From Signal to Model

From Signal to Model

Various signal and image processing methods are employed to reconstruct nerve fibre architecture from images produced during 3D polarised light imaging (PLI) measurements. These methods are validated using analytical and numerical simulations.

In the simplest case, the measurement corresponds to the rotation of polarising filters, resulting in a sinusoidal signal per pixel. The Fourier series is then used to decompose this signal into its principal frequencies, from which the light transmittance, amplitude (retardation) and phase (direction) can be determined. This yields a two-dimensional projection of the birefringent axis.

By using a tilt stage or a tilted light beam, this measurement can be extended into the third dimension. This enables the inclination of the birefringent axis to be determined directly.

To the top

In 3D-PLI, sinusoidal intensity profiles are calculated for every single image pixel. This means that several hundred thousand pixels must be analysed per image. Each measurement comprises nine rotational images and up to four additional tilt measurements.

The generated data is first automatically subjected to a quality check to identify measurement or data issues as early as possible. It is then converted into an HPC-compliant data container and transferred to our database on the JSC systems.

Further processing takes place on the HPC system JURECA. A separate job is generated for each measurement, which processes the respective measurement in parallel. This workflow includes signal analysis, stitching and the calculation of further visualisations.

Depending on the scope of the measurement, a single human brain section generates several hundred gigabytes to several terabytes of data. For a complete human brain, this results in a data volume in the petabyte range. The largest proportion of this consists of raw microscopic data. Once the analysis is complete, this raw data is transferred to magnetic tape for long-term archiving.

The compressed signal data, on the other hand, remains on high-speed storage systems that are directly connected to the HPC system. This ensures that it remains efficiently available for subsequent analyses and visualisations.

To the top

3D Blockface Reconstruction

Before the brain tissue is sectioned, high-resolution blockface images of the tissue surface are acquired. Since these images are free of deformation, they serve as a stable reference for 3D reconstruction. The reconstruction is performed in two steps. First, fiducial markers are automatically segmented, checked, and used for the initial coarse alignment of the individual sections. The registration is then refined through intensity-based alignment, initially using marker information and, if necessary, additionally incorporating tissue information. After visual quality control, outlier sections can be registered again. This results in a more accurate spatial reconstruction of the original tissue block.

Feature Matching

Feature matching methods are used to align tissue sections and block face images initially. This process involves the automatic detection of distinctive anatomical structures or textural features in the images, which are then matched against one another to identify corresponding points. The mathematically linked pairs of points that result form the basis for transforming and roughly fitting the sections into a shared 3D space.

To the top

Segmentation of brain tissue

Automated segmentation of image data is a key step in image processing. First, the brain tissue is isolated from the background of the image to eliminate artefacts and noise for subsequent processes. The method also enables precise distinction between grey and white matter within the tissue. This is achieved using a deep learning pipeline based on U-Net, combined with various encoders and sampling strategies. The model learns to reliably detect complex anatomical patterns and subtle tissue transitions. This provides a robust foundation for 3D reconstructions, serial registration, and detailed brain mapping in brain research. It also allows for the integration of further microscopic and multimodal image data for precise future analyses.

Segmentation of blood vessels

The aim of automatically segmenting and visualising cerebral blood vessels in high-resolution blockface image data is to accurately map and improve understanding of the brain's complex vascular structures, thereby supporting neurobiological analyses. To this end, interactive machine learning methods are employed alongside Ilastik, the Frangi filter and morphological operations such as dilation, erosion and size filtering. Skeletonisation and sphere-based representations are then used to clearly visualise connections and vessel diameters along the course of the vessels. The resulting 3D visualisations and animations facilitate interpretation of the vascular network, forming the basis for further research and improved image analysis.

To the top

2D and 3D Visualisation

We use tile servers to visualise our sections. Depending on the current zoom level of the image section, these enable software such as OpenSeadragon or Neuroglancer to visualise our large datasets.

We also use further in-house visualisations that can display statistical fibre orientations, for example.

Visualization of the fiber architecture

Appropriate visualisation of the 3D PLI data allows for clear, interactive analysis. Tissue structures are depicted using surface and volume rendering, while clipping boxes offer focused insight into specific brain regions. The orientation of nerve fibres is visualised using colour-coded glyphs, with different colours indicating different fibre directions. Lighting and shading enhance spatial perception. To present large amounts of data clearly, vectors are aggregated into supervoxels. Tractography and atlas-based masking also help to examine fibre tracts and anatomical regions in a targeted manner. This makes complex neural structures comprehensible, comparable and scientifically interpretable.

To the top