Transforming SINGLE VIEW MRI images

[cepa995]

Hey there!

Recently I've tried extracting AXIAL view from RSNA MICCAI Brain Tumor Kaggle competition, and storing it as a sequence of DICOM files by using DCMTK img2dcm binary. When I try to feed the data to my network using following set of preprocessing transformations:

preprocess = tio.Compose([
            tio.ToCanonical(),
            tio.HistogramStandardization(landmarks_dict),
            tio.RescaleIntensity((-1, 1)),
            tio.Resample((1.0, 1.0, 1.0)),
            tio.Resize((128, 128, 64)),
        ])

I am getting this WARNING message which I ma not sure why it occurs:

C:\Users\StefanCepa995\anaconda3\envs\kaggle\lib\site-packages\torchio\transforms\preprocessing\spatial\resize.py:55: UserWarning: Output shape (128, 128, 1) != target shape (128, 128, 64). Fixing with CropOrPad

Why is output shape 128, 128, 1? How does keeping only 1 image VIEW affect this?

[fepegar]

Hi, @cepa995. Can you share one image to reproduce this?

Resize is typically not a good choice in medical imaging. Can you try with CropOrPad instead? Resizewill deform your images.

[cepa995]

Ah, never mind. I think you were right @fepegar about Resize. It really ruins everything. I think you solved my issue with CropOrPad. Thanks!

[fepegar]

Thanks for your feedback. I need to add a warning in the docs for Resize!

[cepa995]

Yeah, that would be great! Thanks!

[cepa995]

@fepegar out of curiosity, is there any built-in functionality in torch IO to train NN on single image view like AXIAL? I did code something up myself, but I am wondering if there is already an in-built mechanism to achieve this.

[fepegar]

You could use patch-based training. Take a look at the Sample slices from volumes example in the gallery.