bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022.

The copyright holder for this preprint

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

R EAL -T IME S PATIAL R EGISTRATION FOR 3D H UMAN ATLAS

Lu Chen Dejun Teng

Department of Computer Science School of Computer Science and Technology
Stony Brook University Shandong University, China

Tian Zhu Jun Kong

Department of Applied Mathematics and Statistics Department of Mathematics
Stony Brook University Georgia State University

Bruce W. Herr II Andreas Bueckle

Department of Intelligent Systems Engineering, Department of Intelligent Systems Engineering,
Luddy School of Informatics, Luddy School of Informatics,
Computing, and Engineering, Computing, and Engineering,
Indiana University Bloomington Indiana University Bloomington

Katy Borner Fusheng Wang

Department of Intelligent Systems Engineering, Department of Computer Science
Luddy School of Informatics, Stony Brook University
Computing, and Engineering,
Indiana University Bloomington

A BSTRACT
The human body is made up of about 37 trillion cells (adults). Each cell has its own unique role
and is affected by its neighboring cells and environment. The NIH Human BioMolecular Atlas
Program (HuBMAP) aims at developing a 3D atlas of human body consisting of organs, vessels,
tissues to singe cells with all 3D spatially registered in a single 3D human atlas using tissues obtained
from normal individuals across a wide range of ages. A critical step of building the atlas is to
register 3D tissue blocks in real-time to the right location of a human organ, which itself consists
of complex 3D substructures. The complexity of the 3D organ model, e.g., 35 meshes for a typical
kidney, poses a significant computational challenge for the registration. In this paper, we propose a
comprehensive framework TICKET (TIssue bloCK rEgisTration) for tissue block registration for 3D
human atlas, including (1) 3D mesh pre-processing, (2) spatial queries on intersection relationship
and (3) intersection volume computation between organs and tissue blocks. To minimize search space
and computation cost, we develop multi-level indexing on both the anatomical structure level and
mesh level, and utilize OpenMP for parallel computing. Considering cuboid based shape of the tissue
block, we propose an efficient voxelization-based method to estimate the intersection volume. Our
experiments demonstrate that the proposed framework is practical and efficient. TICKET is now
integrated into the HuBMAP CCF registration portal [1].

Keywords 3D Spatial Queries · Spatial Indexing · Parallel Computing · Human Atlas

1 INTRODUCTION
With the significant development of tissue imaging technologies[2] [3] [4] [5], enormous amounts of information about
our human body are now produced at both cellular and subcellar levels. The Human BioMolecular Atlas Program
(HuBMAP) [6] of National Institutes of Health is a national initiative to develop an open, global framework for
comprehensively mapping of the human body at cellular resolution, in particular, generating foundational 3D tissue
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

atlases of human body, to better understand how the healthy human body works and what changes during aging or
disease. A Human Body Common Coordinate Framework [7], encompassing 3D organization of whole organs and
thousands of anatomical structures, are being built by HuBMAP, where the organs and the anatomical structures are
represented in mesh based 3D models [8].

Figure 1: Tissue block registration. The hilum of kidney, renal column and renal pyramid are anatomical structures
inside the kidney, which may collide with the tissue block.

By collecting 3D tissue blocks from human body, a vast amount of spatial information (cells and biomarkers) will be
derived from various tissue imaging technologies on the tissues (2D and 3D). An essential component of HuBMAP
is tissue block registration [9] (Figure 1), a critical step to enable organ experts to register tissue data to the atlas. It
will also help researchers to explore tissue data in a spatially and semantically explicit manner. In order to explore and
analyze the inner structures inside the tissue block, 3D computational geometry and spatial data management plays a
major role for exploring spatial relationships, e.g. anatomical structures that are colliding with the tissue block, as well
as quantitative measurements, e.g. volume of each anatomical structure inside the tissue block. However, there are
several challenges we need to address:
Complex 3D Structures: 3D anatomical structures have diverse shapes. Some are in regular shapes while others
may have complex structures such as vasculatures and skins as Figure 2 shows. Although minimal bounding boxes
(MBBs) is one of practical approaches to simplify complicated models, MBBs is not effective to represent such complex
3D objects for spatial queries. A new indexing method for primitives of 3D models is necessary to reduce the query
complexity. Besides, the hierarchical organization of human body demands a multi-level indexing approach.
Varying Mesh Quality of 3D models: 3D models of organs come from different sources such as Allen Brain Atlas
[10], the Human Tumor Atlas [11], HuBMAP [6], the Kidney Precision Medicine Project [12] and so on. The mesh
quality varies due to different purposes of the projects and varying requirements. For example, visualization oriented
3D organ models are targeted to visualization only, and may have limitations for spatial processing and queries, for
instance, existence of non-manifold edges [13] and holes, even though smoothness and other post-processing techniques

2
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Figure 2: 3D models of vasculatures

will make the models "look" smooth and perfect. For scientific computing purpose, 3D models for anatomic structures
must be spatially compliant to ensure accurate calculation of spatial relationships and measurements. However, it is
inevitable to lose vertices or edges due to the occlusion, reflectance or raw data preprocessing, rendering holes in the
surface and non-manifold meshes. For example, the mesh in Figure 3a is divided into two parts, i.e., the renal pyramid
and the renal papilla, according to biological functionalities and structures. The holes are automatically generated since
the surface is not closed any more due to the division of the mesh in Figure 3b. Therefore, a hole filling algorithm will
be applied to the 3D organ models as Figure 3 shows in advance to guarantee the precision of tissue block registration.
High Computation Complexity: 3D spatial queries involve computationally intensive geometric operations for
quantitative measurements and identifications of spatial relationships. While a traditional minimal bounding box (MBB)
approach can filter out the candidates dissatisfying the constraints of spatial queries in the filter stage, the refinement
stage still cost too much computation when examining the candidates with respect to their exact geometry to generate
the exact solutions [14] [15]. Due to the complex structures of 3D objects, the refinement stage could dominate the
spatial query cost. Moreover, queries which require quantitative computations such as intersection volume of objects,
have much higher complexity than normal query types.
The above challenges demand a practical and efficient framework that can normalize meshes, exploit indexing techniques
suitable for complicated objects, reduce computational complexity, and mitigate potential I/O cost to tackle the task of
3D tissue block registration.
Our main contributions are summarized as follows:

• We propose a comprehensive framework TICKET for tissue block registration including three steps: (1) Mesh
pre-processing such as remeshing non-manifold meshes and filling the holes to produce watertight meshes,
(2) identify the anatomical structures that are collided with the tissue block, (3) compute the volume of each
anatomical structure inside the tissue block, if intersected;
• We introduce multi-level spatial indexing at the anatomical structure level and mesh level, which improve
query performance compared with the traditional filter-refine paradigm [16]. We take an in-memory approach

3
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

(b) Separate structures before hole fill-

(a) Origin models ing (c) After filling holes
Figure 3: Example of mesh-preprocessing: hole filling

for data management and query to mitigate the I/O cost and reduce latency. Besides, we provide a spatial
query http service for tissue block registration, which can be invoked on-demand when users upload the tissue
blocks.
• We develop an efficient voxelization-based method to estimate the intersection volume between the tissue
block and the organ considering the regular cuboid shape of the tissue block. We utilize parallel programming
to accelerate the volume computation.

The rest of the paper is organized as follows. We first introduce the background in Section 2. The framework of
tissue block registration is demonstrated in Section 3. Next, we introduce multi-level spatial indexing in Section 4.
Section 5 presents the voxelization-based method to estimate the intersection volume as well as the parallel computing
scheme to accelerate the volume computation. The performance of the framework is evaluated in Section 6 followed by
conclusion.

2 BACKGROUND
2.1 Mesh Pre-processing

Mesh pre-processing is a fundamental step of 3D data acquisition. Poor quality meshes may cause errors in spatial
queries or lead inaccuracy intersection volume computation. There are some typical mesh pre-processing techniques
such as mesh self-intersection checking, remeshing and hole filling. Self-intersection checking is intuitive and easy to
implement while hole filling is much more subjective and semantic.
Due to the differences of data representation, current hole filling algorithms [17] can be divided into two categories:
point cloud-based methods and mesh-based methods. Point cloud-based methods were developed in earlier years, while
mesh-based methods gain more attention and play an important role with the rapid development of computer graphics.
Typical workflow of hole filling algorithms can be summarized in two steps: hole boundaries should be first detected
and then holes are filled by generating new meshes. In particular, mesh-based hole filling algorithms are divided into
two categories: volume-based algorithms and surface-based algorithms according to the data used in the phase of filling
holes. Normally, the simple triangulation algorithms work well to triangulate the polygonal hole regions, especially
for the flat and disk-like holes without intense changes to the hole topology and geometry. For large holes that are too
complex to fill using triangulation, various hole filling algorithms are proposed [18] [19] [20] [21]. In this paper, we
adopt the hole filling algorithm [20] to fill all the problematic holes.

2.2 Spatial indexing

Each human organ consists of a set of anatomic structures (AS). For example, a kidney has renal pyramids, renal
papilla, outer cortex, a total of 35 ASs. A brain contains more than 200 anatomical structures. The organ consisting

4
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

of multiple ASs is represented by a 3D model, typically using a mesh based representation, using formats such as
Graphics Language Transmission Format (e.g., GLB) [22] or Object File Format (OFF) [23] . A typical kidney may
be represented by 35 mesh surfaces. To register a human tissue block to an organ, proper indexing is needed for such
complex 3D models to minimize the search space and geometric computations [24]. The key principle of the spatial
indexing is to divide the search space into regions. There are two common types of space decomposition:

• Regular decomposition, also known as space driven index, divides the space in a regular manner independent
of the distribution of the objects (i.e., indirectly related to the objects in the space). Objects are mapped to the
cells according to some geometric criteria. The most common methods are Quad Tree [25] and cell-based
Hashing [26].

• Object Directed Decomposition is also known as data driven index. The division of the index space is
determined directly by objects. The space is divided by one of or a combination of the following properties:
coordinates of data points, the extents, the bounding box or the spatial objects. The most common examples
are K-D tree [27], BSP-tree [28] and R-tree [29]. Although K-D tree and BSP-tree are easier to implement, the
deletion and insertion of them are non-trivial while R-tree is designed for dynamic changing data and support
deletion and insertion.

Traditional filter-refine paradigm with single spatial index built on MBBs of objects is not suitable for complex objects
because the geometric computation during the refinement stage will dominate the spatial query cost. Therefore, building
index on primitives (vertices, edges and faces of meshes) can greatly reduce the complexity.

3 OVERVIEW

3.1 Framework of the Tissue Block Registration

We aim to develop a comprehensive framework for tissue block registration, which supports mesh pre-processing,
spatial queries on intersection relationship and intersection volume computation. We present the overall framework for
the tissue block registration in Figure 4. Initial 3D mesh data is stored in disks in object file format (OFF). We first check
whether the meshes are 2-manifold or not. If not, the medical illustrators of HuBMAP will remesh the non-manifold
meshes manually. The meshes with holes are then pre-processed by the hole filling algorithm [20]. After the hole filling,
the generated patches can be refined and faired [30]. Each individual mesh of anatomical structures will be loaded and
stored in memory. We also create a hash map to store the meshes, where the key is the organ name and the value is a list
of the meshes of the anatomical structures of the organ for an easy search. Next, an anatomical structure level index
will be generated according to the MBBs of the anatomical structures for each organ, and a mesh level index is built for
the mesh of every anatomical structure. In addition, different from traditional intersection volume computation based
on Boolean Operations [31] on meshes, we propose a voxelization-based method to estimate the intersection volume
between the tissue block and the anatomical structure. To accelerate the intersection volume computation, we exploit
parallel programming framework OpenMP to check whether each voxel of the tissue block is inside the polygon mesh
of the anatomical structure in parallel.

3.2 On-demand Spatial Query Service

We provide an on-demand spatial query http service for the ease of use. Users can send a http request with the location
information of the tissue block including translation, rotation and scaling parameters. The server will process the request
accordingly, with the following procedure (the workflow of the tissue registration is also summarized in Algorithm
1): first transform the rotation angle-based representation of the tissue block to a mesh-based representation. Next,
the anatomical structures whose MBBs are not overlapped with the MBBs of the tissue block are filtered out by the
anatomical structure level index. The exact intersected anatomical structures with the tissue block are reported by
using mesh level spatial index. If intersected, we will compute the intersection volume by proposed voxelization-based
intersection volume computation. Otherwise, we will return 0 immediately because they have no intersection.

5
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Figure 4: Framework of 3D tissue block registration

Algorithm 1 Workflow of tissue block registration

for each query for tissue block t do:
transform from angle-based to mesh-based representation of t;
filter with anatomical structure level index;
query exact intersection with anatomical structures using mesh level index.
if t is intersected with certain anatomical structure then
compute the intersection volume using OpenMP specified with resolution
end if
end for

4 MULTI-LEVEL SPATIAL INDEXING

4.1 Anatomical Structure Level Indexing

Figure 1 demonstrates an organ with internal anatomical structures. Different anatomical structures contain different
tissue types and cells, with different functionalities. To fast locate the right positions in the organ for the tissue block

6
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

(a) Anatomical structure level indexing (b) Mesh level indexing

Figure 5: Multi-level spatial indexing

registration, we build an anatomical structure level index according to the MBBs of the anatomical structures. An
R-tree is constructed on the MBBs of the anatomical structures to achieve the dynamic insertion and deletion of the
anatomical structures as Figure 5a presents. If cell models are provided in the future, we can further build spatial index
on cells [32] based on cell types and spatial distirbution patterns. Therefore, from macro level to micro level, we build
multi-level spatial indexes according to the hierarchical organisation of our human body.

4.2 Mesh Level Indexing

For complex structures, traditional filter-refine paradigm suffers from geometric computation in the refinement stage,
because the simplified MBBs cannot filter out most unqualified candidates, such as a tissue block outside but near the
vessel. Therefore, the mesh-level index built on the primitives (faces) of the complicated structure is necessary as Figure
5b. As the resolution of 3D models becomes higher, a mesh-level spatial index can greatly reduce the complexity of
spatial queries from O(N ) to O(log N ) where N is the number of primitives. Therefore, we build a BSP-tree, also
called AABB-tree on the primitives for every static mesh of the anatomical structures.

5 VOXELIZATION-BASED INTERSECTION VOLUME COMPUTATION

The traditional way to compute the intersection volume between two polygon meshes is through Mesh Boolean
Operations [31], which consists of two steps: the first step is to compute the mesh intersection of the two polygon
meshes. The volume of the intersection mesh is computed next. However, the complexity of Boolean Operations
is very high. One of the improvements of Boolean Operations is to use Constructive Solid Geometry (CSG) [33].
Triangle-triangle intersections need to be computed in advance in CSG, which is very time-consuming.
In this paper, we propose an efficient voxelization-based method supporting different resolution of voxelization to
estimate the intersection volume. Furthermore, we use the parallel programming framework OpenMP to make the full
use of the multi-core processor, which is widely used in modern computers.

7
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

5.1 Tissue Representation Transformation

Tissue blocks provided by HuBMAP are described with size, rotation, translation, and scaling parameters, which are
not the standard mesh representation by vertices, edges and faces. Therefore, we need to recover the mesh from the
parametric representation.
Assuming a tissue block is described as:
{dx , dy , dz , rx , ry , rz , tx , ty , tz }
where dx , dy , dz are the lengths of the tissue block in x,y,z dimensions respectively, rx , ry , rz are the rotation angles
and tx , ty , tz are the translations. The initial center of the tissue block is the origin and the default rotation order is xyz
as specified by HuBMAP. Since it is the Euler angle system rather than the fixed angle system that is adopted during
rotation by HuBMAP, the rotation order should be adjusted to zyx instead. Hence, the rotation matrix with the rotation
angles rx , ry , rz can be written as:
R = Rx Ry Rz (1)
where
1 0 0
!
Rx = 0 cos(rx ) sin(rx ) . (2)
0 − sin(rx ) cos(rx )

cos(ry ) 0 sin(ry )
!
Ry = 0 1 0 . (3)
− sin(ry ) 0 cos(ry )

cos(rz ) sin(rz ) 0
!
Rz = − sin(rz ) cos(rz ) 0 . (4)
0 0 1

Any point p = (px , py , pz )T inside or on the boundary of the initial tissue block, i.e.,
dx dx
px ∈ [− , ], (5)
2 2
dy dy
py ∈ [− , ], (6)
2 2
dz dz
pz ∈ [− , ] (7)
2 2
can be transformed by applying the affine transformation: p′ = Rp + T where T = (tx , ty , tz )T , R = Rx Ry Rz . Note
that the edges and faces relations, the size of the tissue block are not changed due to the affine invariance. Through the
above affine transformation, it becomes fairly simple to construct the standard mesh given the initial coordinates of
vertices of the tissue block, the rotation matrix and the translation vector.

5.2 Voxelization

A voxel represents a regular grid in 3D space similar to a pixel in a 2D image. Traditionally, voxelization is applied to
3D models in fixed axes directions. Due to regular cuboid shape of the tissue block, we always voxelize the oriented
tissue block in its coordinate system along three different diagonal directions. For each voxel, we test whether the center
of the voxel lies inside a certain anatomical structure. The estimated intersection volume can be calculated through the
volume sum of the voxels whose center are inside the anatomical structure. Via this approach, the complicated volume
computation problem is reduced to a point location test, which saves the computational cost greatly.
Point location test is a very classic problem in computational geometry and the current implementation is based on the
number of triangles intersected by a ray with the query point as the source. If the number of triangles intersected is
odd, the point is inside the polyhedron. Besides, the AABB tree constructed on each mesh can accelerate the point
location queries because the ray casting is an intersection problem essentially. The reuse of AABB trees significantly
reduces the overhead. To make the full advantage of multi-core processors, we implement the voxelization-based
volume computation in parallel using OpenMP. In addition, to avoid memory overflow, we process the voxelization and
point location test on the fly rather than store all the voxels in the memory.

8
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Although the complexity of point location query is O(log N ), a higher resolution voxelization requires more computa-
tional resources. We define the resolution coefficient k as the number of voxels in one unit in one dimension, which
indicates a tissue block is divided into kdx × kdy × kdz voxels. We assume that the lowest resolution is k = 1 in
which the tissue block is divided into dx × dy × dz voxels. Since dx , dy , dz are integers defined by HuBMAP, we can
ignore the "partial" voxel problem. The voxelization-based intersection volume computation algorithm is summarized
in Algorithm 2.

Algorithm 2 Parallel computing for voxelization-based volume computation

Input: tissue block t, anatomical structure S
Output: the intersection volume between the tissue block and the anatomical structure

Initialize: sum ← 0
# pragma omp parallel for collapse(3) reduction(+:sum)
for i from 1 to kdx do
for j from 1 to kdy do
for k from 1 to kdz do
compute the center c of the voxel located in (i, j, k)
if c is inside S then
sum + = 1
end if
end for
end for
end for

return dx dy dz sum/k 3

6 EXPERIMENTAL RESULTS

All the tests are conducted on a compute node with a 24-core CPU (AMD Ryzen Threadripper 3960 at 3.8GHz). The
node comes with 128GB memory(DDR4 3200) and 2TB SSD (NVMe M.2 PCI-Express 3.0). The OS is Ubuntu 20.04.4
with 5.4.0-124 kernel.
We conduct these tests over the official HuBMAP 3D reference organ models [8] and tissue block dataset [34], which
contains 68 tissue blocks registered in kidneys.
First, we compare the loading time of the models and the time of building multi-level spatial indexes for five organ
models, namely Kidney, Allen Brain, Lung, Heart and Small Intestine. From Figure 6, we can see that large models
such as Allen Brain which contains 282 anatomical structures with total of 324,833 vertices and 656,268 faces, cost
more time to load and build spatial indexes while small models such as Small Intestine which contains 9 anatomical
structures with total of 16998 vertices and 33524 faces, cost less time to load and build spatial indexes. Besides, building
spatial indexes costs only half of the time of loading 3D models to the memory.
Next, we evaluate four different methods (brute force, only building index on MBBs of anatomical structures, only
building spatial index on primitives for each mesh, and building multi-level spatial index both on MBBs of anatomical
structures and primitives) of the spatial query on intersection. Figure 7 shows the total query time of all 68 tissue blocks
of the spatial query on intersection under four different methods. It is obvious that the spatial index on MBBs of ASs,
which follows the traditional filter-refine paradigm, can greatly reduce the query time compared with the brute force
method. Another important fact is that the spatial index built on primitives can further reduce the query time. We also
test the spatial intersection query using PostGIS, which is extremely slow. Therefore, the multi-level spatial indexing
can significantly improve the performance compared with all other methods.
Moreover, we investigate the impact of parallel computing using OpenMP on voxelization-based intersection volume
computation method. From Figure 8 we can conclude that the overhead of cache and synchronization of parallel
computing results in no improvement in terms of the query performance when the resolution is low. The number of
parallel units increases when the resolution increases. Therefore, the query can make full advantage of multi-core
processor, which greatly reduce the query time compared with serial computation.

9
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

0.6 Building spatial index time

Loading time

0.5

0.4
Time(s)

0.3

0.2

0.1

0.0

Kidney Allen Brain Lung Heart Small Intestine

Organs

Figure 6: Loading time and building indexes time of different organs

Non-parallel

6 Parallel

5
Query time(s)

0 1 2 3 4 5 6 7 8 9 10

Resolution

Figure 8: Comparisons of query time between parallel and non-parallel

Finally, we demonstrate the accuracy of the proposed voxelization-based intersection volume computation method. We
use the relative error in most cases where the interaction volume exceeds a threshold. Meanwhile, we use the absolute
error instead when the intersection volume is around 0. As Figure 9 shows, the error rate decreases as the resolution
increases. The error rate is below 0.01 when the resolution is greater than or equal to 8. The outlier with k = 2 may be

10
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

1.4

1.24172
1.2

1.0

Query time (s)

0.8

0.6

0.4

0.2 0.17873

0.00588 0.00483
0.0

Brute force Spatial index Spatail Index Multi-level

on MBBs of AS on primitives spatial indexing

Method

Figure 7: Comparisons of different indexing methods

caused by the irregular shape of the 3D models, which seems to have low error rate by accident. In conclusion, the
proposed intersection volume computation method achieves high performance in terms of both efficiency and accuracy.

0.14

0.12 error rate

0.10
error rate

0.08

0.06

0.04

0.02

0.00

0 2 4 6 8 10

resolution

Figure 9: Error rate of different resolutions

11
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

7 CONCLUSION
We propose a comprehensive framework for 3D tissue block registration for human atlas. In particular, we build
multi-level indexing both on MBBs of anatomical structures and primitives of meshes. In addition, we present a
voxelization-based method for intersection volume computation. Experimental results show that the framework achieves
high accuracy in real-time spatial queries. The 3D registration framework is implemented as a service for HuBMAP,
and will be deployed in Amazon AWS platform for production use in short term future.

8 Acknowledgments
This research is supported in part by grants from National Institutes of Health 1OT2OD033756-01 and 3OT2OD026671-
01S4.

References
[1] Bruce W. Herr II. HuBMAP CCF REGISTRATION. 2022. URL: https://hubmapconsortium.github.io/
ccf-ui/rui/ (visited on 08/30/2022).
[2] GTEx Consortium Lead analysts: Aguet François 1 Brown Andrew A. 2 3 4 Castel Stephane E. 5 6 Davis Joe
R. 7 8 He Yuan 9 Jo Brian 10 Mohammadi Pejman 5 6 Park YoSon 11 Parsana Princy 12 Segrè Ayellet V. 1
Strober Benjamin J. 9 Zappala Zachary 7 8 et al. “Genetic effects on gene expression across human tissues”. In:
Nature 550.7675 (2017), pp. 204–213.
[3] Heeva Baharlou et al. “Mass cytometry imaging for the study of human diseases—applications and data analysis
strategies”. In: Frontiers in immunology 10 (2019), p. 2657.
[4] Yury Goltsev et al. “Deep profiling of mouse splenic architecture with CODEX multiplexed imaging”. In: Cell
174.4 (2018), pp. 968–981.
[5] Laura Kuett et al. “Three-dimensional imaging mass cytometry for highly multiplexed molecular and cellular
mapping of tissues and the tumor microenvironment”. In: Nature Cancer 3.1 (2022), pp. 122–133.
[6] Caltech-UW TMC Cai Long lcai@ caltech. edu 21 b et al. “The human body at cellular resolution: the NIH
Human Biomolecular Atlas Program”. In: Nature 574.7777 (2019), pp. 187–192.
[7] Katy Börner et al. “Construction and Usage of a Human Body Common Coordinate Framework Comprising
Clinical, Semantic, and Spatial Ontologies”. In: arXiv preprint arXiv:2007.14474 (2020).
[8] Bruce W. Herr II. CCF 3D Reference Object Library. 2022. URL: https://hubmapconsortium.github.io/
ccf/pages/ccf-3d-reference-library.html (visited on 08/30/2022).
[9] Katy Borner et al. “Tissue Registration and Exploration User Interfaces in support of a Human Reference Atlas”.
In: bioRxiv (2021).
[10] Song-Lin Ding et al. “Comprehensive cellular-resolution atlas of the adult human brain”. In: Journal of Compar-
ative Neurology 524.16 (2016), pp. 3127–3481.
[11] Sudhir Srivastava et al. “The making of a PreCancer Atlas: promises, challenges, and opportunities”. In: Trends
in Cancer 4.8 (2018), pp. 523–536.
[12] Tarek M El-Achkar et al. “A multimodal and integrated approach to interrogate human kidney biopsies with
rigor and reproducibility: guidelines from the Kidney Precision Medicine Project”. In: Physiological genomics
53.1 (2021), pp. 1–11.
[13] Michael Zilske, Hans Lamecker, and Stefan Zachow. “Adaptive remeshing of non-manifold surfaces”. In: (2007).
[14] Ablimit Aji et al. “Hadoop gis: a high performance spatial data warehousing system over mapreduce”. In:
Proceedings of the VLDB Endowment 6.11 (2013), pp. 1009–1020.
[15] Dejun Teng et al. “3DPro: Querying Complex Three-Dimensional Data with Progressive Compression and
Refinement.” In: EDBT. 2022, pp. 2–104.
[16] Jordan Wood. “Filter and Refine Strategy”. In: Encyclopedia of GIS. Ed. by Shashi Shekhar and Hui Xiong.
Boston, MA: Springer US, 2008, pp. 320–320. ISBN: 978-0-387-35973-1. DOI: 10.1007/978-0-387-35973-
1_415. URL: https://doi.org/10.1007/978-0-387-35973-1_415.
[17] Xiaoyuan Guo, Jun Xiao, and Ying Wang. “A survey on algorithms of hole filling in 3D surface reconstruction”.
In: The Visual Computer 34.1 (2018), pp. 93–103.
[18] James Davis et al. “Filling holes in complex surfaces using volumetric diffusion”. In: Proceedings. First
international symposium on 3d data processing visualization and transmission. IEEE. 2002, pp. 428–441.

12
bioRxiv preprint doi: https://doi.org/10.1101/2022.09.27.509335; this version posted September 28, 2022. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

[19] Tao Ju. “Robust repair of polygonal models”. In: ACM Transactions on Graphics (TOG) 23.3 (2004), pp. 888–
895.
[20] Peter Liepa. “Filling holes in meshes”. In: Proceedings of the 2003 Eurographics/ACM SIGGRAPH symposium
on Geometry processing. 2003, pp. 200–205.
[21] Marco Attene. “A lightweight approach to repairing digitized polygon meshes”. In: The visual computer 26.11
(2010), pp. 1393–1406.
[22] The Khronos Group. glTF. URL: https://www.khronos.org/gltf/.
[23] Randi J Rost. “Off-a 3d object file format”. In: Digital Equipment Corporation Technical Report (1989).
[24] Shaik Abdul Nusrath Begum and KP Supreethi. “A survey on spatial indexing”. In: Journal of Web Development
and Web Designing 3.1 (2018).
[25] Hanan Samet. “The quadtree and related hierarchical data structures”. In: ACM Computing Surveys (CSUR) 16.2
(1984), pp. 187–260.
[26] Ronald Fagin et al. “Extendible hashing—a fast access method for dynamic files”. In: ACM Transactions on
Database Systems (TODS) 4.3 (1979), pp. 315–344.
[27] John T Robinson. “The KDB-tree: a search structure for large multidimensional dynamic indexes”. In: Proceed-
ings of the 1981 ACM SIGMOD international conference on Management of data. 1981, pp. 10–18.
[28] William C Thibault and Bruce F Naylor. “Set operations on polyhedra using binary space partitioning trees”. In:
Proceedings of the 14th annual conference on Computer graphics and interactive techniques. 1987, pp. 153–162.
[29] Antonin Guttman. “R-trees: A dynamic index structure for spatial searching”. In: Proceedings of the 1984 ACM
SIGMOD international conference on Management of data. 1984, pp. 47–57.
[30] Mario Botsch and Olga Sorkine. “On linear variational surface deformation methods”. In: IEEE transactions on
visualization and computer graphics 14.1 (2007), pp. 213–230.
[31] Aristides AG Requicha and Herbert B Voelcker. “Boolean operations in solid modeling: Boundary evaluation
and merging algorithms”. In: Proceedings of the IEEE 73.1 (1985), pp. 30–44.
[32] Yanhui Liang et al. “ISPEED: an efficient in-memory based spatial query system for large-scale 3D data with
complex structures”. In: Proceedings of the 25th ACM SIGSPATIAL International Conference on Advances in
Geographic Information Systems. 2017, pp. 1–10.
[33] Christoph Martin Hoffmann. “Geometric and solid modeling”. In: (1989).
[34] Bruce W. Herr II. HuBMAP CCF EXPLORATION. 2022. URL: https://portal.hubmapconsortium.org/
ccf-eui (visited on 08/30/2022).

13