Counting and identifying stained cells step by step

On Friday, Dr Joaquín de Navascués and I held the first of two Image Analysis Workshops at Cardiff University. We had a good audience with PhD students and researchers. Joaquin led the first session which explored the fundamentals of images and using Image J or Fiji to open and explore those images.

Key points included:

• good images start at the microscope - try not to over saturate your images during collection.
• protect your original files as it can be easy to alter them permanently and lose data.
• digital pictures are arrays of numbers. 
• these numbers can be visualised and transformed in lots of different ways. 

In just under two weeks time, it's my turn. I am going to talk about using R to analyse images. I like R because it allows the development of reproducible, sharable and scalable workflows. The aim of the second workshop is to show an automated workflow that attendees can adapt to their own work if they want.

The key package that is useful for analysing images in R is EBImage developed by members of the EMBL and the EBI. It's a powerful and useful package.

This script below is one that will probably be used during the workshop. It is about using R to count the numbers of cells in an image. I'm going to go through the process step by step. The steps are as follows:

1. Download the image from Github
2. Make the image a little brighter
3. Blur the image using a Gaussian filter
4. Apply an Otsu threshold the image to convert to a binary image
5. Count the connected regions - i.e. the cells.
6. Display the regions in a multi-coloured output. 

Here is a confocal microscope picture and the image with the cell regions identified:

Here is the script that identifies and counts the regions:

SCRIPT START
# aim of this script is to identify and count cells....
library(EBImage)
library(RCurl)

# this is the link to the image
link <- "https://raw.githubusercontent.com/brennanpincardiff/RforBiochemists/master/data/images/Dros_c2.tif"
# the download.file() function downloads and saves the file with the name given
download.file(url=link, destfile="file.tif", mode="wb")

# import image of stained cells into R
c2 <- readImage("file.tif")
# creates an object called c2

# show the image in the R graphics window
display(c2, method = "raster")  #"raster" method means within R

# display a brighter image
display(c2*4, method = "raster") 
# because an image is numbers we just multiply to make it brighter

# check details by writing the name of the object
c2 # shows some information
# it's a greyscale image 

# make a brighter image by multiplying all the values by 2
c2.b <- c2*2

# gaussian blur
c2.b.blur <- gblur(c2.b, sigma = 5)
display(c2.b.blur, method = "raster")

Can you spot the difference?

## threshold using Otsu'smethod 
otsu(c2.b.blur) # gives a threshold value using Otsu algorithm
# value = 0.1777344
c2.b.blur.thres <- c2.b.blur > otsu(c2.b.blur) # apply this value 
display(c2.b.blur.thres, method = "raster")

# we see that the image is starkly black and white
# all pixels have been turned into either 0 or 1 - a binary image.

# generate an image with different values for each connected region
# key here is the bwlabel( ) function
c2.b.blur.thres.cnt <- bwlabel(c2.b.blur.thres)

# show this as a coloured blobs 
display(colorLabels(c2.b.blur.thres.cnt), method = "raster")

# count by giving us the max value in the bwlabel() function
nucNo <- max(bwlabel(c2.b.blur.thres))
# output this number to the Console
nucNo  # count = 41. 

# do we need to brighten the image (probably not in this case)?
c2.blur.thres <- gblur(c2, sigma = 5) > otsu(gblur(c2, sigma = 5))
display(c2.blur.thres, method = "raster")
max(bwlabel(c2.blur.thres))
# answer is 42 - so not a big difference - good staining. 

# what happens if we don't blur?
# we can calculate the Otsu threshold to the original image
otsu(c2)
c2.thres <- c2 > otsu(c2)
display(colorLabels(bwlabel(c2.thres)), method = "raster")

# doesn't look that different to the eye but...
# if we try counting....
max(bwlabel(c2.thres))
# answer is 1398 - lots of dots causing problems... 

# there are other ways to apply thresholds but that's for later
SCRIPT END


Resources:

• Bioconductor link about EBImage include Introduction & Reference Manual
• Supplementary information from a Nature Methods paper (2013) which using EBImage and R to generate lots of nice data - it's very detailed but interesting.