minimal-surface is a Python package which can be utilized for the segmentation of objects in 3D images using as few as 2 points and a 2D slice! This algorithm uses energy minimization to estimate the surface of the object. To achieve fast execution the method was implemented in C++, but can be called from Python through a thin wrapper.
When the pen is mightier than the sword: semi-automatic 2 and 3D image labelling
Réka Hollandi, David Bauer, Akos Diosdi, Bálint Schrettner, Timea Toth, Dominik Hirling, Gábor Hollandi, Maria Harmati, József Molnár, Peter Horvath
bioRxiv 2024.01.15.575658
doi: https://doi.org/10.1101/2024.01.15.575658
It can be installed from PyPI using pip:
python -m pip install minimal-surface
Note: This package is currently available only for Windows. In the future we plan to release it for Linux and Mac systems.
The easiest way to utilize our tool is by using Annotation Toolbox, a napari plugin created for fast 2 and 3D image annotation.
minimal-surface-demo.mp4
In order to run this example, you'll need these packages:
- numpy
- scipy
- scikit-image
- pooch
- matplotlib
import minimal_surface
from skimage.data import cells3d
from scipy import ndimage
import matplotlib.pyplot as plt
import numpy as np
def calc_features(image):
delta = 0.1
max_ = np.quantile(image, .95)
min_ = image.min()
image = np.clip((image-min_)/(max_-min_), 0, 1)
# This feature will consider both image gradient and intensity
# (giving higher values when the gradient is high and the intensity is low)
gradient = ndimage.gaussian_gradient_magnitude(image, (1., 1., 1.))
weights = 1+(delta-1)*image**2
phi = gradient*weights +delta*(1-image**2)
phi = (phi-(min_:=phi.min()))/(phi.max()-min_)
return phi
# Crop the part to the smallest size possible to prevent long computation
data = cells3d()[:, 1, 60:130, 130:190]
# Two points on the surface of the nucleus
p1 = np.asarray([ 20., 60., 35.])
p2 = np.asarray([ 39., 12., 35.])
data = (data - (min_ := data.min())) / (data.max() - min_)
data = ndimage.gaussian_filter(data, 1.)
features = calc_features(data)
alpha = 3e-3
phi = alpha + (1-alpha)*np.exp(-5*features)
phi = phi/phi.max()
calculator = minimal_surface.MinimalSurfaceCalculator()
def segment_slice(image_slice, distance_map):
# This function segments the input image slice.
# Asking for user input is also possible for accurate segmentation.
return image_slice >= 0.24
calculator.set_initial_plane_calculator(segment_slice)
# If you have multiple objects, you can run the first part separately. which requires user input,
# then finish the computationally expensive part together for the whole dataset.
# calculator.calc_eikonal_and_transport_init(phi, data, p1, p2, True)
transport_function = calculator.calculate(phi, data, p1, p2)
print("calculated")
mask = transport_function >= 0.
# The segmentation is contained in mask
plt.imshow(mask[35], cmap="plasma")
plt.show()If you bump into any problems regarding the package, please file an issue.