FireANTs_IMPACT/Model.py

# Copyright (c) 2025 Valentin Boussot
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0
#
# This wrapper does NOT copy any FireANTs source: it only calls the public FireANTs API of the
# separately-installed ``fireants`` wheel (PyPI). FireANTs is distributed under the FireANTs License
# v1.0 and must be cited — see the NOTICE file in this directory for the license, copyright and
# bibliography that ship with this app.

"""FireANTs registration as a self-contained KonfAI model (shared by the FireANTs presets).

Same idiomatic ``add_module`` graph and the same output contract as the ConvexAdam preset
(``MovedImage`` + ``DisplacementField`` on the FIXED grid, split by two ``ChannelSelect``), so the
orchestrator / app.json / ensemble / uncertainty are unchanged. The engine chains FireANTs' own
composable stages (GPU, Riemannian Adam), each seeding the next like ANTs' ``-t`` stages:

    Rigid (MI, centre-of-mass init) -> Affine (MI, seeded by the rigid) -> deformable

The deformable stage is selected by ``deformable_method`` — the ONE knob that specialises this shared
Model.py into the different presets (exactly as ConvexAdam's shared Model.py is specialised by
``stages``):

    "syn"    symmetric diffeomorphic SyN (CC)   — invertible, higher quality, averages cleanly for ensembling
    "greedy" greedy diffeomorphic (CC)          — one-directional, faster / lower VRAM
    "none"   linear only                        — Rigid+Affine, no deformable (the FireANTs_Affine preset)

Masks: the optional Fixed/Moving masks restrict the metric to a region. FireANTs implements this by
carrying the mask as the last image channel and prefixing the metric with ``masked_``; a mask is only
honoured when it actually restricts (some voxels in, some out), so the common mask-free path is
unchanged (an absent optional mask arrives as a whole-image default and is treated as no mask).

The deformable stages produce the single TOTAL displacement field on the fixed grid (the linear
pre-align is baked in via ``init_affine``, ANTs convention); ``none`` uses the affine matrix directly.
``MovedImage`` and the emitted ``DisplacementField`` are rebuilt from that transform with SimpleITK —
the same output path as the ConvexAdam engine — so all presets/engines are interchangeable in an
ensemble. FireANTs' output-transform writer only serialises to a file, so the deformable field is
round-tripped through a temporary NIfTI (no FireANTs internals are reimplemented here).

NOTE: do NOT add ``from __future__ import annotations`` — KonfAI's config engine relies on
runtime-evaluated annotations (``get_origin``); PEP 563 stringized annotations break binding.
"""

import contextlib
import json
import os
import tempfile
from dataclasses import dataclass
from pathlib import Path
from typing import Annotated, Literal

import numpy as np
import SimpleITK as sitk
import torch
from konfai.metric.measure import IMPACTReg
from konfai.network import network
from konfai.utils.config import Choices, Range
from konfai.utils.dataset import Attribute, data_to_image, image_to_data

DIM = 3

# Feature-model registry (models.json): the available IMPACT feature models, fetched from HF (NOT bundled).
# Only consulted by the "impact" deformable metric; ``KONFAI_IMPACT_MODELS_REGISTRY`` (a local path) wins
# for dev/offline. Mirrors the ConvexAdam preset so the same 30-model catalogue and picker are shared.
_IMPACT_MODELS_REGISTRY = "VBoussot/impact-torchscript-models:models.json"

_DISTANCES: dict[str, type[torch.nn.Module]] = {"L1": torch.nn.L1Loss, "L2": torch.nn.MSELoss}


def registry_choices() -> list[str]:
    """The per-model ``ref`` picker's values — model refs (``repo:path``) from the feature-model registry."""
    repo = _IMPACT_MODELS_REGISTRY.split(":", 1)[0]
    return [f"{repo}:{key}" for key in load_models_registry()]


def load_models_registry(ref: str = _IMPACT_MODELS_REGISTRY) -> dict:
    """Load ``models.json`` (available feature models). ``KONFAI_IMPACT_MODELS_REGISTRY`` (local path) wins
    for dev/offline; otherwise ``ref`` is a ``repo:file`` Hugging Face reference (fetched, not bundled)."""
    from huggingface_hub import hf_hub_download

    local = os.environ.get("KONFAI_IMPACT_MODELS_REGISTRY", "")
    if local:
        path = Path(local)
    elif ":" in ref:
        repo, filename = ref.split(":", 1)
        path = Path(hf_hub_download(repo_id=repo, filename=filename, repo_type="model"))  # nosec B615
    else:
        raise ValueError(
            f"models_registry '{ref}' must be a 'repo:file' Hugging Face reference — or set "
            "KONFAI_IMPACT_MODELS_REGISTRY to a local file for offline use."
        )
    return json.loads(path.read_text(encoding="utf-8"))


def _sorted_specs(mapping: dict) -> list:
    """A dict keyed by string indices ('0','1',...) -> its values in numeric order."""
    return [mapping[k] for k in sorted(mapping, key=lambda key: int(key))]


@dataclass
class ModelSpec:
    """One IMPACT feature model in the deformable metric (several are fused). ``ref`` picks the model; the
    rest are its per-model knobs — the same as the ConvexAdam / elastix ``ModelSpec`` except ``voxel_size``
    (an itk-impact resampling knob) has no meaning for FireANTs' geometry-free torch ``custom_loss`` and is
    intentionally absent."""

    ref: Annotated[str, Choices(registry_choices)]
    layers_mask: str = "01"  # per-layer bitmask, one char per model layer ('1' = use, '0' = skip), like elastix
    layers_weight: float = 1.0  # this model's weight in the multi-model fusion
    pca: Annotated[int, Range(0, 100)] = 0  # keep the top-K principal components of the features (0 = no PCA)
    distance: Literal["L1", "L2"] = "L1"


@contextlib.contextmanager
def _no_texpr_fuser():
    """Disable the TensorExpr JIT fuser while IMPACT's TorchScript feature model runs under autograd.

    The IMPACT feature models are TorchScript; run under FireANTs' gradient optimisation the TensorExpr
    fuser trips on shape ops (``aten::size`` INTERNAL ASSERT). Scoped and restored so no other torch/JIT
    user is affected; the modern profiling executor stays on (this is NOT the legacy executor).
    """
    torch._C._jit_set_texpr_fuser_enabled(False)
    try:
        yield
    finally:
        torch._C._jit_set_texpr_fuser_enabled(True)


class _ImpactCore(IMPACTReg):
    """One IMPACT feature model, exposed as a FireANTs ``forward(moved, fixed)``.

    Reuses ``IMPACTReg._compute`` / ``preprocessing`` verbatim — the stats-normalised feature extraction
    (the model wants per-image ``[min, mean, max, std]``) and the per-layer weighted distance — so the
    metric is exactly KonfAI's, not a re-derivation. Only KonfAI's config-binding ``__init__`` and its
    ``Attribute``-based geometry are replaced: FireANTs passes raw tensors at the current pyramid scale, so
    the intensity statistics are computed from those tensors directly. ``pca`` (absent from KonfAI's torch
    ``IMPACTReg``) is added here as a per-layer feature-space reduction matching itk-impact.
    """

    def __init__(self, ref: str, in_channels: int, weights: list[float], distance: str, pca: int) -> None:
        from huggingface_hub import hf_hub_download

        torch.nn.Module.__init__(self)  # bypass IMPACTReg.__init__ (KONFAI_CONFIG_PATH / apply_config binding)
        self.name = "Reg"
        self.in_channels = int(in_channels)
        self.weights = [float(w) for w in weights]
        self.nb_layer = len(self.weights)
        self.loss = _DISTANCES[distance]()
        self.pca = int(pca)  # PCA lives in KonfAI's IMPACTReg._compute (same behaviour as itk-impact)
        self.dim = DIM
        self.shape = None  # score the whole (downsampled) tensor — no ModelPatch tiling
        if ":" in ref:  # a "repo:path" HF reference; otherwise a local model file
            repo, filename = ref.split(":", 1)
            self.model_path = hf_hub_download(repo, filename, repo_type="model")  # nosec B615
        else:
            self.model_path = ref
        self.model = None  # lazy-loaded on the first forward, like IMPACTReg

    @staticmethod
    def _stats(tensor: torch.Tensor) -> dict:
        detached = tensor.detach()
        return {
            "ImageMin": float(detached.min()),
            "ImageMean": float(detached.mean()),
            "ImageMax": float(detached.max()),
            "ImageStd": float(detached.std()),
        }

    def forward(self, moved: torch.Tensor, fixed: torch.Tensor) -> torch.Tensor:  # type: ignore[override]
        if self.model is None:
            self.model = torch.jit.load(self.model_path)  # nosec B614
        self.model.to(moved.device).eval()
        with _no_texpr_fuser():
            loss, true_nb = self._compute(moved, [self._stats(moved)], fixed, [self._stats(fixed)], None)
        return loss / max(true_nb, 1)


class ImpactFeatureLoss(torch.nn.Module):
    """FireANTs ``custom_loss`` = the KonfAI IMPACT metric fused over several feature models.

    ``forward(moved, fixed)`` sums each model's ``layers_weight * IMPACT(model)``. A model's per-layer
    weights come from its ``layers_mask`` bitmask; its input channel count is read from the registry
    (``models.json`` ``numberofchannels``) so it never has to be configured by hand.
    """

    def __init__(self, specs: list["ModelSpec"]) -> None:
        super().__init__()
        registry = load_models_registry()
        self._cores = torch.nn.ModuleList()
        self._model_weights: list[float] = []
        for spec in specs:
            in_channels = int(registry.get(spec.ref.split(":", 1)[-1], {}).get("numberofchannels", 1))
            weights = [1.0 if char == "1" else 0.0 for char in spec.layers_mask]
            self._cores.append(_ImpactCore(spec.ref, in_channels, weights, spec.distance, spec.pca))
            self._model_weights.append(float(spec.layers_weight))

    def forward(self, moved: torch.Tensor, fixed: torch.Tensor) -> torch.Tensor:
        total: torch.Tensor | None = None
        for weight, core in zip(self._model_weights, self._cores, strict=True):
            term = weight * core(moved, fixed)
            total = term if total is None else total + term
        return total


class FireANTsEngine:
    """Register a fixed/moving pair with FireANTs (Rigid -> Affine -> [SyN | Greedy | none]); return
    (moved, dvf) on the fixed grid.

    ``fireants`` is imported lazily inside :meth:`register` so this module can be imported for config
    /signature introspection (SlicerImpactReg reads the tuning knobs off the ``RegistrationNet``
    annotations) on a machine without a GPU or without FireANTs installed.
    """

    def __init__(
        self,
        scales: list[int],
        affine_iterations: list[int],
        deformable_iterations: list[int],
        cc_kernel: int,
        affine_metric: str,
        affine_lr: float,
        deformable_method: str,
        deformable_metric: str,
        deformable_lr: float,
        integrator_n: int,
        smooth_warp_sigma: float,
        smooth_grad_sigma: float,
        seed: int,
        impact_specs: list["ModelSpec"],
    ) -> None:
        self._scales = [int(s) for s in scales]
        self._affine_iterations = [int(i) for i in affine_iterations]
        self._deformable_iterations = [int(i) for i in deformable_iterations]
        self._cc_kernel = int(cc_kernel)
        self._affine_metric = affine_metric
        self._affine_lr = float(affine_lr)
        self._deformable_method = deformable_method
        self._deformable_metric = deformable_metric
        self._deformable_lr = float(deformable_lr)
        self._integrator_n = int(integrator_n)
        self._smooth_warp_sigma = float(smooth_warp_sigma)
        self._smooth_grad_sigma = float(smooth_grad_sigma)
        self._seed = int(seed)
        # IMPACT deformable metric (only used when deformable_metric == "impact"): KonfAI IMPACT feature
        # models drive the SyN/greedy stage instead of the analytic CC/MI/MSE.
        self._impact_specs = impact_specs

    @staticmethod
    def _is_partial_mask(mask: "sitk.Image | None") -> bool:
        """True only for a mask that actually restricts the region — some voxels in, some out. An absent
        optional mask arrives as a whole-image (all-ones) default and an all-zero mask is degenerate; both
        are treated as no mask so the plain (non-masked) metric path is used."""
        if mask is None:
            return False
        arr = sitk.GetArrayViewFromImage(mask)
        return bool((arr > 0).any()) and bool((arr == 0).any())

    @staticmethod
    def _affine_to_sitk(affine_matrix: "torch.Tensor") -> sitk.AffineTransform:
        """FireANTs' physical (LPS) linear matrix -> SimpleITK AffineTransform (fixed -> moving points),
        the same convention FireANTs writes into an ANTs ``0GenericAffine.mat``."""
        matrix = affine_matrix.float().cpu().numpy()[0]
        affine = sitk.AffineTransform(DIM)
        affine.SetMatrix(matrix[:DIM, :DIM].flatten().astype(np.float64))
        affine.SetTranslation(matrix[:DIM, DIM].astype(np.float64))
        return affine

    def _total_field_transform(self, reg) -> sitk.Transform:
        """Optimise a deformable stage and return its TOTAL displacement (affine baked in) as a
        SimpleITK ``DisplacementFieldTransform`` on the fixed grid.

        FireANTs serialises the total field (ANTs convention, fixed grid) only to a file, so it is
        round-tripped through a temporary NIfTI — its public API, no internals reimplemented."""
        reg.optimize()
        with tempfile.TemporaryDirectory() as tmp:
            warp_path = os.path.join(tmp, "total_warp.nii.gz")
            reg.save_as_ants_transforms(warp_path)
            total_field = sitk.ReadImage(warp_path, sitk.sitkVectorFloat64)
        return sitk.DisplacementFieldTransform(total_field)  # consumes total_field

    def register(
        self,
        fixed: sitk.Image,
        moving: sitk.Image,
        device_index: int,
        fixed_mask: sitk.Image | None = None,
        moving_mask: sitk.Image | None = None,
    ) -> tuple[np.ndarray, np.ndarray]:
        """Register ``moving`` onto ``fixed``; return (moved, dvf) as channel-first arrays on the fixed grid."""
        from fireants.io import BatchedImages, Image
        from fireants.io.imagemask import apply_mask_to_image, generate_image_mask_allones
        from fireants.registration.affine import AffineRegistration
        from fireants.registration.rigid import RigidRegistration

        torch.manual_seed(self._seed)
        device = f"cuda:{device_index}" if device_index >= 0 else "cpu"
        # FireANTs' Image ctor accepts a SimpleITK image directly, so the fixed/moving cross into
        # FireANTs in-memory (no file load) with their geometry preserved.
        fixed_img = Image(fixed, device=device)
        moving_img = Image(moving, device=device)

        # Masked metric only when a mask genuinely restricts the region. FireANTs' masked mode wants the
        # mask as the last channel of BOTH images (all-ones where one side has none) and a ``masked_``
        # metric prefix; the plain path is untouched when no real mask is present.
        use_fixed_mask = self._is_partial_mask(fixed_mask)
        use_moving_mask = self._is_partial_mask(moving_mask)
        masked = use_fixed_mask or use_moving_mask
        if masked:
            fmask = Image(fixed_mask, device=device) if use_fixed_mask else generate_image_mask_allones(fixed_img)
            mmask = Image(moving_mask, device=device) if use_moving_mask else generate_image_mask_allones(moving_img)
            fixed_img = apply_mask_to_image(fixed_img, fmask)
            moving_img = apply_mask_to_image(moving_img, mmask)

        bf = BatchedImages([fixed_img])
        bm = BatchedImages([moving_img])
        affine_loss = f"masked_{self._affine_metric}" if masked else self._affine_metric
        deformable_loss = f"masked_{self._deformable_metric}" if masked else self._deformable_metric

        # Linear: Rigid(MI, COM init) -> Affine(MI, seeded by the rigid), mirroring ANTs. The affine
        # seeds the deformable stage (or is the whole transform when deformable_method == "none").
        rigid = RigidRegistration(
            scales=self._scales,
            iterations=self._affine_iterations,
            fixed_images=bf,
            moving_images=bm,
            loss_type=affine_loss,
            optimizer="Adam",
            optimizer_lr=self._affine_lr,
            cc_kernel_size=self._cc_kernel,
            init_translation="cof",
        )
        rigid.optimize()
        rigid_matrix = rigid.get_rigid_matrix().detach()

        affine = AffineRegistration(
            scales=self._scales,
            iterations=self._affine_iterations,
            fixed_images=bf,
            moving_images=bm,
            loss_type=affine_loss,
            optimizer="Adam",
            optimizer_lr=self._affine_lr,
            cc_kernel_size=self._cc_kernel,
            init_rigid=rigid_matrix,
        )
        affine.optimize()
        affine_matrix = affine.get_affine_matrix().detach()

        # Deformable stage (or none). SyN and Greedy share the same constructor surface; both warm-start
        # from the affine so their TOTAL transform already bakes in the linear pre-align.
        if self._deformable_method == "none":
            transform: sitk.Transform = self._affine_to_sitk(affine_matrix)
        else:
            if self._deformable_method == "syn":
                from fireants.registration.syn import SyNRegistration as Deformable
            elif self._deformable_method == "greedy":
                from fireants.registration.greedy import GreedyRegistration as Deformable
            else:
                raise ValueError(
                    f"Unknown deformable_method '{self._deformable_method}' (expected 'syn', 'greedy' or 'none')."
                )
            # "impact" swaps the analytic metric for a KonfAI IMPACT feature loss on the deformable stage
            # (the linear pre-align keeps its own affine_metric); masks do not restrict the IMPACT metric.
            if self._deformable_metric == "impact":
                loss_type: str = "custom"
                custom_loss: torch.nn.Module | None = ImpactFeatureLoss(self._impact_specs)
            else:
                loss_type, custom_loss = deformable_loss, None
            reg = Deformable(
                scales=self._scales,
                iterations=self._deformable_iterations,
                fixed_images=bf,
                moving_images=bm,
                loss_type=loss_type,
                custom_loss=custom_loss,
                cc_kernel_size=self._cc_kernel,
                deformation_type="compositive",
                integrator_n=self._integrator_n,
                smooth_warp_sigma=self._smooth_warp_sigma,
                smooth_grad_sigma=self._smooth_grad_sigma,
                optimizer="Adam",
                optimizer_lr=self._deformable_lr,
                init_affine=affine_matrix,
            )
            transform = self._total_field_transform(reg)

        if torch.cuda.is_available():
            torch.cuda.synchronize()

        # Rebuild moved + DVF from the single transform on the fixed grid — the ConvexAdam output path,
        # so every FireANTs preset emits identical-shaped results.
        moved = sitk.Resample(moving, fixed, transform, sitk.sitkLinear, 0.0, moving.GetPixelID())
        dvf = sitk.TransformToDisplacementField(
            transform,
            sitk.sitkVectorFloat64,
            fixed.GetSize(),
            fixed.GetOrigin(),
            fixed.GetSpacing(),
            fixed.GetDirection(),
        )
        moved_np, _ = image_to_data(moved)
        dvf_np, _ = image_to_data(dvf)
        return moved_np, dvf_np


class FireANTsRegistration(torch.nn.Module):
    """Graph module: (fixed, moving) tensors + their geometry -> moved image + DVF on the fixed grid.

    ``accepts_attributes = True`` opts this module into receiving the per-branch ``Attribute`` list
    alongside the tensors (same convention as the ConvexAdam / elastix engines); registration needs the
    physical geometry, and the mask branches restrict the metric.
    """

    accepts_attributes = True

    def __init__(self, engine: FireANTsEngine) -> None:
        super().__init__()
        self._engine = engine

    def forward(
        self,
        fixed: torch.Tensor,
        moving: torch.Tensor,
        fixed_mask: torch.Tensor,
        moving_mask: torch.Tensor,
        attributes: list[list[Attribute]],
    ) -> torch.Tensor:
        # attributes = [fixed, moving, fixed_mask, moving_mask] branch attrs; each a list[Attribute] over
        # the batch. Returns, per sample, the moved image (1 channel) channel-stacked with the
        # displacement field (DIM channels); downstream ChannelSelect modules split them. A whole-image
        # mask (the default when none is supplied) restricts nothing.
        fixed_attrs, moving_attrs, fmask_attrs, mmask_attrs = attributes
        device_index = fixed.device.index if fixed.device.type == "cuda" else -1
        combined = []
        # FireANTs runs a gradient-based instance optimisation (Riemannian Adam over the warp); the
        # predictor calls forward under torch.inference_mode(), which forbids autograd. The image tensors
        # have already crossed to numpy/SimpleITK here, so re-enable grad for the optimisation.
        with torch.inference_mode(False), torch.enable_grad():
            for b in range(fixed.shape[0]):
                fixed_img = data_to_image(fixed[b].detach().cpu().numpy(), fixed_attrs[b])
                moving_img = data_to_image(moving[b].detach().cpu().numpy(), moving_attrs[b])
                fixed_mask_img = data_to_image(fixed_mask[b].detach().cpu().numpy(), fmask_attrs[b])
                moving_mask_img = data_to_image(moving_mask[b].detach().cpu().numpy(), mmask_attrs[b])
                moved_np, dvf_np = self._engine.register(
                    fixed_img, moving_img, device_index, fixed_mask_img, moving_mask_img
                )
                combined.append(torch.from_numpy(np.concatenate([moved_np, dvf_np], axis=0)))
        return torch.stack(combined, dim=0).to(fixed.device)


class ChannelSelect(torch.nn.Module):
    """Select a channel slice ``[start:stop]`` (splits the registration output into moved / DVF)."""

    def __init__(self, start: int, stop: int) -> None:
        super().__init__()
        self._start = start
        self._stop = stop

    def forward(self, tensor: torch.Tensor) -> torch.Tensor:
        return tensor[:, self._start : self._stop]


class RegistrationNet(network.Network):
    """Pairwise FireANTs registration as an ``add_module`` graph (fixed = branch 0, moving = branch 1,
    fixed mask = 2, moving mask = 3; masks restrict the metric, whole-image = no restriction).

    Outputs on the fixed grid: ``MovedImage`` (moving resampled onto fixed) and ``DisplacementField``
    (the DIM-component displacement field, in mm). Geometry is attached by the predictor via
    ``same_as_group: Volume_0:Fixed``. The knobs below are read straight from these annotations by the
    UI: ``Annotated[.., Range]`` gives numeric spin bounds; ``Literal`` a dropdown. ``deformable_method``
    is the knob that specialises this shared model into each FireANTs preset.
    """

    def __init__(
        self,
        optimizer: network.OptimizerLoader = network.OptimizerLoader(),
        schedulers: dict[str, network.LRSchedulersLoader] = {
            "default:ReduceLROnPlateau": network.LRSchedulersLoader(0)
        },
        outputs_criterions: dict[str, network.TargetCriterionsLoader] = {"default": network.TargetCriterionsLoader()},
        scales: list[int] = [4, 2, 1],
        affine_iterations: list[int] = [200, 100, 50],
        deformable_iterations: list[int] = [200, 100, 50],
        cc_kernel: Annotated[int, Range(1, 21)] = 5,
        affine_metric: Literal["mi", "cc", "mse"] = "mi",
        affine_lr: Annotated[float, Range(0.0, 10.0)] = 0.003,
        deformable_method: Literal["none", "syn", "greedy"] = "syn",
        deformable_metric: Literal["cc", "mi", "mse", "impact"] = "cc",
        deformable_lr: Annotated[float, Range(0.0, 10.0)] = 0.25,
        integrator_n: Annotated[int, Range(1, 100)] = 10,
        smooth_warp_sigma: Annotated[float, Range(0.0, 100.0)] = 0.5,
        smooth_grad_sigma: Annotated[float, Range(0.0, 100.0)] = 1.0,
        seed: int = 42,
        models: dict[str, ModelSpec] = {},
    ) -> None:
        super().__init__(
            in_channels=1,
            optimizer=optimizer,
            schedulers=schedulers,
            outputs_criterions=outputs_criterions,
            dim=3,
        )
        engine = FireANTsEngine(
            scales,
            affine_iterations,
            deformable_iterations,
            cc_kernel,
            affine_metric,
            affine_lr,
            deformable_method,
            deformable_metric,
            deformable_lr,
            integrator_n,
            smooth_warp_sigma,
            smooth_grad_sigma,
            seed,
            _sorted_specs(models),
        )
        self.add_module(
            "Registration", FireANTsRegistration(engine), in_branch=[0, 1, 2, 3], out_branch=["registration"]
        )
        self.add_module("MovedImage", ChannelSelect(0, 1), in_branch=["registration"], out_branch=["moved"])
        self.add_module("DisplacementField", ChannelSelect(1, 4), in_branch=["registration"], out_branch=["dvf"])