Crop cell

Cropping cell from microscopy image with Otsu's Method and OpenCV¶

In [1]:

# Imports
import numpy as np
from skimage.filters import threshold_otsu
from PIL import Image
import matplotlib.pyplot as plt
import skimage
from torch.nn import Conv2d
import torch
import cv2
import time
from random import sample

In [2]:

# Helper functions
def load_and_process_image(path = "26118-I10-4.npz"):
    img_gray = np.load(path)["sample"][:,:,2]
    img_rgb = cv2.cvtColor(img_gray,cv2.COLOR_GRAY2RGB)
    img_gray = img_gray <= threshold_otsu(img_gray, nbins = 1024)
    img_gray = np.where(img_gray, 0, 1)
    
    thresh = cv2.threshold(img_gray.astype(np.uint8),0,1,cv2.THRESH_BINARY)[1]
    plt.imshow(thresh, cmap = "gray")
    result = img_rgb.copy()
    contours = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    contours = contours[0] if len(contours) == 2 else contours[1]
    list_centers = []
    for cntr in contours:
        x,y,w,h = cv2.boundingRect(cntr)
        cv2.rectangle(result, (x, y), (x+w, y+h), (0, 255, 0), 2)
        cv2.circle(result, (int(x+(w/2)), int(y+(h/2))), radius=2, color=(0, 0, 255), thickness=-1)
        list_centers.append((x+(w/2), y+(h/2)))
    return(result, list_centers)


def cell_crop(list_centers, path = "26118-I10-4.npz", h = 64, w = 64):
    img_gray = np.load(path)["sample"]
    list_image = []
    for dye in range(5):
        channel = img_gray[:,:,dye]
        #channel = cv2.cvtColor(channel,cv2.COLOR_GRAY2RGB)
        list_crop = []
        for center in list_centers:
            x = int(center[0]-(w*0.5))
            y = int(center[1]-(h*0.5))
            
            if x>=0 and y>=0 and x+w<=696 and y+h<=520:
                crop_img = channel[y:y+h, x:x+w]
                list_crop.append(crop_img)
            else:
                continue
            #cv2.rectangle(channel, (x, y), (x+width, y+height), (0, 255, 0), 2)
        list_image.append(list_crop)
    return(list_image)

In one of my projects, I needed to crop single cells out of microscopy images for a modelling task. The requirement is to do everything in Python, and not to use third-party softwares.

Here is how I did it, using Otsu's method

Brief intro to Cell Painting Images:¶

For the same view, we 'paint' different cell component with a different dye (of different excitation ranges) to observe morphological changes.

The result is 5 'color channels' for each view (like, say, normal RGB images have 3 channels.)

In [3]:

fig = plt.figure(figsize=(20, 3))
counter=1
for i in range(5):
    fig.add_subplot(1, 6, counter)
    plt.imshow(np.load('26118-I10-4.npz')["sample"][:,:,i], cmap = "gray")
    counter+=1

In [4]:

# Up close the 'Ph_golgi' channel
# Golgi body dyed by Phalloidin/Alexa Fluor 594 conjugate, wheat germ agglutinin (WGA)/Alexa Fluor 594 conjugate
plt.imshow(np.load('26118-I10-4.npz')["sample"][:,:,4], cmap = "gray")

Out[4]:

<matplotlib.image.AxesImage at 0x7fe423080a90>

The mission¶

... is to crop cells in all 5 channels, centered on the nucleus. The cropped cells should align perfectly over 5 channels

The method¶

... is called Otsu's method. It is actually a quite popular for cell cropping.

So how does it work?

https://en.wikipedia.org/wiki/Otsu%27s_method#Algorithm

tl,dr (too long, didn't read): The algorithm returns a single intensity threshold that separate pixels into two classes, foreground and background.

We can use this to 'pop' objects out of the background

1. Pop the nucleus out¶

Channel 'Hoechst' is responsible for dyeing the nucleus

In [5]:

plt.imshow(np.load('26118-I10-4.npz')["sample"][:,:,2], cmap = "gray")

Out[5]:

<matplotlib.image.AxesImage at 0x7fe422dc2a00>

We use this channel to 'pop' the nuclei out from the dark background

In [6]:

result, list_centers = load_and_process_image()

2. Find the centers of the nuclei¶

Based on the 'popped' nuclei, we can draw the bounding box surrounding the nuclei (with openCV), and find the centers of the nuclei (blue dot).

In [7]:

plt.imshow(result, cmap = "gray")

Out[7]:

<matplotlib.image.AxesImage at 0x7fe422cc8100>

These centers will be the center of our 64x64 cropped cell image

3. Crop cells into 64x64 images¶

We need all images to be of the same size for Machine Learning modelling.

In [8]:

list_image = cell_crop(list_centers)

Samples of cropped cells:

In [9]:

random_list = sample(range(len(list_image[0])),6)
fig = plt.figure(figsize=(20, 3))
counter=1
for i in random_list:
    fig.add_subplot(1, 6, counter)
    plt.imshow(list_image[0][i], cmap = "gray")
    counter+=1

In [10]:

#Sample 6 cropped cells from Ph_golgi
fig = plt.figure(figsize=(20, 3))
counter=1
for i in random_list:
    fig.add_subplot(1, 6, counter)
    plt.imshow(list_image[1][i], cmap = "gray")
    counter+=1

In [11]:

#Sample 6 cropped cells from ER_Syto
fig = plt.figure(figsize=(20, 3))
counter=1
for i in random_list:
    fig.add_subplot(1, 6, counter)
    plt.imshow(list_image[2][i], cmap = "gray")
    counter+=1

In [12]:

#Sample 6 cropped cells from Hoechst
fig = plt.figure(figsize=(20, 3))
counter=1
for i in random_list:
    fig.add_subplot(1, 6, counter)
    plt.imshow(list_image[3][i], cmap = "gray")
    counter+=1