Figure 1.
WMSA detection on simulated preterm T2-weighted brain images.
Images at three mid-axial levels show, from left to right: noise-free images with manually drawn WMSA regions in yellow (ground truth); addition of Rician noise (SNR = 15) and INU (20% level); and WMSA detection by our proposed method marked in yellow.
Figure 2.
A flowchart of the generation of individual tissue probability maps.
A target individual anatomy was first normalized to the reference space formed by the very preterm probabilistic atlas using LDDMM and the resultant transformation matrix was saved. The inverse transformation matric was applied to the tissue probability maps to create the desired target individual tissue probability maps.
Figure 3.
Comparison of automated WMSA detection on simulated infant MR images with ground truth.
Quantitatively, automated detection showed very high Dice similarity index values (left) and low false detection rates (right) at each noise level with ground truth.
Figure 4.
Automated WMSA detection at six different axial levels.
Top row: T2-weighted images; bottom row: detected WMSA marked in red. The automated detections closely approximated the visually apparent signal abnormalities.
Figure 5.
Linear regression and Pearson correlation analyses of automated quantified WMSA within different WM regions and Bayley III cognitive and language scores.
R2 denotes linear regression’s coefficient of determination and r denotes Pearson correlation coefficient. Larger WMSA volumes correspond with both lower cognitive and language scores. WMSA regional volume within centrum semiovale is a better predictor of Bayley scores than that within the periventricular WM regions.