DIC to actin training data

It seems like there is no training data from DIC to actin. Is this intentional? Or by mistake?

Hi jianxuchen,

Yes this is normal.
Unfortunately we were unable to acquire actin images in DIC for the training database...

I think the answer is no but just to make sure, we also never have more than one input channel type (DIC or PC or BF) for a given image right? We can have it at different z as a natural way of augmenting the data but never do we have the same image in DIC and PC for example?

This is a very good point thanks for asking.
Indeed there is only one modality for the input images, BF only or PC only or DIC only.

personally, i understand the goal of evaluating generalizability. But, I am not sure how meaningful it is ... In practice, I am not sure if people will train a BF --> actin model and run prediction on DIC images ... Maybe when after the challenge is done, there could be additional analysis to investigate different aspects of generalization.

You're absolutely right.
And that's why we're going to draw up a more specific evaluation and metrics board than simply aggregating the metrics.
Grand Challenge offers us the advantage of not having to normalize the metrics to obtain a ranking - "the mean position" in the leaderboard.
But, don't worry, the result of each evaluation (the predictions.json file) also contains information on z deviations and modalities (BF, PC and DIC).

Thus , we won't just be evaluating overall scores (the dream of widespread uses), but also modality scores and z-deviations scores (the other aspects of generalization you mentioned).

Will we get the 30 images of the phase 1 test set available when phase 2 begins? Maybe we could train on the DIC to actin images in there even if they are very few of them?

As we want to keep the possibility of repeating the challenge, we won't be making the Phase 1 test images available to keep the GT images secret.
If several images had an actin channel, we would have considered making them available,
but, only a few images have actin in phase 1.