Tissue Cytometry

The workflow has six stages, from tissue image to single-cell spatial analysis.

1. Image acquisition. Multiplex fluorescence or brightfield imaging captures several markers on the same section. A brightfield image is split into one image per stain with Color SeparationLoading....
2. Cell detection. NucleiLoading... finds the nuclei and gives each one a unique ID, recorded in a coded imageLoading....
3. Compartments. Within each layerLoading..., measurement masks define the compartments. The Cellular engine makes nuclear, ring and identified masks, Grow (global)Loading... inflates or shrinks the events, and the MembraneLoading... engine detects membrane structures as a mask and a skeleton.
4. Measurement. Standard MeasurementsLoading... computes shape from the coded image and intensity from a Shades image, per cell and, through a Mask, per compartment.
5. Classification. GatesLoading... and cutoffsLoading... on diagrams sort cells into positive and negative populations, the phenotypesLoading....
6. Spatial analysis. Distance mapsLoading..., proximity areasLoading..., Density of EventsLoading... and Phenotype InteractionsLoading... measure how those populations are arranged. Together they are how spatial phenotypingLoading... is done in StrataQuest, which has no engine of that name.

Backward and forward connection close the loop. Choose a gate or quadrant and its cells are outlined on the image; double-click a cell on the image and it is highlighted in the diagrams and Raw Data. Flow cytometry can't offer that link, because its cells are gone.

Source: StrataQuest 8.1 Engines Reference, ch. 7 (pp. 361–362), ch. 10 (p. 382), ch. 12 (p. 400), ch. 14 (p. 405), ch. 24 (p. 438), ch. 26 (p. 449), ch. 29 (p. 465), ch. 34 (pp. 483–484), ch. 36 (p. 500), ch. 41 (pp. 537–539); User Manual §12.1 (pp. 160–167), §12.6 (pp. 177–182), §12.7 (pp. 216–226).

Pawley describes the signal chain from specimen to measurement as preparation, contrast, photon collection, digitization, deconvolution, segmentation and measurement. In tissue cytometry every link counts, and each hands what happened upstream to everything after it. Preparation sets staining quality, photon collection sets precision through Poisson noise, and segmentation determines which pixels belong to which cell. Many of the decisions that limit an analysis are made at acquisition, before any image processing.

Roysam and colleagues, who work on automated analysis of cells in tissue, raise two cautions that apply here. Automated analysis requires better specimen preparation than manual scoring does, because a person adapts to a poor stain and an algorithm doesn't. And segmentation is best validated against more than one human annotator, since annotators disagree too.

To a statistician, typed cell positions in 2D form a marked spatial point process, and there is a toolkit for those. It includes local intensity (the Gaussian-weighted density of Density of EventsLoading... is one form), distance to landmarks (distance maps), neighbor counts within a distance (Phenotype Interactions), and second-order summaries such as the pair correlation function and Ripley's K, which compare observed neighbor counts with a random arrangement. StrataQuest computes the first three; the fourth runs on coordinates exported from Raw DataLoading....

Flow cytometry quantifies each dissociated cell precisely and keeps no spatial information. Tissue cytometry keeps it: you learn not only that 15% of the cells are CD8+ T cells, but whether they sit in the tumor or at its margin. In colorectal cancer, Galon and colleagues (2006) reported that the type, density and location of T cells in the tumor center and invasive margin were associated with patient outcome.

Take a colorectal cancer biopsy through the workflow.

1. Acquisition. A seven-channel multiplex IF panel images DAPI plus six markers: CD3, CD8, FOXP3, PD-L1, CK and CD20.
2. Detection. NucleiLoading... runs on the DAPI channel, and every nucleus becomes an event.
3. Measurement. The six markers measured in three compartments (nucleus, ring and identified mask) give 18 intensity measurements per cell, plus area and shape.
4. Phenotyping. Gating sorts the cells into seven populations: PD-L1+ tumor, PD-L1− tumor, CD3+CD8+ T, CD3+CD8− T, CD3+FOXP3+ regulatory T, CD20+ B cells and other.
5. Spatial analysis. Proximity AreasLoading... around the tumor region give CD3+ and CD8+ counts per band, Phenotype Interactions counts CD8+ T cells within 30 µm of each tumor cell, and Density of Events maps crowding.

From one section you get the cell types, their proportions per region, and how they are arranged. Whether an arrangement differs from chance is a further statistical step on the exported data.