TIL and Non-TIL Synthetic Image Generation

Data

The raw PNG patches of TIL and Non-TIL have been stored at the SharmaLab storage device /labs/sharmalab/TilGAN_Monjoy/Data

create_h5_file_TilGAN.py - the FIRST script in the pipeline

create_h5_file_TilGAN.py is the first script of this repository. This script has been used for generating a single .h5 file from image patches.

model_train_TilGAN.py - the SECOND script in the pipeline

model_train_TilGAN.py is the second script in the pipeline. This script has been used for TilGAN model training. This script takes .h5 files as input.

model_test_TilGAN.py - the THIRD script in the pipeline

model_test_TilGAN.py is the third script in the pipeline. This script has been used for TilGAN model testing and generating synthetic patches. To run this script successfully, you need trained models, which will be generated by model_train_TilGAN.py. So, if trained models are not available, run model_train_TilGAN.py before running model_test_TilGAN.py. If trained models are available, then you no need to run model_train_TilGAN.py again.

TIL and Non-TIL classification

The scripts inside the GAN_classification folder have been used for TIL and Non-TIL image classification. For this classification, 90% synthetic and 10% real image patches were used for model training. The model's performance has been tested fully on the real images. We didn't use the same dataset, which was used for synthetic image generation. The classification performance has been tested fully on a separate set of images. Now, we will discuss the script related to TIL and Non-Til classification.

model_train.py - the FIRST script in the classification pipeline

model_train.py is used for classification model training.

model_predict.py - the SECOND script in the pipeline

model_predict.py has been used for model prediction.

TilGAN is a generative adversarial network, which has been used to generate synthetic/fake TIL patches from the Lung cancer pathology images. This code has also been used to generate synthetic COVID-19 dataset. In both dataset, this code is performing well.

Trained TilGAN models and synthetic images can be downloaded using this url:

We perform image synthesis with TilGAN, which is constructed using different baseline architectures, such as Pathology GAN, BigGAN, a cycle-consistency GAN, and a relativistic average GAN. This repository may not reflect the actual architecture of our TilGAN paper.