In collaboration with the BRAIN Inititative Cell Census Network (BICCN), the Brain Imaging Library (BIL), the Penn Image Computing and Science Lab (PICSL), we are running the first Allen Institute for Brain Science CCF registration competition!

A brief description of the competition motivation, format, and next steps is below Motivation

The Allen Mouse Common Coordinate Framework (CCF) defines anatomical locations within an annotated population averaged mouse brain. This resource is already valuable to the field, providing a 3D reference framework for the mouse brain. Enhanced impact is possible if new 3D datasets can be routinely aligned into the CCF space. This enables comparison of results across experiments with a common and objective reference for mouse brain anatomy. Large-scale resources can be built systematically by leveraging automated alignment into a common spatial framework.

Achieving this vision requires robust alignment of new image data into the CCF space, regardless of acquisition modality. The field has produced many options for this task, but unbiased comparison of these techniques does not exist. To benchmark various approaches, and to encourage researchers to share available resources, we propose the Allen Institute for Brain Science organize a competition between CCF alignment tools sourced from the global neuroscience and image processing community.

The Allen Institute for Brain Science will provide the following image datasets:

The Allen Institute uses fluorescence micro-optical sectioning tomography (fMOST) to collect three-dimensional whole brain fluorescence microscopy volumes at sub-micron resolution. Embedded mouse brain tissue is repeatedly sectioned at 1 µm depth while images are acquired at the tissue block face (0.35 µm x 0.35 µm x 1.0 µm native voxel size). The resulting datasets contain endogenous signal from the tissue as well as sparse cells expressing fluorescent proteins throughout their entire cell volume. Expert analysists reconstruct 3D computer models of these labeled cells from the fMOST datasets. Alignment of cells from many different brains into the CCF reference space permits comparisons of neuronal anatomy, projection patterns, and diversity of connection types within and across regions of origin for the labeled cells. Example results from this pipeline can be seen in the preprint here.

Current practice is to first generate a multi-resolution data pyramid from the native fMOST dataset. A resolution approximating that of the CCF reference volume is used for alignment. The subsequent steps remove imaging artifacts and the align using pairs of points chosen by an expert annotator in the experimental and reference space. This operation works sufficiently well for a small number of specimens, but scalability is limited by the need for expert annotation of each specimen. Automating this step would be extremely valuable for this pipeline and for others using fluorescence microscopy to study the mouse brain.

Current challenges for automating fMOST - CCF alignment include:

• Striping artifacts from the fMOST sectioning procedure
• Tissue damage from embedding and sectioning, leading to separations between adjacent anatomical regions
• Enlarged ventricles in the fMOST datasets due to the sample preparation procedure
• Anatomical heterogeneity between specimens at the scale of the cell reconstruction features

(To-do: add details - Julie)

The datasets can be access through BIL:

(To-do: BIL to setup site in BIL ) (To-do: Rusty to upload data) (To-do: instruction on how to download - BIL)

The data is in the following format:

(To-do: PICSL team to define format and convert the data)

For each dataset, a set of landmarks covering the whole extent of the brain will be indentified by expert anatomists. Type of landmarks include:

• Pairs of corresponding points between the image dataset and the CCF
• Points that are within in specific structures in the CCF
• Points that are on the surface between two structures in the CCF

(To-do: Penn team to define the landmark format) (To-do: Lydia/Julie to figure out fMOST data) (To-do: Rusty/Josh to format the OPT data)

Example landmarks will be provided to participants as reference. The rest are withheld and used for scoring the registration quality.

To submit a result, a participant will

• start a pull request
• create a new submission page using the template
• provide information about their team and methods
• provide links to the registration code (source code or product page)
• For each imageset:
  • Scripts and parameters that was used to produce
  • Upload the deformation field that maps a image point to the CCF in the format described below
  • Add location of the deformation field
• Submit pull request

Once we received the pull request, we will compute registration score and post in on the scoreboard.

(To-do: Penn team to define the deformation field format)

(To-do: Penn team to code registration scoring method ) (To-do: Penn / BIL team to setup how to run the scoring at BIL) (To-do: Rusty to update the score board)

Competition start is tentatively scheduled for summer 2020. We will release details regarding datasets, scoring, and entrance details as they become available.

We will have a workshop at the Neuroinformatics 2020 Congress in Seattle in Aug 17-18, 2020