Add vista3d inferers (#8021)

Fixes # .

### Description

A few sentences describing the changes proposed in this pull request.

### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [ ] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [ ] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [ ] In-line docstrings updated.
- [ ] Documentation updated, tested `make html` command in the `docs/`
folder.

---------

Signed-off-by: heyufan1995 <heyufan1995@gmail.com>
Signed-off-by: Yiheng Wang <vennw@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Yiheng Wang <vennw@nvidia.com>
Co-authored-by: Yiheng Wang <68361391+yiheng-wang-nv@users.noreply.github.com>
Co-authored-by: YunLiu <55491388+KumoLiu@users.noreply.github.com>

Author: 4 people

docs/source/apps.rst

@@ -248,6 +248,22 @@ FastMRIReader
 ~~~~~~~~~~~~~
 .. autofunction:: monai.apps.reconstruction.complex_utils.complex_conj
 
+`Vista3d`
+---------
+.. automodule:: monai.apps.vista3d.inferer
+.. autofunction:: point_based_window_inferer
+
+.. automodule:: monai.apps.vista3d.transforms
+.. autoclass:: VistaPreTransformd
+    :members:
+.. autoclass:: VistaPostTransformd
+    :members:
+.. autoclass:: Relabeld
+    :members:
+
+.. automodule:: monai.apps.vista3d.sampler
+.. autofunction:: sample_prompt_pairs
+
 `Auto3DSeg`
 -----------
 .. automodule:: monai.apps.auto3dseg

monai/apps/generation/maisi/utils/__init__.py renamed to monai/apps/vista3d/__init__.py

@@ -0,0 +1,177 @@
+# Copyright (c) MONAI Consortium
+# 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.
+
+from __future__ import annotations
+
+import copy
+from collections.abc import Sequence
+from typing import Any
+
+import torch
+
+from monai.data.meta_tensor import MetaTensor
+from monai.utils import optional_import
+
+tqdm, _ = optional_import("tqdm", name="tqdm")
+
+__all__ = ["point_based_window_inferer"]
+
+
+def point_based_window_inferer(
+    inputs: torch.Tensor | MetaTensor,
+    roi_size: Sequence[int],
+    predictor: torch.nn.Module,
+    point_coords: torch.Tensor,
+    point_labels: torch.Tensor,
+    class_vector: torch.Tensor | None = None,
+    prompt_class: torch.Tensor | None = None,
+    prev_mask: torch.Tensor | MetaTensor | None = None,
+    point_start: int = 0,
+    center_only: bool = True,
+    margin: int = 5,
+    **kwargs: Any,
+) -> torch.Tensor:
+    """
+    Point-based window inferer that takes an input image, a set of points, and a model, and returns a segmented image.
+    The inferer algorithm crops the input image into patches that centered at the point sets, which is followed by
+    patch inference and average output stitching, and finally returns the segmented mask.
+
+    Args:
+        inputs: [1CHWD], input image to be processed.
+        roi_size: the spatial window size for inferences.
+            When its components have None or non-positives, the corresponding inputs dimension will be used.
+            if the components of the `roi_size` are non-positive values, the transform will use the
+            corresponding components of img size. For example, `roi_size=(32, -1)` will be adapted
+            to `(32, 64)` if the second spatial dimension size of img is `64`.
+        sw_batch_size: the batch size to run window slices.
+        predictor: the model. For vista3D, the output is [B, 1, H, W, D] which needs to be transposed to [1, B, H, W, D].
+            Add transpose=True in kwargs for vista3d.
+        point_coords: [B, N, 3]. Point coordinates for B foreground objects, each has N points.
+        point_labels: [B, N]. Point labels. 0/1 means negative/positive points for regular supported or zero-shot classes.
+            2/3 means negative/positive points for special supported classes (e.g. tumor, vessel).
+        class_vector: [B]. Used for class-head automatic segmentation. Can be None value.
+        prompt_class: [B]. The same as class_vector representing the point class and inform point head about
+            supported class or zeroshot, not used for automatic segmentation. If None, point head is default
+            to supported class segmentation.
+        prev_mask: [1, B, H, W, D]. The value is before sigmoid. An optional tensor of previously segmented masks.
+        point_start: only use points starting from this number. All points before this number is used to generate
+            prev_mask. This is used to avoid re-calculating the points in previous iterations if given prev_mask.
+        center_only: for each point, only crop the patch centered at this point. If false, crop 3 patches for each point.
+        margin: if center_only is false, this value is the distance between point to the patch boundary.
+    Returns:
+        stitched_output: [1, B, H, W, D]. The value is before sigmoid.
+    Notice: The function only supports SINGLE OBJECT INFERENCE with B=1.
+    """
+    if not point_coords.shape[0] == 1:
+        raise ValueError("Only supports single object point click.")
+    if not len(inputs.shape) == 5:
+        raise ValueError("Input image should be 5D.")
+    image, pad = _pad_previous_mask(copy.deepcopy(inputs), roi_size)
+    point_coords = point_coords + torch.tensor([pad[-2], pad[-4], pad[-6]]).to(point_coords.device)
+    prev_mask = _pad_previous_mask(copy.deepcopy(prev_mask), roi_size)[0] if prev_mask is not None else None
+    stitched_output = None
+    for p in point_coords[0][point_start:]:
+        lx_, rx_ = _get_window_idx(p[0], roi_size[0], image.shape[-3], center_only=center_only, margin=margin)
+        ly_, ry_ = _get_window_idx(p[1], roi_size[1], image.shape[-2], center_only=center_only, margin=margin)
+        lz_, rz_ = _get_window_idx(p[2], roi_size[2], image.shape[-1], center_only=center_only, margin=margin)
+        for i in range(len(lx_)):
+            for j in range(len(ly_)):
+                for k in range(len(lz_)):
+                    lx, rx, ly, ry, lz, rz = (lx_[i], rx_[i], ly_[j], ry_[j], lz_[k], rz_[k])
+                    unravel_slice = [
+                        slice(None),
+                        slice(None),
+                        slice(int(lx), int(rx)),
+                        slice(int(ly), int(ry)),
+                        slice(int(lz), int(rz)),
+                    ]
+                    batch_image = image[unravel_slice]
+                    output = predictor(
+                        batch_image,
+                        point_coords=point_coords,
+                        point_labels=point_labels,
+                        class_vector=class_vector,
+                        prompt_class=prompt_class,
+                        patch_coords=unravel_slice,
+                        prev_mask=prev_mask,
+                        **kwargs,
+                    )
+                    if stitched_output is None:
+                        stitched_output = torch.zeros(
+                            [1, output.shape[1], image.shape[-3], image.shape[-2], image.shape[-1]], device="cpu"
+                        )
+                        stitched_mask = torch.zeros(
+                            [1, output.shape[1], image.shape[-3], image.shape[-2], image.shape[-1]], device="cpu"
+                        )
+                    stitched_output[unravel_slice] += output.to("cpu")
+                    stitched_mask[unravel_slice] = 1
+    # if stitched_mask is 0, then NaN value
+    stitched_output = stitched_output / stitched_mask
+    # revert padding
+    stitched_output = stitched_output[
+        :, :, pad[4] : image.shape[-3] - pad[5], pad[2] : image.shape[-2] - pad[3], pad[0] : image.shape[-1] - pad[1]
+    ]
+    stitched_mask = stitched_mask[
+        :, :, pad[4] : image.shape[-3] - pad[5], pad[2] : image.shape[-2] - pad[3], pad[0] : image.shape[-1] - pad[1]
+    ]
+    if prev_mask is not None:
+        prev_mask = prev_mask[
+            :,
+            :,
+            pad[4] : image.shape[-3] - pad[5],
+            pad[2] : image.shape[-2] - pad[3],
+            pad[0] : image.shape[-1] - pad[1],
+        ]
+        prev_mask = prev_mask.to("cpu")  # type: ignore
+        # for un-calculated place, use previous mask
+        stitched_output[stitched_mask < 1] = prev_mask[stitched_mask < 1]
+    if isinstance(inputs, torch.Tensor):
+        inputs = MetaTensor(inputs)
+    if not hasattr(stitched_output, "meta"):
+        stitched_output = MetaTensor(stitched_output, affine=inputs.meta["affine"], meta=inputs.meta)
+    return stitched_output
+
+
+def _get_window_idx_c(p: int, roi: int, s: int) -> tuple[int, int]:
+    """Helper function to get the window index."""
+    if p - roi // 2 < 0:
+        left, right = 0, roi
+    elif p + roi // 2 > s:
+        left, right = s - roi, s
+    else:
+        left, right = int(p) - roi // 2, int(p) + roi // 2
+    return left, right
+
+
+def _get_window_idx(p: int, roi: int, s: int, center_only: bool = True, margin: int = 5) -> tuple[list[int], list[int]]:
+    """Get the window index."""
+    left, right = _get_window_idx_c(p, roi, s)
+    if center_only:
+        return [left], [right]
+    left_most = max(0, p - roi + margin)
+    right_most = min(s, p + roi - margin)
+    left_list = [left_most, right_most - roi, left]
+    right_list = [left_most + roi, right_most, right]
+    return left_list, right_list
+
+
+def _pad_previous_mask(
+    inputs: torch.Tensor | MetaTensor, roi_size: Sequence[int], padvalue: int = 0
+) -> tuple[torch.Tensor | MetaTensor, list[int]]:
+    """Helper function to pad inputs."""
+    pad_size = []
+    for k in range(len(inputs.shape) - 1, 1, -1):
+        diff = max(roi_size[k - 2] - inputs.shape[k], 0)
+        half = diff // 2
+        pad_size.extend([half, diff - half])
+    if any(pad_size):
+        inputs = torch.nn.functional.pad(inputs, pad=pad_size, mode="constant", value=padvalue)  # type: ignore
+    return inputs, pad_size

monai/apps/vista3d/inferer.py

@@ -0,0 +1,172 @@
+# Copyright (c) MONAI Consortium
+# 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.
+
+from __future__ import annotations
+
+import copy
+import random
+from collections.abc import Callable, Sequence
+from typing import Any
+
+import numpy as np
+import torch
+from torch import Tensor
+
+__all__ = ["sample_prompt_pairs"]
+
+ENABLE_SPECIAL = True
+SPECIAL_INDEX = (23, 24, 25, 26, 27, 57, 128)
+MERGE_LIST = {
+    1: [25, 26],  # hepatic tumor and vessel merge into liver
+    4: [24],  # pancreatic tumor merge into pancreas
+    132: [57],  # overlap with trachea merge into airway
+}
+
+
+def _get_point_label(id: int) -> tuple[int, int]:
+    if id in SPECIAL_INDEX and ENABLE_SPECIAL:
+        return 2, 3
+    else:
+        return 0, 1
+
+
+def sample_prompt_pairs(
+    labels: Tensor,
+    label_set: Sequence[int],
+    max_prompt: int | None = None,
+    max_foreprompt: int | None = None,
+    max_backprompt: int = 1,
+    max_point: int = 20,
+    include_background: bool = False,
+    drop_label_prob: float = 0.2,
+    drop_point_prob: float = 0.2,
+    point_sampler: Callable | None = None,
+    **point_sampler_kwargs: Any,
+) -> tuple[Tensor | None, Tensor | None, Tensor | None, Tensor | None]:
+    """
+    Sample training pairs for VISTA3D training.
+
+    Args:
+        labels: [1, 1, H, W, D], ground truth labels.
+        label_set: the label list for the specific dataset. Note if 0 is included in label_set,
+            it will be added into automatic branch training. Recommend removing 0 from label_set
+            for multi-partially-labeled-dataset training, and adding 0 for finetuning specific dataset.
+            The reason is region with 0 in one partially labeled dataset may contain foregrounds in
+            another dataset.
+        max_prompt: int, max number of total prompt, including foreground and background.
+        max_foreprompt: int, max number of prompt from foreground.
+        max_backprompt: int, max number of prompt from background.
+        max_point: maximum number of points for each object.
+        include_background: if include 0 into training prompt. If included, background 0 is treated
+            the same as foreground. Always be False for multi-partial-dataset training. If needed,
+            can be true for finetuning specific dataset, .
+        drop_label_prob: probability to drop label prompt.
+        drop_point_prob: probability to drop point prompt.
+        point_sampler: sampler to augment masks with supervoxel.
+        point_sampler_kwargs: arguments for point_sampler.
+
+    Returns:
+        label_prompt: [B, 1]. The classes used for training automatic segmentation.
+        point: [B, N, 3]. The corresponding points for each class.
+        Note that background label prompt requires matching point as well ([0,0,0] is used).
+        point_label: [B, N]. The corresponding point labels for each point (negative or positive).
+        -1 is used for padding the background label prompt and will be ignored.
+        prompt_class: [B, 1], exactly the same with label_prompt for label indexing for training loss.
+        label_prompt can be None, and prompt_class is used to identify point classes.
+    """
+    # class label number
+    if not labels.shape[0] == 1:
+        raise ValueError("only support batch size 1")
+    labels = labels[0, 0]
+    device = labels.device
+    unique_labels = labels.unique().cpu().numpy().tolist()
+    if include_background:
+        unique_labels = list(set(unique_labels) - (set(unique_labels) - set(label_set)))
+    else:
+        unique_labels = list(set(unique_labels) - (set(unique_labels) - set(label_set)) - {0})
+    background_labels = list(set(label_set) - set(unique_labels))
+    # during training, balance background and foreground prompts
+    if max_backprompt is not None:
+        if len(background_labels) > max_backprompt:
+            random.shuffle(background_labels)
+            background_labels = background_labels[:max_backprompt]
+
+    if max_foreprompt is not None:
+        if len(unique_labels) > max_foreprompt:
+            random.shuffle(unique_labels)
+            unique_labels = unique_labels[:max_foreprompt]
+
+    if max_prompt is not None:
+        if len(unique_labels) + len(background_labels) > max_prompt:
+            if len(unique_labels) > max_prompt:
+                unique_labels = random.sample(unique_labels, max_prompt)
+                background_labels = []
+            else:
+                background_labels = random.sample(background_labels, max_prompt - len(unique_labels))
+    _point = []
+    _point_label = []
+    # if use regular sampling
+    if point_sampler is None:
+        num_p = min(max_point, int(np.abs(random.gauss(mu=0, sigma=max_point // 2))) + 1)
+        num_n = min(max_point, int(np.abs(random.gauss(mu=0, sigma=max_point // 2))))
+        for id in unique_labels:
+            neg_id, pos_id = _get_point_label(id)
+            plabels = labels == int(id)
+            nlabels = ~plabels
+            plabelpoints = torch.nonzero(plabels)
+            nlabelpoints = torch.nonzero(nlabels)
+            # final sampled positive points
+            num_pa = min(len(plabelpoints), num_p)
+            # final sampled negative points
+            num_na = min(len(nlabelpoints), num_n)
+            _point.append(
+                torch.stack(
+                    random.choices(plabelpoints, k=num_pa)
+                    + random.choices(nlabelpoints, k=num_na)
+                    + [torch.tensor([0, 0, 0], device=device)] * (num_p + num_n - num_pa - num_na)
+                )
+            )
+            _point_label.append(
+                torch.tensor([pos_id] * num_pa + [neg_id] * num_na + [-1] * (num_p + num_n - num_pa - num_na)).to(
+                    device
+                )
+            )
+        for _ in background_labels:
+            # pad the background labels
+            _point.append(torch.zeros(num_p + num_n, 3).to(device))  # all 0
+            _point_label.append(torch.zeros(num_p + num_n).to(device) - 1)  # -1 not a point
+    else:
+        _point, _point_label = point_sampler(unique_labels, **point_sampler_kwargs)
+        for _ in background_labels:
+            # pad the background labels
+            _point.append(torch.zeros(len(_point_label[0]), 3).to(device))  # all 0
+            _point_label.append(torch.zeros(len(_point_label[0])).to(device) - 1)  # -1 not a point
+    if len(unique_labels) == 0 and len(background_labels) == 0:
+        # if max_backprompt is 0 and len(unique_labels), there is no effective prompt and the iteration must
+        # be skipped. Handle this in trainer.
+        label_prompt, point, point_label, prompt_class = None, None, None, None
+    else:
+        label_prompt = torch.tensor(unique_labels + background_labels).unsqueeze(-1).to(device).long()
+        point = torch.stack(_point)
+        point_label = torch.stack(_point_label)
+        prompt_class = copy.deepcopy(label_prompt)
+        if random.uniform(0, 1) < drop_label_prob and len(unique_labels) > 0:
+            label_prompt = None
+            # If label prompt is dropped, there is no need to pad with points with label -1.
+            pad = len(background_labels)
+            point = point[: len(point) - pad]  # type: ignore
+            point_label = point_label[: len(point_label) - pad]
+            prompt_class = prompt_class[: len(prompt_class) - pad]
+        else:
+            if random.uniform(0, 1) < drop_point_prob:
+                point = None
+                point_label = None
+    return label_prompt, point, point_label, prompt_class