Extracting motion parameters in registration

[OscarWC]

When completing registration, rather than warping the moving image to get a predicted image, is there a way to get the motion parameters for affine transformations between moving and fixed images through GlobalNet or any other network. Thank you in advance for the help.

[wyli]

I think you can follow this approach in general https://discuss.pytorch.org/t/how-can-i-extract-intermediate-layer-output-from-loaded-cnn-model/77301

for globalnet:

import torch
from monai.networks.blocks import Warp
from monai.networks.nets import GlobalNet

activation = {}


def get_activation(name):
    def hook(model, input, output):
        activation[name] = output.detach()

    return hook


input_param = {
    "image_size": (16, 16),
    "spatial_dims": 2,
    "in_channels": 1,
    "num_channel_initial": 16,
    "depth": 1,
    "out_kernel_initializer": "kaiming_uniform",
    "out_activation": None,
    "pooling": True,
    "concat_skip": True,
    "encode_kernel_sizes": 3,
}
net = GlobalNet(**input_param)
net.output_block.fc.register_forward_hook(get_activation("fc"))
warp_layer = Warp()
img = torch.randn((1, 1, 16, 16))
result = net(img)
warped = warp_layer(img, result)

print(activation)

[OscarWC]

Thank you very much for this, it has given me results using my own data. I was hoping you could explain to me what the numbers correspond to for each 6-parameter tensor I get back for each image. Many thanks.

[wyli]

sure, these parameters are used by the affine transform head:

MONAI/monai/networks/nets/regunet.py

Lines 236 to 244 in 1b0808d

	class AffineHead(nn.Module):
	def __init__(self, spatial_dims: int, image_size: list[int], decode_size: list[int], in_channels: int):
	super().__init__()
	self.spatial_dims = spatial_dims
	if spatial_dims == 2:
	in_features = in_channels * decode_size[0] * decode_size[1]
	out_features = 6
	out_init = torch.tensor([1, 0, 0, 0, 1, 0], dtype=torch.float)
	elif spatial_dims == 3:

[OscarWC]

@wyli apologies I am still a bit confused on what the output as 'out_init' actually means for each of the 6 output parameters for a 2D case (or even for the 12 parameters in a 3D case).

[wyli]

now it's also addressed by #6459