Skip to content

ai4life-opencalls/AI4Life-MDC24-example-submission

Repository files navigation

Denoising challenge AI4Life-MDC24


Example submission container

Welcome to the AI4Life-MDC24 denoising challenge! Check The Challenge Page for all the details about the challenge.

On this page, you can find an example submission for the Grand Challenge platform.

Submission checklist

  • Make sure that the INPUT_PATH and OUTPUT_PATH in your code are correct. More info below.
  • Check that all the requirements are contained in the requirements.txt and Dockerfile.
  • Test your container locally with test_run.sh. Check how much memory (RAM) your algorithm requires.
  • Create a gzip archive from your image, run docker save example-algorithm | gzip -c > example-algorithm.tar.gz
  • You are ready to submit! Go to this page for further instructions on how to submit to the Grand Challenge.

How to use this code

  1. Example input image is stored in the test/input/images/image-stack-structured-noise folder.
  2. Look through the contents of the inference.py script.
  3. Run test_run.sh to build and test the container execution.
  4. The resulting image should appear in the test/output/images/image-stack-denoised folder.

What is in the example?

Here, we are showcasing an example pytorch model and its inference. The model contains only a Gaussian Blur operator. The model is packaged into jit. See create_model.py for details.

We use a light python container with pytorch-cpu for this example, see Dockerfile. You can also use GPU version!

The container runs inference.py script, which loops through the noisy images in the INPUT_PATH and applies the model to them individually. The result denoised images are then saved into OUTPUT_PATH folder.

Data location

For datasets containing structured noise:

  1. The input images in the container are stored as /input/images/image-stack-structured-noise/<uuid>.tif
  2. The output of your algorithm should be saved as /input/images/image-stack-denoised/<uuid>.tif

For datasets containing unstructured noise:

  1. The input images in the container are stored as /input/images/image-stack-unstructured-noise/<uuid>.tif
  2. The output of your algorithm should be saved as /input/images/image-stack-denoised/<uuid>.tif

For more details about the datasets, check out Data description page!

Useful links

To learn more about Docker and how to write Dockerfile check out the Official documentation.

Make sure to check Grand Challenge documentation and forum with any questions you may have.

For any other questions or issues, create a topic on the challenge forum or drop us an email through the Email organizers button on the challenge page.

Thank you for participating, and we are looking forward to receiving your submission!

About

An example of submission code and container for AI4Life-MDC24 challenge

Resources

License

Stars

Watchers

Forks