3D Patch extraction from 3D input

[saqibqamar]

I want to extract 3D patches with shape is 32x32x32 from a 3D input. I have given example of loads images from directory and gets shapes of image axes. Please let me know how to extract 3D patches samples from this input
`
from medpy.io import load
import numpy as np
import os
import h5py

data_path = ../....
for i in range(10):

    subject_name = 'subject-%d-' % i
    f = os.path.join(data_path, subject_name + 'C.hdr')
    img, header = load(f)
    inputs = img.astype(np.float32)
      
    A = inputs.shape[0]     #142
    B = inputs.shape[1]     # 176
    C = inputs.shape[2]     # 181
    D = np.arange(A*B*C).reshape(A,B,C)
    print (D.shape)

`
How could I use the function to create patches with size of 32x32x32 from this input? Please reply.. Thanks

[loli]

Well, extracting patches should be quite straight forward with numpy (remember: each image returned by medpy is, in fact, a numpy array).

You must first consider how to treat incomplete patches. The image won't always have dimension that are multiples of 32 (e.g., shape=363636). What do you do then? You must decide either to accept smaller left-over patches (e.g., shape=444). Or, alternatively, you can pad the patches. Or, third option, cropthe images to a size that is a multiple of 32 in all dimensions.

Next, should the patches be overlapping? And if yes, how far?

You can get some inspiration from Google (https://www.google.de/search?q=numpy+array+patches+extract+3d)

Medpy provides a few patch iterators for special cases: http://loli.github.io/medpy/iterators.html
For you, the CentredPatchIterator might be interesting: equally sized patches, automatically padded and non-overlapping. Us eit like

for patch i CentredPatchIterator(image, (32,32,32), cval=0):
   do_something_with_patch()

I hope I could be of help. If you have further questions, please write directly to my email, since the issue tracker is intended for bugs only.