Walkthrough: Point Cloud to Surface Registration - rg2/jhmr-v2 GitHub Wiki

Point Cloud to Surface Registration

The xreg-sur-regi tool performs a registration between a point cloud and a surface using the iterative most-likely point (ICP) approach. The method optimizes over a rigid tranformation, but may also attempt to recover scale. ICP is intialized using the identity transformation by default, but two FCSV files may be provided in order to perform a paired-point registration to provide a better initialization. A KD-Tree of the surface is constructed in order to efficiently perform closest-point lookups.

Example

This example will simulate the case of an intraoperative registration of the pelvis by digitizing points on the pelvis surface with a tracked tool. Thepelvis.nii.gz volume created in the volume cropping example and the pelvis_sur.ply mesh created in the mesh creation example will be reused in this example.

First, load the pelvis volume into 3D Slicer and launch a volume rendering with a threshold displaying the bone surface. Next, use the fiducials module (point picker) to identify the following bilateral landmarks:

  • Anterior Superior Iliac Spine (ASIS)
  • Anterior Inferior Iliac Spine (AIIS)
  • Pubic Symphisis (PS) A suffix of "-l" is used to indicate left side and "-r" is used to indicate the right side. The following screenshot demonstrates this process:

Example Pelvis Preop Landmarks

Save this FCSV file as pelvis_preop_lands.fcsv. These will simulate preoperatively identified landmarks that will be useful for initializing an intraoperative ICP registration.

Next, we shall simulate the process of digitizing these landmarks intraoperatively for a surgery on the left side. In a new markup fiducials list, choose the ASIS-l, AIIS-l and PS-l again, but choose the landmarks slightly imprecisely and not exactly where the preoperative landmarks were selected. This is shown in the following screenshot:

Example Pelvis Intraop Boney Landmarks

Since the surgery is on the left side, we have assumed that an incision was made, soft-tissues dissected, and a tracked pointer tool is able to sufficiently reach the boney anatomy. Next, we shall simulate a digitization of the ASIS-r on the skin surface. Lower the volume rendering threshold and choose a point roughly corresponding to the ASIS-r.

Example Pelvis Intraop Skin Landmark

Since the volume was previously cropped, much of the skin surface is not visible. However, this will suffice for illustrating this example.

We shall now create a simulation of the clinicial moving the tracked pointer across the pelvis surface. Change the volume rendering back to show the bone surface, create a new markup fiducials list, and start digitizing points along the left iliac crest, left iliac spine, and left ilium wing. An example of this is shown below:

Example Pelvis Intraop Point Cloud

Next, use the transforms module in Slicer to apply an arbitrary rigid transform to the intraoperative landmarks and point cloud as shown below:

Example Pelvis Intraop Transform

Use the "harden" transform button on the fiducials and save the intraoperative landmarks to disk as pelvis_intraop_lands.fcsv (an example is here) and pelvis_intraop_point_cloud.fcsv (an example is here).

We are now ready to run the ICP registration:

xreg-sur-regi pelvis_sur.ply pelvis_intraop_point_cloud.fcsv pelvis_regi.h5 --mesh-lands pelvis_preop_lands.fcsv --pts-lands pelvis_intraop_lands.fcsv

The output registration transform will be saved as pelvis_regi.h5. Passing the verbose flag (-v) will enable printing of some other diagnostic information of the method, of which an example snippet is listed below:

reading mesh from disk...
  complete.
Point Cloud is FCSV: Yes
reading point cloud from disk...
  complete.
converting FCSV point cloud RAS --> LPS...
reading mesh landmarks...
    AIIS-r:   +95.306000 ,  -104.043999 ,  -608.679993
    ASIS-r:  +107.676003 ,   -85.442001 ,  -572.283020
    AIIS-l:  -106.779999 ,  -129.621994 ,  -608.471008
      PS-r:    +2.256000 ,   -97.639000 ,  -643.765015
      PS-l:    -8.283000 ,  -100.521004 ,  -644.674011
    ASIS-l:  -122.205002 ,  -114.559998 ,  -578.065979
reading point cloud landmarks...
    ASIS-r:  -655.932007 ,  +360.757996 ,   +36.963001
    AIIS-l:  -608.034973 ,  +288.065002 ,  -183.858994
      PS-l:  -675.512024 ,  +312.785004 ,  -109.380997
    ASIS-l:  -576.088013 ,  +264.449005 ,  -170.052994
landmarks RAS --> LPS...
estimating initial transformation using corresponding paired points...
running ICP...
ICP init...
  Creating KD-Tree for target surface...
    KD-Tree created.
Allocating additional working buffers...
ICP init complete!
Starting Point to Surface ICP: Num Pts 45
   Surface Num Verts 302318, Num Tri Faces 604788
  Initial Xform:
-0.329926  0.509905 -0.794447   311.505
-0.260543 -0.858066 -0.442538   -38.787
 -0.90734 0.0609827   0.41595   32.7704
        0         0         0         1
    Stop Thresh: 0.999
  Compute Scale: No
   Outlier Det.: No
  Iteration 0
    Xform:
-0.329926  0.509905 -0.794447   311.505
-0.260543 -0.858066 -0.442538   -38.787
 -0.90734 0.0609827   0.41595   32.7704
        0         0         0         1
    Mean Dist: 3.99371
  Iteration 1
    Xform:
-0.338177  0.538369 -0.771878   326.728
-0.266438 -0.841416 -0.470138  -35.3461
-0.902578 0.0466676  0.427989   28.3553
        0         0         0         1
    Mean Dist: 3.26624
    Delta Mean Dist. Ratio: 0.817846
  Iteration 2
    Xform:
-0.343989  0.558381 -0.754906   337.537
-0.269822 -0.828842 -0.490119  -33.1895
-0.899371 0.0350941  0.435776   24.7853
        0         0         0         1
    Mean Dist: 2.67226
    Delta Mean Dist. Ratio: 0.818147

.
.
.

  Iteration 52
    Xform:
-0.300638  0.530685 -0.792458   292.454
-0.246907 -0.845877 -0.472789  -49.0581
-0.921224 0.0535248  0.385333   35.9088
        0         0         0         1
    Mean Dist: 0.282814
    Delta Mean Dist. Ratio: 0.998813
  Iteration 53
    Xform:
-0.300588   0.53067 -0.792488   292.414
-0.246999 -0.845879 -0.472736  -48.9918
-0.921216 0.0536451  0.385336    35.949
        0         0         0         1
    Mean Dist: 0.282509
    Delta Mean Dist. Ratio: 0.998925
  Iteration 54
    Xform:
-0.300548  0.530636 -0.792526   292.373
-0.247109 -0.845892 -0.472656  -48.9141
-0.921199 0.0537845  0.385356   35.9935
        0         0         0         1
    Mean Dist: 0.282327
    Delta Mean Dist. Ratio: 0.999354
    termination criteria met: mean distance has stopped decreasing
writing regi transform to disk...
exiting...

Now transform the intraoperative point cloud using the registration transformation:

xreg-xform-fcsv pelvis_intraop_point_cloud.fcsv pelvis_regi.h5 pelvis_intraop_point_cloud_regi.fcsv

This transformed point cloud, pelvis_intraop_point_cloud_regi.fcsv, may then be loaded into 3D Slicer and overlaid on the bone surface, shown below:

Example Pelvis Regi. Point Cloud