MLMI Practical Course: OCT Image Generation via Diffusion Models

Repository for OCT Project MLMI Practical Course from Summer Semester 2022 at TUM.

Current code is an extension of guided-diffusion.

Scope of the project

The project aims to build a neural network to extract meaningful features from OCT Imaging.

The goal is to further use this trained model in other applications, such as automatic report generation based on images alone.

Our effort relies upon implementing Diffusion Models to generate new OCT images, with the expectation that the extracted features of the neural network are good enough to make meaning of new unseen data. If this is accomplished, other downstream tasks such as the automatic report generation might be feasible with use of this neural network.

How to run

Note: Tested only on Windows OS.

Requirements

• python >= 3.8
• Python environment manager, such as pipenv
• pip >= 22.0.4
• Python MPI: Instructions here

Step by step

1. Clone repo locally

git clone https://github.com/murilobellatini/mlmi-oct-diffusion.git

2. Move to local repo root

cd ./mlmi-oct-diffusion

3. Initiate python environment (example for pipenv below)

pipenv shell

4. Install dependencies

pip install -r requirements.txt

5. Either run notebooks or cli commands below.
6. Optional: For notebooks it might be required to install kernel profile. If so, it can be done with code below.

jupyter kernelspec install-self 

Tip

Try overfitting your first model ;) How? Take a look.

1. Put around 10 image samples on the ./data/raw folder
2. Execute the method guided_diffusion.src.resize_images on that folder (without output_dir param)
3. Run the test_train.bat
4. Grab a cup coffee and relax ☕