diff --git a/src/cellmap_data/dataset.py b/src/cellmap_data/dataset.py
index c0b3630..6b19a6d 100644
--- a/src/cellmap_data/dataset.py
+++ b/src/cellmap_data/dataset.py
@@ -406,18 +406,27 @@ def get_crop_class_matrix(self) -> np.ndarray:
     # Class counts
     # ------------------------------------------------------------------
 
-    @property
+    @cached_property
     def class_counts(self) -> dict[str, Any]:
-        """Aggregate per-class foreground voxel counts from all target sources."""
+        """Aggregate per-class foreground voxel counts from all target sources.
+
+        Returns a dict with:
+        - ``"totals"``: per-class foreground voxel counts at training resolution.
+        - ``"totals_total"``: per-class total voxel counts (full array size) at
+          training resolution.
+        """
         totals: dict[str, int] = {}
+        totals_total: dict[str, int] = {}
         for cls in self.classes:
             src = self.target_sources.get(cls)
             if src is not None:
                 counts = src.class_counts
                 totals[cls] = counts.get(cls, 0)
+                totals_total[cls] = src.total_voxels
             else:
                 totals[cls] = 0
-        return {"totals": totals}
+                totals_total[cls] = 0
+        return {"totals": totals, "totals_total": totals_total}
 
     # ------------------------------------------------------------------
     # Misc
diff --git a/src/cellmap_data/empty_image.py b/src/cellmap_data/empty_image.py
index 3869efc..062f268 100644
--- a/src/cellmap_data/empty_image.py
+++ b/src/cellmap_data/empty_image.py
@@ -62,6 +62,10 @@ def bounding_box(self) -> None:
     def sampling_box(self) -> None:
         return None
 
+    @property
+    def total_voxels(self) -> int:
+        return 0
+
     @property
     def class_counts(self) -> dict[str, int]:
         return {self.label_class: 0}
diff --git a/src/cellmap_data/image.py b/src/cellmap_data/image.py
index a2ae5bc..8fb113d 100644
--- a/src/cellmap_data/image.py
+++ b/src/cellmap_data/image.py
@@ -553,6 +553,26 @@ def class_counts(self) -> dict[str, int]:
             logger.warning("class_counts failed for %s: %s", self.path, exc)
             return {self.label_class: 0}
 
+    @cached_property
+    def total_voxels(self) -> int:
+        """Total number of voxels in the data volume at training resolution.
+
+        Derived from the cached :attr:`bounding_box` (world-space extent of the
+        dataset) divided by the training-resolution voxel size, so no additional
+        zarr I/O is needed beyond what is already cached.
+        """
+        try:
+            total = 1
+            for ax, (start, end) in self.bounding_box.items():
+                axis_voxels = int(round((end - start) / self.scale[ax]))
+                if axis_voxels < 1:
+                    axis_voxels = 1
+                total *= axis_voxels
+            return total
+        except Exception as exc:
+            logger.warning("total_voxels failed for %s: %s", self.path, exc)
+            return 0
+
     def _scale_count(self, s0_count: int, s0_idx: int = 0) -> int:
         """Scale a voxel count from s0 resolution to training resolution."""
         try:
diff --git a/src/cellmap_data/multidataset.py b/src/cellmap_data/multidataset.py
index a9cc49c..41170da 100644
--- a/src/cellmap_data/multidataset.py
+++ b/src/cellmap_data/multidataset.py
@@ -58,23 +58,36 @@ def class_counts(self) -> dict[str, Any]:
 
         Sequential scan (parallelism offers no benefit over NFS; see
         project MEMORY.md notes on ``CellMapMultiDataset.class_counts``).
+
+        Returns a dict with:
+        - ``"totals"``: per-class foreground voxel counts.
+        - ``"totals_total"``: per-class total voxel counts (full array sizes).
         """
         totals: dict[str, int] = {cls: 0 for cls in self.classes}
+        totals_total: dict[str, int] = {cls: 0 for cls in self.classes}
         for ds in tqdm(self.datasets, desc="Counting class voxels", leave=False):
-            ds_counts = ds.class_counts.get("totals", {})
+            ds_counts = ds.class_counts
             for cls in self.classes:
-                totals[cls] += ds_counts.get(cls, 0)
-        return {"totals": totals}
+                totals[cls] += ds_counts.get("totals", {}).get(cls, 0)
+                totals_total[cls] += ds_counts.get("totals_total", {}).get(cls, 0)
+        return {"totals": totals, "totals_total": totals_total}
 
     @property
     def class_weights(self) -> dict[str, float]:
-        """Per-class sampling weight: ``bg_voxels / fg_voxels``."""
-        counts = self.class_counts["totals"]
-        total_voxels = sum(counts.values())
+        """Per-class sampling weight: ``bg_voxels / fg_voxels``.
+
+        Background voxels for each class are derived from the actual data
+        volume size (``totals_total``) minus foreground voxels, so the ratio
+        correctly reflects the class imbalance within each volume.
+        """
+        counts = self.class_counts
+        fg_counts = counts["totals"]
+        total_counts = counts["totals_total"]
         weights: dict[str, float] = {}
         for cls in self.classes:
-            fg = counts.get(cls, 0)
-            bg = total_voxels - fg
+            fg = fg_counts.get(cls, 0)
+            total = total_counts.get(cls, 0)
+            bg = max(total - fg, 0)
             weights[cls] = float(bg) / float(max(fg, 1))
         return weights
 
diff --git a/tests/test_image.py b/tests/test_image.py
index da7be0a..92bd769 100644
--- a/tests/test_image.py
+++ b/tests/test_image.py
@@ -261,3 +261,11 @@ def test_class_counts_keys(self, tmp_path):
         counts = img.class_counts
         assert "mito" in counts
         assert counts["mito"] >= 0
+
+    def test_total_voxels_equals_array_size(self, tmp_path):
+        shape = (10, 10, 10)
+        data = np.zeros(shape, dtype=np.uint8)
+        data[2:5, 2:5, 2:5] = 1
+        path = create_test_zarr(tmp_path, shape=shape, data=data)
+        img = CellMapImage(path, "mito", [8.0, 8.0, 8.0], [4, 4, 4])
+        assert img.total_voxels == int(np.prod(shape))
diff --git a/tests/test_multidataset.py b/tests/test_multidataset.py
index bb4b95c..29f02ee 100644
--- a/tests/test_multidataset.py
+++ b/tests/test_multidataset.py
@@ -62,7 +62,40 @@ def test_class_counts_keys(self, tmp_path):
         multi = CellMapMultiDataset([ds1], CLASSES, INPUT_ARRAYS, TARGET_ARRAYS)
         counts = multi.class_counts
         assert "totals" in counts
+        assert "totals_total" in counts
         assert all(c in counts["totals"] for c in CLASSES)
+        assert all(c in counts["totals_total"] for c in CLASSES)
+
+    def test_class_counts_total_equals_volume_size(self, tmp_path):
+        """totals_total should reflect the actual array volume, not sum of fg counts."""
+        shape = (8, 8, 8)
+        voxel_size = [8.0, 8.0, 8.0]
+        ds1 = _make_ds(tmp_path, "d1", shape=shape, voxel_size=voxel_size)
+        multi = CellMapMultiDataset([ds1], CLASSES, INPUT_ARRAYS, TARGET_ARRAYS)
+        counts = multi.class_counts
+        expected_total = int(np.prod(shape))
+        for cls in CLASSES:
+            assert counts["totals_total"][cls] == expected_total, (
+                f"totals_total[{cls!r}] should equal array volume {expected_total}, "
+                f"got {counts['totals_total'][cls]}"
+            )
+
+    def test_class_weights_bg_uses_volume_size(self, tmp_path):
+        """bg in class_weights should be total_voxels - fg, not sum(fg) - fg."""
+        shape = (8, 8, 8)
+        voxel_size = [8.0, 8.0, 8.0]
+        ds1 = _make_ds(tmp_path, "d1", shape=shape, voxel_size=voxel_size)
+        multi = CellMapMultiDataset([ds1], CLASSES, INPUT_ARRAYS, TARGET_ARRAYS)
+        counts = multi.class_counts
+        weights = multi.class_weights
+        total = int(np.prod(shape))
+        for cls in CLASSES:
+            fg = counts["totals"][cls]
+            expected_bg = total - fg
+            expected_weight = float(max(expected_bg, 0)) / float(max(fg, 1))
+            assert (
+                abs(weights[cls] - expected_weight) < 1e-6
+            ), f"class_weights[{cls!r}] mismatch: expected {expected_weight}, got {weights[cls]}"
 
     def test_get_crop_class_matrix_shape(self, tmp_path):
         ds1 = _make_ds(tmp_path, "d1")