Add Fréchet Radiomics Distance (FRD) to MONAI metrics (#8643)

monai/metrics/__init__.py

@@ -18,6 +18,7 @@
 from .cumulative_average import CumulativeAverage
 from .f_beta_score import FBetaScore
 from .fid import FIDMetric, compute_frechet_distance
+from .frd import FrechetRadiomicsDistance, get_frd_score
 from .froc import compute_fp_tp_probs, compute_fp_tp_probs_nd, compute_froc_curve_data, compute_froc_score
 from .generalized_dice import GeneralizedDiceScore, compute_generalized_dice
 from .hausdorff_distance import HausdorffDistanceMetric, compute_hausdorff_distance

monai/metrics/frd.py

@@ -0,0 +1,71 @@
+# 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
+
+from monai.metrics.fid import get_fid_score
+from monai.metrics.metric import Metric
+
+__all__ = ["FrechetRadiomicsDistance", "get_frd_score"]
+
+
+class FrechetRadiomicsDistance(Metric):
+    """
+    Fréchet Radiomics Distance (FRD). Computes the Fréchet distance between two
+    distributions of radiomic feature vectors, in the same way as the Fréchet
+    Inception Distance (FID) but for radiomics-based features.
+
+    Unlike FID, FRD uses interpretable, clinically relevant radiomic features
+    (e.g. from PyRadiomics) and works for both 2D and 3D images, with optional
+    conditioning by anatomical masks. See Konz et al. "Fréchet Radiomic Distance
+    (FRD): A Versatile Metric for Comparing Medical Imaging Datasets."
+    https://arxiv.org/abs/2412.01496
+
+    This metric accepts two groups of pre-extracted radiomic feature vectors with
+    shape (number of samples, number of features). The same Fréchet distance
+    formula as in FID is applied to the mean and covariance of these features.
+
+    Args:
+        y_pred: Radiomic feature vectors for the first distribution (e.g. from
+            generated or reconstructed images), shape (N, F).
+        y: Radiomic feature vectors for the second distribution (e.g. from real
+            images), shape (N, F).
+
+    Returns:
+        Scalar tensor containing the FRD value.
+    """
+
+    def __call__(self, y_pred: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+        return get_frd_score(y_pred, y)
+
+
+def get_frd_score(y_pred: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+    """Computes the FRD score from two batches of radiomic feature vectors.
+
+    The implementation reuses the same Fréchet distance as FID; only the
+    semantics (radiomic features vs. deep features) differ.
+
+    Args:
+        y_pred: Feature vectors for the first distribution, shape (N, F).
+        y: Feature vectors for the second distribution, shape (N, F).
+
+    Returns:
+        Scalar tensor containing the Fréchet Radiomics Distance.
+
+    Raises:
+        ValueError: When either tensor has more than 2 dimensions. Inputs must have
+            shape (number of samples, number of features).
+    """
+    if y_pred.ndimension() > 2 or y.ndimension() > 2:
+        raise ValueError("Inputs should have (number images, number of features) shape.")
+    return get_fid_score(y_pred, y)

tests/metrics/test_compute_frd_metric.py

@@ -0,0 +1,49 @@
+# 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 unittest
+
+import numpy as np
+import torch
+
+from monai.metrics import FIDMetric, FrechetRadiomicsDistance
+from monai.utils import optional_import
+
+_, has_scipy = optional_import("scipy")
+
+
+@unittest.skipUnless(has_scipy, "Requires scipy")
+class TestFrechetRadiomicsDistance(unittest.TestCase):
+    def test_results(self):
+        x = torch.Tensor([[1, 2], [1, 2], [1, 2]])
+        y = torch.Tensor([[2, 2], [1, 2], [1, 2]])
+        results = FrechetRadiomicsDistance()(x, y)
+        np.testing.assert_allclose(results.cpu().numpy(), 0.4444, atol=1e-4)
+
+    def test_frd_matches_fid_for_same_features(self):
+        """FRD uses the same Fréchet formula as FID; same inputs give same value."""
+        y_pred = torch.Tensor([[1.0, 2.0], [1.0, 2.0], [1.0, 2.0]])
+        y = torch.Tensor([[2.0, 2.0], [1.0, 2.0], [1.0, 2.0]])
+        frd_score = FrechetRadiomicsDistance()(y_pred, y)
+        fid_score = FIDMetric()(y_pred, y)
+        np.testing.assert_allclose(frd_score.cpu().numpy(), fid_score.cpu().numpy(), atol=1e-6)
+
+    def test_rejects_high_dimensional_input(self):
+        """FrechetRadiomicsDistance raises ValueError when inputs have ndimension() > 2."""
+        high_dim = torch.ones([3, 3, 144, 144])
+        with self.assertRaises(ValueError):
+            FrechetRadiomicsDistance()(high_dim, high_dim)
+
+
+if __name__ == "__main__":
+    unittest.main()