Enforcing unit8 and float32 types in the prediction files

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "raidionicsseg"
-version = "1.5.0"
+version = "1.5.1"
 description = "Raidionics segmentation and classification back-end with ONNX runtime"
 readme = "README.md"
 license = { text = "BSD-2-Clause" }

raidionicsseg/Utils/io.py

@@ -53,16 +53,22 @@ def dump_predictions(
     try:
         naming_suffix = "pred" if parameters.predictions_reconstruction_method == "probabilities" else "labels"
         class_names = parameters.training_class_names
-
+        modified_header = nib_volume.header.copy()
+        if parameters.predictions_reconstruction_method != "probabilities":
+            modified_header.set_data_dtype(np.uint8)
+            assert predictions.dtype == np.uint8
+        else:
+            modified_header.set_data_dtype(np.float32)
+            assert predictions.dtype == np.float32
         if len(predictions.shape) == 4:
             first_class = 0 if parameters.training_activation_layer_type == "sigmoid" else 1
             for c in range(first_class, predictions.shape[-1]):
-                img = nib.Nifti1Image(predictions[..., c], affine=nib_volume.affine, header=nib_volume.header)
+                img = nib.Nifti1Image(predictions[..., c], affine=nib_volume.affine, header=modified_header)
                 predictions_output_path = os.path.join(storage_path, naming_suffix + "_" + class_names[c] + ".nii.gz")
                 os.makedirs(os.path.dirname(predictions_output_path), exist_ok=True)
                 nib.save(img, predictions_output_path)
         else:
-            img = nib.Nifti1Image(predictions, affine=nib_volume.affine, header=nib_volume.header)
+            img = nib.Nifti1Image(predictions, affine=nib_volume.affine, header=modified_header)
             predictions_output_path = os.path.join(storage_path, naming_suffix + "_" + "argmax" + ".nii.gz")
             os.makedirs(os.path.dirname(predictions_output_path), exist_ok=True)
             nib.save(img, predictions_output_path)

tests/generic_tests/test_inference_segmentation_reconstruction.py

@@ -148,6 +148,7 @@ def test_inference_segmentation_reconstruction_method(test_dir, tmp_path):
                 segmentation_gt_nib.header.get_zooms()[0:3]) * 1e-3
             logging.info(f"Volume difference: {abs(pred_volume - gt_volume)}\n")
             assert abs(pred_volume - gt_volume) < 0.1, "Ground truth and prediction volumes are very different"
+            assert segmentation_pred_nib.get_data_dtype() == np.float32, "Predictions is not of type float32"
         except Exception as e:
             logging.error(f"Error during inference Python package test with: {e} \n {traceback.format_exc()}.\n")
             if os.path.exists(tmp_test_input_fn):

tests/generic_tests/test_inference_segmentation_simple.py

@@ -87,6 +87,7 @@ def test_inference_cli(test_dir, tmp_path):
         segmentation_gt = nib.load(segmentation_gt_filename).get_fdata()[:]
         assert np.array_equal(segmentation_pred,
                               segmentation_gt), "Ground truth and prediction arrays are not identical"
+        assert nib.load(segmentation_pred_filename).get_data_dtype() == np.uint8, "Tresholded predictions is not of type uint8"
     except Exception as e:
         logging.error(f"Error during inference CLI test with: {e}\n {traceback.format_exc()}.\n")
         raise ValueError("Error during inference CLI test.\n")
@@ -155,6 +156,8 @@ def test_inference_package(test_dir, tmp_path):
             segmentation_pred = nib.load(segmentation_pred_filename).get_fdata()[:]
             segmentation_gt = nib.load(segmentation_gt_filename).get_fdata()[:]
             assert np.array_equal(segmentation_pred, segmentation_gt), "Ground truth and prediction arrays are not identical"
+            assert nib.load(
+                segmentation_pred_filename).get_data_dtype() == np.uint8, "Tresholded predictions is not of type uint8"
         except Exception as e:
             logging.error(f"Error during inference Python package test with: {e} \n {traceback.format_exc()}.\n")
             if os.path.exists(tmp_test_input_fn):