Robust consensus nuclear and cell segmentation

Abstract

Cell segmentation is a crucial step in numerous biomedical imaging endeavors-so much so that the community is flooded with publicly available, state-of-the-art segmentation techniques ready for out-of-the-box use. Assessing the strengths and limitations of each method on a tissue sample set and then selecting the optimal method for each research objective and input image are time-consuming and exacting tasks that often monopolize the resources of biologists, biochemists, immunologists, and pathologists, despite not being the primary goal of their research projects. In this work, we present a segmentation software wrapper, coined CellSampler, which runs a selection of established segmentation methods and then combines their individual segmentation masks into a single optimized mask. This so-called "uber mask" selects the best of the established masks across local neighborhoods within the image, where both the neighborhood size and the statistical measure used to define what qualifies as "best" are user-defined.

Keywords: bioinformatics; computer vision; imaging mass cytometry; multiplexed imaging; single-cell segmentation.

FIGURE 1

Schematic overview of the CellSampler pipeline. The input image is processed using various state-of-the-art segmentation algorithms, producing segmentation masks, which are then combined into a single optimal mask: the “uber mask.” The inclusion of an unknown method ‘X’ is to highlight the fact that CellSampler can, and will, be extended with additional segmentation methods.

FIGURE 2

Ten example images (nuclear channel) from the BRCA2 dataset.

FIGURE 3

Top row: the cell detections selected from each method using the uber mask for our example image. Middle row: the uber mask contours in red, overlaid on the input image. From left to right: the uber mask made using 1) the reciprocal of the natural logarithm of the standard deviations of the nucleus areas, 2) the number of nuclei detected, and 3) the Jaccard score as the score of merit. Bottom row: the ground-truth mask.

FIGURE 4

Top left: comparison between the median scores for all the nuclear masks of the IMC dataset. Top right: the percentage of split and merge errors in our 258 images. Bottom row: the same as the top left but as a bar chart, where the error bars are the mean absolute deviation values for the scores.

FIGURE 5

Top left: comparison between the median scores for all the nuclear masks created for the fluorescent imaging dataset of 49 images. Top right: the percentage of split and merge errors in our 49 images. Bottom row: the same as the top left but as a bar chart, where the error bars are the mean absolute deviation values for the scores.

FIGURE 6

Two IMC images of mouse tissue: lung tissue on the left and ovarian tissue on the right.

FIGURE 7

Specific region of the lung IMC sample. The nuclei segmented through Cellpose v2, Mesmer, watershed, and the uber mask are shown as red contours overlaid onto the nuclear channel data.

FIGURE 8

Specific region of the lung IMC sample. The top left image shows the uber mask nuclear detections as contours over the original image. The contours are colored according to the method that detected them. The top right, bottom left, and bottom right images show the Cellpose v2, Mesmer, and watershed detections in purple, teal, and light brown, respectively.

FIGURE 9

Specific region of the ovarian IMC sample. The nuclei segmented through Cellpose, Cellpose v2, Mesmer, Mesmer v2, and the uber mask are shown as red contours overlaid onto the nuclear channel data.

FIGURE 10

Specific region of the ovarian IMC sample. The top left image shows the uber mask nuclei detections as contours over the original image. The contours are colored according to the method that detected them. The top right, middle left, middle right, and bottom left images show the Cellpose, Cellpose v2, Mesmer, and Mesmer v2 detections in light blue, purple, teal, and light green, respectively.

FIGURE 11

Examples of some diagnostic plots produced for the segmentation of the IMC mouse-lung tissue. The top row shows the histogram distribution of nuclear areas. On the bottom row, the left-hand plot shows the 50 nearest-neighbor nuclear density plot, and the right-hand plot shows the brightest nuclei in the CD68 channel in cyan and the CD11b channel in magenta, overlaid onto all the detected nuclei (shown in black). The figures in the bottom row were generated using the UM(2) segmentations.