4609: Add AUC-Margin Loss for AUROC optimization

monai/losses/__init__.py

@@ -12,6 +12,7 @@
 from __future__ import annotations
 
 from .adversarial_loss import PatchAdversarialLoss
+from .aucm_loss import AUCMLoss
 from .barlow_twins import BarlowTwinsLoss
 from .cldice import SoftclDiceLoss, SoftDiceclDiceLoss
 from .contrastive import ContrastiveLoss

monai/losses/aucm_loss.py

@@ -0,0 +1,207 @@
+# 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 torch
+import torch.nn as nn
+from torch.nn.modules.loss import _Loss
+
+from monai.utils import LossReduction
+
+
+class AUCMLoss(_Loss):
+    """
+    AUC-Margin loss with squared-hinge surrogate loss for optimizing AUROC.
+
+    The loss optimizes the Area Under the ROC Curve (AUROC) by using margin-based constraints
+    on positive and negative predictions. It supports two versions: 'v1' includes class prior
+    information, while 'v2' removes this dependency for better generalization.
+
+    Reference:
+        Yuan, Zhuoning, Yan, Yan, Sonka, Milan, and Yang, Tianbao.
+        "Large-scale robust deep auc maximization: A new surrogate loss and empirical studies on medical image classification."
+        Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.
+        https://arxiv.org/abs/2012.03173
+
+        Implementation based on: https://github.com/Optimization-AI/LibAUC/blob/1.4.0/libauc/losses/auc.py
+
+    Example:
+        >>> import torch
+        >>> from monai.losses import AUCMLoss
+        >>> loss_fn = AUCMLoss()
+        >>> input = torch.randn(32, 1, requires_grad=True)
+        >>> target = torch.randint(0, 2, (32, 1)).float()
+        >>> loss = loss_fn(input, target)
+    """
+
+    def __init__(
+        self,
+        margin: float = 1.0,
+        imratio: float | None = None,
+        version: str = "v1",
+        reduction: LossReduction | str = LossReduction.MEAN,
+    ) -> None:
+        """
+        Args:
+            margin: margin for squared-hinge surrogate loss (default: ``1.0``).
+            imratio: the ratio of the number of positive samples to the number of total samples in the training dataset.
+                If this value is not given, it will be automatically calculated with mini-batch samples.
+                This value is ignored when ``version`` is set to ``'v2'``.
+            version: whether to include prior class information in the objective function (default: ``'v1'``).
+                'v1' includes class prior, 'v2' removes this dependency.
+            reduction: {``"none"``, ``"mean"``, ``"sum"``}
+                Specifies the reduction to apply to the output. Defaults to ``"mean"``.
+                Note: This loss is computed at the batch level and always returns a scalar.
+                The reduction parameter is accepted for API consistency but has no effect.
+
+        Raises:
+            ValueError: When ``version`` is not one of ["v1", "v2"].
+            ValueError: When ``imratio`` is not in [0, 1].
+
+        Example:
+            >>> import torch
+            >>> from monai.losses import AUCMLoss
+            >>> loss_fn = AUCMLoss(version='v2')
+            >>> input = torch.randn(32, 1, requires_grad=True)
+            >>> target = torch.randint(0, 2, (32, 1)).float()
+            >>> loss = loss_fn(input, target)
+        """
+        super().__init__(reduction=LossReduction(reduction).value)
+        if version not in ["v1", "v2"]:
+            raise ValueError(f"version should be 'v1' or 'v2', got {version}")
+        if imratio is not None and not (0.0 <= imratio <= 1.0):
+            raise ValueError(f"imratio must be in [0, 1], got {imratio}")
+        self.margin = margin
+        self.imratio = imratio
+        self.version = version
+        self.a = nn.Parameter(torch.tensor(0.0))
+        self.b = nn.Parameter(torch.tensor(0.0))
+        self.alpha = nn.Parameter(torch.tensor(0.0))
+
+    def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            input: the shape should be B1HW[D], where the channel dimension is 1 for binary classification.
+            target: the shape should be B1HW[D], with values 0 or 1.
+
+        Returns:
+            torch.Tensor: scalar AUCM loss.
+
+        Raises:
+            ValueError: When input or target have incorrect shapes.
+            ValueError: When input or target have fewer than 2 dimensions.
+            ValueError: When target contains non-binary values.
+        """
+        if input.ndim < 2 or target.ndim < 2:
+            raise ValueError("Input and target must have at least 2 dimensions (B, C, ...)")
+        if input.shape[1] != 1:
+            raise ValueError(f"Input should have 1 channel for binary classification, got {input.shape[1]}")
+        if target.shape[1] != 1:
+            raise ValueError(f"Target should have 1 channel, got {target.shape[1]}")
+        if input.shape != target.shape:
+            raise ValueError(f"Input and target shapes do not match: {input.shape} vs {target.shape}")
+
+        input = input.flatten()
+        target = target.flatten()
+
+        if input.numel() == 0:
+            raise ValueError("Input and target must contain at least one element.")
+
+        if not torch.all((target == 0) | (target == 1)):
+            raise ValueError("Target must contain only binary values (0 or 1)")
+
+        pos_mask = (target == 1).float()
+        neg_mask = (target == 0).float()
+
+        mean_pos_sq = (input - self.a) ** 2
+        mean_neg_sq = (input - self.b) ** 2
+
+        # Note:
+        # v1 uses global expectations (normalized by total number of samples),
+        # following the original LibAUC implementation.
+        # v2 uses class-conditional expectations (normalized by number of samples
+        # in each class), implemented via non-zero averaging.
+        # These behaviors differ and should not be unified.
+        if self.version == "v1":
+            p = float(self.imratio) if self.imratio is not None else float(pos_mask.mean().item())
+            p1 = 1.0 - p
+
+            mean_pos = self._global_mean(mean_pos_sq, pos_mask)
+            mean_neg = self._global_mean(mean_neg_sq, neg_mask)
+
+            interaction = self._global_mean(p * input * neg_mask - p1 * input * pos_mask, pos_mask + neg_mask)
+
+            loss = (
+                p1 * mean_pos
+                + p * mean_neg
+                + 2 * self.alpha * (p * p1 * self.margin + interaction)
+                - p * p1 * self.alpha**2
+            )
+
+        else:  # v2
+            mean_pos = self._class_mean(mean_pos_sq, pos_mask)
+            mean_neg = self._class_mean(mean_neg_sq, neg_mask)
+
+            mean_input_pos = self._class_mean(input, pos_mask)
+            mean_input_neg = self._class_mean(input, neg_mask)
+
+            loss = (
+                mean_pos + mean_neg + 2 * self.alpha * (self.margin + mean_input_neg - mean_input_pos) - self.alpha**2
+            )
+
+        return loss
+
+    def _global_mean(self, tensor: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
+        """
+        Compute the global mean of a masked tensor.
+
+        This computes the mean over all elements, where values outside the mask
+        are zeroed out. The result is normalized by the total number of elements,
+        not by the number of masked elements.
+
+        This corresponds to a global expectation:
+            E[mask * tensor]
+
+        Args:
+            tensor: Input tensor.
+            mask: Binary mask tensor of the same shape as ``tensor``.
+
+        Returns:
+            Scalar tensor representing the global mean.
+        """
+        masked = tensor * mask
+        if masked.numel() == 0:
+            return torch.zeros((), dtype=tensor.dtype, device=tensor.device)
+        return masked.mean()
+
+    def _class_mean(self, tensor: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
+        """
+        Compute the class-conditional mean of a masked tensor.
+
+        This computes the mean over only the masked (non-zero) elements, i.e.,
+        the result is normalized by the number of masked elements.
+
+        This corresponds to a class-conditional expectation:
+            E[tensor | mask]
+
+        Args:
+            tensor: Input tensor.
+            mask: Binary mask tensor of the same shape as ``tensor``.
+
+        Returns:
+            Scalar tensor representing the class-conditional mean.
+            Returns 0 if no elements are selected by the mask.
+        """
+        denom = mask.sum()
+        if denom.item() == 0:
+            return torch.zeros((), dtype=tensor.dtype, device=tensor.device)
+        return (tensor * mask).sum() / denom