Semi-automatic segmentation of the fetal brain from Magnetic Resonance Imaging

Abstract

AbstractTemplate-based segmentation techniques have been used for targeting deep brain structures in fetal MR images. In this study, two registration algorithms were compared to determine the optimal strategy of segmenting subcortical structures in T2-weighted images acquired during the third trimester of pregnancy. Adult women with singleton pregnancies (n=9) ranging from 35-39 weeks gestational age were recruited. Fetal MRI were performed on 1.5 T and 3 T scanners. Automatic fetal brain segmentation and volumetric reconstruction algorithms were performed on all subjects using the NiftyMIC software. An atlas of cortical and subcortical structures (36 weeks’ gestation) was registered into native space using ANTs (Automatic Normalization Tools) and FLIRT (FMRIB’s linear image registration tool). The cerebellum and thalamus were manually segmented. Dice coefficients were calculated to validate the reliability of automatic methods and to compare the performance between ANTs (nonlinear) and FLIRT (affine) registration algorithms compared to the gold-standard manual segmentations. The Dice-kappa values for the automated labels were compared using the Wilcoxon test. Comparing cerebellum and thalamus masks against the manually segmented masks, the median Dice-kappa coefficients for ANTs and FLIRT were 0.76 (interquartile range [IQR]= 0.56-0.83) and 0.65 (IQR=0.5-0.73), respectively. The Wilcoxon test (Z=4.9, P<0.01) indicated that the ANTs registration method performed better than FLIRT for the fetal cerebellum and thalamus. We found that a nonlinear registration method, provided improved results compared to an affine transformation. Nonlinear registration methods may be preferable for subcortical segmentations in MR images acquired in third-trimester fetuses.