Cell segmentation is critical for isolating single cell expression data from a tissue matrix. The CODEX Processor uses inputs from various parameters under the segmentation tab to tailor the segmentation to different cell types and signal intensities.
This technical note summarizes the results and experimental procedures used for optimizing the accuracy of CODEX current segmentation algorithm.
For this purpose, a 6-cycle CODEX run in a 3x3-tiles region of human tonsil FFPE tissue section was acquired using different exposure times for DAPI nuclear staining (5, 10 and 20 ms). A region of interest (ROI) of a representative tile was chosen for optimization to allow visual assessment of the segmentation accuracy (Figure 1).
Figure 1: A region is selected for optimizing the segmentation parameters Changing Parameters Affects Segmentation Outcome
In order to elucidate how different segmentation parameters and exposure times affect the segmentation outcome, different segmentation runs using the CODEX Processor were performed changing the value of only one parameter at a time while keeping all others as default (default values are reported in Figure 2).
Figure 2: Default parameters for segmentation Segmentation runs were also performed on images of the same ROI obtained with different exposure times using only default values. Figure 3 (A - E) shows the trend of the different parameters with respect to the cell count. It can be observed that the radius, the minimum cutoff and the size cutoff strongly influence the cell count, which tend to decrease as any of these values increase. Conversely, the relative cutoff and differences in exposure times have a negligible effect on the cell count calculated by the segmentation algorithm. The trend observed in the graph is depend on the sample.
Figure 3: Variation of the cell count as a function of one segmentation parameter on the 10ms DAPI image (A - D) and of the exposure time for segmentations with default values (E). Optimization of Segmentation Parameters
The DAPI image at 10ms exposure time was segmented using different parameters. Examples of segmentations runs be found in the Appendix. Segmentation accuracy was evaluated by visual identification of over- and under-segmented cells and of false segments (see Figure 5). The total cell count was also compared to the one calculated by visual inspection. For this dataset, the radius and the minimum mutoff values were determined to be critical parameters, and the segmentation was improved by using the following parameters:
Overlays between DAPI nuclear staining and segmentation masks obtained with default and optimized values are reported in Figure 4A and 4B, while the trends visually observed and used to evaluate the segmentation accuracy are reported in Figure 5.
Figure 4A: Superposition between DAPI staining and segmentations masks obtained with default parameters Figure 4B. Superposition between DAPI staining and segmentations masks obtained with optimized parameters. Figure 5: Cell counts: Comparison of the number of cells obtained by manual count and by segmentation using default and optimal parameters (A) Number of over- and under-segmented cells and number of missed segmentation events using different segmentation settings (B). Optimal parameters were also used to segment images of DAPI staining with different exposure times, and, analogous to segmentations done with default settings, the difference in exposure time did not cause important variations of the cell counts (Figure 6).
Figure 6: Comparison of cell counts calculated by visual inspection and by applying optimal segmentation parameters determined for 10ms to DAPI images obtained at different exposure times. This study demonstrates the parameter optimization process for segmentation using the CODEX Processor for the analysis of a human FFPE tonsil tissue section. While default parameter settings might be sufficient for some tissue types, altered parameters are often required to minimize inaccurate segmentation events including over-segmentation, under-segmentation and missed segmentation events.
