on nonparametric Markov (NPM) statistics of tissue intensities in a supervised-learning framework, which is robust to low and variable tissue contrast in neonatal brain MR images. In addition, the NPM model requires a small amount of manual segmentation, and is tolerant to mislabeling in the training data. The template-based tissue probability map (TPM) is incorporated to compensate the NPM model only in areas with insufficient contrast. Because of the smaller areas involved, the TPM is relatively easy to create. A spatially adaptive generative model is learned in an iterative scheme to adapt to region-specific tissue contrast, mainly because of region-based white matter myelination and MR field inhomogeneity. Furthermore, spatial coherence of the segmentation is enforced via a Markov random field model in the label space, which is optimized by an efficient graph-cut algorithm.;The effectiveness of the proposed method, together with each of its components, was evaluated both qualitatively and quantitatively, by using both clinical neonatal brain-MR images and standard evaluation data for brain tissue segmentation. We have applied the proposed method to clinical studies of abnormal brain development caused by congenital heart defects (CHDs). Our results show significant reduction of gray matter and white matter volumes, in comparison with normative data, which suggest the preoperative risks related to the effects of CHDs may be essential for adverse brain development in these patients. |
