Fig 1.
T1-weighted MRI for adult (a) and infant (b) brains.
Table 1.
Summary of brain segmentation related work.
Fig 2.
Proposed segmentation framework.
Fig 3.
The calculated shape probability for the CSF(a), GM(b), and WM(c).
Fig 4.
Normalized empirical density using the LCDG model for an infant subject (a), and an adult one (b).
Note that dashed = empirical, red = CSF component, green = GM component, blue = WM component.
Fig 5.
Samples of the second- (a), third- (b), and fourth-order (c) cliques for the 26-neighborhood (graph cliques are shown in different colors for visualization purpose).
Table 2.
Summary of databases used to validate the proposed method.
Fig 6.
Segmentation results; for CSF (a), GM (b), and WM (c); projected onto axial plane for a six-month-old subject from the IBIS database for infants: Segmentation using our proposed method (first row); using the iBEAT method (second row); and Ground truth (third row).
Fig 7.
Segmentation results; for CSF (a), GM (b), and WM (c); projected onto axial plane for a additional six-month-old subject from the IBIS database for infants: Segmentation using our proposed method (first row); using the iBEAT method (second row); and Ground truth (third row).
Fig 8.
Segmentation results; for CSF (a), GM (b), and WM (c); projected onto axial plane for a nine-year-old subject from the KKI database: Segmentation using our proposed method (first row); using the FSL method (second row); using the FreeSurfer method (third row); and Ground truth (fourth row).
Fig 9.
Segmentation results; for CSF (a), GM (b), and WM (c); projected onto axial plane for a 16-year-old subject from the UCLA database: Segmentation using our proposed method (first row); using the FSL method (second row); using the FreeSurfer method (third row); and Ground truth (fourth row).
Fig 10.
Segmentation results; for CSF (a), GM (b), and WM (c); projected onto axial plane for a sample from the NYU database: Segmentation using our proposed method (first row); using the FSL method (second row); using the FreeSurfer method (third row); and Ground truth (fourth row).
Table 3.
Accuracy of our segmentation approach using Dice Similarity Coefficient (DSC)(%), the modified Hausdorff Distance (MHD)(mm), and Absolute Brain Volume Difference (ABVD) (%) for the WM, GM, and CSF of the IBIS database.
Metrics are represented as Mean±Standard Deviation. Results for the proposed approach are shown using both the second- and higher-order MGRF model. Age of this group is 6 months.
Table 4.
Accuracy of our segmentation approach using Dice Similarity Coefficient (DSC)(%), the modified Hausdorff Distance (MHD)(mm), and Absolute Brain Volume Difference (ABVD) (%) for the WM, GM, and CSF of the KKI database.
Metrics are represented as Mean±Standard Deviation. Results for the proposed approach are shown using both the second- and higher-order MGRF model. Age range of this group is 8–13 years.
Table 5.
Accuracy of our segmentation approach using Dice Similarity Coefficient (DSC)(%), the modified Hausdorff Distance (MHD)(mm), and Absolute Brain Volume Difference (ABVD) (%) for the WM, GM, and CSF of the NYU and UCLA databases.
Metrics are represented as Mean±Standard Deviation. Results for the proposed approach are shown using both the second- and higher-order MGRF model. Age range of this group is 6.5–39.1 years.
Table 6.
Summary of the time required by the proposed approach and other approaches for segmenting a brain subject.