Case Study 1: Counting Cells from Microscopic Images

2019

Date Source: 

The image set BBBC005v1 (https://data.broadinstitute.org/bbbc/BBBC005/) is from the Broad Bioimage Benchmark Collection [Ljosa et al., Nature Methods, 2012].

Organizer: 

Dr. Pingzhao Hu, Qian Liu, Department of Biochemistry and Medical Genetics, University of Manitoba; Dr. Kathryn Morrison, Precision Analytics Inc. and McGill University.

The emerging of high-throughput microscopic imaging modality triggers the “big data” problem even in a single experiment. It is no longer possible to manually analyze the microscopic images even for estimating a simple feature like cell count. Therefore, the needs for automated cell counting become crucial.

BBBC005v1 is a simulated microscopic image dataset. Each of the images is in an 8-bit TIF format with size 696 x 520 pixels. It provides in-focus (clear) and out-of-focus (blur) synthetic images. Each of these images was simulated for a given cell count. Gaussian filters were applied to simulate the images at different blur levels. The nuclei and cell body areas of the images were matched to the average nuclei and cell areas from a real microscopic dataset. 

Images with three (F1, F23, F48) of the 16 blur levels of the BBBC005v1 dataset were selected as our raw image data for this case study. Both images with cell body stain and nuclei stain are included. This includes a total of 3,600 microscopic images.
 

Research Question: 

Students are invited to develop statistical and computational methods to estimate the cell counts in the images. 
 

Variables: 

Competition Strategies

These 3,600 images were randomly assigned to a training set with 2,400 images and a test set with 1,200 images. The true cell count for each of the 2,400 images in the training set is provided while the true cell count for each of the 1,200 images in the test set is blinded. Students are required to develop statistical and computational methods to build a prediction model (s) based on the training set, which will be applied to predict the cell counts in the test set.

Students are required to submit their predicted cell counts for the 1,200 images in the test set to Dr. Pingzhao Hu at pingzhao.hu@umanitoba.ca before or at May 20, 2019. The performance of the prediction results will be evaluated based on the root mean square error (RMSE):

$RMSE = \sqrt{mean(c-\hat{c})^2}$

Here, c = [x1, x2, …, xn] represent the true cell counts of the 1,200 images in the test set, and ĉ = [y1, y2, …, yn] represent the predicted cell counts of the same 1,200 images.

Each team is also required to make a poster presentation in the SSC 2019 Case Studies Competition. The model prediction will take 60% and poster presentation will take 40% of the total mark of a team. The final score will be marked as 0.6/rank of the model predictions + 0.4/rank of the poster presentations.

 

Data Access: 

How to download the data sets: The data set can be downloaded from here: 

   https://www.dropbox.com/sh/buofl2fhvyfi5bc/AAArMZbeKncXfz64kcY17l0pa?dl=...  (400MB file)

Please email pingzhao.hu@umanitoba.ca if you have any questions about the data set. 


Organizer: 

Dr. Pingzhao Hu
Department of Biochemistry and Medical Genetics/Department of Computer Science
University of Manitoba

Division of Biostatistics, University of Toronto
e-mail: pingzhao.hu@umanitoba.ca
 


The "Data Files_Question1_SSC2019CaseStudy.zip" file includes the following folders and files:


train folder

The training set includes 2,400 images, which were randomly selected from 3 levels of blur (F1, F23, F48) and two types of stains (w1 – cell body stain, w2-nuclei stain). There are 400 images in each of the six combinations of blur levels and types of stains: F1_w1,F1_w2,F23_w1,F23_w2,F48_w1,F48_w2. File train_label.csv include more details of the images.

test folder

The test set includes 1,200 images, which were randomly selected from 3 levels of blur (F1, F23, F48) and two types of stains (w1 – cell body stain, w2-nuclei stain). There are 200 images in each of the six combinations of blur levels and types of stains: F1_w1,F1_w2,F23_w1,F23_w2,F48_w1,F48_w2. File test_label.csv include more details of the images.

Description of the Files

train_label.csv

Rows include the information for individual images in the training set. The meanings of the four columns are:

  • image_name:  for example,  if an image is named as A01_C1_F1_s01_w1.TIF, which can be explained as
    • A01: The simulated 384-well plate format. Rows are named A-P and columns 1-24. You don't need this information.
    • C1: The number of cells simulated in the image (1-100).
    • F1: The amount of focus blur applied (1, 23, 48).
    • s01: Number of samples (1-25). You don't need this information.
    • w1: 1 = cell body stain, 2 = nuclei stain.
  • cell_count: from 1 to 100.
  • blur_level: The amount of focus blur applied (1, 23, 48).
  • stain: 1 = cell body stain, 2 = nuclei stain.

test_label.csv

Rows include the information for individual images in the test set. The meanings of the three columns are the same as those in the train_label.csv file. We blinded the cell count information for each of the images in the test set.

We also provided the R and python codes to read the images for data analysis.

code.R

R code for reading the pixel values of the images and their label information.

  • Please change the directory to your working path.
  • This R code may take more than 15 minutes to read all the images in your computing environment. Feel free to edit this or use your own methods to read the pixel values of the images in your computing environment.

code.py

Python code for reading the pixel values of the images and their label information.

  • Please change the directory to your working path.
  • This Python code may take more than 15 minutes to read all the images in your computing environment. Feel free to edit this or use your own methods to read the pixel values of the images in your computing environment.