TissueImageAnalytics · shaneahmed · Feb 20, 2026 · Jan 8, 2026 · Jan 8, 2026 · Jan 8, 2026
diff --git a/examples/08-nucleus-instance-segmentation.ipynb b/examples/08-nucleus-instance-segmentation.ipynb
diff --git a/tests/engines/test_multi_task_segmentor.py b/tests/engines/test_multi_task_segmentor.py
@@ -158,7 +158,7 @@ def test_mtsegmentor_patches(remote_sample: Callable, track_tmp_path: Path) -> N
             fields=fields,
             expected_counts=expected_counts,
             task_name=task_name,
-            class_dict=mtsegmentor.model.class_dict,
+            class_dict=mtsegmentor._get_model_attr("class_dict"),
         )
 
 
@@ -252,6 +252,7 @@ def test_single_task_mtsegmentor(
 
     assert len(output_ann) == 3
 
+    class_dict_ = mtsegmentor._get_model_attr("class_dict")
     assert_annotation_store_patch_output(
         inputs=inputs,
         output_ann=output_ann,
@@ -260,7 +261,7 @@ def test_single_task_mtsegmentor(
         fields=["box", "centroid", "contours", "prob", "type"],
         expected_counts=expected_counts_nuclei,
         task_name=None,
-        class_dict=mtsegmentor.model.class_dict["nuclei_segmentation"],
+        class_dict=class_dict_["nuclei_segmentation"],
     )
 
     zarr_file = track_tmp_path / "patch_output_annotationstore" / "output.zarr"
@@ -486,7 +487,7 @@ def test_wsi_segmentor_annotationstore(
         verbose=False,
     )
 
-    class_dict = mtsegmentor.model.class_dict
+    class_dict = mtsegmentor._get_model_attr("class_dict")
 
     # Return Probabilities is False
     output = mtsegmentor.run(

diff --git a/tiatoolbox/models/engine/engine_abc.py b/tiatoolbox/models/engine/engine_abc.py
@@ -730,7 +730,7 @@ def save_predictions(
             # scale_factor set from kwargs
             scale_factor = kwargs.get("scale_factor", (1.0, 1.0))
             # class_dict set from kwargs
-            class_dict = kwargs.get("class_dict", self.model.class_dict)
+            class_dict = kwargs.get("class_dict", self._get_model_attr("class_dict"))
 
             return dict_to_store_patch_predictions(
                 processed_predictions,
@@ -1080,6 +1080,9 @@ def _update_ioconfig(
             if stride_shape is not None:
                 ioconfig.stride_shape = stride_shape
             if input_resolutions is not None:
+                for input_resolutions_ in input_resolutions:
+                    if input_resolutions_["units"] == "baseline":
+                        ioconfig = ioconfig.to_baseline()
                 ioconfig.input_resolutions = input_resolutions
 
             return ioconfig

diff --git a/tiatoolbox/models/engine/multi_task_segmentor.py b/tiatoolbox/models/engine/multi_task_segmentor.py
@@ -899,7 +899,7 @@ def post_process_wsi(  # skipcq: PYL-R0201
         """
         probabilities = raw_predictions["probabilities"]
 
-        tile_h, tile_w = self.ioconfig.tile_shape
+        tile_h, tile_w = self._ioconfig.tile_shape
 
         trigger_tile_proc = any(
             p.shape[0] > tile_h or p.shape[1] > tile_w for p in probabilities
@@ -1087,16 +1087,16 @@ def _process_tile_mode(
               modified in this method; it only derives task-centric outputs.
 
         """
-        highest_input_resolution = self.ioconfig.highest_input_resolution
+        highest_input_resolution = self._ioconfig.highest_input_resolution
         wsi_reader = self.dataloader.dataset.reader
 
         # assume ioconfig has already been converted to `baseline` for `tile` mode
         wsi_proc_shape = wsi_reader.slide_dimensions(**highest_input_resolution)
 
         # * retrieve tile placement and tile info flag
         # tile shape will always be corrected to be multiple of output
-        tile_info_sets = self._get_tile_info(wsi_proc_shape, self.ioconfig)
-        ioconfig = self.ioconfig.to_baseline()
+        tile_info_sets = self._get_tile_info(wsi_proc_shape, self._ioconfig)
+        ioconfig = self._ioconfig.to_baseline()
 
         tile_metadata = _build_tile_tasks(
             tile_info_sets=tile_info_sets,
@@ -1109,7 +1109,7 @@ def _process_tile_mode(
         # Calculate batch size for dask compute
         vm = psutil.virtual_memory()
         bytes_per_element = np.dtype(probabilities[0].dtype).itemsize
-        tile_elements = np.prod(self.ioconfig.tile_shape)
+        tile_elements = np.prod(self._ioconfig.tile_shape)
         prod_dim2 = math.prod(p.shape[2] for p in probabilities if len(p.shape) > 2)  # noqa: PLR2004
         tile_memory = len(probabilities) * tile_elements * prod_dim2 * bytes_per_element
         # available memory
@@ -1625,7 +1625,7 @@ def _save_predictions_as_annotationstore(
         # scale_factor set from kwargs
         scale_factor = kwargs.get("scale_factor", (1.0, 1.0))
         # class_dict set from kwargs
-        class_dict = kwargs.get("class_dict")
+        class_dict = kwargs.get("class_dict", self._get_model_attr("class_dict"))
 
         # Need to add support for zarr conversion.
         save_paths = []
@@ -1839,7 +1839,7 @@ def save_predictions(
 
         # This runs dask.compute and returns numpy arrays
         # for saving annotationstore output.
-        class_dict = kwargs.get("class_dict", self.model.class_dict)
+        class_dict = kwargs.get("class_dict", self._get_model_attr("class_dict"))
         if len(self.tasks) == 1 and class_dict is not None:
             kwargs["class_dict"] = class_dict[next(iter(self.tasks))]
         else:

diff --git a/tiatoolbox/models/engine/nucleus_detector.py b/tiatoolbox/models/engine/nucleus_detector.py
@@ -425,10 +425,10 @@ def post_process_wsi(
         # min_distance and postproc_tile_shape cannot be None here
         min_distance = kwargs.get("min_distance")
         if min_distance is None:
-            min_distance = self.model.min_distance
+            min_distance = self._get_model_attr("min_distance")
         postproc_tile_shape = kwargs.get("postproc_tile_shape")
         if postproc_tile_shape is None:
-            postproc_tile_shape = self.model.postproc_tile_shape
+            postproc_tile_shape = self._get_model_attr("postproc_tile_shape")
 
         # Add halo (overlap) around each block for post-processing
         depth_h = min_distance
@@ -600,7 +600,7 @@ class names.
             # class_dict set from kwargs
             class_dict = kwargs.get("class_dict")
             if class_dict is None:
-                class_dict = self.model.output_class_dict
+                class_dict = self._get_model_attr("output_class_dict")
 
             out = self._save_predictions_annotation_store(
                 processed_predictions,

diff --git a/tiatoolbox/models/engine/semantic_segmentor.py b/tiatoolbox/models/engine/semantic_segmentor.py
@@ -696,7 +696,7 @@ def save_predictions(
         # scale_factor set from kwargs
         scale_factor = kwargs.get("scale_factor", (1.0, 1.0))
         # class_dict set from kwargs
-        class_dict = kwargs.get("class_dict", self.model.class_dict)
+        class_dict = kwargs.get("class_dict", self._get_model_attr("class_dict"))
 
         # Need to add support for zarr conversion.
         save_paths = []