Tomographic brain imaging with nucleolar detail and automatic cell counting

Abstract

Brain tissue evaluation is essential for gaining in-depth insight into its diseases and disorders. Imaging the human brain in three dimensions has always been a challenge on the cell level. In vivo methods lack spatial resolution, and optical microscopy has a limited penetration depth. Herein, we show that hard X-ray phase tomography can visualise a volume of up to 43 mm(3) of human post mortem or biopsy brain samples, by demonstrating the method on the cerebellum. We automatically identified 5,000 Purkinje cells with an error of less than 5% at their layer and determined the local surface density to 165 cells per mm(2) on average. Moreover, we highlight that three-dimensional data allows for the segmentation of sub-cellular structures, including dendritic tree and Purkinje cell nucleoli, without dedicated staining. The method suggests that automatic cell feature quantification of human tissues is feasible in phase tomograms obtained with isotropic resolution in a label-free manner.

Figure 1. 2D-3D registration and comparison of registered CT slice vs. histological slice.

(a) Identification of the histological section within the CT data set. (b) CT slice and selected region mapped to (c) histological section. The automatic 2D-3D registration provides the identification of the Stratum granulosum and Stratum moleculare with comparable intensities, as well as the location of the Purkinje cells (see also Supplementary Video 1).

Figure 2. Localisation of the Purkinje cells using histological findings and its extension into the third dimension.

(a) 3D view of Purkinje cells with a tomogram slice resembling a 3D extension of histology and (b–g) their identification. (b) Registered phase contrast image. (c) Phase contrast image coloured similar to H&E staining. (d) Purkinje cell segmentation mask. (e) Phase contrast image with separately coloured Purkinje cells to resemble f, H&E histology. The staining is insufficient to visualise the nucleolus, due to post mortem autolysis. g, Biospy - H&E section with nucleoli visible. The scale bar corresponds to 100 µm (see also Supplementary Video 2).

Figure 3. Purkinje cell layer identification in a volume of 0.4 mm × 1.0 mm × 0.5 mm.

(a) Objects detected by the Frangi-based filter include vessels and Corpora amylacea. (b) After deselecting extremely tubular and spherical structures, the error rate was determined based on approximately 100 objects. (c) The Purkinje cell layer is coloured according to local cell density, shown with cell locations (red). The cell density was evaluated with respect to surface area and varied between 90 and 290 cells per mm². Average cell density was 177 cells per mm with respect to the Purkinje cell layer (see also Supplementary Video 3).

Figure 4. Purkinje layers coloured according to the local density of Purkinje cells and their location in a volume of 43 mm³.

(a) The average detected density was 165 cells per mm², related to the manifold and volumetric cell density to 116 mm⁻³ of (b) localised Purkinje cells. Segmented objects were filtered in sparse distributions not connected to the Purkinje layer, which is derived and described implicitly using a level set approach. It features disconnectivities at the border of the dataset. The axes of the scale bar correspond to 1 mm (see also Supplementary Video 4).

Figure 5. Purkinje cell including the main dendritic tree.

(a) Representation as a 3D surface and (b) a 2D phase tomogram slice in comparison to (c) a H&E stained histological image of a Purkinje cell. The images verify clear correspondence with respect to cell morphology and the nucleolus, and it is less visible in the nucleus. In the 3D view the red-coloured surface represents the outer contour of the segmented cell and the blue one the nucleolus. The dimensions of the region of interest, namely 382.5 × 247.5 × 54.0 μm³, result in 56.1 million isotropic voxels with an edge length of 0.45 μm (see also Supplementary Video 5).