High resolution structural imaging of the brain after premature birth can be used to quantify tissue volume globally and locally and also to characterize brain shape and cortical development [1]. Studies of brain injury after premature birth are challenging, because of the differences between individual clinical cases (eg severity of prematurity, type and location of injury and haemorrhages etc). These mean that larger multi-site studies may be needed to be able to capture a consistent picture of different effects on the brain. Such larger multi-site studies need to be robust to the varying image quality available at non-research sites. In particular, dedicated neonatal imaging technology is not always available even at modern hospitals. One of the key factors is the use of older scanners with coils not specifically aimed at imaging the smaller premature neonatal brain. Such imaging can result in severe intensity variation across the field of view and also fundamental differences in contrast to noise which can cause conventional image analysis pipelines to fail. Here we describe and examine the use of a hybrid tissue segmentation pipeline that combines an initial template based bias correction step [3] prior to an atlas based age specific Expectation-Maximization (EM) segmentation technique [4]. We show that this can significantly improve automated segmentation quality for imaging studies where image quality is not well controlled.

Four T1-weighted coronal brain MR images (TR=50ms,TE=9.2ms,3D flash, voxel size 1x1x1mm) of premature neonates aged around 40GW at time of scan were acquired with a 3T Siemens Verio system (B17 software) using a 15 channel TX/RX adult knee coil to provide improved signal to noise in the premature neonates. The clinical imaging procedure used resulted in significant local variation in signal intensity in the resulting images (see figure 1).

Automated tissue segmentation: A spatio-temporal atlas was previously constructed from 32 high quality research scans [2], manually traced into 7 brain tissue types: grey matter (GM), white matter (WM), ventricles (VENT), deep grey matter (DGM), sulcal CSF (sCSF), brain-stem (BS) and cerebellum (CBL). To segment each clinical test scan an age-specific anatomy was first generated from the atlas. Automated tissue labelling was carried out using the patch-based augmentation of atlas-based EM (PBAEM) segmentation framework [4] using a variability constrained patch-based search and incorporates a bias estimation within the tissue labelling framework. This approach has been found robust to conventional imaging but provided significant errors when presented with extreme local signal variations in this data as shown in Figure 2. To address this we employed a hybrid bias correction: This made use of an initial template based bias estimate [3] to an age specific MRI synthesized from the atlas and non-rigidly aligned to the subject MRI. The relative distortion between the subject and age specific template was then estimated by fitting a 5th order polynomial to the intensities in the reference and subject data, and then estimating the relative bias field by dividing these two polynomial fields. This coarse bias estimate was then used to rescale the intensities in the subject MRI to remove relative variation in intensities. To further refine this estimate a second non-rigid registration was calculated between this new MRI and the age specific template and the relative bias re-calculated and used to re-correct the subject MRI. This second pass correction was then used as input to the atlas-based EM segmentation.

Validation: To evaluate the quality of the final segmentation a careful manual editing of the initial segmentation was carried out to create full 3D tissue segmentations of each of the 4 test subjects

DICE similarity coefficients (DSC) between the hybrid segmentation pipeline and the manually annotated segmentations were calculated, and compared to the DSC of the original segmentation method. The average DSC increased in all the tissue classes across the four scans (Table 2). A similar comparison is made with the Hausdorff distance (HD) and a decrease is seen in most tissue classes. Indicating the hybrid segmentation pipeline performed better compared to the original (Figure 2).Experimental results show that the proposed method provides a robust process pipeline for premature neonatal brain MRIs that differ greatly in image quality, especially ones with severe bias fields due that can occur when dedicated neonatal head coils are unavailable. This enables automated and consistent analysis on large-scale multi-center cross-acquisition datasets, which will allow us to better quantify early human brain development.