Representation of Structural Relationships in the
Foundational Model of Anatomy
José L.V. Mejino, Jr. M.D.,1 Natalya F. Noy Ph.D., 2 Mark Musen M.D., Ph.D., 2 and Cornelius Rosse, M.D., D.Sc. 1
Structural Informatics Group, Department of Biological Structure,
University of Washington, Seattle, WA1 and
Section on Medical Informatics, Stanford University School of Medicine,
Stanford, CA
2
ABSTRACT
Previous
attempts at the symbolic representation of anatomical relationships have been
largely limited to partonomy. We propose an ontology of anatomical relationships
and illustrate the inheritance of structural attributes in the Digital Anatomist
Foundational Model. Our purpose is to generate a sharable resource that can
support inference about the structural organization of the body.
INTRODUCTION
The main objective of
the terminologies correlated by UMLS is to serve as repositories of terms that
can be reused with consistency by a variety of applications.1 In
general, current biomedical and educational applications are designed to present
hard-coded, didactic information, or they support low-level, look-up functions
with no, or at best limited, capabilities for inference. The semantic structure
of today's controlled medical terminologies (CMTs) seems adequate for the needs
of such contemporary applications. Next-generation applications, however, will
have to incorporate increasing levels of intelligence in order to meet the
demands of the evolving environment in education and the practice of the various
health professions. Such knowledge-based applications call for the
representation of much deeper and richer knowledge than that retrievable from
today's CMTs. Since these CMTs primarily target clinical medicine, they are
deficient in basic science concepts necessary to support reasoning. Moreover,
since relationships between concepts constitute an important dimension of
knowledge, next-generation knowledge sources must model comprehensively not only
the concepts but also the relationships that characterize a particular field of
basic science. Therefore, there is
a need to generate enabling knowledge sources at least in those domains that
generalize to diverse fields of education and clinical practice. Anatomy is such
a fundamental domain.
We
are developing the Foundational Model of anatomy (FM)2,3 as an
evolving resource for knowledge-based applications4. Our intent is
that the FM should furnish, at the highest level of granularity, not only
anatomical concepts but also the relationships that comprehensively describe the
structural organization of the body. Figure 1 illustrates that three of the four
components of the FM (described elsewhere in these proceedings5,6)
are, in fact, based on different
classes of relationships:
Fm
= (AO, ASA, ATA, Mk)
(1)
AO,
the Anatomy Ontology, is a type hierarchy based on the IS-A relationship; ASA, the anatomical structural abstraction, based
on 'structural relationship', is the subject of this report; ATA, the anatomical
transformation abstraction, is based on relationships that describe the
morphological transformation of anatomical entities during pre- and postnatal
development.
Our
reports in these proceedings,5,6 as well as elsewhere,2,3,7-10
are primarily concerned with the classification of physical anatomical entities
(material objects, spaces, surfaces, lines and points). In this communication
our objective is to illustrate the importance of anatomical relationships for
the symbolic modeling of structural knowledge, a dimension unique to anatomy
among the biomedical sciences.
ANATOMICAL STRUCTURAL ABSTRACTION
The
FM is being developed as an anatomical enhancement of UMLS. Its classes and
relationships extend the specificity of UMLS semantic types and relationships.
We have proposed a scheme for representing anatomical structural relationships
in terms of interacting networks:3,10
ASA
= (DO, Pn, Bn, SAn)
(2)
Where:
DO =
Dimensional ontology
Pn =
Part-of network
Bn =
Boundary network
SAn = Spatial Association
network
DO
is a type hierarchy of geometric objects and shapes, in terms of which the three
networks of ASA may be described at an abstract level. Of the ASA networks, SAn
itself consists of a number of subnets corresponding to the descendants of the
'Spatial association relationship' class shown in Figure 2. Since other reports
tend to deal with part-whole relationships11,12, in this
communication we illustrate spatial associations, which have received less
attention. The descendants of this relationship class correspond to a number of
axes or viewpoints in terms of which anatomical spatial associations, such as
location, may be conceptualized. We illustrate the symbolic modeling of
'Anatomical adjacency', which poses a particular challenge, since adjacency is
not simply a relationship between two objects; rather it has attributes, such as
directional vector and right or left laterality.
An
anatomical structure, such as the esophagus, or a part of it, inherits its shape
from the DO class 'conventional hollow cylinder'. This shape specifies the set
of adjacency relationships that is allowed for this shape class. Figure 3 shows
these relationships graphically in terms of a qualitative radial coordinate
system. In Figure 4 the qualitative coordinate system for cylinder are
superimposed and centered on the esophagus in a section of the male Visible
Human at the level of the eighth thoracic vertebra. In Figure 5 the adjacencies
of 'T8 part of the esophagus' are expressed symbolically in terms of these
qualitative coordinates. Although some of these adjacency relationships remain
constant, others change from one vertebral level to the next. The AO of the FM
represents each vertebral level of the esophagus as a discrete subzone, which
permits the symbolic modeling of the changing adjacency relationships of the
esophagus as it "passes" from the neck to the abdomen.
The
spatial knowledge captured by the adjacency relationships shown in Figure 5 is
of importance to a student dissecting the esophagus for the first time and also
to a surgeon planning to remove a lymph node adjacent to the esophagus through a
mediastinoscope. The FM can provide knowledge of adjacency relationships
appropriate for applications developed for each of these types of users.
Moreover, since we can represent inverse values for these relationships, and
make inferences based on their transitivity, the FM could support inference
required for answering user-generated spatial
queries at different levels of complexity.
Figures 4 and 5 invite comment about the relative usefulness of geometric and qualitative coordinates for representing such structural attributes as location and adjacency. The relationships expressed in terms of qualitative coordinates could be derived from the quantitative geometric matrix of the Visible Human data set. These geometric coordinates, however, would have to be expressed as qualitative coordinates in order to make them intelligible in anatomical discourse. Geometric coordinates are valid only for one instance, whereas anatomical qualitative coordinates describe relationships that hold true in all members of a species. Only those structures can be referenced by geometric coordinates that are visible with a particular imaging modality. Qualitative coordinates, on the other hand, can describe the relationship of invisible structures to visible ones, as illustrated in Figure 5 by the esophageal plexus, fibrous pericardium and mediastinal pleura; none of these structures can be identified in the image of the anatomical section. Moreover, inference required for reasoning about structural relationships within the body must make use of qualitative coordinates. Therefore, the symbolic representation of structural relationships in terms of qualitative coordinates is an important component of the FM.
UWDA
and FM. We began the development of
the University of Washington Digital Anatomist (UWDA) vocabulary
(the initial iteration of the
FM) as an anatomical
enhancement of UMLS13. Initially we were less concerned with the
variety of anatomical relationships than with the classification and
comprehensiveness of anatomical concepts. The authoring tool we developed was
designed to generate parallel hierarchies (directed acyclic graphs) which were
based on IS-A, PART-OF, BRANCH-OF and TRIBUTARY-OF
relationships. As we populated classes of 'Organ part' in the IS-A
hierarchy, for example, we also aligned the concepts along the transitive PART-OF
relationship in another hierarchy. However, such a link-centric view and
representation of anatomy proved to be inadequate once we began to appreciate
the complexity of relationships that were necessary for comprehensively
describing the anatomy of the body. The need for such a comprehensive, reusable
resource led to the Foundational Model, a conceptualization of the physical
organization (structure) of the human body.
Approximately
50,000 concepts in the AO of the FM are accessible through the UWDA vocabulary
of UMLS, providing a comprehensive controlled terminology for macroscopic
anatomy. Our current work entails the instantiation of the ASA networks of these
concepts. The association of such multi-dimensional relationships with
anatomical concepts calls for a node-centric view of anatomy, which is beyond
the capacity of the link-centric representation we implemented. The frame-based
knowledge acquisition system Protégé-200014 has the requisite
expressivity and scalability for comprehensively
modeling anatomical relationships encompassed by the ASA.
The same will be true for ATA relationships, once we begin the
implementation of developmental transformations.
Protégé-2000
correlates four ontologies within the FM: the large 'Anatomical entity' ontology
(the AO) and the smaller 'Dimensional
object', 'Physical state' and 'Anatomical entity metaclass' ontologies. The
first three ontologies provide the values for anatomical relationships, whereas
the metaclass ontology assures the inheritance of the attributes of concepts
represented in the AO. The 'Anatomical entity metaclass' ontology contains high
level templates for the classes in the AO, which instantiate the template. Each
template is a frame composed of a set of slots; each slot corresponds to a
defining or other attribute manifested by a particular AO class. We define a
hierarchy of templates and the attributes are inherited through the hierarchy.
For instance, each concept has a UWDA-ID number and a preferred name, and it may
have one or more synonyms. Therefore the template for 'Anatomical entity', which
is the root of the AO, includes slots for each of these attributes. The
templates of all descendants of the root inherit these slots. When a new concept
is entered in the AO, values must be assigned to each of these slots (Figure 5).
Figure
5 presents the frame for 'T8 part of esophagus', which is highlighted in the AO
(left pane). The right pane shows some of the values for the slots of the 'Zone
of esophagus template' pertinent to the ‘T8 part of esophagus’. Although
they would have different values, the same kinds of slots specify the anatomical
relationships of cervical or thoracic parts of the esophagus as those of the
parts that correspond to vertebral levels. The slots for the first two rows of
values are inherited from the template of the root. The slot 'has intrinsic 3D
shape' is first introduced in the 'Anatomical structure' template as a defining
attribute. This attribute distinguishes 'Anatomical structure' from 'Body
substance' (see Fig. 1.), as explained in a companion report in these
proceedings.5 The intrinsic shape slot is inherited by all templates
corresponding to descendants of 'Anatomical structure' in the AO, and its values
are provided by the DO.
The
'has orientation' template slot, however, has to be introduced in the template
of 'Physical anatomical entity', since not only anatomical structures but also
surfaces and lines (classified as non-material physical anatomical entities)
have orientation (for class hierarchy, see Fig.1). For the same reason, the
location attribute is also introduced in the 'Physical anatomical entity
template'. On the other hand, slots for the different classes of location
relationships (see Fig. 2) must be introduced in the templates of selected
descendant classes of 'Physical anatomical entity'. For instance, the 'contains'
slot is introduced in the template of 'Anatomical space', whereas its inverse
'contained in' is inserted in the 'Material physical anatomical entity
template', since both anatomical structures and body substances can be contained
in anatomical spaces. Therefore, in the frame of 'T8 part of esophagus',
'posterior mediastinum' can be a value for the relationship 'contained in',
since 'posterior mediastinum' is classified as a
'Compartment', which is a subclass of 'Anatomical space'.
Adjacencies
also refer to location (see Fig.2). However, they are attributed relationships
and therefore 'has adjacency' itself is modeled as a frame with its own slots.
For a conventional cylinder, these slots correspond to the radial coordinates
shown in Figure 3. 'Zone of esophagus template' inherits these slots from the
'Conventional cylinder template'. The adjacency relationships of 'T8 part of
esophagus' are displayed in Figure 5 as the values of these radial coordinates.
These values are the immediate adjacencies of this zone of the esophagus. The
concepts that correspond to each of these values (e.g., fibrous pericardium,
azygos vein, thoracic aorta) also contain adjacency relationships in their own
frames. 'T8 part of esophagus' is a value for one of the adjacency coordinate
slots along a vector reciprocal to that which relates 'T8 part of esophagus' to
the concept. For example in the frame of Azygos vein the value for 'left
anterior' coordinate of adjacency would be 'T8 part of esophagus'. Since
adjacency relationships are transitive along a coordinate vector, it may be
inferred from immediate adjacencies that T8 vertebra (labeled ‘T8’ in Fig 4)
is located 'right posterior' in relation to ‘T8 part of esophagus'. It is
these interacting adjacency relationships that constitute the adjacency subnet
of the SAn component of the Anatomical Structural Abstraction.
We
selected 'T8 part of esophagus' to illustrate the symbolic representation of
detailed structural relationships within the Foundational Model of anatomy. We
are validating the templates we have developed so far by assessing the extent to
which they generalize to different classes of anatomical entities. Our
contention is that a logical and comprehensive symbolic model of anatomical
structure will be indispensable to programs and applications that call for
reasoning about the human body. By making the FM available as the UWDA
vocabulary of UMLS, we hope to obviate the need for ad hoc, repetitive and inconsistent modeling of anatomy by
developers of educational or clinical applications who require detailed
knowledge of specific parts of the body. Our objective is to enrich this
resource by a comprehensive representation of structural relationships, which
are an integral component of reasoning about the human body by both humans and
machines.
Acknowledgments
This work was supported in part
by contract LM03528 and grant LM06822, National Library of Medicine.
References
1.
Lindberg DA, Humphreys BL, McCray AT. The Unified Medical Language
System. Methods Inf Med 1993;32:281-91.
2.
Rosse C, Mejino JL, Modayur BR, Jakobovits R, Hinshaw KP, Brinkley JF.
Motivation and organizational principles for anatomical knowledge
representation: the Digital Anatomist symbolic knowledge base. J Am Med Inform
Assoc 1998;5:17-40.
3.
Rosse C, Shapiro LG, Brinkley JF. The Digital
Anatomist Foundational Model: principles for defining and structuring its
concept domain. Proc AMIA Symp 1998;820-4.
4.
Brinkley JF, Wong BA, Hinshaw KP, Rosse C. Design of an anatomy
information system. IEEE Comp Graphics Applic 1999;3:38-48.
5.
Michael J, Mejino JLV, Rosse C. The role of definitions in biomedical
concept representation. Proc AMIA Symp 2001. Submitted.
6.
Martin RF, Mejino JLV, Bowden DM, Brinkley JF, Rosse C. Foundational
model of neuroanatomy: its implications for the Human Brain Project. Proc AMIA
Symp 2001. Submitted.
7.
Mejino JL, Rosse C. The potential of the Digital Anatomist Foundational
Model for assuring consistency in UMLS sources. Proc AMIA Symp 1998;825-9.
8.
Mejino JL, Rosse C. Conceptualization of anatomical spatial entities in
the Digital Anatomist Foundational Model. Proc AMIA Symp 1999;112-6.
9.
Agoncillo AV, Mejino JL, Rosse C. Influence of the Digital Anatomist
Foundational Model on traditional
representations of anatomical concepts. Proc AMIA Symp 1999;2-6.
10.
Neal PJ, Shapiro LG, Rosse C. The Digital Anatomist structural
abstraction: a scheme for the spatial description of anatomical entities. Proc
AMIA Symp 1998;423-7.
11.
Rogers J, Rector A. GALEN's model of parts and wholes: experience and
comparisons.
Proc
AMIA Symp 2000;714-8.
12.
Schulz S, Romacker M, Hahn U. Part-whole reasoning in medical ontologies
revisited- introducing SEP triplets into classification-based description
logics. ProcAMIA Symp.1998:830-4.
13.
Rosse C, Ben Said M, Eno KR, Brinkley JF. Enhancements of anatomical
information in UMLS knowledge sources. Proc AMIA Symp 1995;873-877.
14.
Noy NF, Fergerson RW, Musen MA. The knowledge model of Protégé-2000:
combining interoperability and flexibility. In: Proceedings of the 12th
International Conference on Knowledge Engineering and Knowledge Management
(EKAW-2000) 2000.Juan-les-PinsFrance,Springer.