Submissions

You say:

When submitting, participants will have two choices: Submit open source code (Recommended) or only submit the inference part including the saved trained model (in h5 format) (in which case it will not be eligible for prizes)

I don't understand, does that mean that the training will run on your server if we submit open source code? Or do we train the model locally and only upload the inference part with open-sourced weights? But in that case that looks a lot like option 2.

Also can someone confirm:

In the training we have different z-focus images for the input but for the training we will only have one image at one given z-focus which can be suboptimal (what you call "deviations of the focus plane")?

Yes, you're right, we're going to clarify it further. Indeed, training is local. So, as you say, participants have to submit the model with the inference weights in docker format. However, only those who provide their code in addition (not necessarily in the docker) will be eligible for prizes.

Is that clearer for you?

Regarding your other request, I'm not sure I understand the last question. For training purposes, there could be, for example, image_1_BF_z0, image_1_BF_z1, ...., image_1_BF_z10, image_1_Nucleus, image_1_Tubulin. The aim is to learn how to get as close as possible to the best focal plane (= the sharpest image plane) in fluorescence, because predicting blurred images is not really useful for biologists. How and what would be "suboptimal"?

Ok I understand for the submission thank you very much.

For the other part, I was just asking for clarifications because it says that in the test set we will only have one image at one given z-focus instead of the stacks we will get in the training images (I now realise there was a typo in my initial post sorry).

So for testing purposes, there would be, for example, image_1_BF_zk and that's it?

Hi, I invite you to go to the database page https://lightmycells.grand-challenge.org/database/ in the Test section,
we've made the effort to explain everything on this page, and you have examples.

You can then see that there is no z indication for the tests, all information will be kept by the Light my cells team until the end.
So you can simply have image_1_BF.ome.tif, image_23_DIC.ome.tif, and so on.

Please let us know if anything is unclear or confusing.

Hello,

I'd like to inquire about a statement regarding the inability to predict an organelle, necessitating the creation of a black image, as mentioned in the database description.

If we take "Image X2" as representative of the test database, would the optimal prediction be as follows:

For image_X2_PC_z0.ome.tif and image_X2_PC_z3.ome.tif, would it require four black images? For image_X2_PC_z1.ome.tif, would it require images_X2_Nucleus.ome.tif, along with three black images? And for image_X2_PC_z2.ome.tif, would it require images_X2_Tubulin.ome.tif, along with three black images?

Or is the optimal prediction for each file image_X2_PC_z*.ome.tif composed of images_X2_Nucleus.ome.tif, images_X2_Tubulin.ome.tif, and two black images?

Hi pyatkovsky15022001,

Thanks for this relevant question.
I'll try to clarify.

All the Z stacks are not longer considered in all Test datasets. So, whatever the initial Z the input images will be named with a "z0".

Taking your example, if Image_X2 is a representative image of one of a test dataset and its modality is phase contrast -
then, whatever the initial z (z12 for example) ;
you have to predict from image_17_PC_z0.ome.tiff the 4 outputs channels : images_17_Nucleus.ome.tiff, images_17_Mitochondria.ome.tiff, images_17_Tubulin.ome.tiff and images_17_Actin.ome.tiff.

We invited participants to generate black images as some may have difficulty predicting Actin and Tubulin, for example.
I insist that if you don't have 4 outputs, the algorithm docker will fail.

In this case from image_17_PC_z0.ome.tiff you can predict images_17_Nucleus.ome.tiff, images_17_Mitochondria.ome.tiff, and create 2 black images images_17_Tubulin.ome.tiff and images_17_Actin.ome.tiff.

In this example image_17_PC_z0.ome.tiff was the z12 of Image_X2 in phase contrast_.
Then, for the z5 of Image_X2 in phase contrast you may have image_23_PC_z0.ome.tiff in one of the Test datasets and have to predict the 4 channels outputs : images_23_Nucleus.ome.tiff, images_23_Mitochondria.ome.tiff, images_23_Tubulin.ome.tiff and images_23_Actin.ome.tiff.

Is that clearer?

I've updated the Test section in the Database page. If it is not clear I will make other updates.