Segment human cells (in mitosis)

In this example, we analyze a microscopy image of human cells. We use data provided by Jason Moffat [1] through CellProfiler.

import matplotlib.pyplot as plt
import numpy as np
from scipy import ndimage as ndi

from skimage import (
    color, feature, filters, measure, morphology, segmentation, util
from import human_mitosis

image = human_mitosis()

fig, ax = plt.subplots()
ax.imshow(image, cmap='gray')
ax.set_title('Microscopy image of human cells stained for nuclear DNA')
Microscopy image of human cells stained for nuclear DNA

We can see many cell nuclei on a dark background. Most of them are smooth and have an elliptical shape. However, we can distinguish some brighter spots corresponding to nuclei undergoing mitosis (cell division).

Another way of visualizing a grayscale image is contour plotting:

fig, ax = plt.subplots(figsize=(5, 5))
qcs = ax.contour(image, origin='image')
ax.set_title('Contour plot of the same raw image')
Contour plot of the same raw image

The contour lines are drawn at these levels:

array([  0.,  40.,  80., 120., 160., 200., 240., 280.])

Each level has, respectively, the following number of segments:

[len(seg) for seg in qcs.allsegs]
[0, 320, 270, 48, 19, 3, 1, 0]

Estimate the mitotic index

Cell biology uses the mitotic index to quantify cell division and, hence, cell proliferation. By definition, it is the ratio of cells in mitosis over the total number of cells. To analyze the above image, we are thus interested in two thresholds: one separating the nuclei from the background, the other separating the dividing nuclei (brighter spots) from the non-dividing nuclei. To separate these three different classes of pixels, we resort to Multi-Otsu Thresholding.

thresholds = filters.threshold_multiotsu(image, classes=3)
regions = np.digitize(image, bins=thresholds)

fig, ax = plt.subplots(ncols=2, figsize=(10, 5))
ax[1].set_title('Multi-Otsu thresholding')
Original, Multi-Otsu thresholding

Since there are overlapping nuclei, thresholding is not enough to segment all the nuclei. If it were, we could readily compute a mitotic index for this sample:


Whoa, this can’t be! The number of dividing nuclei


is overestimated, while the total number of cells


is underestimated.

fig, ax = plt.subplots(ncols=3, figsize=(15, 5))
ax[2].set_title('All nuclei?')
ax[1].set_title('Dividing nuclei?')
Original, Dividing nuclei?, All nuclei?

Count dividing nuclei

Clearly, not all connected regions in the middle plot are dividing nuclei. On one hand, the second threshold (value of thresholds[1]) appears to be too low to separate those very bright areas corresponding to dividing nuclei from relatively bright pixels otherwise present in many nuclei. On the other hand, we want a smoother image, removing small spurious objects and, possibly, merging clusters of neighboring objects (some could correspond to two nuclei emerging from one cell division). In a way, the segmentation challenge we are facing with dividing nuclei is the opposite of that with (touching) cells.

To find suitable values for thresholds and filtering parameters, we proceed by dichotomy, visually and manually.

Dividing nuclei

We are left with


dividing nuclei in this sample.

Segment nuclei

To separate overlapping nuclei, we resort to Watershed segmentation. To visualize the segmentation conveniently, we colour-code the labelled regions using the color.label2rgb function, specifying the background label with argument bg_label=0.

distance = ndi.distance_transform_edt(cells)

local_max_coords = feature.peak_local_max(distance, min_distance=7)
local_max_mask = np.zeros(distance.shape, dtype=bool)
local_max_mask[tuple(local_max_coords.T)] = True
markers = measure.label(local_max_mask)

segmented_cells = segmentation.watershed(-distance, markers, mask=cells)

fig, ax = plt.subplots(ncols=2, figsize=(10, 5))
ax[0].imshow(cells, cmap='gray')
ax[0].set_title('Overlapping nuclei')
ax[1].imshow(color.label2rgb(segmented_cells, bg_label=0))
ax[1].set_title('Segmented nuclei')
Overlapping nuclei, Segmented nuclei

Additionally, we may use function color.label2rgb to overlay the original image with the segmentation result, using transparency (alpha parameter).

color_labels = color.label2rgb(segmented_cells, image, alpha=0.4, bg_label=0)

fig, ax = plt.subplots(figsize=(5, 5))
ax.set_title('Segmentation result over raw image')
Segmentation result over raw image

Finally, we find a total number of


cells in this sample. Therefore, we estimate the mitotic index to be:

print(cleaned_dividing.max() / segmented_cells.max())

Total running time of the script: ( 0 minutes 0.829 seconds)

Gallery generated by Sphinx-Gallery