Neural networks

Under construction

In the meantime, please refer to the API Reference of deepali.networks.

Basic building blocks and (sub-)networks of learned image registration models.

For spatial transformation models used in both classic non-learning based registration and learned image registration, where transformation parameters are inferred by a neural network from the input data instead of being optimized directly, see the spatial library.

Commonly, the neural network model infers the transformation parameters from the input data. These parameters are then used to evaluate and apply the spatial transformation. For this, the params attribute of parameteric transformations can be set to a neural network instance of type torch.nn.Module. The input data of the so parametrized spatial transformation is then set as SpatialTransform.condition(), which constitutes the input of the neural network.