[MedicalSeg] Fix lung precision problem (#2029)

Author: shiyutang

contrib/MedicalSeg/configs/lung_coronavirus/vnet_lung_coronavirus_128_128_128_15k.yml

@@ -5,4 +5,4 @@ model:
   elu: False
   in_channels: 1
   num_classes: 3
-  pretrained: null
+  pretrained: https://bj.bcebos.com/paddleseg/dygraph/lung_coronavirus/vnet_lung_coronavirus_128_128_128_15k/pretrain/model.pdparams

contrib/MedicalSeg/medicalseg/core/val.py

@@ -139,7 +139,7 @@ def evaluate(model,
                 save_array(
                     save_path=os.path.join(
                         save_dir,
-                        str(iter) + "_" + str(np.mean(per_channel_dice))),
+                        str(iter)),
                     save_content={
                         'pred': pred.numpy(),
                         'label': label.numpy(),

contrib/MedicalSeg/medicalseg/utils/download.py

@@ -118,7 +118,7 @@ def download_file_and_uncompress(url,
                                  extrapath=None,
                                  extraname=None,
                                  print_progress=True,
-                                 cover=False,
+                                 cover=True,
                                  delete_file=True):
     if savepath is None:
         savepath = "."

contrib/MedicalSeg/tools/prepare.py

@@ -139,9 +139,7 @@ def uncompress_file(self, num_files, form):
         """unzip all the file in the root directory"""
         files = glob.glob(os.path.join(self.dataset_root, "*.{}".format(form)))
 
-        assert len(files) == num_files, print(
-            "The file directory should include {} compressed files, but there is only {}"
-            .format(num_files, len(files)))
+        assert len(files) == num_files, "The file directory should include {} compressed files, but there is only {}".format(num_files, len(files))
 
         for f in files:
             extract_path = os.path.join(self.raw_data_path,
@@ -162,24 +160,23 @@ def load_medical_data(f):
 
         # validate nii.gz on lung and mri with correct spacing_resample
         if filename.endswith((".nii", ".nii.gz", ".dcm")):
-            itkimage = sitk.ReadImage(f)
-            if itkimage.GetDimension() == 4:
-                slicer = sitk.ExtractImageFilter()
-                s = list(itkimage.GetSize())
-                s[-1] = 0
-                slicer.SetSize(s)
-                for slice_idx in range(itkimage.GetSize()[-1]):
-                    slicer.SetIndex([0, 0, 0, slice_idx])
-                    sitk_volume = slicer.Execute(itkimage)
-                    images.append(sitk_volume)
+            if "radiopaedia" in filename or "corona" in filename:
+                 f_nps = [nib.load(f).get_fdata(dtype=np.float32)]
             else:
-                images = [itkimage]
-            images = [sitk.DICOMOrient(img, 'LPS') for img in images]
-            f_nps = [sitk.GetArrayFromImage(img) for img in images]
-
-            # if previous line not swap to xyz
-            if f_nps[0].shape[0] != f_nps[0].shape[1]:
-                f_nps = [np.transpose(f_np, [1, 2, 0]) for f_np in f_nps] # swap to xyz 
+                itkimage = sitk.ReadImage(f)
+                if itkimage.GetDimension() == 4:
+                    slicer = sitk.ExtractImageFilter()
+                    s = list(itkimage.GetSize())
+                    s[-1] = 0
+                    slicer.SetSize(s)
+                    for slice_idx in range(itkimage.GetSize()[-1]):
+                        slicer.SetIndex([0, 0, 0, slice_idx])
+                        sitk_volume = slicer.Execute(itkimage)
+                        images.append(sitk_volume)
+                else:
+                    images = [itkimage]
+                images = [sitk.DICOMOrient(img, 'LPS') for img in images]
+                f_nps = [sitk.GetArrayFromImage(img) for img in images]
 
         elif filename.endswith(
             (".mha", ".mhd", "nrrd"

contrib/MedicalSeg/tools/preprocess_utils/values.py

@@ -64,7 +64,7 @@ def normalize(image, min_val=None, max_val=None):
     return image
 
 
-def HUnorm(image, HU_min=-1000, HU_max=600, HU_nan=-2000):
+def HUnorm(image, HU_min=-1200, HU_max=600, HU_nan=-2000):
     """
     Convert CT HU unit into uint8 values. First bound HU values by predfined min
     and max, and then normalize. Due to paddle.nn.conv3D doesn't support uint8, we need to convert
@@ -76,11 +76,12 @@ def HUnorm(image, HU_min=-1000, HU_max=600, HU_nan=-2000):
     HU_nan: float, value for nan in the raw CT image.
     """
 
-    if not isinstance(image, np.ndarray):
+    if not isinstance(image, np.ndarray): 
         image = np.array(image)
     image = np.nan_to_num(image, copy=False, nan=HU_nan)
 
-    image = (image - HU_min) / (HU_max - HU_min)
-    np.clip(image, 0, 1, out=image)
+    # normalize to [0, 1]
+    image = (image - HU_min) / ((HU_max - HU_min) / 255)
+    np.clip(image, 0, 255, out=image)
 
     return image