search

HenriquesLab / ZeroCostDL4Mic

ZeroCostDL4Mic: A Google Colab based no-cost toolbox to explore Deep-Learning in Microscopy
MIT License
553 stars 129 forks source link

StarDist 2D notebooks does not segmentate but outputs raw input #14

Closed DaniK-BIP closed 4 years ago

DaniK-BIP commented 4 years ago

Hello,

First, thank you and your team for making such a nice platform for ease of use in microscopy deep-learning and research.

The StarDist 2D notebook that i have been using works well, however, when i want the network to analyse my own dataset, it outputs the raw input. Besides this, it also outputs a .rar with black images, where the network tries to find a cell.

For this, i have performed no edits to the StarDist 2D notebook, and trained the network with 1000 epochs at standard settings. I have tried different bit versions of the same file, which was originally 16-bit, now adapted to 32-bit.

Below i have included my system information and my data and it's results, zipped. My data input into the notebook as a .tiff file, not a .rar. I also included the original 16-bit .tiff file before it was

https://drive.google.com/file/d/1WMlHVQOQOGWQiGH1NjtFnR9NUJUAtuoU/view?usp=sharing Systeminfo.txt

guijacquemet commented 4 years ago

Hi, Thanks a lot for reaching out! It indeed look like there is a bug with the image being exported. We will fix this shortly. However, the .zip file should contain the ROI of the detected object and this appears to work fine. Looking at your images and the ROI being detected, did you train the network using your own images ? or did you try to apply on your images a network trained with the one we provide? Cheers Guillaume

DaniK-BIP commented 4 years ago

Hello Guillaume, I have trained the network the training data provided in the ZeroCostDL4mic wiki for StarDist 2D. From the data i have provided you, are you achieving the desired result? I use Fiji to open the files, i assume the .tiff and ROI files have to be combined, but i have not yet been able to do this through Fiji

Cheers, Dani

guijacquemet commented 4 years ago

Hi, No, the segmentation looks horrible, but this is because the data we provide is too different from the one you want to segment. You need to create a training set and train Stardist using your own images. In Fiji, you can open the ROI exported by the stardist segmentation by just dragging and dropping the .Zip file. You can also open the .Zip file directly from Fiji ROI manager.

Cheers Guillaume

DaniK-BIP commented 4 years ago

Hello, I have gotten the ROI to overlay onto the brightfield image, portaying 4 cells in each corner. Regarding making our own training set, is this done using, for example, using the StarDist 2D plugin available in Fiji?

Cheers, Dani

guijacquemet commented 4 years ago

To create your own training set you can follow our suggestions here: https://github.com/HenriquesLab/ZeroCostDL4Mic/wiki/Stardist Or follow the suggestions provided by the Stardist authors: https://github.com/mpicbg-csbd/stardist All the best Guillaume

guijacquemet commented 4 years ago

Hi, I have now fixed the bug you reported! thanks again for reaching out Cheers Guillaume