Single plane label-free prediction (fnet) with ZeroCostDL4Mic


I have been looking into the Label-free prediction (fnet) in ZeroCostDL4Mic with Jupyter Notebooks for Google Colab .

I am interested in predicting the position of nuclei from single plane bright-field images. However, according to the documentation The patch sizes of Label-free prediction (fnet) are hard-coded within the network itself. Currently the patches are 64x64x32 (x,y,z). So, the dataset used for training need to be at the very minimum these sizes, otherwise the notebook will crash.

Is there a workaround that would allow to use single-plane images ? Duplicating the image 32 times in a stack ? Filling the stack with blank images ?

@Ricardo_Henriques @gregjohnso
Thanks a lot, I love the ZeroCostDL4Mic project !

1 Like

I see that the documentation has been updated and now:

[...] Stacks with fewer than 32 slices will only successfully train the model when they contain 16, 8 or 4 slices and other values will throw a tensordimensions error. [...]

So having 4 slice stacks is a bit more manageable.

But is there a trick to convert a single plane into a synthetic 3D stack and make fnet work? Gaussian blur on Z axis?

Thank you

Hi @LPUoO,

If that’s ok, open the request as an issue or feature request in the ZeroCostDL4Mic. It makes it easier for us to keep track of features we need to add or bugs to zap. Plus we’re very active at giving feedback there.

Thanks for using ZeroCostDL4Mic =).

All the best,