using Subject metadata in torchio.transforms

[Jarartur]

Hello,
I am using a 2D histology dataset from here and my task is registration.
I am loading my image pairs and saving them as two ScalarImages in a Subject. All images have the same pixel size in mm so to do registration I need to pad them to be the same shape.
I searched through the docs but I am still confused as to using other metadata from a Subject instance. Can I for example use it like that:

data = pd.read_csv('summary_file.csv')
subject = tio.Subject(
        src=tio.ScalarImage(data['src_path'][0]),
        trg=tio.ScalarImage(data['trg_path'][0]),
        desired_shape=data['shapes'][0]
        )
transform = tio.Compose([
    tio.CropOrPad(subject['desired_shape'])
])
augmented_subject = tranform(subject)

This example works but only because I am using one subject and not a list. My question is can I somehow reference desired_shape key in every iteration on a list of subjects.

[fepegar]

Do you need a different desired_shape for each subject? If you do, maybe you can create a custom transform.

def CustomCrop(subject):
    crop_pad = tio.CropOrPad(subject['desired_shape'])
    return crop_pad(subject)

[Jarartur]

That being said I learned a lot on how torchio works and I think I can still employ it with my custom preprocessing. Also, in the future I will certainly look more into how to preprocess them using more standard ways.

[Jarartur]

Here is the example I meant after applying Resample('src')
And from what I understand in case where 'src' shape would be i.e. 14000x7000 and 'trg' shape 8000x 15000 Resample('src') would resample 'trg' image in 14000x7000 shape which would greatly reduce its information on tissue structure.

[fepegar]

CropOrPad is designed to be safe against misaligned medical images. In your case, images are misaligned anyway because that's the point of the task, so this issue can be ignored, I suppose. It is a little bit difficult to discuss these issues because you are not using medical images (which have certain associated spatial metadata).

My first question is: why do the width/height ratios of both images, 14/7 and 8/15, differ so much?

[Jarartur]

It happens when one image is rotated aprox. 90 degrees which is in the range of the task. Authors then crop the image to conserve space as one of those images currently takes 75-100 Mb.
As the model would require enormous receptive field for a 90-degree rotation I pre-align them manually and then make non-rigid registration on the patches of data.
Also, I can't simply save them aligned as there are ~200 images but 480 pairs so I need to pre-algin them on load time.
So, my problems come from the non-standard dataset attributes, and they are solved by custom loading them.

[Jarartur]

But I think most of my problems will be solved by creating a custom transform which shouldn't be that hard to do as I need to adapt my previous code to make use of the subject object. If I can use things like OpenCV in the tio.Compose then it should be straight forward now that I know how everything works.