updating json to match dockerhub, fixing examples, removing warnings

Author: kmader

CTBoneSegmenter.json

@@ -11,8 +11,8 @@
   "docker":
     {
     "dockerhub_repository": "4quant/bone-segmenter-cpu",
-    "digest": "sha256:43f13084e4d75508a0bc08b4a8c55880b4a5defb90043cb82d6a39ac25adabbf",
-    "size": "4.6 GB"
+    "digest": "sha256:41020bcd68cc5854aa1757b452defc52e9035d39ed19edc1b33439d327dd99b5",
+    "size": "3.6 GB"
     }
   ,
   "training":

Dockerfile.cpu

@@ -36,13 +36,13 @@ RUN pip install --upgrade pip && \
     conda install -c menpo opencv=2.4.11 && \
     conda clean -yt
 
-RUN mkdir /home/deepinfer/.keras &&
+RUN mkdir /home/deepinfer/.keras && \
     mkdir /home/deepinfer/.keras/models && \
     echo { '"image_dim_ordering"': '"th"', '"epsilon"': 1e-07, '"floatx"': '"float32"', '"backend"': '"theano"'}>/home/deepinfer/.keras/keras.json
 
 
 RUN mkdir -p /home/deepinfer/github && \
-    mkdir -p /home/deepinfer/github/bone-segmenter
+    mkdir -p /home/deepinfer/github/bone-segmenter && \
     mkdir /home/deepinfer/data
 
 WORKDIR "/home/deepinfer/github/bone-segmenter"

README.md

@@ -17,9 +17,13 @@ docker build -t bone-segmenter-cpu -f Dockerfile.cpu .
 
 ##### Running the docker
 ```
-docker run -t -v [Absolute PATH to the Project Folder]/Prostate-Segmenter/prostatesegmenter/data/test/:/home/deepinfer/data deepinfer/prostate-segmenter-cpu --InputVolume /home/deepinfer/data/input.nrrd --OutputLabel /home/deepinfer/data/label_predicted_test.nrrd
+docker run -t -v [Absolute PATH to the Project Folder]/Bone-Segmenter/data/test/:/home/deepinfer/data 4Quant/bone-segmenter-cpu --InputVolume /home/deepinfer/data/input.nrrd --OutputLabel /home/deepinfer/data/label_predicted_test.nrrd
 ```
 
+Windows Example
+```
+docker run -t -v "C:\Users\Kevin Mader\Dropbox\4Quant\Projects\Bone-Segmenter":/home/deepinfer/data bone-segmenter-cpu --InputVolume /home/deepinfer/data/CTChest.nrrd --OutputLabel /home/deepinfer/data/label_predicted_test.nrrd
+```
 
 # Results
 

boneseg/segmenter.py

@@ -1,7 +1,7 @@
 import SimpleITK as sitk
 import numpy as np
 
-from prostatesegmenter.utils import reshape_volume
+from boneseg.utils import reshape_volume
 
 
 class Segmenter(object):
@@ -11,22 +11,20 @@ def __init__(self, model, threshold=0.5):
 
     def __predict(self, vol_nda):
         predicted_label = self.model.predict_test(input_test_slices_arr=vol_nda)
-        predicted_label[predicted_label > self.threshold] = 1
-        predicted_label[predicted_label <= self.threshold] = 0
         return predicted_label
 
     def segment_bone_volume(self, input_volume_path, output_mask_path, rows, cols):
         vol = sitk.ReadImage(input_volume_path)
-        vol = sitk.Cast(sitk.RescaleIntensity(vol), sitk.sitkUInt8)
         vol_nda = sitk.GetArrayFromImage(vol)
         original_shape = vol_nda.shape
         vol_nda = reshape_volume(vol_nda, rows, cols)
         vol_nda = np.expand_dims(vol_nda, axis=1)
         label_nda = self.__predict(vol_nda)
         label_nda = label_nda[:, 0, :, :]
         # TODO: Check the indices for shape!
-        label_nda = reshape_volume(label_nda, original_shape[1], original_shape[2])
-        label = sitk.GetImageFromArray(label_nda)
+        # since the zoom function interpolates, threshold it again
+        label_nda = reshape_volume(label_nda, original_shape[1], original_shape[2])>0.5
+        label = sitk.GetImageFromArray(label_nda.astype(np.uint8))
         label.CopyInformation(vol)
         writer = sitk.ImageFileWriter()
         writer.Execute(label, output_mask_path, True)

boneseg/utils.py

@@ -1,3 +1,4 @@
+from __future__ import print_function, absolute_import, division
 import numpy as np
 import os
 import sys
@@ -25,11 +26,12 @@ def dice_coef_loss(y_true, y_pred):
     return 1 - dice_coef(y_true, y_pred)
 
 
+
 def reshape_volume(nda_vol, rows, cols):
-    from cv2 import resize, INTER_CUBIC
-    nda_vol_reshaped = np.ndarray((nda_vol.shape[0], rows, cols), dtype=np.uint8)
-    for i in range(len(nda_vol)):
-        nda_vol_reshaped[i] = resize(nda_vol[i], (cols, rows), interpolation= INTER_CUBIC)
+    from scipy.ndimage import zoom
+    nda_vol_reshaped = np.ndarray((nda_vol.shape[0], rows, cols), dtype=np.float32)
+    zoom_fact = (1, rows/nda_vol.shape[1], cols/nda_vol.shape[2])
+    zoom(nda_vol, zoom_fact, output = nda_vol_reshaped)
     return nda_vol_reshaped
 
 
@@ -80,4 +82,4 @@ def dl_progress(count, block_size, total_size):
             os.remove(fpath)
         raise e
 
-    return fpath
+    return fpath

figures/boneseg_predictions.png

@@ -1,3 +1,11 @@
+from __future__ import print_function, absolute_import, division
+# Silence warnings
+import warnings
+warnings.simplefilter(action="ignore", category=FutureWarning)
+warnings.simplefilter(action="ignore", category=UserWarning)
+warnings.simplefilter(action="ignore", category=RuntimeWarning)
+warnings.simplefilter(action='ignore', category=DeprecationWarning)
+
 import getopt
 import os
 import sys
@@ -31,13 +39,13 @@ def main(argv):
         print('usage: fit.py -InputVolume <InputVolumePath> -OutputLabel <OutputLabelPath>')
         sys.exit()
     if os.path.isfile(InputVolume) and os.path.isdir(os.path.dirname(OutputLabel)):
-        print("Making the model.")
+        print("Building the model.")
         model = unet1.model(weights=True, summary=False)
-        cnn = CNNModel(data_streamer=ds, model=model)
+        cnn = CNNModel(model=model, mean_val = 0, max_val = 300)
         rows = 128
         cols = 128
         #
-        print("Starting the segmenter.")
+        print("Resizing images and segmenting")
         sg = Segmenter(cnn)
         sg.segment_bone_volume(InputVolume, OutputLabel, rows, cols)
     else:

fit.py

@@ -1,2 +1,2 @@
 #!/bin/bash
-python fit.py --InputVolume prostatesegmenter/data/test/Case1-t2ax-intraop.nrrd --OutputLabel prostatesegmenter/data/test/Case1-t2ax-intraop_label.nrrd
+python fit.py --InputVolume CTChest.nrrd --OutputLabel test_out.nrrd