Merge pull request #15 from IvanKuchin/development

save checkpoint on each epoch

config.py

@@ -18,8 +18,8 @@
 # https://radiopaedia.org/articles/windowing-ct?lang=us
 # Option 2) 3D Slicer preset for abdominal CT
 # W/L: 350/40, which makes the pancreas range from -310 to 390
-PANCREAS_MIN_HU = -524
-PANCREAS_MAX_HU = 1024
+PANCREAS_MIN_HU =  -310 # -512
+PANCREAS_MAX_HU = 400 # 1024
 
 IMAGE_DIMENSION_X = 160
 IMAGE_DIMENSION_Y = IMAGE_DIMENSION_X

dataset/pomc_dataset.py

@@ -147,7 +147,7 @@ def consistency_check(self, data, label, data_metadata, label_metadata):
 
  if data.shape != label.shape:
  print("ERROR: data shape(", data.shape, ") is not equal to the label shape(", label.shape, ")")
- # return False
+ return False
 
  if not self._point_inside_box(data_metadata["min"], data_metadata["max"], label_metadata["space origin"]):
  print("ERROR: label space origin(", label_metadata["space origin"], ") is outside the data box(", data_metadata["min"], data_metadata["max"], ")")

predict.py

@@ -18,10 +18,19 @@ def __read_dcm_slices(self, dcm_folder):
  print("ERROR: no dcim files found")
  return slices
 
+ # Stored Values (SV) are the values stored in the image pixel data attribute.
+ # Representation value should be calculated as:
+ # Rescaled value = SV * Rescale Slope + Rescale Intercept
+ # https://dicom.innolitics.com/ciods/digital-x-ray-image/dx-image/00281052
+ def __dcim_slice_stored_value_to_rescaled_value(self, slice):
+ rescale_intercept = slice.RescaleIntercept if hasattr(slice, "RescaleIntercept") else 0
+ rescale_slope = slice.RescaleSlope if hasattr(slice, "RescaleSlope") else 1
+ return slice.pixel_array * rescale_slope + rescale_intercept
+
  def __get_pixel_data(self, dcm_slices):
  result = np.array([])
  if len(dcm_slices):
- result = np.stack([_.pixel_array for _ in dcm_slices], axis = -1)
+ result = np.stack([self.__dcim_slice_stored_value_to_rescaled_value(_) for _ in dcm_slices], axis = -1)
  else:
  print("ERROR: dcim list is empty")
 
@@ -85,6 +94,12 @@ def __get_affine_matrix(self, dcm_slices):
  affine[1, 1] = dcm_patient_orientation[4] * dcm_pixel_spacing[1]
  affine[2, 1] = dcm_patient_orientation[5] * dcm_pixel_spacing[1]
 
+ affine[0, 3] = dcm_patient_position[0]
+ affine[1, 3] = dcm_patient_position[1]
+ affine[2, 3] = dcm_patient_position[2]
+
+ affine[2, 2] = dcm_slice_thickness
+
  # --- inverse axes X and Y. This was found experimental way
  # --- could be wrong ... 
  affine[0, 0] = -affine[0, 0]
@@ -95,11 +110,8 @@ def __get_affine_matrix(self, dcm_slices):
  affine[1, 1] = -affine[1, 1]
  affine[2, 1] = -affine[2, 1]
 
- affine[2, 2] = dcm_slice_thickness
-
- affine[0, 3] = dcm_patient_position[0]
- affine[1, 3] = dcm_patient_position[1]
- affine[2, 3] = dcm_patient_position[2]
+ affine[0, 3] = -affine[0, 3]
+ affine[1, 3] = -affine[1, 3]
 
  return affine
 

train_segmentation.py

@@ -54,8 +54,13 @@ def main():
  model = craft_network(config.MODEL_CHECKPOINT)
  # predict_on_random_data(model)
 
- checkpoint_cb = tf.keras.callbacks.ModelCheckpoint(config.MODEL_CHECKPOINT, monitor = config.MONITOR_METRIC,
- mode = config.MONITOR_MODE, verbose = 2, save_best_only = True)
+ checkpoint_cb = tf.keras.callbacks.ModelCheckpoint(
+ config.MODEL_CHECKPOINT,
+ monitor = config.MONITOR_METRIC,
+ mode = config.MONITOR_MODE,
+ verbose = 2,
+ # save_best_only = True
+ )
  csv_logger = tf.keras.callbacks.CSVLogger(get_csv_dir(), separator = ',', append = True)
  tensorboard_cb = tf.keras.callbacks.TensorBoard(get_tensorboard_log_dir())
  reduce_lr_on_plateau = tf.keras.callbacks.ReduceLROnPlateau(factor = 0.1,