resizing the image instead of using CropOrPad?

Quick question, Is it possible to to resize the image instead of using CropOrPad ? like torchvision.transforms.Resize transform ? I basically want to make use of interpolation methods to assign new target shape.

[fepegar]

Resizing an image means resampling it, if you think about it. You’re reducing the number of samples (pixels/voxels) while keeping the field of view by reducing the sampling rate (e.g. keeping every other voxel in an image). This is the equivalent of using resample and specify a larger target spacing than the original one.

In computer vision, you typically don’t care about the pixel size because it doesn’t really have a meaning in most cases. In medical images, however, there is inherent spatial information for each voxels and that needs to be taken into account when processing the images.

Maybe section 2 of the preprint will help you understand all these issues: TorchIO: a Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning.

[fepegar]

Glad to hear that! But beware, you are actually losing a bit of information when using 2 bytes for your float variables in the convolutions instead of 4.