Denoising Diffusion Probabilistic Model for Generating Histopathology Images

The source code for the conference paper designated Denoising Diffusion Probabilistic Model for Generating Histopathology Images has been shared in this GitHub repository.

I will explain different folders and files available in this repository concisely to make things easy for the users.

• Let's focus on the P1 folder first:

1. Part1_KPC_DDPM_Linear_Training.ipynb: This file hosts the code for training the unconditional DDPM using a linear variance scheduler.
2. Part2_KPC_DDPM_Linear_Sampling.ipynb: This file features the code to randomly generate histopathology samples using the trained unconditional DDPM via a linear variance scheduler.
3. Part3_KPC_DDPM_Linear_FID.ipynb: This file provides the implementation for calculating the FID distance using the original images and the images generated with a trained unconditional DDPM using a linear variance scheduler.
4. Part4_KPC_DDPM_Linear_Forward_Process.ipynb: This file features the code to obtain results following the forward process using a linear variance scheduler.
5. __init__.py: This is an empty file which is required to mark a directory as a Python package.
6. hist_ddpm_20240507.150.tsv: This file contains the the unconditional DDPM training metrics using a linear variance scheduler.

• Now, let's focus on the P2 folder:

1. Part1_KPC_DDPM_Cosine_Training.ipynb: This file hosts the code for training the unconditional DDPM using a cosine variance scheduler.
2. Part2_KPC_DDPM_Cosine_Sampling.ipynb: This file features the code to randomly generate histopathology samples using the trained unconditional DDPM via a cosine variance scheduler.
3. Part3_KPC_DDPM_Cosine_FID.ipynb: This file provides the implementation for calculating the FID distance using the original images and the images generated with a trained unconditional DDPM using a cosine variance scheduler.
4. Part4_KPC_DDPM_Cosine_Forward_Process.ipynb: This file features the code to obtain results following the forward process using a cosine variance scheduler.
5. __init__.py: This is an empty file which is required to mark a directory as a Python package.
6. hist_ddpm_20240508.150.tsv: This file contains the the unconditional DDPM training metrics using a cosine variance scheduler.

• Next, we focus on the P3 folder:

1. Part1_KPC_DDPM_Sigmoid_Training.ipynb: This file hosts the code for training the unconditional DDPM using a sigmoid variance scheduler.
2. Part2_KPC_DDPM_Sigmoid_Sampling.ipynb: This file features the code to randomly generate histopathology samples using the trained unconditional DDPM via a sigmoid variance scheduler.
3. Part3_KPC_DDPM_Sigmoid_FID.ipynb: This file provides the implementation for calculating the FID distance using the original images and the images generated with a trained unconditional DDPM using a sigmoid variance scheduler.
4. Part4_KPC_DDPM_Sigmoid_Forward_Process.ipynb: This file features the code to obtain results following the forward process using a sigmoid variance scheduler.
5. __init__.py: This is an empty file which is required to mark a directory as a Python package.
6. hist_ddpm_20240509.150.tsv: This file contains the the unconditional DDPM training metrics using a sigmoid variance scheduler.

• The folder Figures has 5 folders inside it: Figure 1, Figure 2, Figure 3, Figure 4, and Figure 5. These 5 folders have the figures that can be seen in the conference paper.

• The folder MODULES has three .py files: ATTEND.py, DENOISING_DIFFUSION_PYTORCH.py, and __init__.py. ATTEND.py stores some helper functions. DENOISING_DIFFUSION_PYTORCH.py offers the building blocks required to construct the DDPM architecture. __init__.py is an empty file.

Finally, Inception_Score.ipynb shows the code for calculating the inception score using images geenrated with all the schedulers (linear, cosine, and sigmoid).

We haven't been able to provide the following things:

1. model_ddpm_20240507.150.ckpt: The model checkpoint for the unconditional DDPM trained with a linear variance scheduler.
2. model_ddpm_20240508.150.ckpt: The model checkpoint for the unconditional DDPM trained with a cosine variance scheduler.
3. model_ddpm_20240509.150.ckpt: The model checkpoint for the unconditional DDPM trained with a sigmoid variance scheduler.
4. ORG: The directory that contains 11,000 original images required for computing FID distance and inception score.
5. GEN_Linear: The directory that contains 11,000 generated images obtained using a linear variance scheduler and required for computing FID distance and inception score.
6. GEN_Cosine: The directory that contains 11,000 generated images obtained using a cosine variance scheduler and required for computing FID distance and inception score.
7. GEN_Sigmoid: The directory that contains 11,000 generated images obtained using a sigmoid variance scheduler and required for computing FID distance and inception score.

These 7 files and directories are too big and can't be shared in GitHub due to the limitations it has.

This repository uses MIT License. Read the terms and conditions from LICENSE text file.

Read the conference paper from here: KPC_DDPM.