Is nn-Unet perform well on smaller datasets? What would be the minimum number of data needed?

[iromimp]

Hello,

I do have a really small dataset where I have around 8 patients which I have to work on. I really do want to work on nn-unet and just wondering whether running an nn-Unet model on that dataset is worth getting an idea. What would be the minimal number of data needed to get a good performance in nn-Unet? I really appreciate it if you have any guidance or suggestions to come up with a workaround.

Thank You

[FabianIsensee]

The amount of training data needed strongly depends on the segmentation problem. The more difficult it is and the more diverse the appearance of the images/target structures the more training cases you need. This is true for all segmentation methods, not just nnU-Net.
The only way to find out if it's enough is to run it and look at the results
Best,
Fabian

[iromimp]

Hi,

Yes. 0.39 is for "all". I couldn't figure out any reason for poor performance.

Thank You
Best Regards
Iromi

[Joeycho]

OrderedDict([('train', array(['ANON50135', 'ANON51274', 'ANON67365', 'GH_SABR_03'], dtype='<U10')), ('val', array(['ANON82986'], dtype='<U10'))])

Hi, so the above is matched with fold 3, train and validation list. Maybe you can try the exact same training set like the above, but train with 'all', and test on 'ANON82986'? The result should be similar as described in the 2nd progress plot.

If the result is actually similar with 1st one, I think maybe we can try to expose some training set to validation set, too and do 5 folding..

[iromimp]

Thank you for the suggestion. I will do the same and let you know the results.

Thanks again

[iromimp]

Hi,

Sorry for the late reply. I have trained a model with fold 3 but "all". But I got a dice score of 0. But the progress plot is better as below.

Then I tried the training with data augmentation (nnUNetTrainerV2_DA5.py). Yet I am getting a dice score as 0.
Could you please let me know what tools you used to convert the Dicom data to nifti?
I would like to know if is there any possibility for you to share the dataset you used for me to go through the experiment.

Thank You
Best Regards
Iromi

[Joeycho]

Hi Iromi,

Unfortunately, the dataset is not owned by me, and as you might imagine.. it is somehow.. confidential. I cannot share it to you...

From DICOM to nifti, I've used pydicom libarary to read dicom folder, and save it as nifti with sitk.

import SimpleITK as sitk
from pydicom.filereader import dcmread

sitk.ReadImage with option "GDCMImageIO", or sitk.ImageSeriesReader().

I think there are many other tools too.

Regarding the above graph, it seems fine!. Metric goes upto 0.9, that means training is working, although you use 'all' setting, which means exposing validation set in training. Maybe then.. Could you try another fold(4), and predict one of data, which was training data in fold 3, but not in this new fold (4).

Or another check point is just visualize those data(one data in test (dice 0), and one data from training set). and see whether they are similar or not.. If they look quite different, then it makes sense.. Source data and target data comes from different distribution which makes the task difficult..