Gradcam 3D Visualisation

[Meddebma]

Hello,

I was trying to implement occlusion sensitivity map and Gradcam for my classification task but I've got two Problems:
1- the displayed slice was in a different reconstruction as the original one(displayed in Check Data Shape section), how can I make the same reconstruction?

2- for the Gradcam I got this error:
AttributeError: 'numpy.ndarray' object has no attribute 'detach'

here is the code:
https://github.com/Meddebma/pyradiomics/blob/master/3D_Classification.ipynb

Thank you very much

[Nic-Ma]

Hi @Meddebma ,

Thanks for your interest here.
Let me try to reproduce your issue locally and get back to you as soon as possible.

Thanks.

[Meddebma]

@rijobro
thank you, It worked this time!
Somehow I feel that the results of cam and occlusion sensitivity are divergent, could it be the case? And is it possible to transpose the original image with the gradcam?
https://github.com/Meddebma/pyradiomics/blob/master/3D_Classification_Dict.ipynb

[rijobro]

GradCAM and occlusion sensitivity are different, so aren't guaranteed to always highlight the same differences. In your case, you only used one example, so it's hard to figure out any trends based on that. Try adding more rows to your visualisation (and use random numbers rather than val_ds[2]).

Your confusion matrix also highlights that your network isn't currently doing very well, so you probably want to improve that before the gradcam and occlusion sensitivity become useful. Just to explain that a little further, there's no point asking your network "why did you make that decision?" if it's currently not making very educated guesses.

[Meddebma]

@rijobro
Thank you I did that and it worked well!
I am not sure that the correct labels belong to the correct patients, how can I best find out if it fits? can I display the name of the nifti files with the labels? the labels are in a csv with the names of the nifti files

[rijobro]

I looked through your code and it looks ok. This would be ok:

for data in data_loader:
    x, y = data["img"], data["label"]

because you get both the label and the image at the same time.

If you have an error it would therefore be at the reading of your excel sheet versus the glob for your files. This starts before the MONAI workflow, so you'll have to check this yourself. For example, when you run sorted(glob(os.path(...)) are the files returned in the order you that you expect?

[Meddebma]

@rijobro Okey I double checked and it looks fine with sorted(glob...) :)
By the way in the confusion matrix, I wanted to implement other metrics using this:
cmm = ConfusionMatrixMetric (include_background=True, metric_name=["sensitivity","specificity","negative predictive value"])
how can I display the results?