A Glimpse into StrataQuest 8 Features: Neighborhood Analysis in Colorectal Cancer

Tissue cytometry provides a holistic approach, which includes whole-slide automatic scanning and consequent image analysis, for deciphering the tumor microenvironment. With more than 20 years of experience in tissue cytometry, TissueGnostics is at the forefront of medtech innovations, and is launching suite 8 of the image analysis software, StrataQuest, in 2024. Apart from the new user-friendly interface, there is a number of innovative features which greatly reduce the time spent on analysis, increase the scientific output, and provide data to the researcher right at the fingertips. In the newest version of StrataQuest, contextual image analysis software, it is now possible to perform in-depth neighborhood analysis among multiple cellular phenotypes.

 This case study provides essential information on the newest features, including a practical example to inform your decision. The entire image analysis pipeline is fully automated and only needs minimal user input.

The image below shows a TMA core of an inflamed colon sample acquired with TissueFAXS Spectra (TissueGnostics), capable of multispectral imaging and spectral unmixing. The sample was stained for several markers: DAPI (nuclei), CD4 (T helper cells), PD1 (immune checkpoint molecule, participates in downregulating of the immune response), CD8 (cytotoxic T cells, CTLs), and CK (epithelial cell marker).

Figure 1. Spectral Unmixing.

One of the first steps in an image analysis workflow, especially when dealing with multiplexed immunofluorescence, is spectral unmixing. StrataQuest’s Spectral Unmixing is an engine which separates the measured spectrum (mixed signal pixel) into a collection of constituent spectra. Standard dye spectra are available in the software (spectral database) for usage. This step is necessary to “clean” the image by removing background and channel bleed through (Figure 1).

 Figure 2. Nuclei segmentation.

Another critical step is nuclei segmentation, where it is possible to choose either classical segmentation algorithm or a deep neural network (DNN) model. The latter often performs better with weakly stained or densely packed nuclei. Based on detected marker intensity, cellular phenotypes can now be defined by gating a population on a scattergram and selecting appropriate thresholds (Figure 3). For mutually exclusive phenotypes (e.g. CD4+ and CD8+) it is possible to calculate intensity ratios and select a cut-off to more precisely divide the phenotypes (A) that might be detected as double-positive. The selected population is then used as an input to define a marker-positive population (B). By using the backward connection function, marker-positive cells can be visualized in red contour on the tissue image and marker-negative in green (C).

Figure 3. Gating strategy and phenotyping.

Further, the whole tissue is detected and the classifier engine can be applied to detect certain morphological structures, such as colon crypts. To train the engine, the user marks a few examples for each class that needs to be detected (e.g. colon crypts, stroma, and background). The classifier trains on that information and outputs a model which can be applied on tissue samples. If necessary, the model can be retrained and improved. Distance or proximity maps can be generated to define how many and which cellular phenotypes are in certain distance from a detected tissue structure.

Figure 4. Tissue classification.

In this project, the focus was given to the distribution of CD4+ T helper cells across the tissue and their proximity to the colon crypts. Figure 5 below shows the amount and distribution of CD4+ T cells in the ranges of 0-25 µm (20% of CD4+ T cells) and 25-50 µm (40%) from the crypts. The distances can be custom defined for any given cell type.

Figure 5. Proximity measurements.

Aside from these measurements, available in previous versions of StrataQuest, the new Phenotypes Interactions engine allows easy exploring of cellular phenotypes in tissues. After selecting two phenotypes of interest and the desired distance, the algorithm generates an image and also allows to export the numerical data from a scattergram.

An example of such an image is shown in Figure 6. CD4+ T cells are outlined in green contour (left) and CD8+ T cells in yellow contour (right) and each cell is connected to a CK+ cell in maximum 30 µm range visualized by red lines. In this example the distance of 30 µm between two cells will be considered as an interaction, but the range can be changed according to specific needs. Corresponding scattergrams report the number of interactions (e.g. number of CK cells in 30 µm range from a CD4+ or CD8+ T cell) grouped in ranges from 1-4, 5-8 or 9+ interactions per cell. This way, it can be determined that most cytotoxic CD8+ T cells (70% of all CD8+ T cells) have at least one interaction with a CK+ cell, and 37% of all CD8+ T cells interact with 9 or more CK+ epithelial cells. In this range the average interaction number reaches 15. For CD4+ helper T cells, only half of them have at least one interaction with a CK+ cell, and the proportion of cells is relatively equally distributed among different interaction ranges. On average, one CD4+ T cell is surrounded by four CK+ cells in 30 µm range, whereas one CD8+ T cell by seven CK+ cells, which may indicate that there could be some kind of attraction mechanism which specifically moves CD8+ T cells closer to CK+ cells.

Figure 6. Phenotype interactions.

There are two more tools to closely investigate marker expression distribution across cellular phenotypes available in StrataQuest 8 shown in Figure 7, violin plots (A) and 3D diagrams (B). For example, it is possible to compare the expression pattern of PD1 among all cells, CD4+PD1+ T helper cells, CD8+PD1+ CTLs and CK+PD1+ epithelial cells. A 3D diagram featuring PD1 (y-axis), CD4 (x-axis), and CD8 (z-axis) mean intensities can be freely rotated to explore the amount and distribution of double-positive cells across two phenotypes. These types of diagrams, which are directly available in the software, reduce the time spent in statistical and data visualization programs and provide immediate access to data to navigate further decisions.

Figure 7. Phenotype distributions.

StrataQuest 8 also supports manifold learning techniques, also known as dimensionality reduction: t-SNE, UMAP, and SONG (A). Nonparametric t-SNE (t-distributed Stochastic Neighbor Embedding) and UMAP (Uniform Manifold Approximation and Projection) are widely used for analysis of large, high-dimensional datasets, commonly derived from single-cell RNAseq or multi/hyperplex staining. SONG algorithm (Self Organizing Nebulous Growths) was published in 2020 by Senanayake and colleagues and represents a parametric nonlinear dimensionality reduction technique that supports incremental data visualization, i.e., addition of new data while preserving the structure of the existing visualization. Each technique may have certain advantages or disadvantages for some applications. Depending on the user’s specific needs, StrataQuest offers to visualize image analysis data using all of the mentioned methods.

Within these diagrams (Figure 8) it is possible to gate the cell populations of interest and trace back to their spatial distribution. The example (B) shows two gated populations on a t-SNE plot and gated cells on the tissue sample in red contour. This way, a population of interest can be immediately selected and traced back to its location.

Figure 8. Phenotype extraction.

All the data and results, including images, graphs, and numeric data, can be exported in commonly used formats, such as png, jpg, tiff, Excel, PDF, etc. StrataQuest 8 is a sophisticated tool for fully automated high-end image analysis combined with comprehensive data mining solutions and can be used to investigate cellular relationships. Contact our team today for more information.