The CI-deepimagej-bioimage-io project enables continuous integration between DeepImageJ and the BioImage Model Zoo, allowing seamless integration of pre-trained deep learning models from the BioImage Model Zoo into DeepImageJ's framework.

• Introduction
• Getting Started
  • Prerequisites
  • Installation
  • Goals
    • Model Card
    • Local Testing
• System Overview
  • CI stages
  • Manually tested models
• Implementation Details
  • Python CI
  • Clojure CI (Deprecated)
• Contribute
• License
• Resources

DeepImageJ is an open-source plugin for ImageJ that integrates deep learning models into the ImageJ ecosystem. The BioImage Model Zoo is a repository of pre-trained deep learning models designed for biological image analysis. This project aims to bridge the gap between the two, allowing users to easily utilize models from the BioImage Model Zoo within DeepImageJ's environment.

The CI process ensures that only models known to work properly with DeepImageJ are made available to users, avoiding surprises when downloading models that may not function as expected.

• DeepImageJ: Make sure you have DeepImageJ installed in your ImageJ/Fiji setup.
• BioImage Model Zoo: You should have access to the BioImage Model Zoo repository.

Clone this repository:

git clone https://github.com/deepimagej/CI-deepimagej-bioimage-io.git

Models that pass the CI will display on the webpage similarly to the figure below. In the Test Summary section of the model card, clicking on [more details] will reveal additional information about the CI process.

On the other hand, models that fail will display the image below on its model card.

The error information can be helpful, as it specifies the where the testing process failed.

• Change directory to python_CI_scripts/ or to java_CI_scripts/ to your own paths.
• Follow the instructions from the Readme.md there.

The CI system operates as follows:

• The input of the system is the specification of the models to test, provided in JSON format.
• The output is a test summary for each of the input models, saved as YAML files.
• The test summary is marked as "pass" if the model was able to run on DeepImageJ headless mode and produce the correct output.
• If the model fails the testing process, the test summary contains information about the encountered error.
• All test summaries are saved in the gh-pages branch of this repository.
• That branch contains also additional information of the latest CI run, such as:
  • a detailed report with the results for every model report.json.
  • a brief summary of the results Readme.md).
  • detailed information of models separated by error type in the folder errors_info/
  • the logs of the headless fiji run, standard output and standard error

The CI process is divided into three stages: init, download, and reproduce, as shown in the figure above. Each stage tests the models for specific criteria, and the corresponding errors are detected and reported in the test summary.

1. init: The init stage parses the input and prepares the models to be tested.
2. download: The download stage involves the download of the models from the bioimage.io repository.
3. reproduce: The reproduce stage runs the models inside Fiji, with DeepImageJ in headless mode. It then compares the output images with the expected output to validate the model's compatibility with DeepImageJ.

After the 3 stages are completed and all the test summaries generated, a report is automatically created with the results of the testing for every model.

The blue words from the figure represent code name-spaces that implement the functionality. Some code focus only on the requirements for one of the stages, whereas others are needed all throughout the CI.

Manually tested models are represented by the ladder in the figure. These models are handled during the init stage and bypass all the other stages. A test summary that pass is generated for these models. The steps to specify models that have been manually tested are:

1. Run model inference manually with DeepImageJ on Fiji's graphical user interface
2. Check if the output of the inference is similar to the provided sample output. This fiji script (in the resources/ folder) allows you to compare 2 images and provides the same metrics that the automatic CI uses.
3. If steps 1. and 2. were correct, you should add an entry of the model (resource_id and version_id) to the list in manually_tested.json located here.

All models in that list are considered manually tested and will (by)pass the CI.

The CI is implemented in Python and resides in the python folder of this GitHub repository. Detailed information on the Python CI process can be found in the corresponding Readme.

The deprecated Clojure CI implementation is located in the clojure folder. It is no longer in use.

Contributions are welcome! If you find any issues or have suggestions for improvements, please feel free to create an issue or submit a pull request.

This project is licensed under the BSD 2-Clause License. See the LICENSE file for more details.

• Bioimageio Wiki
• Bioimageio Issue