diff --git a/flash/image/segmentation/data.py b/flash/image/segmentation/data.py
index f96573e262..8ee8382002 100644
--- a/flash/image/segmentation/data.py
+++ b/flash/image/segmentation/data.py
@@ -63,7 +63,8 @@
 
 if _TORCHVISION_AVAILABLE:
     import torchvision
-    from torchvision.datasets.folder import has_file_allowed_extension, IMG_EXTENSIONS
+    import torchvision.transforms.functional as FT
+    from torchvision.datasets.folder import default_loader, has_file_allowed_extension, IMG_EXTENSIONS
 else:
     IMG_EXTENSIONS = None
 
@@ -148,7 +149,7 @@ def load_sample(self, sample: Mapping[str, Any]) -> Mapping[str, Union[torch.Ten
         img_labels_path = sample[DefaultDataKeys.TARGET]
 
         # load images directly to torch tensors
-        img: torch.Tensor = torchvision.io.read_image(img_path)  # CxHxW
+        img: torch.Tensor = FT.to_tensor(default_loader(img_path))  # CxHxW
         img_labels: torch.Tensor = torchvision.io.read_image(img_labels_path)  # CxHxW
         img_labels = img_labels[0]  # HxW
 
@@ -163,7 +164,7 @@ def load_sample(self, sample: Mapping[str, Any]) -> Mapping[str, Union[torch.Ten
     @staticmethod
     def predict_load_sample(sample: Mapping[str, Any]) -> Mapping[str, Any]:
         img_path = sample[DefaultDataKeys.INPUT]
-        img = torchvision.io.read_image(img_path).float()
+        img = FT.to_tensor(default_loader(img_path)).float()
 
         sample[DefaultDataKeys.INPUT] = img
         sample[DefaultDataKeys.METADATA] = {
@@ -195,7 +196,7 @@ def load_sample(self, sample: Mapping[str, str]) -> Mapping[str, Union[torch.Ten
         img_path = sample[DefaultDataKeys.INPUT]
         fo_sample = _fo_dataset[img_path]
 
-        img: torch.Tensor = torchvision.io.read_image(img_path)  # CxHxW
+        img: torch.Tensor = FT.to_tensor(default_loader(img_path))  # CxHxW
         img_labels: torch.Tensor = torch.from_numpy(fo_sample[self.label_field].mask)  # HxW
 
         sample[DefaultDataKeys.INPUT] = img.float()
@@ -209,7 +210,7 @@ def load_sample(self, sample: Mapping[str, str]) -> Mapping[str, Union[torch.Ten
     @staticmethod
     def predict_load_sample(sample: Mapping[str, Any]) -> Mapping[str, Any]:
         img_path = sample[DefaultDataKeys.INPUT]
-        img = torchvision.io.read_image(img_path).float()
+        img = FT.to_tensor(default_loader(img_path)).float()
 
         sample[DefaultDataKeys.INPUT] = img
         sample[DefaultDataKeys.METADATA] = {