Hello,
I’m transferring a cell segmentation pipeline from python/skimage to Julia/Images.jl, and ran into some function calls that weren’t straightforward to translate. In particular:
- autolevel_percentile
- skimage.morphology.disk
- skimage.feature.match_template
- peak_local_max
The first two I can write if need be. match_template and peak_local_max, with its footprint option, would seem to be much more involved, so wanted to ask if anyone was aware of pertinent Julia options?
Also, if anyone has other comments, feedback or code to point me to that may be relevant, even if taking a different approach, that would be great! I did see Celltracker possibilities? - #4 by mschauer, which is helpful albeit in a lower density regime.
import os
import numpy as np
import scipy
from scipy import ndimage
import skimage
from skimage.morphology import disk, watershed
from skimage.filter import rank
from skimage.filter.rank import otsu, autolevel_percentile, enhance_contrast_percentile
from skimage.filter import threshold_otsu
from skimage.feature import peak_local_max
from skimage.exposure import adjust_gamma
def segment_nuclei(img):
maxval = np.max(img)
img_gamma = np.power(img,0.45) * maxval / np.power(maxval,0.45)
img_gamma = img_gamma.astype('uint16')
# use image histogram to filter out background
nbins = 200
hist, bin_edges = np.histogram(np.ravel(img_gamma), bins=nbins)
hthres = bin_edges[np.argmax(hist) + 6]
#print "Histogram threshold: " + str(float(hthres) / np.max(img_gamma))
fg_mask = (img_gamma >= hthres)
# autolevel_percentile
selem = disk(4)
img_level = autolevel_percentile(img_gamma, selem=selem, p0=.1, p1=.90)
img_level *= fg_mask # apply otsu mask
# otsu threshold 2 to binary image
gthres = threshold_otsu(img_level)
gthres*=1.2
img_gthres = img_level >= gthres
# find cell centers
kern = disk(2)
distance = ndimage.distance_transform_edt(img_gthres)
feat_img = skimage.feature.match_template(img_level, kern, pad_input=True, mode='constant', constant_values=0)
local_maxi = peak_local_max(feat_img, indices=False, footprint=disk(4), exclude_border=False, threshold_rel=.4)
# apply watershed
markers = ndimage.label(local_maxi)[0]
labels = watershed(-feat_img, markers, mask=img_gthres)
# exclude regions that are too small, too large
roi_props = skimage.measure.regionprops(labels, intensity_image=None, cache=True)
bKeep = np.array([(rp['area']>5) & (rp['area']<70) & (rp['eccentricity']<.9) for rp in roi_props]).astype(bool)
labels = np.zeros_like(labels)
from itertools import compress
for nLabel, rp in enumerate(compress(roi_props,bKeep)):
labels[rp['coords'][:,0], rp['coords'][:,1]] = nLabel
roi_props = skimage.measure.regionprops(labels, intensity_image=None, cache=True)
# filter rois according to properities
label_img = skimage.color.label2rgb(labels, image=None, colors=None, alpha=1, bg_label=0, bg_color=[0,0,0], image_alpha=.5)
intermediates = {}
intermediates['img_gamma'] = img_gamma
intermediates['img_level'] = img_level
intermediates['img_gthres'] = img_gthres
intermediates['img_gthres'] = img_gthres
rois = labels
roi_image = label_img
return roi_props, rois, roi_image, intermediates
GOAL: to segment all of these cells 
Edit: some tips here for peak_local_max, but 1D case only. Maybe findlocalmaxima does what I need, but it’s not clear from documentation if something like min_distance can be achieved.
anyone for adding some.