Classification of subcellular protein locations in microscopic images using random forest machine learning: comparison of different feature extraction methods

Group members: Zeng Fung Liew, Yixing Lu

Abstract

We plan to do a multi-class image classification data analysis project. The dataset is obtained from https://kodu.ut.ee/~leopoldp/2016_DeepYeast/. It is a set of single cell images extracted by Pärnamaa et. al (2017) from the original image library described in Chong et. al (2015). Each sample is a 64*64 pixel RGB image which contains two channels: red channel (RFP) shows the cell outline and green channel (GFP) shows the different subcellular location of protein of interest. There are a total of 12 different subcellular protein location classes, with some class imbalance present. The classes are disjoint, meaning one image will be assigned to only one class. Single cell images will be used as training and test set for the project since cell segmentation is not the focus for the current project. The objective of the project is to classify the subcellular protein locations based on single cell fluorescent images.

The pipeline for our project goes as follows: preprocessing → model fitting → model ensembling. For preprocessing methods, we planned to work with feature detection and extraction (such as Scale-Invariant Feature Transform (SIFT), Wavelet Transforms, Wavelet Scattering Transforms, Haralick Features etc) and some dimension reduction techniques (such as PCA). The Python libraries that we will mainly use will therefore be Scikit-Learn, OpenCV, and PyWavelets. Ideally, we hope to obtain results that can rival the models discussed in the original paper, yet much more achievable on regular computers without GPUs, ie. less computationally expensive.

Table of Contents

1. Setup
2. Feature extraction codes
3. Data analysis and model fitting notebook
4. Report

Setup

The setup for reproducing our work follows 3 main parts:

• Julia installation (our project imports Julia code via PyJulia)
• Dataset organization
• Activating our given environment using environment.yml

Part 1: Installing Julia

This step is relatively straightforward. One can download Julia from their official page. At the time of writing, the stable version of Julia is v1.6.1.

The purpose of this step is to allow for the use of PyJulia later on. PyJulia is a library in Python that calls the codes from Julia. Once Julia is downloaded, you are set. There is no prior knowledge in Julia programming required as wrapper functions have been written in place for the purpose of this project.

Part 2: Organizing dataset

1. Download the dataset and all its required files. Here are the links to download them:

   • Image data
   • HOwt_doc.txt
   • HOwt_train.txt
   • HOwt_val.txt
   • HOwt_test.txt

2. The downloaded data is originally very messy and some data reorganization is required. To easily reorganize the data, first clone this repository. This can be done using git clone in a folder you want to store this project as shown below.

$ git clone [email protected]:zengfung/MicroscopyImageClassification.git

3. Move all the downloaded files from step 1 into the data/ folder.

4. Open up terminal, navigate to the data/ directory using the cd command and unzip the main.tar.gz file using the following command. Once this is done, you will see 11 folders named plate01/ to plate11/ in your data/ directory.

# Linux
$ tar xvzf main.tar.gz

# Mac OS
$ gunzip -c foo.tar.gz | tar xopf -

# Windows
$ tar xzvf main.tar.gz

5. Run codes/reorganize-data.py. Wait approximately 5 minutes and the folders will be reorganized into the path data/label/filename.png while the original folders data/plate/filename.png will be deleted. Additionally, a CSV file labels-files.csv will be created to track the filenames, paths, labels, and label index.

$ python codes/reorganize-data.py

Part 3: Activating environment

Our project requires the use of various libraries. Depending on the versions used, some updated versions of certain libraries are not compatible with running this project, and therefore we recommend that one creates a conda environment that match ours.

Note: Part 2 is assumed to be completed before continuing to Part 3.

1. Navigate to the repo directory. You'll notice a file named environment.yml. Create the environment from the environment.yml file:

conda env create -f environment.yml

2. The environment name for this conda environment is cells. This environment has to be activated every time the project is run. To activate this environment, hit:

conda activate cells

3. You are done. You can verify that cells is activated using:

conda env list