Does Deep learning based analysis require image registration?

Image registration is commonly required in bio-medical images analysis. To allign and to get best features. Whereas, deep learning based artitecture tries to find best features own its own.

My question is if we are dealing with bio-medical lower abdominal CT scan images using deep learning based classification, does registration is still required for training datasets? Since we use data augmentation in deep learning for more variability of training data in that case what will be the benefit of image registration? In my opinion after augmentation of images there seems to be no useful application of image registration?

Answer 1

[Priysha Aggarwal]

Image registration can help you solve cases of affine transformations in your CT scan images. If your deep learning architechture for classification fits on a training set which does not have images with different rotations and translations etc., image registration of all input images might help to handle different possible transformations possible in images in testing set.