Cortical Surface Segmentation in Infants by Coupled Surfaces Deformation across Feature Field

Understanding the development of the human brain is challenging because imaging an immature brain encounters several difficulties. First, partial volume effects due to the small size of the brain associated with an already complex pattern of gyrification (Fig. 1) hamper cortex edges detection. Second, the GMWM contrast is weak due to unmyelinated white matter. Finally, the human brain undergoes big and fast changes during the first months of post-natal life (e.g., the cranial perimeter increases by 0.5cm per week). However, these changes are not homogeneous across the brain, some areas showing intense myelination and synaptogenesis (e.g., visual and motor areas) while others have a more protracted development (e.g., frontal areas) [1]. This maturation inhomogeneity produces important variation in tissue intensity on T2 Magnetic Resonance (MR) images. As shown in Figure 1, WM is much darker in myelinated areas than in non-myelinated ones. These specific characteristics of the infant brain explain why segmentation methods designed for T1 MR images of the adult brain are not optimal. To our knowledge, none of the well-known brain softwares, such as FreeSurfer, Caret, BrainSuite and BrainVisa, have produced as yet automatic reconstruction of the infant’s cortical surface. As for T2-specific methods, atlas-based segmentations of the cortex [2, 3] require accurate and robust brain templates. However, variations of tissue contrast are rapid and asynchronous during the first months of life. We have not used brain atlas because we believe that multiple atlases would be required to capture the anatomical variability of infant brains. Xue and colleagues [4] have recently developed the only, to our knowledge, atlas-free automatic method for use in preterms and newborns. Their approach is mainly based on local estimates of tissue intensity to deal with intensity fluctuations across the brain. However, the GM-WM contrast quickly decreases in areas being myelinated during the first months of life. Thus, the GM-WM interface would not be properly detected in those regions using tissue intensity alone. Here we propose features to ameliorate this problem.