Merge pull request #13 from zivy/resize

resizing an image while retaining physical location.

Author: blowekamp

SimpleITK/utilities/__init__.py

@@ -21,6 +21,7 @@
 from .make_isotropic import make_isotropic
 from .fft import fft_based_translation_initialization
 from .overlay_bounding_boxes import overlay_bounding_boxes
+from .resize import resize
 
 try:
     from ._version import version as __version__
@@ -34,6 +35,7 @@
     "make_isotropic",
     "fft_based_translation_initialization",
     "overlay_bounding_boxes",
+    "resize",
     "__version__",
 ]
 

SimpleITK/utilities/resize.py

@@ -0,0 +1,110 @@
+import SimpleITK as sitk
+from typing import Sequence
+from typing import Union
+from math import ceil
+import collections
+
+
+def _issequence(obj):
+    if isinstance(obj, (bytes, str)):
+        return False
+    return isinstance(obj, collections.abc.Sequence)
+
+
+def resize(
+    image: sitk.Image,
+    new_size: Sequence[int],
+    isotropic: bool = True,
+    fill: bool = True,
+    interpolator=sitk.sitkLinear,
+    fill_value: float = 0.0,
+    use_nearest_extrapolator: bool = False,
+    anti_aliasing_sigma: Union[None, float, Sequence[float]] = None,
+) -> sitk.Image:
+    """
+    Resize an image to an arbitrary size while retaining the original image's spatial location.
+
+    Allows for specification of the target image size in pixels, and whether the image pixels spacing should be
+    isotropic. The physical extent of the image's data is retained in the new image, with the new image's spacing
+    adjusted to achieve the desired size. The image is centered in the new image.
+
+    :param image: A SimpleITK image.
+    :param new_size: The new image size in pixels.
+    :param isotropic: If False, the original image is resized to fill the new image size by adjusting space. If True,
+    the new image's spacing will be isotropic.
+    :param fill: If True, the output image will be new_size, and the original image will be centered in the new image
+    with constant or nearest values used to fill in the new image. If False and isotropic is True, the output image's
+    new size will be calculated to fit the original image's extent such that at least one dimension is equal to
+    new_size.
+    :param fill_value: Value used for padding.
+    :param interpolator: Interpolator used for resampling.
+    :param use_nearest_extrapolator: If True, use a nearest neighbor for extrapolation when resampling, overridding the
+    constant fill value.
+    :param anti_aliasing_sigma: If zero no antialiasing is performed. If a scalar, it is used as the sigma value in
+     physical units for all axes. If None or a sequence, the sigma value for each axis is calculated as
+     $sigma = (new_spacing - old_spacing) / 2$ in physical units. Gaussian smoothing is performed prior to resampling
+     for antialiasing.
+    :return: A SimpleITK image with desired size.
+    """
+    new_spacing = [
+        (osz * ospc) / nsz
+        for ospc, osz, nsz in zip(image.GetSpacing(), image.GetSize(), new_size)
+    ]
+
+    if isotropic:
+        new_spacing = [max(new_spacing)] * image.GetDimension()
+    if not fill:
+        new_size = [
+            ceil(osz * ospc / nspc)
+            for ospc, osz, nspc in zip(image.GetSpacing(), image.GetSize(), new_spacing)
+        ]
+
+    center_cidx = [0.5 * (sz - 1) for sz in image.GetSize()]
+    new_center_cidx = [0.5 * (sz - 1) for sz in new_size]
+
+    new_origin_cidx = [0] * image.GetDimension()
+    # The continuous index of the new center of the image, in the original image's continuous index space.
+    for i in range(image.GetDimension()):
+        new_origin_cidx[i] = center_cidx[i] - new_center_cidx[i] * (
+            new_spacing[i] / image.GetSpacing()[i]
+        )
+
+    new_origin = image.TransformContinuousIndexToPhysicalPoint(new_origin_cidx)
+
+    input_pixel_type = image.GetPixelID()
+
+    if anti_aliasing_sigma is None:
+        # (s-1)/2.0 is the standard deviation of the Gaussian kernel in index space, where s downsample factor defined
+        # by nspc/ospc.
+        anti_aliasing_sigma = [
+            max((nspc - ospc) / 2.0, 0.0)
+            for ospc, nspc in zip(image.GetSpacing(), new_spacing)
+        ]
+    elif not _issequence(anti_aliasing_sigma):
+        anti_aliasing_sigma = [anti_aliasing_sigma] * image.GetDimension()
+
+    if any([s < 0.0 for s in anti_aliasing_sigma]):
+        raise ValueError("anti_aliasing_sigma must be positive, or None.")
+    if len(anti_aliasing_sigma) != image.GetDimension():
+        raise ValueError(
+            "anti_aliasing_sigma must be a scalar or a sequence of length equal to the image dimension."
+        )
+
+    if all([s > 0.0 for s in anti_aliasing_sigma]):
+        image = sitk.SmoothingRecursiveGaussian(image, anti_aliasing_sigma)
+    else:
+        for d, s in enumerate(anti_aliasing_sigma):
+            if s > 0.0:
+                image = sitk.RecursiveGaussian(image, sigma=s, direction=d)
+
+    return sitk.Resample(
+        image,
+        size=new_size,
+        outputOrigin=new_origin,
+        outputSpacing=new_spacing,
+        outputDirection=image.GetDirection(),
+        defaultPixelValue=fill_value,
+        interpolator=interpolator,
+        useNearestNeighborExtrapolator=use_nearest_extrapolator,
+        outputPixelType=input_pixel_type,
+    )

test/test_utilities.py

@@ -1,5 +1,8 @@
+import math
+
 import SimpleITK as sitk
 import SimpleITK.utilities as sitkutils
+from numpy.testing import assert_allclose
 
 
 def test_Logger():
@@ -88,3 +91,201 @@ def test_overlay_bounding_boxes():
         scalar_hash == "d7dde3eee4c334ffe810a636dff872a6ded592fc"
         and rgb_hash == "d6694a394f8fcc32ea337a1f9531dda6f4884af1"
     )
+
+
+def test_resize():
+    original_image = sitk.Image([128, 128], sitk.sitkUInt8) + 50
+    resized_image = sitkutils.resize(image=original_image, new_size=[128, 128])
+    assert resized_image.GetSize() == (128, 128)
+    assert resized_image.GetSpacing() == (1.0, 1.0)
+    assert resized_image.GetOrigin() == (0.0, 0.0)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint((-0.5, -0.5)) == (
+        -0.5,
+        -0.5,
+    )
+
+    resized_image = sitkutils.resize(image=original_image, new_size=[64, 64])
+    assert resized_image.GetSize() == (64, 64)
+    assert resized_image.GetSpacing() == (2.0, 2.0)
+    assert resized_image.GetOrigin() == (0.5, 0.5)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint((-0.5, -0.5)) == (
+        -0.5,
+        -0.5,
+    )
+
+    resized_image = sitkutils.resize(
+        image=original_image, new_size=[64, 128], fill=False
+    )
+    assert resized_image.GetSize() == (64, 64)
+    assert resized_image.GetSpacing() == (2.0, 2.0)
+    assert resized_image.GetOrigin() == (0.5, 0.5)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint((-0.5, -0.5)) == (
+        -0.5,
+        -0.5,
+    )
+
+    resized_image = sitkutils.resize(
+        image=original_image, new_size=[64, 128], isotropic=False
+    )
+    assert resized_image.GetSize() == (64, 128)
+    assert resized_image.GetSpacing() == (2.0, 1.0)
+    assert resized_image.GetOrigin() == (0.5, 0.0)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint((-0.5, -0.5)) == (
+        -0.5,
+        -0.5,
+    )
+
+
+def test_resize_3d():
+    original_image = sitk.Image([128, 128, 128], sitk.sitkUInt8) + 50
+    resized_image = sitkutils.resize(image=original_image, new_size=[128, 128, 128])
+    assert resized_image.GetSize() == (128, 128, 128)
+    assert resized_image.GetSpacing() == (1.0, 1.0, 1.0)
+    assert resized_image.GetOrigin() == (0.0, 0.0, 0.0)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint(
+        (-0.5, -0.5, -0.5)
+    ) == (
+        -0.5,
+        -0.5,
+        -0.5,
+    )
+
+    resized_image = sitkutils.resize(image=original_image, new_size=[64, 64, 64])
+    assert resized_image.GetSize() == (64, 64, 64)
+    assert resized_image.GetSpacing() == (2.0, 2.0, 2.0)
+    assert resized_image.GetOrigin() == (0.5, 0.5, 0.5)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint(
+        (-0.5, -0.5, -0.5)
+    ) == (
+        -0.5,
+        -0.5,
+        -0.5,
+    )
+
+    resized_image = sitkutils.resize(image=original_image, new_size=[64, 32, 64])
+    assert resized_image.GetSize() == (64, 32, 64)
+    assert resized_image.GetSpacing() == (4.0, 4.0, 4.0)
+    assert resized_image.GetOrigin() == (-62.5, 1.5, -62.5)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint(
+        (-0.5, -0.5, -0.5)
+    ) == (
+        -64.5,
+        -0.5,
+        -64.5,
+    )
+
+    resized_image = sitkutils.resize(
+        image=original_image, new_size=[64, 64, 32], fill=False
+    )
+    assert resized_image.GetSize() == (32, 32, 32)
+    assert resized_image.GetSpacing() == (4.0, 4.0, 4.0)
+    assert resized_image.GetOrigin() == (1.5, 1.5, 1.5)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint(
+        (-0.5, -0.5, -0.5)
+    ) == (
+        -0.5,
+        -0.5,
+        -0.5,
+    )
+
+    resized_image = sitkutils.resize(
+        image=original_image, new_size=[32, 64, 64], isotropic=False
+    )
+    assert resized_image.GetSize() == (32, 64, 64)
+    assert resized_image.GetSpacing() == (4.0, 2.0, 2.0)
+    assert resized_image.GetOrigin() == (1.5, 0.5, 0.5)
+    assert resized_image.TransformContinuousIndexToPhysicalPoint(
+        (-0.5, -0.5, -0.5)
+    ) == (
+        -0.5,
+        -0.5,
+        -0.5,
+    )
+
+
+def test_resize_fill():
+    original_image = sitk.Image([16, 32], sitk.sitkFloat32) + 1.0
+
+    resized_image = sitkutils.resize(
+        image=original_image, new_size=[32, 32], fill=True, fill_value=10.0
+    )
+    assert resized_image.GetSize() == (32, 32)
+    assert resized_image.GetSpacing() == (1.0, 1.0)
+    assert resized_image.GetOrigin() == (-8.0, 0.0)
+    assert resized_image[0, 0] == 10.0
+    assert resized_image[15, 15] == 1.0
+    assert resized_image[31, 31] == 10.0
+
+    resized_image = sitkutils.resize(
+        image=original_image,
+        new_size=[32, 32],
+        fill=True,
+        use_nearest_extrapolator=True,
+    )
+    assert resized_image.GetSize() == (32, 32)
+    assert resized_image.GetSpacing() == (1.0, 1.0)
+    assert resized_image.GetOrigin() == (-8.0, 0.0)
+    assert resized_image[0, 0] == 1.0
+    assert resized_image[15, 15] == 1.0
+    assert resized_image[31, 31] == 1.0
+
+
+def test_resize_anti_aliasing():
+    original_image = sitk.Image([5, 5], sitk.sitkFloat32)
+    original_image[2, 2] = 1.0
+
+    resized_image = sitkutils.resize(
+        image=original_image,
+        new_size=[3, 3],
+        interpolator=sitk.sitkNearestNeighbor,
+        anti_aliasing_sigma=0,
+    )
+    assert resized_image.GetSize() == (3, 3)
+    assert_allclose(resized_image.GetSpacing(), (5 / 3, 5 / 3))
+    assert_allclose(resized_image.GetOrigin(), (1 / 3, 1 / 3))
+    assert resized_image[0, 0] == 0.0
+    assert resized_image[1, 1] == 1.0
+    assert resized_image[1, 0] == 0.0
+    assert resized_image[0, 1] == 0.0
+
+    resized_image = sitkutils.resize(
+        image=original_image,
+        new_size=[3, 3],
+        interpolator=sitk.sitkNearestNeighbor,
+        anti_aliasing_sigma=None,
+    )
+    assert resized_image.GetSize() == (3, 3)
+    assert_allclose(resized_image.GetSpacing(), (5 / 3, 5 / 3))
+    assert_allclose(resized_image.GetOrigin(), (1 / 3, 1 / 3))
+    assert math.isclose(resized_image[1, 1], 0.960833, abs_tol=1e-6)
+    assert resized_image[1, 0] == resized_image[0, 1]
+    assert resized_image[0, 0] == resized_image[2, 2]
+
+    resized_image = sitkutils.resize(
+        image=original_image,
+        new_size=[3, 3],
+        interpolator=sitk.sitkNearestNeighbor,
+        anti_aliasing_sigma=0.5,
+    )
+    assert resized_image.GetSize() == (3, 3)
+    assert_allclose(resized_image.GetSpacing(), (5 / 3, 5 / 3))
+    assert_allclose(resized_image.GetOrigin(), (1 / 3, 1 / 3))
+    assert math.isclose(resized_image[1, 1], 0.621714, abs_tol=1e-6)
+    assert math.isclose(resized_image[0, 0], 0, abs_tol=1e-6)
+    assert math.isclose(resized_image[1, 0], resized_image[0, 1], abs_tol=1e-8)
+    assert math.isclose(resized_image[0, 0], resized_image[2, 2], abs_tol=1e-8)
+
+    resized_image = sitkutils.resize(
+        image=original_image,
+        new_size=[3, 3],
+        interpolator=sitk.sitkNearestNeighbor,
+        anti_aliasing_sigma=[1.0, 0.0],
+    )
+    assert resized_image.GetSize() == (3, 3)
+    assert_allclose(resized_image.GetSpacing(), (5 / 3, 5 / 3))
+    assert_allclose(resized_image.GetOrigin(), (1 / 3, 1 / 3))
+    assert math.isclose(resized_image[1, 1], 0.400101, abs_tol=1e-6)
+    assert math.isclose(resized_image[0, 0], 0, abs_tol=1e-6)
+    assert resized_image[0, 0] == resized_image[2, 2]
+    assert resized_image[1, 0] == 0.0
+    assert resized_image[1, 2] == 0.0