Building a segmentation pipeline

In this module we extend our thresholding widget into a full segmentation pipeline: we add morphological cleaning, label connected components, compute quantitative features (area, centroid), display them on a Points layer, and finally refine the result with a watershed step seeded by manual point annotations.

1. Thresholding + morphology + labels¶

Let’s build on the previous widget. We add two new parameters — min_hole_size and min_obj_size — and return intermediate layers so you can see each processing step.

04_labels.py

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
from skimage import data, filters, morphology, measure
import napari
import numpy as np
from typing import Annotated


def threshold_and_label(
    layer: napari.layers.Image,
    sigma: Annotated[float, {'widget_type': 'FloatSlider', 'min': 0, 'max': 2, 'step': 0.1}] = 0.5,
    threshold: Annotated[float, {'widget_type': 'FloatSlider', 'min': 0, 'max': 1, 'step': 0.05}] = 0.3,
    min_hole_size: Annotated[int, {'widget_type': 'Slider', 'min': 0, 'max': 1000, 'step': 50}] = 0,
    min_obj_size: Annotated[int, {'widget_type': 'Slider', 'min': 0, 'max': 1000, 'step': 50}] = 0,
) -> list[napari.types.LayerDataTuple]:
    """Apply a gaussian filter, threshold, and compute labels on a napari Image."""
    if not layer:
        return
    norm = (layer.data - np.min(layer.data)) / np.max(layer.data)
    blur = filters.gaussian(norm, sigma=sigma)
    blobs = blur >= threshold

    filled = morphology.remove_small_holes(blobs, max_size=min_hole_size)
    cleaned = morphology.remove_small_objects(filled, max_size=min_obj_size)
    labels = measure.label(cleaned)

    return [
        (blur, {'name': 'blur'}, 'image'),
        (blobs, {'name': 'blobs'}, 'image'),
        (filled, {'name': 'filled'}, 'image'),
        (cleaned, {'name': 'cleaned'}, 'image'),
        (labels, {'name': 'result'}, 'labels'),
    ]


viewer = napari.Viewer()
image = data.cells3d()[30, 1]  # 2d
image_layer = viewer.add_image(image)

viewer.window.add_function_widget(threshold_and_label, magic_kwargs={'auto_call': True})

Play with min_hole_size and min_obj_size to clean up the binary mask before labeling. You can toggle visibility (eye icon in the layerlist) of the intermediate layers to see what each step does.

from napari.utils import nbscreenshot

t = viewer.window.dock_widgets['threshold and label']
t.sigma.value = 1
t.threshold.value = 0.15
t.min_hole_size.value = 471
t.min_obj_size.value = 79
t()
nbscreenshot(viewer)

2. Quantitative features + Points layer¶

Let’s compute region properties (area, centroid) for each detected object and display them. We use skimage.measure.regionprops_table and attach the results as layer features, then add a Points layer with the centroids.

05_features_and_points.py

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
from skimage import data, filters, morphology, measure
import napari
import numpy as np
from typing import Annotated


def threshold_and_label(
    layer: napari.layers.Image,
    sigma: Annotated[float, {'widget_type': 'FloatSlider', 'min': 0, 'max': 2, 'step': 0.1}] = 0.5,
    threshold: Annotated[float, {'widget_type': 'FloatSlider', 'min': 0, 'max': 1, 'step': 0.05}] = 0.3,
    min_hole_size: Annotated[int, {'widget_type': 'Slider', 'min': 0, 'max': 1000, 'step': 50}] = 0,
    min_obj_size: Annotated[int, {'widget_type': 'Slider', 'min': 0, 'max': 1000, 'step': 50}] = 0,
) -> list[napari.types.LayerDataTuple]:
    """Apply a gaussian filter, threshold, and compute labels on a napari Image.

    Label properties (area and centroid) are also computed and exposed via layer
    `features`. Centroids are also shown in a Points layer.
    """
    norm = (layer.data - np.min(layer.data)) / np.max(layer.data)
    blur = filters.gaussian(norm, sigma=sigma)
    blobs = blur >= threshold

    filled = morphology.remove_small_holes(blobs, max_size=min_hole_size)
    cleaned = morphology.remove_small_objects(filled, max_size=min_obj_size)
    labels = measure.label(cleaned)

    props = measure.regionprops_table(labels, properties=['label', 'area', 'centroid'])
    props['index'] = props.pop('label')
    centroids = np.array([props[f'centroid-{i}'] for i in range(layer.ndim)]).T

    return [
        (blur, {'name': 'blur'}, 'image'),
        (blobs, {'name': 'blobs'}, 'image'),
        (filled, {'name': 'filled'}, 'image'),
        (cleaned, {'name': 'cleaned'}, 'image'),
        (labels, {'name': 'result', 'features': props}, 'labels'),
        (centroids, {'name': 'centroids', 'features': props}, 'points'),
    ]


viewer = napari.Viewer()
image = data.cells3d()[30, 1]  # 2d
image_layer = viewer.add_image(image)

viewer.window.add_function_widget(threshold_and_label, magic_kwargs={'auto_call': True})

The number of layers is starting to be high and knowing exactly which parameter to adjust can be tricky without an overview of all the steps. Let’s enable the grid view to spread out each individual layer into its own viewbox. We can also enable the layer name overlay on each layer, to make it easier to know what’s what.

viewer.canvas.grid.enabled = True
for layer in viewer.layers:
    layer.name_overlay.visible = True
viewer.reset_view()

Sometimes the label boundaries are not perfect — especially when objects touch. We can refine them using a watershed segmentation, seeded by points we place manually.

For this, let’s implement a separate function, since this requires manual intervention and is also too computationally expensive to run continuously.

06_watershed.py

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
from skimage import data, filters, morphology, measure, segmentation
import napari
import numpy as np
from typing import Annotated
from scipy import ndimage as ndi


def threshold_and_label(
    layer: napari.layers.Image,
    sigma: Annotated[float, {'widget_type': 'FloatSlider', 'min': 0, 'max': 2, 'step': 0.1}] = 0.5,
    threshold: Annotated[float, {'widget_type': 'FloatSlider', 'min': 0, 'max': 1, 'step': 0.05}] = 0.3,
    min_hole_size: Annotated[int, {'widget_type': 'Slider', 'min': 0, 'max': 1000, 'step': 50}] = 0,
    min_obj_size: Annotated[int, {'widget_type': 'Slider', 'min': 0, 'max': 1000, 'step': 50}] = 0,
) -> list[napari.types.LayerDataTuple]:
    """Apply a gaussian filter, threshold, and compute labels on a napari Image."""
    if not layer:
        return
    norm = (layer.data - np.min(layer.data)) / np.max(layer.data)
    blur = filters.gaussian(norm, sigma=sigma)
    blobs = blur >= threshold

    filled = morphology.remove_small_holes(blobs, max_size=min_hole_size)
    cleaned = morphology.remove_small_objects(filled, max_size=min_obj_size)
    labels = measure.label(cleaned)

    props = measure.regionprops_table(labels, properties=['label', 'area', 'centroid'])
    props['index'] = props.pop('label')
    centroids = np.array([props[f'centroid-{i}'] for i in range(layer.ndim)]).T

    return [
        (blur, {'name': 'blur'}, 'image'),
        (blobs, {'name': 'blobs'}, 'image'),
        (filled, {'name': 'filled'}, 'image'),
        (cleaned, {'name': 'cleaned'}, 'image'),
        (labels, {'name': 'result', 'features': props}, 'labels'),
        (centroids, {'name': 'centroids', 'features': props}, 'points'),
    ]


def watershed(
    markers: napari.layers.Points,
    labels: napari.layers.Labels,
) -> list[napari.types.LayerDataTuple]:
    """Improve Labels using watershed and seeds from a Points layer."""
    if not markers or not labels:
        return
    base_labels = labels.data != 0
    distance_field = ndi.distance_transform_edt(base_labels)

    # generate seeds for the watershed algorithm from point markers
    markers_array = np.zeros_like(base_labels, dtype=bool)
    markers_array[tuple(markers.data.astype(int).T)] = True
    markers = ndi.label(markers_array)[0]

    watershedded = segmentation.watershed(-distance_field, markers, mask=base_labels)

    return [
        (distance_field, {'name': 'distance field'}, 'image'),
        (watershedded, {'name': 'watershed'}, 'labels'),
    ]


viewer = napari.Viewer()
image = data.cells3d()[30, 1]  # 2d
image_layer = viewer.add_image(image)

viewer.window.add_function_widget(threshold_and_label, magic_kwargs={'auto_call': True})
viewer.window.add_function_widget(watershed)

t = viewer.window.dock_widgets['threshold and label']
t.sigma.value = 1
t.threshold.value = 0.15
t.min_hole_size.value = 471
t.min_obj_size.value = 79
t()

# done this manually so we can show what it looks like in the screenshot
manual_seeds = np.array(
    [
        [ 13.96171682,  26.20149842],
        [ 30.29515511,  80.64629274],
        [ 10.33206386, 155.05417831],
        [ 41.18411398, 178.64692252],
        [ 39.3692875 , 222.20275798],
        [ 70.22133761, 111.49834286],
        [ 75.66581705,  58.86837501],
        [ 93.81408182, 160.49865775],
        [ 95.6289083 , 249.42515514],
        [144.62922319, 242.16584923],
        [144.62922319, 193.16553434],
        [150.07370262, 115.12799581],
        [173.66644682,  49.79424263],
        [220.85193524,  84.2759457 ],
        [242.62985296, 142.35039297],
        [246.25950592, 211.31379911],
        [219.03710876, 245.79550218],
        [200.57463293, 171.9494017 ],
        [137.28352217,  33.42546117]
    ]
)

l = viewer.add_points(manual_seeds)
w = viewer.window.dock_widgets['watershed']
w.markers.value = l
w()
nbscreenshot(viewer)

Recap¶

1. Added morphological cleaning (remove_small_holes, remove_small_objects) and connected-component labeling.
2. Computed region properties (area, centroid) and displayed them as layer features and a Points layer.
3. Added a watershed refinement widget that uses manual point markers to split touching objects.

Next up: mouse callbacks for interactive label inspection, and taking our pipeline into 3D!

Building a segmentation pipeline

1. Thresholding + morphology + labels¶

2. Quantitative features + Points layer¶

3. Watershed refinement¶

Recap¶