Docker Testing Phase Now Open

Dear Challenge Participants,

We apologize for the delay in opening the Docker Testing Phase.

The Docker Testing Phase is now open and will remain open until the challenge concludes. Algorithm submissions can be created at Docker Testing Phase Submissions.

We also provide an example grand-challenge algorithm at COSAS Test Algorithm. The related code can be found at our GitHub repository.

When you submit your Docker, the dataset will be automatically mounted at /input/images/adenocarcinoma-image in the Docker, and the output mask should be saved in the directory /output/images/adenocarcinoma-mask, which may need to be created manually. Please ensure the output directory exists before saving your image in your code. The output mask should have the same size and format (.mha) as the input images. The methods for reading and writing these files can be found in our GitHub repository.

Best regards,

COSAS Challenge Team

Dear challenge organizers,

Thank you for providing the Docker testing code. I was wondering: Why are the test images provided in .mha format when the training dataset consists of .png images? My current code is adapted to the .png format as in- and output and I would need to convert the .mha images to .png and back. As I have never worked with .mha images before, I am worried that this back-and-forth conversion will introduce some inconsistencies (e.g. scale the values from 0..255 to 0..1), which might considerably affect model performance.

If, for some reason, there is a need to provide the test images as .mha could you maybe provide some of the training images also as .mha to test the proper conversion?

Thank you for your support. Kind regards, Frauke

Dear Frauke,

Thank you for your inquiry and for participating in the challenge.

The reason for providing the test images in .mha format is that all test images posted to the grand challenge platform are automatically converted to this format. Unfortunately, this means we have no choice but to provide the test images in .mha.

However, you don’t need to worry about inconsistencies. The pixel values when reading .mha images with SimpleITK, as shown in the testing code, are the same as when reading .png images with cv2.imread, both ranging from 0 to 255 with BGR channel. This should help maintain consistency in your model’s performance.

If you need further assistance or additional resources, please feel free to reach out.

Kind regards,

Xi Long COSAS Challenge Team

Dear Frauke,

I have uploaded three MHA images and their corresponding PNG images from the train set of Task 2 to the GitHub. I hope these resources will be helpful for your analysis.

Best regards,

COSAS Challenge Team