problem with predictions

[Noy987]

I've ran the script on my slice with the datasets - pnn_v2_scoring_rank_learning, pnn_v2_fasterrcnn_640 - and i got only a few PNN marked, do you know why? Is it connected to this error in .hydra\config.yaml or do you suggest doing self training? If so, how is it done? Would really appreciate the help.

I also looked at the self-training models and saw that there are many datasets, but i only need PNN and PV, should i download only the PNN one in the pic below? Thanks!

[LeonardoLupori]

Hi!

Thanks for writing feedback :) I have a few questions to clarify the situation and we will try our best to help:

1. What's the scale of the images you use for predictions? (in terms of micrometers per pixel)

I'm asking because the model was trained with relatively high-resolution images with the same scale, so it expects PNNs to be of a certain size in pixels. Small deviations from that scale should not be a big problem. Big deviations in the direction of more resolution can be solved by simply downsampling the images, while unfortunately if your resolution is too low there's not much you can do. You can try a desperate attempt by digitally upscaling the images but we did not test whether upscaled, interpolated images can be used effectively.

3. When you run predictions when do you get that error? Do you get it every time you run predict.py? If the software crashes, how are you able to get predictions?

Thanks a lot :)

[Noy987]

Thanks you for the answer :)

1. I saw there is a limitation for 178,956,970 pixels so i downsized my slice to a matching size, these are the details: 5484*4256=23,339,904 pixels 8-bit 22MB width (pixel) 2742 height (pixel) 2128 Depth (images) 1

how do i know the micrometers per pixel of the slice? To my opinion, the PNN are still visiable enough, would love to her what you think and change the size accoringly. Here is a zoom in of the pic:

3. Yes, i get that error everytime. The software doesn't crash, it gives an output. here it is:

I tried to downsize the image less so we can check the best quality within the limitation of the code and got more patches, but unfortunately the localizations.csv file isn't created (it's not in the folder). These are the dimensions: 15185*11784=178,940,040 pixel

Thank you!!

[fabiocarrara]

About the FileNotFound error, check if, when unzipping the pretrained model, you also unzipped the .hydra/ folder and it is present in the pnn_v2_fasterrcnn_640/ folder. Files prepended with the dot character are often considered hidden files, and you might have skipped it while unzipping. The folder contains a configuration file with the correct parameters of the pretrained model. Not loading that file could cause the software to malfunction.

[Noy987]

Thank you, i will try that.

Regarding the scale of the images - I checked with the lab, and wanted to ask what is the size and resolution in your imaging? We can match ours accordingly. Thanks!

[LeonardoLupori]

Sorry for the delay

Regarding the scale of the images - I checked with the lab, and wanted to ask what is the size and resolution in your imaging? We can match ours accordingly. Thanks!

• In our work images had pixel size of 0.645 um (pixels are square, so same number for both x and y).

As a troubleshooting step, you can also try to download the raw images dataset in the paper at this link and try to predict one of the raw images here: https://zenodo.org/records/7419283 If that does not work, then we know that whatever is not working correctly is not in the images but in the prediction process

Let us know if this helps :)

[Noy987]

Thank you for the feedback! I will try it out!