niftynet.network.interventional_dense_net module

class INetDense(decay=0.0, smoothing=0, disp_w_initializer=None, disp_b_initializer=None, acti_func='relu', multi_scale_fusion=True, name='inet-dense')[source]

Bases: niftynet.network.base_net.BaseNet

### Description

The network estimates dense displacement fields from a pair of moving and fixed images:

Hu et al., Label-driven weakly-supervised learning for multimodal deformable image registration, arXiv:1711.01666 https://arxiv.org/abs/1711.01666

Hu et al., Weakly-Supervised Convolutional Neural Networks for Multimodal Image Registration, Medical Image Analysis (2018) https://doi.org/10.1016/j.media.2018.07.002

see also:
https://github.com/YipengHu/label-reg

### Building blocks [DOWN CONV] - Convolutional layer + Residual Unit + Downsampling (Max pooling) [CONV] - Convolutional layer [UP CONV] - Upsampling + Sum + Residual Unit [FUSION] - Multi-scale displacement fields fusion [DISPtoDEF] - (Smoothing if required) Conversion to deformation field (adding base grid)

### Diagram INPUT PAIR –> [DOWN CONV] [UP CONV] –> [CONV] –[FUSION] –> [DISPtoDEF] –> DENSE FIELD

| |
[DOWN CONV] [UP CONV] –> [CONV] —–|
| |
[DOWN CONV] [UP CONV] –> [CONV] —–|
| |
[DOWN CONV] [UP CONV] –> [CONV] —–|
| |

——– [CONV]—————— [CONV]——-

### Constraints
  • input spatial rank should be either 2 or 3 (2D or 3D images only)
  • fixed image size should be divisible by 16
__init__(decay=0.0, smoothing=0, disp_w_initializer=None, disp_b_initializer=None, acti_func='relu', multi_scale_fusion=True, name='inet-dense')[source]
Parameters:
  • decay – float, regularisation decay
  • smoothing – float, smoothing factor for dense displacement field
  • disp_w_initializer – initialisation of the displacement fields
  • disp_b_initializer – initialisation of the displacement fields
  • acti_func – activation function to use
  • multi_scale_fusion – True/False indicating whether to use multiscale feature fusion.
  • name – layer name
layer_op(fixed_image, moving_image, base_grid=None, is_training=True, **unused_kwargs)[source]
Parameters:
  • fixed_image – tensor, fixed image for registration (defines reference space)
  • moving_image – tensor, moving image to be registered to fixed
  • base_grid – initial identity or affine displacement field
  • is_training – boolean, True if network is in training mode
Returns:

estimated dense displacement fields