ioneliabuzatu / celllinesclassificationcompetition

Learn to classify unknown cell lines into 9 types. Competiton at JKU 2022. 2nd place.

Cell Lines Classifier

Challenge at JKU - AI in Life Science S2021

Classification of 9 unknown cell lines given microscopy images. The classes to distinguish are: PC-3, U-251 MG, HeLa, A549, U-2 OS, MCF7, HEK 293, CACO-2 and RT4. Each cell consists of 3 seperate images showing different parts (nucleus, microtubules, endoplasmic reticulum) of the same cell.

Example of staining different parts of a cell where the rgb image is the result of the 3 seprate channels combined together:

Usage of this repo

1. preprocessing.py Preprocess data:
   1. uncompress the files and then run the below cmd for the train images to remove leading zeros

   for FILE in `ls`; do mv $FILE `echo $FILE | sed -e 's:^0*::'`; done
   

   2. Change the path in the config.py as you need and then run python preprocessing.py this will save the rgb images.
2. train.py Train model by running python train.py
3. Inference.py Use the saved checkpoint to generate csv with the predicted classes of the testset.

Report link

Results

Model	Accuracies (validation)
AlexNet	90%
VGG19	92%
VGG19 (bn)	95%
Resnet18	92%
Resnet34	96%
Resnet50	96%
ResnetWide50	96%
Resnet101	96%
Resnet152	95%

Files needed:

• images_train.tar: Training set (#28896 images) containing three 64x64 pixel grayscale images per sample in png format. Each sample has a unique ID. The three images per sample represent nucleus ("_blue.png"), microtubules ("_red.png") and endoplasmic reticulum ("_yellow.png").
• images_test.tar: Test set #20607 images (for public + private leaderboard) in same format as training set.
• y_train.csv: ID of the sample and corresponding label (cell line) for the training set.
• sample_submission.csv: The format in which the predictions must be submitted.