Add tool to convert monochrome 1 images to monochrome 2

README.md

@@ -17,6 +17,7 @@ Inference model integration SDK
 - [Testing the inference server](#testing-the-inference-server)
   - [To send an inference request to the mock inference server](#to-send-an-inference-request-to-the-mock-inference-server)
   - [Running Unit Tests](#running-unit-tests)
+- [Handling different photometric interpretations of input files](#handling-different-photometric-interpretations-of-input-files)
 - [Nifti image format support](#nifti-image-format-support)
 - [Secondary capture support](#secondary-capture-support)
 
@@ -319,6 +320,25 @@ python3 -m unittest
 
 You must have Python 3.6+ installed.
 
+## Handling different photometric interpretations of input files
+
+DICOM files can have multiple photometric interpretations (RGB, YBR, grayscale, etc.).
+It is the responsibility of the model developer to handle all the possible interpretations supported for the input image types (CT, Xray, etc.).
+
+In `utils/image_conversion.py` there is a function that converts MONOCHROME1 (0 is white) image files to MONOCHROME2 (0 is black).
+You can use it like this:
+
+```python
+for dcm_file in dicom_instances:
+    dcm = pydicom.read_file(dcm_file)
+    dcm = convert_monochrome_1to2(dcm)
+
+    # If you need to convert the pydicom object back to a BytesIO object:
+    # dcm_bytes = convert_dataset_to_bytes(dcm)
+```
+
+Other conversions might be necessary and might be added in the future.
+
 ## Nifti image format support
 
 In the `utils/image_conversion.py` there are a few functions that can be helpful if your model accepts Nifti files as input or generates Nifti output files.

inference-test-tool/utils.py

@@ -85,6 +85,7 @@ def sort_images(images):
         np.dot(np.subtract(np.array(item2.position), np.array(pos)), direction)))
 
 def get_pixels(dicom_file):
+    """ Gets RGB pixels from a DICOM file. If they were 16 bit ints then they will be converted to uint8. """
     pixels = dicom_file.pixel_array
     if dicom_file.PhotometricInterpretation == 'PALETTE COLOR':
         pixels = apply_color_lut(pixels, dicom_file)

utils/image_conversion.py

@@ -1,7 +1,8 @@
 import numpy as np
 import SimpleITK as sitk
-from pydicom import dcmread
-from pydicom.filebase import DicomBytesIO
+from pydicom import dcmread, dcmwrite
+from pydicom.filebase import DicomBytesIO, DicomFileLike
+from io import BytesIO
 
 ARTERYS_PROBABILITY_MASK='probability_mask'
 ARTERYS_BINARY='binary'
@@ -62,3 +63,32 @@ def get_masks_from_nifti_file(nifti_file, data_type=ARTERYS_PROBABILITY_MASK, nu
         return np.array(output)
 
     return [arr]
+
+
+def convert_monochrome_1to2(dcm):
+    """If the DICOM file `dcm` is in MONOCHROME1 then convert it to MONOCHROME2 """
+    if dcm.PhotometricInterpretation == 'MONOCHROME1':
+        pixels = dcm.pixel_array
+
+        # handle different dtypes
+        if pixels.dtype in [np.uint8, np.int8, np.uint16, np.int16]:
+            pixels = np.invert(pixels)            
+        else:
+            raise RuntimeError("Non integer image data types currently not supported.")
+
+        # Update original DICOM file
+        dcm.PixelData = pixels.tobytes()
+        dcm.PhotometricInterpretation = 'MONOCHROME2'
+
+    return dcm
+
+
+def convert_dataset_to_bytes(dataset):
+    """ Get a Bytes object from a DICOM dataset.
+        Adapted from https://pydicom.github.io/pydicom/dev/auto_examples/memory_dataset.html
+    """
+    with BytesIO() as buffer:
+        memory_dataset = DicomFileLike(buffer)
+        dcmwrite(memory_dataset, dataset)
+        memory_dataset.seek(0)
+        return memory_dataset.read()