Allen Institute for Neural Dynamics + Kitware Collab: SmartSPIM images

Modality	Subject ID	Neuroglancer	ITK-VTK-Viewer	Viewer Notes	Registration Notes	Additional Notes
SmartSPIM	640390	link		Ex_445_Em_469 channel (NIFTI), CCF v3.1 atlas (precomputed)	Ex_445_Em_469 intensity image registered to CCF v3.1 average template
SmartSPIM	631680	link		Ex_647_Em_690 channel (NIFTI), CCF v3.1 atlas (precomputed)	Ex_647_Em_690 intensity image registered to CCF v3.1 average template
SmartSPIM	652506	link		Channels Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 (NIFTI), CCF v3.1 atlas (precomputed)	Ex_561_Em_593 registered to CCF v3.1, transform result applied to other channels with transformix
SmartSPIM	652506	link		Unregistered Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 (OME-Zarr)	N/A
SmartSPIM	652506	link		Unregistered Ex_561_Em_593 (NIFTI), VIIn, MH expert annotations (NIFTI), labels resampled with transformix and "FinalBSplineInterpolationOrder" == "0"	N/A
SmartSPIM	652506	link		Ex_561_Em_593 (NIFTI), VIIn, MH expert annotations (NIFTI)	Ex_561_Em_593 registered to CCF v3.1, labels resampled with transformix and "FinalBSplineInterpolationOrder" == "0"
SmartSPIM	652506	link		Ex_561_Em_593 (OME-Zarr), VIIn, MH expert segmentation (OME-Zarr)	Ex_561_Em_593 registered to CCF v3.1, labels resampled with transformix and "FinalBSplineInterpolationOrder" == "0"
SmartSPIM	652506	link		Ex_561_Em_593 (OME-Zarr), VIIn, MH expert segmentation (OME-Zarr), CCF v3.1 10um atlas (precomputed)	Ex_561_Em_593 registered to CCF v3.1, labels resampled with transformix and "FinalBSplineInterpolationOrder" == "0"
SmartSPIM	652506	link		7n, MHN, Fr, aco expert segmentations (NIFTI), CCF v3.1 10um (precomputed), CCF v3.1 10um contours (NIFTI)	Ex_561_Em_593 registered to CCF v3.1, labels resampled with transformix and "FinalBSplineInterpolationOrder" == "0"
SmartSPIM	652506	link		Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 channels (OME-Zarr), CCF v3.1 10um (precomputed), CCF v3.1 10um contours (NIFTI)	Each channel (Level 4) independently registered to CCF v3.1 25um with nonunit origin.	Previous registrations targeted CCF with unit origin. Images are aligned in Neuroglancer for reference, but may be disoriented in world space.
SmartSPIM	652506	link	link distances	VIIn, MH, fr, aco expert annotations corresponding to Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 channels (precomputed). Neuroglancer: Label volume MPR view with CCF v3.1 10um (precomputed), CCF v3.1 10um contours (NIFTI) ITK-VTK-Viewer: Mesh geometry comparison	Each channel (Level 4) independently registered to CCF v3.1 25um with nonunit origin. Binary labels composed into one label image and resampled with "itk::ResampleImageFilter" and "itk::LabelImageGenericInterpolateImageFunction".	Images are aligned in Neuroglancer for reference, but may be disoriented in world space.
SmartSPIM	652506	link	link distances	VIIn, MH, fr, aco expert annotations corresponding to Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 channels (precomputed). Neuroglancer: Label volume MPR view with CCF v3.1 10um (precomputed), CCF v3.1 10um contours (NIFTI) ITK-VTK-Viewer: Mesh geometry comparison	Each channel (Level 4) independently registered to CCF v3.1 10um with nonunit origin. Binary labels composed into one label image and resampled with "itk::ResampleImageFilter" and "itk::LabelImageGenericInterpolateImageFunction".	Images are aligned in Neuroglancer for reference, but may be disoriented in world space.
SmartSPIM	652506	link	link distances	VIIn, MH, fr, aco expert annotations corresponding to Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 channels (precomputed). Neuroglancer: Label volume MPR view with CCF v3.1 10um (precomputed), CCF v3.1 10um contours (NIFTI) ITK-VTK-Viewer: Mesh geometry comparison	Each channel (Level 4) independently registered to CCF v3.1 25um with nonunit origin. `antspyx` Python package used with `SyN` preset (symmetric normalization, affine + deformable transformation, mutual information metric) and scales [100,10,0], with resulting transforms composed into an `itk::CompositeTransform`. Binary labels composed into one label image and resampled with "itk::ResampleImageFilter" and "itk::LabelImageGenericInterpolateImageFunction".	Images are aligned in Neuroglancer for reference, but may be disoriented in world space.
SmartSPIM	652506	link	link distances	VIIn, MH, fr, aco expert annotations corresponding to Ex_445_Em_469, Ex_488_Em_525, Ex_561_Em_593 channels (precomputed). Neuroglancer: Label volume MPR view with CCF v3.1 10um (precomputed), CCF v3.1 10um contours (NIFTI) ITK-VTK-Viewer: Mesh geometry comparison	Each channel (Level 3) independently registered to CCF v3.1 10um with nonunit origin. Binary labels composed into one label image and resampled with "itk::ResampleImageFilter" and "itk::LabelImageGenericInterpolateImageFunction".	Images are aligned in Neuroglancer for reference, but may be disoriented in world space.

Experiments Comparison

Neuroglancer	ITK-VTK-Viewer	Notes
link	link	Comparison of results that maximize dice overlap for individual labels (as of 06.05.23 experiments): 1. CCF label baseline 2. ANTs L4 Ex_561_Em_593 registration (aco: 0.223) 3. ITKElastix L4 Ex_488_Em_525 registration (MH: 0.691, fr: 0.473) 4. ITKElastix L4 Ex_561_Em_593 registration (VIIn: 0.255)

References

CCF v3.1 "Average Template" Intensity Image	Source	Neuroglancer	ITK-VTK-Viewer
CCF v3.1 atlas (compacted)	Source	Neuroglancer	ITK-VTK-Viewer
SmartSPIM Samples Origin	s3://aind-open-data/
Label Overlap Agreement Scores	Spreadsheet

Additional Notes

• Input source data is obtained as follows:
  1. Data is acquired via the SmartSPIM light sheet microscope as TIFF stacks. Acquisition metadata is composed in an `acquisition.json` file.
  2. Data is stitched via TeraStitcher, converted to OME-Zarr format, and hosted on AWS S3.
• Unless specified otherwise, the following defaults are employed:
  • SmartSPIM intensity images are selected at Level 4 in scale space, meaning each axis has been 2x downsampled four times. The resulting image is (1 / 2 ^ 4) the size of the acquisition volume along each axis or (1 / 2 ^ (3 * 4)) the size of the acquisition volume in total voxel volume.
  • Common Coordinate Framework v3.1 images are selected with 25um spacing, nonunit direction, and unit origin.
  • The registration pipeline executes as follows:
    1. SmartSPIM intensity image is overlaid on CCF image via resampling or origin update in ITK
    2. SmartSPIM intensity image is mapped to CCF image space via ITKElastix optimization of several transform stages:
       1. Rigid (rotation and translation only)
       2. Affine (rotation, translation, scale, skew)
       3. B-Spline deformations with control points at 500 micron spacings
• SmartSPIM mesh geometries are generated for surface rendering in ITK-VTK-Viewer via the following process:
  1. Registration yields a transform from SmartSPIM to CCF space.
  2. We apply the registration transform to resample a composited expert label image from SmartSPIM to CCF space.
  3. To get a binary mesh without scalar data, we convert the new label image to a binary image and apply ITK::BinaryMask3DMeshSource to obtain a mesh geometry.
  4. To get surface metrics, we do the following for each label value:
     1. Get a binary image for the given label value from the CCF label atlas and use ITK::ApproximateSignedDistanceMapImageFilter to create a corresponding signed distance image.
     2. Create a binary image for the given label value from the resampled SmartSPIM label image and apply ITK::BinaryMask3DMeshSource to obtain a mesh geometry for the label region.
     3. Probe the CCF label signed distance image with the transformed SmartSPIM label mesh by visiting each mesh vertex and sampling the distance value at that point in physical space.
     4. Attach the probed distance value at each vertex to the label mesh as scalar data and save to disk.
     5. Evalute registration alignment by considering the distribution of signed distances over the 3D label mesh surface with 1D statistics.