BUG: issue when segmentation touched the image border

The surface computation used in the surface_hausdorff_distance yielded
incorrect results when the segmentation was touching the image
border. In this case the surface is outside the image extent, so when
voxelized the expected voxels were not marked as surface. By padding
the image we resolve the problem.

resolves #453

Author: zivy

Python/34_Segmentation_Evaluation.ipynb

@@ -490,6 +490,18 @@
     "    \"\"\"\n",
     "    Compute symmetric surface distances and take the maximum.\n",
     "    \"\"\"\n",
+    "    # When the segmentation touches the image border we have corner case for the computation.\n",
+    "    # The surface/contour computation yields a surface that is outside the image bounds and\n",
+    "    # is thus not included in the computation. This is not correct for our use case. By padding\n",
+    "    # the segmentations we ensure that we obtain the surface as expected and the distance computations\n",
+    "    # are correct. Padding has no effect on results if the original segmentations did not touch the image\n",
+    "    # border.\n",
+    "    dim = reference_segmentation.GetDimension()\n",
+    "    reference_segmentation = sitk.ConstantPad(\n",
+    "        reference_segmentation, [1] * dim, [1] * dim\n",
+    "    )\n",
+    "    seg = sitk.ConstantPad(seg, [1] * dim, [1] * dim)\n",
+    "\n",
     "    reference_surface = sitk.LabelContour(reference_segmentation)\n",
     "    reference_distance_map = sitk.Abs(\n",
     "        sitk.SignedMaurerDistanceMap(\n",