Segmentation of the developing fetal brain from MRI data is challenging due to partial volume effects, low contrast across brain regions and heterogeneous myelination states of the developing white matter [1,2]. Segmentation of the boundary between the cortical plate and the subplate is needed for the study of normal fetal brain development and neurodevelopmental disorders [1,2]. The most commonly applied segmentation techniques are based on training datasets or a reference atlas [3,4,5], for example the widely used Freesurfer segmentation tool [5], however such information may not be available for brains at specific developmental time-points in all species.

We present an atlas-free method for segmentation of the cortical plate/subplate interface in fetal sheep brain MRI images using cues derived from diffusion-weighted image data. Our method is applied to extracted brains that are embedded in agar prior to scanning; while the segmentation of extracted brains may sound straight-forward, in practice it is challenging due background inhomogeneities and the folded nature of the cortex. The background is heterogeneous due to fixative fluid remaining around the brain and filling the sulci prior to embedding in agar. Our method exploits the relatively high fractional anisotropy (FA) of the cortex in the developing brain, and uses tractography and cortical thickness to identify developing fibre tracts that also display relatively high FA and otherwise confound the cortical plate segmentation. With strong image features derived from the diffusion-weighted data, segmentation proceeds via a standard level set technique.

The proposed algorithm consists of three steps (Fig. 1). 1. A coarse outer mask of the brain is computed by morphological opening and closing operations on the average of diffusion-weighted images. This provides a boundary along the inferior surface of the extracted brains where the cortical plate is absent. 2. Streamlines tractography (MRtrix) is performed to extract streamlines that both exceed a length threshold and an FA threshold.These long tracts do not intersect the cortical plate, as the subplate is a region of low FA that therefore serves to isolate the cortical plate from the streamlines underneath.The lengths of tracts that originate in the cortex are limited to the cortical thickness, due to the cortical plate anisotropy being perpendicular to the surface. The locations of the streamlines are masked to a representative background level in the FA images. 3. Level set contour evolution based on region competition is applied to the computed diffusion-weighted cue (ITK-SNAP) [6]. The contour at time $$$t$$$ is formulated as $$$c_{t}=(\alpha\widetilde{g_{I}}-\beta \kappa)\overline{n}$$$, where the first term is the external force computed from the gradient magnitude, $$$g_{I}$$$, of the FA image, $$$I$$$, the second term is the internal force base on contour mean curvature, $$$\kappa$$$. Both components are applied along the normal vector of the contour, $$$\overline{n}$$$.The tuning parameters $$$\alpha$$$ and $$$\beta$$$ are chosen to be 1 and 0.75, respectively. The contour evolution initiates from a small sphere situated at the center of the brain and runs for $$$1000$$$ iterations.Fetal sheep brains of 90 days gestational age (dga) were perfusion fixed, extracted and suspended in agar gel. Samples were scanned on a Brüker 4.7T small animal MRI system equipped with a BGA-12S gradient set. Diffusion-weighted images were acquired using a DTI-EPI acquisition with TE/TR=65/3000ms, b=6000s/mm², 81 gradient directions, 10 b0-images, resolution of 250µm-isotropic, matrix size 140$$$\times$$$104$$$\times$$$76, $$$\delta$$$=7ms, $$$\Delta$$$=20ms.The non-diffusion weighted images (Fig.2a) for an exemplar 90-dga brain demonstrate the background inhomogeneity, evident in the white border (fixative) around the cortical surface and partially filling the sulci, surrounded by agar gel (gray). The coarse mask (algorithm step 1) is applied to the FA images (Fig.2b). The segmentation of the cortical plate/subplate boundary proceeds after masking using streamlines tractography results (Fig. 2c). The automated boundary detection is visually highly consistent with manual ground-truth results (Fig. 2d). Validation of the segmentation results is presented in Fig. 3. Surface distance between automated and manually extracted surfaces show minimal variation.The segmentation error is 0.29$$$\pm$$$0.31mm, highlighting the accuracy of our method.We have proposed a segmentation method for extracting the cortical plate/subplate boundary from fetal brain MRI data.The proposed method is independent of a reference atlas, and so it is applicable to both normal and pathological brains of any species. In future work, we plan to increase the efficiency of the method through inclusion of further diffusion-weighted imaging cues such as the principal tensor direction to lead the contour evolution.