Add Hausdorff loss (#6994)

Integrating an existing implementation publicly available on GitHub by
Patryk Rygiel into the MONAI framework.

Fixes #6993.

### Description

Hausdorff distance is widely used in evaluating medical image
segmentation methods. Adding an objective/loss function directly to
optimize this distance can be instrumental in optimizing this score [1].

<center><img
src="https://upload.wikimedia.org/wikipedia/commons/2/21/Hausdorff_distance_sample.svg"
alt="Hausdorff Distance"></center>

An existing implementation is [publicly available on
GitHub](https://github.com/PatRyg99/HausdorffLoss) by [Patryk
Rygiel](https://github.com/PatRyg99), although it might take some
engineering work to integrate within the MONAI framework.

I was able to train a 3D segmentation model successfully using MONAI's
SwinUNETR with this implementation and I would like to contribute the
code for this loss so that it's more widely available and easier to
integrate with MONAI-based repositories.

**References**:
[1] Karimi, D., & Salcudean, S. E. (2019). [Reducing the Hausdorff
distance in medical image segmentation with convolutional neural
networks](https://ieeexplore.ieee.org/abstract/document/8767031?casa_token=OGugo3jEuXQAAAAA:y5fp0CuZiEWlbx8FS4XKHu43Wi_W7albLeIAjsnZ1c-IlTpJmEFJUgoMlKl_taSYuBpcPyCvtTMe).
IEEE Transactions on medical imaging, 39(2), 499-513.

### 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).
- [x] New tests added to cover the changes.
- [x] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [x] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [x] In-line docstrings updated.
- [x] Documentation updated, tested `make html` command in the `docs/`
folder.

---------

Signed-off-by: Imad Toubal <imad.toubal@gmail.com>

docs/source/losses.rst

@@ -73,6 +73,11 @@ Segmentation Losses
 .. autoclass:: ContrastiveLoss
     :members:
 
+`HausdorffDTLoss`
+~~~~~~~~~~~~~~~~~
+.. autoclass:: HausdorffDTLoss
+    :members:
+
 Registration Losses
 -------------------
 

monai/losses/__init__.py

@@ -33,6 +33,7 @@
 from .ds_loss import DeepSupervisionLoss
 from .focal_loss import FocalLoss
 from .giou_loss import BoxGIoULoss, giou
+from .hausdorff_loss import HausdorffDTLoss, LogHausdorffDTLoss
 from .image_dissimilarity import GlobalMutualInformationLoss, LocalNormalizedCrossCorrelationLoss
 from .multi_scale import MultiScaleLoss
 from .perceptual import PerceptualLoss

monai/losses/hausdorff_loss.py

@@ -0,0 +1,242 @@
+# 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.
+
+# Hausdorff loss implementation based on paper:
+# https://arxiv.org/pdf/1904.10030.pdf
+
+# Repo: https://github.com/PatRyg99/HausdorffLoss
+
+
+from __future__ import annotations
+
+import warnings
+from typing import Callable
+
+import numpy as np
+import torch
+from torch.nn.modules.loss import _Loss
+
+from monai.metrics.utils import distance_transform_edt
+from monai.networks import one_hot
+from monai.utils import LossReduction
+
+
+class HausdorffDTLoss(_Loss):
+    """
+    Compute channel-wise binary Hausdorff loss based on distance transform. It can support both multi-classes and
+    multi-labels tasks. The data `input` (BNHW[D] where N is number of classes) is compared with ground truth `target`
+    (BNHW[D]).
+
+    Note that axis N of `input` is expected to be logits or probabilities for each class, if passing logits as input,
+    must set `sigmoid=True` or `softmax=True`, or specifying `other_act`. And the same axis of `target`
+    can be 1 or N (one-hot format).
+
+    The original paper: Karimi, D. et. al. (2019) Reducing the Hausdorff Distance in Medical Image Segmentation with
+    Convolutional Neural Networks, IEEE Transactions on medical imaging, 39(2), 499-513
+    """
+
+    def __init__(
+        self,
+        alpha: float = 2.0,
+        include_background: bool = False,
+        to_onehot_y: bool = False,
+        sigmoid: bool = False,
+        softmax: bool = False,
+        other_act: Callable | None = None,
+        reduction: LossReduction | str = LossReduction.MEAN,
+        batch: bool = False,
+    ) -> None:
+        """
+        Args:
+            include_background: if False, channel index 0 (background category) is excluded from the calculation.
+                if the non-background segmentations are small compared to the total image size they can get overwhelmed
+                by the signal from the background so excluding it in such cases helps convergence.
+            to_onehot_y: whether to convert the ``target`` into the one-hot format,
+                using the number of classes inferred from `input` (``input.shape[1]``). Defaults to False.
+            sigmoid: if True, apply a sigmoid function to the prediction.
+            softmax: if True, apply a softmax function to the prediction.
+            other_act: callable function to execute other activation layers, Defaults to ``None``. for example:
+                ``other_act = torch.tanh``.
+            reduction: {``"none"``, ``"mean"``, ``"sum"``}
+                Specifies the reduction to apply to the output. Defaults to ``"mean"``.
+
+                - ``"none"``: no reduction will be applied.
+                - ``"mean"``: the sum of the output will be divided by the number of elements in the output.
+                - ``"sum"``: the output will be summed.
+            batch: whether to sum the intersection and union areas over the batch dimension before the dividing.
+                Defaults to False, a loss value is computed independently from each item in the batch
+                before any `reduction`.
+
+        Raises:
+            TypeError: When ``other_act`` is not an ``Optional[Callable]``.
+            ValueError: When more than 1 of [``sigmoid=True``, ``softmax=True``, ``other_act is not None``].
+                Incompatible values.
+
+        """
+        super(HausdorffDTLoss, self).__init__(reduction=LossReduction(reduction).value)
+        if other_act is not None and not callable(other_act):
+            raise TypeError(f"other_act must be None or callable but is {type(other_act).__name__}.")
+        if int(sigmoid) + int(softmax) > 1:
+            raise ValueError("Incompatible values: more than 1 of [sigmoid=True, softmax=True, other_act is not None].")
+
+        self.alpha = alpha
+        self.include_background = include_background
+        self.to_onehot_y = to_onehot_y
+        self.sigmoid = sigmoid
+        self.softmax = softmax
+        self.other_act = other_act
+        self.batch = batch
+
+    @torch.no_grad()
+    def distance_field(self, img: np.ndarray) -> np.ndarray:
+        """Generate distance transform.
+
+        Args:
+            img (np.ndarray): input mask as NCHWD or NCHW.
+
+        Returns:
+            np.ndarray: Distance field.
+        """
+        field = np.zeros_like(img)
+
+        for batch in range(len(img)):
+            fg_mask = img[batch] > 0.5
+
+            if fg_mask.any():
+                bg_mask = ~fg_mask
+
+                fg_dist = distance_transform_edt(fg_mask)
+                bg_dist = distance_transform_edt(bg_mask)
+
+                field[batch] = fg_dist + bg_dist
+
+        return field
+
+    def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            input: the shape should be BNHW[D], where N is the number of classes.
+            target: the shape should be BNHW[D] or B1HW[D], where N is the number of classes.
+
+        Raises:
+            ValueError: If the input is not 2D (NCHW) or 3D (NCHWD).
+            AssertionError: When input and target (after one hot transform if set)
+                have different shapes.
+            ValueError: When ``self.reduction`` is not one of ["mean", "sum", "none"].
+
+        Example:
+            >>> import torch
+            >>> from monai.losses.hausdorff_loss import HausdorffDTLoss
+            >>> from monai.networks.utils import one_hot
+            >>> B, C, H, W = 7, 5, 3, 2
+            >>> input = torch.rand(B, C, H, W)
+            >>> target_idx = torch.randint(low=0, high=C - 1, size=(B, H, W)).long()
+            >>> target = one_hot(target_idx[:, None, ...], num_classes=C)
+            >>> self = HausdorffDTLoss(reduction='none')
+            >>> loss = self(input, target)
+            >>> assert np.broadcast_shapes(loss.shape, input.shape) == input.shape
+        """
+        if input.dim() != 4 and input.dim() != 5:
+            raise ValueError("Only 2D (NCHW) and 3D (NCHWD) supported")
+
+        if self.sigmoid:
+            input = torch.sigmoid(input)
+
+        n_pred_ch = input.shape[1]
+        if self.softmax:
+            if n_pred_ch == 1:
+                warnings.warn("single channel prediction, `softmax=True` ignored.")
+            else:
+                input = torch.softmax(input, 1)
+
+        if self.other_act is not None:
+            input = self.other_act(input)
+
+        if self.to_onehot_y:
+            if n_pred_ch == 1:
+                warnings.warn("single channel prediction, `to_onehot_y=True` ignored.")
+            else:
+                target = one_hot(target, num_classes=n_pred_ch)
+
+        if not self.include_background:
+            if n_pred_ch == 1:
+                warnings.warn("single channel prediction, `include_background=False` ignored.")
+            else:
+                # If skipping background, removing first channel
+                target = target[:, 1:]
+                input = input[:, 1:]
+
+        if target.shape != input.shape:
+            raise AssertionError(f"ground truth has different shape ({target.shape}) from input ({input.shape})")
+
+        device = input.device
+        all_f = []
+        for i in range(input.shape[1]):
+            ch_input = input[:, [i]]
+            ch_target = target[:, [i]]
+            pred_dt = torch.from_numpy(self.distance_field(ch_input.detach().cpu().numpy())).float()
+            target_dt = torch.from_numpy(self.distance_field(ch_target.detach().cpu().numpy())).float()
+
+            pred_error = (ch_input - ch_target) ** 2
+            distance = pred_dt**self.alpha + target_dt**self.alpha
+
+            running_f = pred_error * distance.to(device)
+            reduce_axis: list[int] = torch.arange(2, len(input.shape)).tolist()
+            if self.batch:
+                # reducing spatial dimensions and batch
+                reduce_axis = [0] + reduce_axis
+            all_f.append(running_f.mean(dim=reduce_axis, keepdim=True))
+        f = torch.cat(all_f, dim=1)
+        if self.reduction == LossReduction.MEAN.value:
+            f = torch.mean(f)  # the batch and channel average
+        elif self.reduction == LossReduction.SUM.value:
+            f = torch.sum(f)  # sum over the batch and channel dims
+        elif self.reduction == LossReduction.NONE.value:
+            # If we are not computing voxelwise loss components at least make sure a none reduction maintains a
+            # broadcastable shape
+            broadcast_shape = list(f.shape[0:2]) + [1] * (len(ch_input.shape) - 2)
+            f = f.view(broadcast_shape)
+        else:
+            raise ValueError(f'Unsupported reduction: {self.reduction}, available options are ["mean", "sum", "none"].')
+
+        return f
+
+
+class LogHausdorffDTLoss(HausdorffDTLoss):
+    """
+    Compute the logarithm of the Hausdorff Distance Transform Loss.
+
+    This class computes the logarithm of the Hausdorff Distance Transform Loss, which is based on the distance transform.
+    The logarithm is computed to potentially stabilize and scale the loss values, especially when the original loss
+    values are very small.
+
+    The formula for the loss is given by:
+        log_loss = log(HausdorffDTLoss + 1)
+
+    Inherits from the HausdorffDTLoss class to utilize its distance transform computation.
+    """
+
+    def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute the logarithm of the Hausdorff Distance Transform Loss.
+
+        Args:
+            input (torch.Tensor): The shape should be BNHW[D], where N is the number of classes.
+            target (torch.Tensor): The shape should be BNHW[D] or B1HW[D], where N is the number of classes.
+
+        Returns:
+            torch.Tensor: The computed Log Hausdorff Distance Transform Loss for the given input and target.
+
+        Raises:
+            Any exceptions raised by the parent class HausdorffDTLoss.
+        """
+        log_loss: torch.Tensor = torch.log(super().forward(input, target) + 1)
+        return log_loss