Motivation

Because the results of segmentation are not perfect, some of the cells that are detected are not real cells. Ideally, it would be possible to eliminate these “false cells” from downstream analysis to achieve more accurate results.

The Cell QC workflow aims to filter out “false cells”.

<aside> 💡 The goal of Cell QC should be to eliminate the most incorrectly segmented cells while keeping all correctly detected cells using the metrics that have been computed. Thus, the QC parameters should be adjusted accordingly. There are no default settings for Cell QC.

</aside>

Methodology

“False cells” often fall into a few general categories that can be easily filtered out:

• Abnormal cell size: “false cells” often are either too large or too small to be real cells;

  

  Nuclear segmentation outlines (white lines) and DAPI channel (blue) of a human skin sample. In this largely acellular region, many of the detected cell nuclei are abnormally large or small, and do not correspond to real cells.

• Abnormal signal sum: “false cells” often have very low signal across biomarker channels because no biomarkers are localized to the cell;

  

  Human rectum sample displaying CODEX stains for PanCK (green), CD45 (red), and Vimentin (yellow), with the nuclear segmentation mask (white lines). The large detected cells at the bottom of the image have low signal for each display channel and are not true cells.

• Abnormal DNA signal: “false cells” often either have very low signal, which means that they are just background noise, or very high signal, which tends to indicate a staining artifact;

  

  Human rectum sample with Hoechst nuclear staining (blue) and nuclear segmentation mask (white lines), same ROI as previous image. Some of the detected cells have very low DNA signal, including the bottom two large cells. These are not real cells and should be omitted from downstream analysis.

• Low signal variability across biomarker channels: “false cells” often have low signal variability due to autofluorescence, which means that a similar expression pattern is detected across many acquisition channels, and/or non-specific staining of acellular structures;

  

  An image of a mouse pancreas displaying staining for DAPI (blue), CD45 (red), PanCK (green), and aSMA (yellow). The cells in the center of the image are highly positive for all of the biomarkers displayed and have low signal variability in general. This is likely a staining artifact.

• Abnormal location: some “false cells” are detected far away from the main tissue mass, which usually indicates an acquisition or edge noise artifact.

  

  Nuclear segmentation mask (white lines) and the DAPI channel (blue) for a single TMA core. The left side of this image has spatial outliers. While there may be real cells there, they may not be relevant for analysis or skew the results of the analysis.

During Cell QC, metrics for cell size, signal sum, DNA signal strength, signal variability, and distance to the nth nearest cell are computed, so that cells with these types of abnormalities can be easily filtered out.

The Cell QC Extension is a tool that allows filtering based on the first four metrics. To filter out detected cells based on location/manually drawn ROIs, Cell QC will need to be performed through Workbench (an example of this workflow is available in the workbench notebooks in the analysis toolkit, linked here).

Best Practices

Since the cell size, biomarker and nuclear signal strengths, and physical arrangement of the tissue samples varies a lot depending on tissue type, sample preparation, instrumentation, and acquisition setting, there are no default settings for Cell QC.

One common way of performing Cell QC filtering is setting a simple cutoff such that 99%, 95%, or 90% of cells fall within for each metric pass. In other words, cells that are below the 0.5th, 2.5th, or 5th percentile and above the 99.5th, 97.5th, or 95th percentile will be filtered out. You can decide on the cutoff percentage ahead of time by visually assessing the segmentation run to get a feel for the proportion of incorrectly detected cells. Most images and segmentation runs do benefit from a more tailored approach, but it can take time to find the most optimized parameters.

Regardless, it is always advised to push Cell QC pass/fail annotations into the Visualizer to generate an overlay, and check that the results are consistent with your observations.

Expected Outcomes

From the Cell QC filtering process, cells will have an additional property as having “passed” or “failed” QC. Thus, when querying the data, you will have the option to only include cells which pass Cell QC in downstream analysis (i.e., exclude failed cells). If the image and segmentation quality is good and the QC parameters are chosen well, most “false cells” should fail QC while true cells should pass QC.