1-s2 0-S1532046401910305-main

Journal of Biomedical Informatics 34, 321–347 (2001)

doi:10.1006/jbin.2001.1030

Symbolic Representation of Anatomical Knowledge:Concept Classification and Development Strategies

P. Cerveri*,†,1 and F. Pinciroli*

*Bioengineering Department, Politecnico di Milano, Piazza Leonardo da Vinci, 32, I-20133 Milan, Italy; and†Centro di Bioingegneria, Politecnico di Milano–Fondazione Pro Juventute Don Carlo Gnocchi ONLUS,Via Capecelatro, 66, 20148 Milan, Italy

Received July 25, 2001; published online May 23, 2002

In this paper a novel approach to anatomy knowledge representationis described. The focus of the present research has been on the develop-ment of a representational framework where the conceptual level hasbeen implemented by using hierarchical and nonhierarchical conceptualnetworks. This has allowed handling the demand for multiple viewsof anatomy (systemic and topographical views). The terminologicallevel of the knowledge representation has been implemented by usinga compositional strategy which has avoided the explicit storage of theterms used to express composite concepts. Hierarchical relations andcomposite concept representations have required supervision of both
the inheritance and concept reconstruction. For this purpose heuristicknowledge has been stored in terms of consistency rules in the knowl-edge base. As proof of the capability of this system, we show how the knowledge base has been used to provide symbolic access to spatialinformation consisting of a reduced set of images from the VisibleHuman Dataset. q 2001 Elsevier Science (USA)
1. INTRODUCTION

Despite a long history of research in the field of medicalinformatics, the design and implementation of accurate,

1To whom correspondence and reprint requests should be addressed.E-mail: [email protected].

1532-0464/01 $35.00 321q 2001 Elsevier Science (USA)All rights reserved.

complete, and flexible representations of biomedical knowl-edge remain an as yet unfulfilled goal. Among the causesof this deficiency, one reason appears to be the minimalattention paid, by both knowledge engineers and softwaredevelopers, to the problem of adequately modeling humananatomy knowledge in computer-based systems. In fact, alarge number of the assertions formulated in all biomedicaldomains make use of anatomical concepts. Clinical treat-ments, diseases, biochemical processes, and surgical inter-ventions all imply the generation of statements in which theinvolved concepts refer to body locations, organs, or genericbody components at macroscopic as well as microscopiclevels. Endowing a computer system with the ability tounderstand biomedical statements and perform reasoningabout that content—what we intend to clarify here withreasoning—is not a trivial task. Let us consider the followingphrases as typical examples of clinical statements involvinganatomical concepts:

• Cirrhosis affects the liver or more precisely a part of
the liver, manifests in its interior, and cannot affect bodyparts other than liver,
• Disease affects an organ or organ part, manifests in theinterior or exterior part of the organ, and always affects thesame organ type.

The first statement refers to a specific disease affecting

322

a particular organ (real concepts). In contrast, the seconddeals with different levels of knowledge involving concep-tual categories (abstract concepts). While the first statementimplies specific knowledge of anatomical entities (specialanatomy), the second requires some a priori assumptionsused to classify concepts into categories (general anatomy).With the aim of emulating the human mind’s capabilityto understand both levels of knowledge, a computer-basedsystem should be provided with four main frameworks: (1)a concept classification, (2) data structures (knowledge base)into which concepts are mapped, (3) a terminological sourceto map concepts to language, and (4) a software engine toperform reasoning (knowledge-based querying and informa-tion reconstruction).

Terminological sources, widely used in biomedicine, rep-resent a low-level attempt to organize biomedical knowledgeinto a computer-based system. The aim of such sources istwofold: (a) they endeavor to achieve standardization of theterms used in a specific domain of knowledge; (b) theyattempt to provide a symbolic representation of underlyingconcepts [2–4]. A terminology is a low-level system thataggregates terms according to simple alphabetical rules;however, no assumptions about the conceptual organizationare made. By assigning unique reference codes to terms, aterminology can be made into a coding system. Furthermore,it can be made into a thesaurus by distinguishing betweenpreferred terms and synonyms. Any preferred term consti-tutes the main reference for a concept, whereas the corres-ponding synonyms are sometimes used to designate the sameconcept. By providing definitions for the terms, a thesaurusassumes the characteristics of a vocabulary. However, avocabulary is a static system and cannot be queried to vali-date any semantic statement.

With regard to anatomy, several terminological sourceswhich collect terms used by experts in specific clinical envi-ronments exist. Depending on the medical field in whichthey have been used, the same terms can sometimes referto different anatomical concepts—leading to inconsistencybetween different sources, or multiple different terms canrefer to a single concept–causing redundancy. The UnifiedMedical Language System (UMLS) [5] from the NationalLibrary of Medicine was conceived as an attempt to builda standardization interface among terminological sources,
with the aim of removing term redundancy. As a result, theMetathesaurus, the main knowledge source of UMLS, makesavailable alphanumeric codes that act as links between con-cepts and terms in individual terminological sources. TheUMLS does not provide a proper terminological source it-self, but rather adopts those of the terminology providers.Therefore, while eliminating redundancy, it does not resolve
CERVERI AND PINCIROLI

term inconsistency. Only moving from a terminological toa conceptual level can address the problem of term inconsis-tency.

Terminological sources as referred to here are consideredto be simple coding systems containing lists of coded termsbut without any explicit relations among the correspondingconcepts. Actually, the available terminological sources pro-vide term organization but little concept classification.Semantic networks or conceptual graphs are the primaryvisualization tools used to represent concepts and their inter-relations [6, 7]. In semantic networks, knowledge is repre-sented by nodes (concepts) and arcs (semantic relations). Inparticular, semantic relations divide into hierarchical andnonhierarchical relations. Starting with most generic con-cepts, concept trees can be obtained in which specializationincreases as the leaves of the tree are approached. For exam-ple, relating concepts through relations of genus (::is-a-kind-of ) allows the generation of a taxonomic tree. Equivalently,a partonomy tree can be obtained by linking conceptsthrough a part-of relation. Nonhierarchical relations can beused to model other concept attributes.

With respect to concept classification, the UMLS providesa knowledge source named the Semantic Network. However,its model takes into account only a few semantic typesthat do not allow for anatomical entities to be adequatelyrepresented. In SNOMED [8], only strict concept classifica-tion is provided. A simple nonpersistent strict hierarchybased on topographical properties of anatomical objects isrealized by using significant alphanumeric codes assignedto terms. Although it collects a relatively large number ofanatomical terms, this source exhibits insufficient flexibilityfor accurate anatomy knowledge representation. The ReadCodes Project [9] pursues a more flexible approach in whichconcepts, separated and linked to terms by nonsignificantcodes, have been classified into categories and structuredinto a taxonomic semantic network. The system, focusingon anatomical structures, has been conceived to expressseveral taxonomic views but does not provide either parto-nomic views or establish nonhierarchical relations betweenconcepts. For example, the cribriform plate, being a part ofthe lateral mass (left and right) of the ethmoid bone, is codedgenerically as a bone structure of the cranium without anyfurther specification. However, while expressing adequate
representational power, albeit with some inconsistencies inthe anatomical representation, the project has the great ad-vantage of addressing the problem of compositional organi-zation of concepts. In antithesis to enumeration, compositionimplies that terms indicating composite concepts are implic-itly stored in the knowledge base.
The GALEN project [10–12] adopts, as other systems do,

our system can be progressively improved.Moreover, among our aims, we did not set the goal of

SYMBOLIC REPRESENTATION OF ANATOMICAL KNOWLEDGE

enumeration for representing anatomical entities, but alsoprovides a compositional framework (the CORE model)manageable by a flexible language named GRAIL [13].These two modules guarantee the ability to derive, from areduced set of concepts, articulated concepts and phrasesthat are consistent and nonredundant due to automated rules.However, the anatomical terminology expressed in GALENhas the drawback of assuming anatomical entities as sitesof disease processes, which reduces the ability to representfully detailed levels of anatomical objects.

Unlike the project of Rosse’s group [14], none of theabove initiatives has systematically addressed the problemof building anatomical concept classification in the realmof pure anatomy. The aim of that project has consisted ofdefining a foundational model for anatomy able to connecta concept classification to a source of terms and to accommo-date a large typology of anatomical views. Much effort hasbeen spent to develop an accurate anatomical terminology(about 25,000 terms), and concept definitions have beengenerated consistently with concept classification properties.The approach pursued to collect concepts has been the enu-merative strategy, and both multiple hierarchies and nonhier-archical relations between concepts have been underesti-mated.

The approach of the Hohne group [15] was focused onthe representation of head anatomical structures and theirinterrelations. The great advantage of the methodology ap-plied is that it addresses the problem of the partonomic andnonhierarchical relations. This methodology has involvedthe use of nonhierarchical relations to describe characteris-tics of nerves and vessels. However, multiple concept classi-fications and inheritance have been disregarded.

We note here that no report in the cited literature hasgiven consideration to inheritance as a basis for the classifi-cation. In principle, the property of inheritance of hierarchi-cal relations guarantees that attributes can be automaticallypassed down from parent to child (concept of monotonicinheritance). This allows distribution of semantic attributesof the concepts along one or more hierarchical networksconstituted by conceptual categories. In practice, featureinheritance is very composite and full of exceptions andpeculiarities (nonmonotonic inheritance). This implies thatat each level of a hierarchy the attempt to subsume attributes
must be accurately validated by consistency rules. This issuehas received little attention to date.
In light of these considerations, our work has been focusedon developing an anatomical knowledge model, based onconcept hierarchies, endowed with an information recoveryengine to constrain inheritance. In particular, it differs fromprevious approaches with respect to the following features:

323

Both hierarchical and nonhierarchical relations betweenconcepts have been used.

Polyhierarchies have been introduced to allow maxi-mum expressiveness as required by anatomical knowledge.

A reconstruction information engine has been used toobtain explicit knowledge from its implicit storage.

A supervised inference algorithm has been developedto reconstruct knowledge consistently blocking undue inheri-tance.

Conscious of the fact that we cannot coordinate all ana-tomical information content, we focused our conceptualframework on an orientation to beginners and assumed that

refining any specific terminology. Rather we attempted toeliminate term redundancy by allowing a real-time connec-tion to the UMLS server able to map our terminology toother terminological sources. In the following sections weexplain these characteristics.

2. MODEL AND METHODS FOR ANATOMICALREPRESENTATION

2.1. Basis of Knowledge Representation

The aim of any representation is to organize the entitiesof a domain of knowledge according to some principles orcommitments that specify how to look at the attributes ofdomain entities. In particular, these principles force one todistinguish what the significant information content entitiesexhibited and what must be accurately modeled from what isless relevant, and thus what can be disregarded or representedwith less precision. This formal operation of conceptualiza-tion, supervised by some a priori assumptions (commit-ments), is named concept classification. In other words,concept classification foresees a subdivision into general andspecific categories based on similarity and discriminatingproperties (intrinsic attributes) of domain objects(individuals/instances). Thus, representing knowledge con-
sists of structuring concepts by expressing semantic con-straints, namely semantic relations that have been identified.
In general, a relation is a function of one or more argu-ments that can be specified in a multivalued variable. Forexample, the concept muscle may be linked to the conceptarm, with possible values that can include the flexion, adduc-tion, torsion, etc., attributes. Equivalently, the concept skull

324

may be linked to the concept bone through the composed-by relation. Binary or two-entity relations are specific rela-tions that join two concepts in a semantic statement. Asimple binary relation can be expressed as A::r B, which isasserting that entity A is related to B by relation r. Underthis assumption, most of the properties of a concept A canbe represented by a list of statements {A::r1 B, A::r2 C,A::r3 D, etc.} and graphically represented by a semanticnetwork. The more that relations between concepts are iden-tified, the more sophisticated the representation becomes.In particular, the use of binary relations allows creation ofinterconnected networks of concepts that are easily delivera-ble in a computer data structure. In Appendix A we reportthe definitions of the main semantic relationships referredto in the paper. However, they have standard definitions thatcan be retrieved by the UMLS knowledge base server.

The simplest anatomical conceptualization through a bi-nary relation can be expressed by a statement such ashand::is-a-kind-of body part. Automatically the definition ofthe concept hand can be built as ‘hand is a body part.’ Theproblem here is what we mean by the concept body part.As reported by the UMLS, the following definition seemsto be acceptable for the meaning of body part: “A collectionof cells and tissues which are localized to a specific area orcombine and carry out one or more specialized functions ofan organism. This ranges from gross structures to smallcomponents of complex organs. These structures are rela-tively localized in comparison to tissues.” According to thisdefinition, anatomical physical objects ranging from grossstructures to organs and even small components of organscan be grouped together. As a consequence, any part of thebody could be qualified as a body part, which does notappear to be useful. We are not in agreement with thisdefinition because we think that the category body partshould be assigned to only physical anatomical objects thatcan be externally discernible on the body, having virtual(external and internal) boundaries. The head can surely bean instance of the body part category: it is externally andinternally separated from the thorax by the neck. In contrast,the concept right ventricle should not be subsumed by thebody part category because an external subdivision of thebody that contains it does not exist. However, in abstract
terms it is a part of the body in the sense that the bodyincludes it. In light of these considerations we propose twodistinct definitions for the concept body part: one refers to theabstract concept as an aggregate of properties of anatomicalentities that may pertain to content, surface, localization, orfunction, with at least one of them present; the other refersto the concrete concept of an anatomical structure as an

aggregate of heterogeneous physical structures (organs, or-gan parts, and tissues) externally demarcated by skin subdivi-sion. The latter has been used as a category in our representa-tion and includes instances such as head, neck, shoulder,arm, and forearm.

2.2. Hierarchical Relations and Granularity Level

As stated, a classification consists of concept networkswhere relations link categories to categories and categoriesto individuals. In the case of monotonic inheritance, thetransitive property of hierarchical relations (genus and parti-tion) allows the distribution of concept properties acrossseveral hierarchical levels. In particular, each branch in atree of categories represents a specification level at whicha concept can be expressed. This aspect is strictly connectedto the topic of granularity level of the description. Let usexplain in more detail.

Medical and in particular anatomical information can bedescribed and used according to a particular context of dis-course. Detailed communications need to deal with a com-prehensive representation of concepts; generic discourses aresufficiently accommodated by the coarse granularity level ofthe concepts used. For example, while discussing a seriousinjury, two clinical specialists might use a phrase like “asevere bilateral open vertical fracture of the sacrum with acomplete cauda equina lesion.” In contrast, while communi-cating the patient’s condition to relatives, they would proba-bly explain it simply as a bone fracture in the pelvis. Theuse of fine-grained concepts allows precise identification ofobjects but requires extensive knowledge to be understood,whereas coarse-grained ones, while losing details, guaranteeimmediate and intuitive insight. With this perspective, ourrepresentation has been developed with the aim of manipu-lating concepts at different levels of abstraction. For exam-ple, (see Fig. 1) the left and right ventricles can be catego-rized as the first approximation through the parent categoryventricle. Moving to higher levels of abstraction, the ventri-cle can be classified as heart cavity and organ cavity.

In addition to the abstraction level, the detail level charac-terizes granularity of a concept. In particular for an anatomi-cal concept, this refers to the description of entity subparts.
The concept heart can be described as an anatomical entitywith two main subdivisions, a left side and a right side. Ifmore detail is needed, its definition can be refined to encom-pass information stating that the left side of the heart com-prises the left ventricle and the left atrium, which are con-nected through the mitral valve. Furthermore, this definitioncan be improved. Both concept description and classification


are based on some choice of granularity. The wider thegranularity levels, the more clear and useful the representa-tion, and the more it can capture the complexity of theknowledge domain. However, the wider the granularity lev-els, the more fragmented the knowledge and the more diffi-cult it is to manage. To conclude, we note that a directrelation exists both between abstraction level and taxonomicclassification and between detail level and partonomic classi-fication. Moreover, hierarchical networks should be flexiblein the sense of allowing one to collapse several abstractionlevels, disregarding intermediate concepts, so that, for exam-ple, an individual can be viewed as directly linked to a highlevel category.

2.2.1. Is-a-kind-of relation. Let us now focus on theproblem of the definition of categories by considering whatstandard anatomical books provide about the topic [17].Some books present anatomy by defining the systems thatconstitute the human body (cardiovascular system, nervoussystem and so on) in such a way that the reader accessesknowledge through the systemic view. This represents acommitment assumed by book authors. In contrast, otherbooks [17] provide a topographical view of anatomical con-tent by describing the human body as constituted by regionsand parts. As a result, merged views, which are not explicitlytaken into account, involve effort by the reader to understand.Another notable aspect is the definition and description ofconcepts that books present. When describing a concept likethe heart, one can ask what this concept is intrinsically. Ingeneral, each knowledge source provides an ad hoc answerto such questions. Organ, body part, muscle, involuntarymuscle, and main structure of the cardiovascular systemseem to be possible classes for heart. Yet, their meaningsand their relationships may vary from one source to another.To overcome such inconsistency, we first focused our devel-opment on building high-level categories. It is well knownthat building a taxonomic tree requires linking of conceptsby ::is-a-kind-of relations. All items in macroscopic anatomycan be easily assigned to three main categories: space, physi-cal structure, and substance. From this reduced set of genericcategories, anatomical concepts have been vertically speci-fied through several abstraction levels (subsumption opera-
tion). The generation of horizontal categories, at a predeter-mined abstraction level, has been driven, where possible,by the principle of null intersection (complementary) classes.
Given the above classification, for example, the heartcould clearly be associated with the physical structure cate-gory because of its intrinsic characteristics. However, toclassify the heart, common sense suggests that organ is a

325

category more specialized than physical structure. The defi-nition of the fundamental category organ has been obtainedaccording to the following statement: it is an anatomicalstructure; it is distinct both morphologically and functionallyfrom other such units; it cannot be divided into organs; itis composed of tissue; it expresses several functions; it canbe dense; it can contain hollow space; it can contain andproduces substances; it can perform actions. Under theseconsiderations, the following definition seems adequate: or-gan is a minimal self-contained anatomical structure able toexpress functional attributes. It is constituted by tissue thatspecifies its shape and its structural properties.

The following list is the formal representation of sucha definition:

organ::is-a-kind-of anatomical structureorgan::expresses functionorgan::performs actionorgan::is-constituted-by tissueorgan part::is-part-of organfull organ::is-a-kind-of organcontainer organ::is-a-kind-of organorgan::is-part-of body system

Based on the structural differences that organs exhibit,two major categories, container organ and full organ, havebeen generated. Container organs typically can enclose liq-uid, gas, or solid substance and necessarily are constitutedby cavities. In contrast, full organs accomplishing the func-tion of body sustainer (as bones), body motion (muscle),substance producer (liver, pancreas) are uniformly filled (ap-proximately) by tissue. As an example of taxonomy, Fig. 2shows the semantic network for the category container organ.

Note that we have identified three main subcategories ofthe category container organ: gas container, liquid container,and solid container organ. For container organ we haveidentified as relevant the morphological property of beingof tubular shape which has led to the generation of thecategory tubular organ. This has led us to allow for thecategory tubular organ to have more than one parent concept.This is called a multiple-child/multiple-parent hierarchy(polyhierarchy), in contrast to strict hierarchies in which achild concept can have only one parent. As will becomeclearer in the following sections, polyhierarchies make the
use of automated inheritance more problematic. Further-more, it is important to emphasize here that the classificationshown in Fig. 2 does not exhaust all structural and functionalproperties of container organs. For example, the conceptlung has structural and functional properties that are nottaken into account.
Based on specific macroscopic structural and functional

FIG. 1. Granularity: Abstraction and detail.

properties of full organs, we have designed categories likeparenchymatous organ, glandular organ, bones, muscle,and joint, and for each of them subcategories have beengenerated.

For example, the concept bone has been subdivided intofour main classes according to morphological properties (i.e.,irregular bone, long bone, flat bone, and short bone). Each ofthese classes has been specialized according to topographical(i.e., cranial bone, carpal bone, vertebra) membership crite-ria. Finally, each single bone has been assigned to the cate-gory that defines it more specifically (see Fig. 3).

Similar criteria have been adopted to classify muscles andjoints. In this case a strict hierarchy is not sufficient toexpress all properties that anatomical objects exhibit. Forexample, rib should be classified for its morphology as beinga long bone. However, it shares an internal structure similarto flat bone (lack of marrow). In this case rib cannot havetwo parents (long and flat bone) because in the definitionof category flat bone we did not take into account a structuralcriterion but only a morphological one. This property of the
rib will be expressed differently as discussed here.
The above problem can be better illustrated in the effortto represent in the network the properties of the pancreas.It should be classified both as a parenchymatous organ likethe liver and as a glandular organ like the thyroid gland. Inaddition, due to its specific functional properties, the pan-creas could be classified as an endocrine gland (see Fig. 4).


A similar issue relates to the classification of the lung as acontainer organ but also assigning the property that it isconstituted by parenchymal tissue. In this case, however,the inheritance between lung concept and full organ conceptis to be explicitly blocked.

An alternative solution consists of reducing a polyhierar-chy into two or more strict hierarchies where inheritanceis not a requirement. This approach can be illustrated byclassifying muscles. At the first level, they can be classifiedaccording to fiber property, smooth and striated. For thesecond level, striated muscles have been specialized intocardiac muscle and skeletal muscle. Also it is reasonable toclassify muscles into voluntary and involuntary. However,note that cardiac muscle and skeletal muscle are to be classi-fied respectively as involuntary and voluntary, while sharingstriated fibers. The attempt to take into account the firstrepresentation (smooth and striated muscle) for voluntaryand involuntary muscle categories by associating them with

326

both smooth and striated muscle would lead to the use of aconstrained inheritance. This would produce a conditionwhere cardiac muscle inherits both smooth and striatedmuscle properties via involuntary category. Therefore, thisexample shows that the genus relation can be characterizedby a set of distinct contexts specifying multiple taxonom-ical views.

FIG. 2. The semantic network for category container organ. Anycontainer organs can have a tubular shape. Note that this classificationdoes not exhaust all structural and functional properties of organs. Forexample, lung concept has structural and functional properties that arenot taken into account here.

bones which have the prevalence of a diameter with respect to theother two. Flat bones have two diameters more relevant than the last

made formal by explicitly using a conceptual relation such

one. Short bones have three similar diameters. Irregular bones haveno regular shape. This represents a particular view over the propertiesof bones. For example, rib is classified for its morphology as being along bone. However, it shares an internal structure similar to flatbones (lack of marrow). This property must be expressed by alternateclassification (context).

2.2.2. Is-part-of relation. The part-of relation plays aparticular role equally important to that of the ::is-a-kind-of relation in the field of semantic networks applied to ana-tomical knowledge representation. In general, two subtypesof partonomic relations exist that respectively refer to physi-cal and conceptual partition. The first type can be equiva-lently expressed by the is-consisting-of relation (cell::is-part-of tissue, tissue::is-part-of organ). The second type isbased on the fact that an anatomical entity considered as awhole can be conceptually (arbitrarily) subdivided into twoor more parts (organ part::is-part-of organ, shaft of femur::is-part-of femur, nose::is-part-of head, finger::is-part-of hand).Equivalently, nose and mouth are parts of the head; larynx,pharynx, and trachea are parts of the neck; bronchus, bron-chiole, alveolus, and lung are parts of thorax (topographicalview). As an example, Fig. 5 shows the partonomy of thefemur.

Moreover, the conceptual subdivision into parts of ana-tomical structures can be expressed at multiple granularitylevels (detail levels). For example, the category long bone
can be further divided into (distal head) diaphysis, (proximalhead) epiphysis, and shaft. Inheritance will imply that alllong bones (like the femur) shall have necessarily thesethree parts.
Apart from subdivision of anatomical structures into parts,the partonomy relation has been used to express functionalmembership (functional view). For example, the partonomy

327

of the respiratory system can be accommodated by simplylinking the nose, mouth, larynx, trachea, bronchus, bronchi-ole, alveolus, lung, and pharynx to the involved system.

2.3. Nonhierarchical Relations

As shown, the genus and partition properties that conceptsexhibit can be represented as hierarchical networks. In con-trast, conceptual, physical, spatial, and functional relationsthat are intrinsically nontransitive have their semantic repre-sentation into nonhierarchical tree of concepts. Conceptualrelations between categories express an obvious logic arisingfrom the considerations used to build the categories. Forexample, taking into account that some organs can containspace and recognizing that this feature is a key discriminatingproperty has led to the generation of the category containerorgan. Necessarily, all container organs will define cavitiesor, more precisely, organ cavities. This condition can be


FIG. 3. Taxonomy of the concept bone. Morphology is the criterionused to provide first level classification of bones. Long bones are the

as container organ::defines organ cavity where statementsorgan cavity::is-a-kind-of body cavity, body cavity::is-a-kind-of body space, and body space::is-a-kind-of anatomicalspace have been established.

This then implies that a consistent definition for organcavity could be that it is a body space defined by a container

FIG. 4. Polyhierarchies for concept organ. The lung is classifiedas both a gas container organ and a parenchimatous organ. By applyingmonotonic inheritance, lung inherits incorrectly characteristics fromboth container and full organ. For lung concept, inheritance fromparenchimatous and full organ is blocked.

LI
328 CERVERI AND PINCIRO
FIG. 5. The partonomy for the concept femur.

FIG. 6. The branching-of relation used to represent the arterialbranch of the upper extremity.

organ. As in the case of stomach, it has been classified ascontainer organ and inheritance sanctions that it must beendowed with an organ cavity.

Physical relations are used to represent attributes and com-mon characteristics shared by two anatomical entities. Herethe most relevant physical relations we used in the represen-tation are ::is-connected-to, ::is-contained-in, ::branching-of, ::consisting-of, ::interconnecting, and ::is-tributary-of.With the ::consisting-of relation (structurally made of, inwhole or in part, some physical units, material or matter), thephysical properties of anatomical categories and individualstructures can be modeled as:

bone::consisting-of bone tissue
muscle::consisting-of muscle tissue
For the periosteum, endosteum, lamellar bone tissue, andspongiosa, kind-of bone tissues, the inheritance propertyguarantees that long bone and the other subcategories willbe constituted by these kinds of bone tissues. Equivalently,because the heart can be classified as a muscle besides as

329

container organ, it can be assumed that it is constituted bymuscle tissue.

The relation of branching between two anatomical con-cepts can be represented for example as:

artery::branching-of arteryfemoral artery::branching-of external iliac arterypharyngeal nerve::branching-of vagus nerve

The relation is-tributary-of can be used to describe a ve-nous tree. For example, the following relations hold:

vein::is-tributary-of veinsplenic vein::is-tributary-of portal vein

Figure 6 shows a ::branching-of-based network for theaxillary artery, which is the main arterial branch of the upperextremity. Note that ::branching-of relation is not transitive;therefore, the tree in Fig. 6 is not hierarchical. Moreover,that schema does not account for spatiality that must beexplicitly expressed by spatial relations, all subsumed by::spatially-related-to relation. These relations refer to proper-ties of adjacency, relative position, path, location, etc. (::be-ing-adjacent-to, ::connecting, ::entering, ::passing-through,::being-behind, ::traversing, ::surrounding, ::being-discon-nected-from, ::being-externally-connected-to, ::being-par-tially-overlapped-by).

In the case of muscles, both origin and insertion into thebones are relevant spatial feature expressed by two relations:muscle::having-origin-in bone part and muscle::having-insertion-in bone part. For example, psoas major has originin the transverse processes of the lumbar vertebrae and bodyof the 12th thoracic vertebra and insertion in the middlesurface of the lesser trochanter of the femur. Moreover, bothinsertion and origin can be further specified. Therefore, thestatement “the biceps brachii through its long head has originat the supraglenoid tubercle of the scapula and through itsshort head at the coracoid process of the scapula. It hasinsertion in the tuberosity of the radius” can be representedas muscle part::having-origin-in bone part.

Although the origin and the insertion of the muscle deter-mine the movement of the underlying bones, spatial proper-ties do not rely on either function or action. From the parentrelation ::functionally-related-to, the following set of rela-tions is relevant for anatomical knowledge representation


::interacting-with, ::innervating, ::supplying-blood-to, ::se-creting, ::acting-on (generic), ::acting-as-flexor-of, ::acting-as-lateral-rotator-of, ::causing-contraction, etc.

Figure 7 shows a representation of the muscle of theshoulder and arm where we model partonomic relationsalong with specific actions and innervations.

A specific functionality of artery consists of supplying

330 AND PINCIROLI

:
FIG. 7. The muscles of the shoulder and arm
blood to tissues of the anatomical structures, whereas nervesprovide functional connection between the central nervoussystem and the body parts. Therefore,

artery::supplying-blood-to organ,renal artery::supplying-blood-to kidney,

Modeling partonomy, action, and innervations.

CERVERI

5th lumbar nerve::innervating inferior gemellus(muscle)

are valid relations.An action is a specific function that an anatomical entity

performs on one or more entities such as rising, pulling,

asy
FIG. 8. The partonomy of the lungs. The
flexing, extending, rotating (medially or laterally), ab-ducting, and adducting. For the muscles, these actions iden-tify their main functionality as in the following example:

muscle::acting-on body partpsoas major::acting-as-flexor-of thigh,obturator internus::acting-as-lateral-rotator-of thigh.

As an example, four main relations can express muscleactions on the foot as reported in Table 1. Particular attentionis to be paid to relationships that involve more than twoanatomical objects. For example, representing the case thatventricle and atrium are connected to each other through avalve involves a relation among tree entities that a binaryrelation cannot accomplish. To include this case we estab-lished a new nonhierarchical relationship named association,through which n independent concepts are related to one
another.
2.4. Knowledge Base Development

Categories (semantic types) define “general” anatomy,whereas the specific anatomical structures (individuals) de-fine “special” anatomy. The relationship between category

mmetry between left and right is evident.

and individual (physical objects) can be better understoodby the following example. Given the category bone, whichidentifies an element of general anatomy, the category longbone is a specialization of such a category, yet still belongingto general anatomy. In contrast, femur, which is an individualof the category long bone, is an element of special anatomy.Provided that certain nonhierachical and hierarchical rela-tions between categories (general anatomy) and other catego-ries (general anatomy) and between categories (general anat-omy) and individuals (special anatomy) are established, thetransitive property of hierarchical relations guarantees thatthe parent attributes are inherited by the offspring. The fol-lowing example clarifies this point:

::is-a-kind-of : long bone (general)::is-a-kind-of bone

SYMBOLIC REPRESENTATION OF ANATOMICAL KNOWLEDGE 331

(general)::consisting-of : bone (general)::is-constituted-by spon-

giosa (general)::is-part-of : epiphysis (general)::is-part-of long bone

(general)::is-a-kind-of : femur (special)::is-a-kind-of long bone

(general)

Tibialis anterior Peroneus longus Tibialis anterior Peroneus tertiusExtensor hallucis longus Peroneus brevis Tibialis posterior Peroneus longusExtensor digitorum lungus Gastrocnemius Extensor hallucis longus Peroneus brevisPeroneus tertius Soleus Flexor hallucis longus Extensor digitorum longus

PlantarisTibialis posteriorFlexor hallucis longusFlexor digitorum longus

::is-a-kind-of : femur, left (special)::is-a-kind-of femur(special)

::is-part-of : femur (special) is-part-of skeleton(special)

From the above, one can automatically infer that the leftfemur has an epiphysis (it is a long bone) and a spongiosa(as all the bones). If the model is accurately verified, thenthis approach avoids redundancy caused when relations car-rying the same meaning are duplicated across different ab-straction levels. This applies also to hierarchical partonomictrees and, for example, if the concept finger is modeled aspart of the concept hand and, in turn, hand is part of upperextremity, then there is no need to define finger explicitlyas part of upper extremities.

Starting from the above described model and accordingto point 2 of the first paragraph in the introduction we havebuilt a knowledge base that maps concepts to alphanumericcodes. Then our efforts have been focused on the develop-ment of a terminological system to map concepts to lan-guage. Toward this aim a terminological system has beenconstructed according to a compositional strategy. This strat-egy implies that only atomic concepts are explicitly storedin the knowledge base. In particular an atomic concept is aconcept whose term, used to express it, is constituted by asingle item, e.g., bone, organ, muscle, tissue, head. In con-trast, nonatomic concepts are expressed by two or moreterms. For instance, the concepts full organ, long bone, and
striated muscle will be explicitly stored in the knowledgebase as an alphanumeric code but the corresponding termsused to express them will be implicitly stored. The term fullis an attribute for the concept organ: associating the termfull to the term organ we generate the term full organ thatis used to reference the concept full organ. In the nextparagraphs we will describe how articulated terms like
Flexor digitorum longus

epiphysis of the left femur can be represented and how toconnect them to the corresponding concepts.

With this approach, by using atomic concepts and distrib-uted attributes, there is no need to explicitly represent termsthat differ in minor ways from each other. If the upperextremity is identified as having left and right laterality, thenonly arm, forearm, and hand concepts need to be explicitlystored as terms in the knowledge base. By using inference,automated reconstruction can transforms implicit knowledge(the fact that arm, forearm, and hand exist in left and rightside) into explicit information. Similarly, the partonomy ofthe ethmoid bone, which is an unpaired bone in the cranium,constituted by a perpendicular plate and two main lateralmasses (left and right) each containing other subcomponents(cribriform plate, lamina orbitalis, middle nasal concha) canbe incrementally represented in the knowledge base as re-ported in Table 2, where explicit and implicit relations havebeen defined:

The ethmoid bone concept, represented as the compositionderived by the anatomical concept bone with ethmoid attri-bute, is explicitly associated in the partonomy to the compos-ite concept lateral mass having left and right side attributes.Then cribriform plate is explicitly linked to lateral mass sothat left and right lateral masses implicitly inherit the partcribriform plate. Equivalently, an explicit link between lam-ina orbitalis and lateral mass makes possible an implicit linkto left and right lateral mass.

332 CERVERI AND PINCIROLI

TABLE 1

Summary of Muscle Action on the Foot

::acting-as-abductor-with- ::acting-as-abductor-with-::acting-as-plantar-flexor-of ::acting-as-dorsal-flexor-of inversion-of inversion-of

In a compositional framework nonhierarchical relationscan also benefit from inheritance. Taking into account afunctional relation like acting as flexor of, the psoas ma-jor muscle is a part of the thigh that is a paired body regionso that “psoas major::acting as flexor of thigh” statementis implicitly duplicated for both right and left psoas major.

Despite its appealing power as a means for knowledge

Ethmoid bone Composite ::is-a-kind-of Implicit Bone AtomicLateral mass Composite ::is-a-kind-of Implicit Mass AtomicLateral mass Composite ::is-part-of Explicit Ethmoid bone Composite
Left lateral mass Composite ::is-part-ofRight lateral mass Composite ::is-part-ofCribriform plate Composite ::is-part-ofLamina orbitalis Composite ::is-part-ofLamina orbitalis Composite ::is-part-of
extrapolation, the compositional approach has a sensiblelimitation: it cannot eliminate completely heuristic knowl-edge about the domain. In order for composite concepts tobe consistently reconstructed, the knowledge base must beprovided with specific control operators to prevent the gener-ation of unreal concepts. To make this point clear, let usconsider the lungs, the partonomic subdivision of which intolobes and bronchopulmonary segments is depicted in Fig. 8.

It is known that the right lung is morphologically slightlydifferent from the left lung because of the different numberof lobes and segments. In fact, the left lung lacks a centralsubdivision (middle lobe) that is present in the right lung.The compositional representation of this information contenttakes into account the concepts lung, lobe, and segment,whereas the laterality attribute (left, right, anterior, superior,medial, etc.) is compositionally represented, as will be clari-fied below. However, a nonsupervised reconstruction is notable to account for heuristic facts such as lung asymmetryand would generate erroneous concepts. By assuming thatthe concept lung exists in left and right instances and isdivided into three different pulmonary lobes (superior, mid-dle, and inferior), the reconstruction engine would assignall three lobes to both right and left lungs erroneously. Thisimplies the need for reconstruction to be driven by a heuristicrule breaking the link between left lung and middle lobe.

As a consequence, inheritance in a knowledge base devel-oped through a compositional approach has two separatefeatures: one related to semantic level as described in Section2.2.1 (no matter how concepts are stored) and the other
related to how concepts have been represented throughterms. An enumerative strategy avoids this last problem byexplicitly storing composite concepts through correspondingdefining terms: the concept head of the left femur will havean entry in a concept structure through a concept code andan entry in the term structure containing the string “head ofthe left femur.” In contrast, in our compositional strategy
Implicit Ethmoid bone CompositeImplicit Ethmoid bone CompositeExplicit Lateral mass CompositeExplicit Lateral mass CompositeImplicit Left lateral mass Composite

only the code of that concept has been explicitly stored ina data structure that we have named Composition, whereasthe corresponding term has been compositionally repre-sented. A devoted structure in the knowledge base has beendesigned to implicitly define composite concepts.

Let us clarify this process by an example. Head and neckare terms that are separately coded with respect to conceptsthat can be indicated through these terms. For example,concept “head” as a body part will have a concept code ina data structure named BasicAnatomicalConcept that willbe linked to other concept codes to define its semantics. Inaddition, it will be associated to a term code with the string“head.” In contrast, the concept “head of the proximal epiph-ysis of the left femur,” being a composite concept, will havea code in the Composition data structure. Similarly, epiphysisis an anatomical concept and part of femur and proximalepiphysis is one of its composite concepts. Table 3 showshow the concept “head of the proximal epiphysis of leftfemur” has been represented in our compositional frame-work.

The top of Table 3 shows the composition of the conceptas a join between a root concept and an attribute. An attributecan be either a feature or concept For example, the termthat refers to the concept left femur has been built by compos-ing the term femur and the laterality feature left. At thebottom of Table 3, we report the related explicit and implicitsemantics. In particular, this development strategy impliesthat the concept “head of proximal epiphysis of left femur”has been explicitly stored as an alphanumeric code, whereas


TABLE 2

Explicit and Implicit Relations Used to Define the Partonomy of the Ethmoid Bone

Concept name Concept type Relation name Relation type Concept name Concept type

the corresponding term has not. The following steps arerequired to iteratively reconstruct that term:

(1) Given the concept code, the procedure recovers thekernel concept code (head of proximal epiphysis of femur),that is in turn a composite concept (first row);

(2) The concept code of head of proximal epiphysis of

Head of proximal epiphysis of femur is part of proximal epiphysis of femurHead of proximal epiphysis of left femur is part of proximal epiphysis of left femur

n roby

Head of proximal epiphysis of femur is part of femur(Intermediate level in the hierarchy has been hidden

Head of proximal epiphysis of left femur is part of left femur(Intermediate level in the hierarchy has been hidden

Note. Concept composition implies implicitly genus relation betweethe knowledge base implementation we have constrained this property

femur is retrieved (second row) as composition of the aboveconcept with femur attribute;

(3) The root concept head of proximal epiphysis is re-trieved (third row) as the composition of the above conceptwith proximal attribute;

(4) Equivalently, the root concept head of epiphysis isretrieved (fourth row);

(5) The anatomical concept head is retrieved (fifth row);(6) The reconstruction engine generates the needed term

head of proximal epiphysis of left femur.

The relevant semantics are reconstructed by using inheri-tance from the implicit representation as depicted at the
bottom of Table 3. In particular, the semantics can be recon-structed according to the following steps:
(a) Given that head of epiphysis of femur::is-part-ofepiphysis of femur (see bottom of Table 3) and proximalepiphysis is compositionally a kind of epiphysis (see topof Table 3), the relation head of proximal epiphysis offemur::is-part-of proximal epiphysis of femur holds;

ot composite concept and concept (left femur is-a-kind-of femur). Ina devoted flag (see Table 4) in Composition data structure.

(b) Given that head of proximal epiphysis of left femuris compositionally a kind of head of proximal epiphysis offemur, the relation head of proximal epiphysis of left fe-mur::is-part-of proximal epiphysis of left femur holds.

The knowledge base has been implemented by utilizinga relational model and stored in a database named Anatomy-Knowledge. The first data structures we developed in Anato-myKnowledge were BasicAnatomicalConcept, Definition,and Term table, the latter containing only anatomical termssuch as organ, bone, long, and short without any explicitconceptual reference. Moreover, we have classified distinctterms as either preferred or synonym. The BasicAnatomi-


TABLE 3

Terminological Composition of Concept “Head of Proximal Epiphysis of Left Femur”

Terminology

Concept code Concept type Kernel concept code Kernel concept type Attribute code Attribute type

Head of proximal epiphysis of Composite Head of proximal epiphysis Composite Left Featureleft femur of femur

Head of proximal epiphysis of Composite Head of proximal epiphysis Composite Femur Conceptfemur

Head of proximal epiphysis Composite Head of epiphysis Composite Proximal FeatureHead of epiphysis Composite Head Atomic Epiphysis ConceptLeft femur Composite Femur Atomic Left Featureproximal epiphysis Composite Epiphysis Atomic Proximal FeatureEpiphysis of femur Composite Epiphysis Atomic Femur ConceptProximal epiphysis of femur Composite Proximal epiphysis Composite Femur Concept

SemanticsExplicit

Epiphysis of femur is part of femurHead of epiphysis of femur is part of epiphysis of femur

ImplicitLeft femur is-a-kind-of femurproximal epiphysis of femur is part of femurproximal epiphysis of left femur is part of left femur

calConcept table (see Table 4 and Fig. 9) has been used tolink anatomical concepts to terms and definitions (organ,bone). Nonatomic concepts (composite) like long bone, shortbone, metatarsal bone, and metacarpal bone have been im-plicitly stored as a link through two separate data structures:the table Attribute containing a term code corresponding tothe attribute value (long, short, . . . , red, green, . . . , lateral,

(artery) (vessel) kind-of)N000w21367 CC000s0453 C000002314 r01a (is-a-

(systemic (artery) kind-of)
artery)
…N3w3128897 C000001231 CC00003212 r01a (is-a-

femur (long bone) kind-of)
… … … …
medial, . . .) and a specification term code corresponding toattribute types such as size, dimension, color, shape, density,and the table Composition storing composite concept codes.

335

Semantic relations among concepts have been distributedacross two different data structures. In the Composition tablethe field Semantics (Boolean value) has been constructed tospecify that the composite concept assumes semantic distinc-tion from the parent concept (see Table 4). In Table 4 Net-work two generic concepts are explicitly linked by a relationwith an associated context.

Figure 9 shows the relational schema that we designedto implement the knowledge base. In particular, note thedata structure we named CrossMapping. In this table wematched the codes of the concepts used with the correspond-ing UMLS concept codes. This allows the reconstructionstage to automatically obtain from the UMLS server (http://umlsks.nlm.nih.gov/) the concept information taken frommajor digital biomedical terminological sources.

2.6. Information Reconstruction and Inheritance

To retrieve understandable information from the knowl-edge base the following main issues have been taken intoaccount: (a) Most terminological knowledge has been im-plicitly stored (compositional approach), as a result of whicha reconstruction procedure is required; (b) semantics of con-cepts can be retrieved by applying inheritance, but must becarefully verified through a supervision based on storedheuristic knowledge; (c) while being easily manageable, therelational data model used to implement the knowledge baseand SQL-based database query engine do not provide eitheradequate data structures or methods to efficiently retrievestored information; (d) information must be presented to theuser in a suitable way.

In order to address these problems and fit the require-ments, we conceived and developed our system in the follow-ing way:

The application has been subdivided into two parts: aserver side and a client side.

The server side performs the following actions:

Receives user requests from client interfaceBuilds queriesRuns queries on the databasesReconstructs informationSends result to client interface (data and images)


TABLE 4

The Main Relational Tables of the AnatomyKnowledge Database

BasicAnatomicalConcept table

ConceptCode TermCode DefinitionCode

C000000128 T000002342 (bone) D00ewf43r3C000008221 T0000r4823 (hand) D0045187g4

Attribute table

AttributeCode Attribute TypeCode Attribute ValueCode

A01a (dimension) (long)A01b (shape) (short)… … …B01a (laterality) (left)B01b (laterality) (right)C01a (laterality) (superior)C01b (laterality) (anterior)C01c (laterality) (lateral)C01a (laterality) (superior)P P P

Composition table

Composite RootConceptConceptCode Code AttributeCode Semantic

P P P PCC00003212 C000000128 A01a TRUE

(long bone) (bone)CC00002821 C000008221 B01a TRUE

(left hand) (hand)P P

Network table

ParentNetworkID ConceptCode ConceptCode RelationCode

N00423e001 C000000128 C000000006 r01a (is-a-(bone) (organ) kind-of)

N004f34002 C000002314 C000000349 r01a (is-a-

The client side has the following functionalities:

Data displayManaging user requestsThe server side collects data from two database systemsAnatomicalKnowledge relational databaseImageMap database

… … …CC00rs000 (lung) CC0000023 (full organ) (is-kind-of)

ascending and/or descending) for a specified concept in aspecified context;

336

The ImageMap database, accounting for coded imagedata, has been used to verify the usability of the developedknowledge base.

The server-side software framework has been developedaccording to the object oriented paradigm and implementedin the Java language, using a set of classes as TermClass,ConceptClass, DescriptionClass, RelationClass, Con-ceptTreeClass, ConceptTreeArrayClass, each characterizedby a set of operators (see Appendix B) that generates asafe SQL statement, issues a corresponding query to theAnatomyKnowledge database, and formats the results for thedefined data structures. For example, GetAscendingTree, amethod of ConceptTreeArrayClass class, acts as a recon-struction operator by receiving as input a TermClass objectcorresponding to an anatomical concept, a RelationClassobject specifying the involved relation, and if necessary auser-needed depth level, and giving back as output an arrayof the parent concept trees. Similarly, GetDescendingTreegives back as output an offspring concept tree. By oppor-tunely composing the operators, the reconstruction engineis able to explicitly obtain a wide range of anatomical infor-mation in the form of constrained views. As mentionedabove, the key point is that the reconstruction engine, auto-matically applying inheritance (inference), cannot ignore thepossibility of generating unreal concepts. Let us clarify thisissue by describing the reconstruction procedure for the par-tonomy of the lung and left lung concepts based on theGetDescendingTree operator. In generating the lung parto-nomic tree, the reconstruction engine automatically recoversthe abstraction level of the corresponding concept and thecorresponding detail level. In this case the concept lung hasbeen explicitly linked (Network table) to the lobe conceptand lobe to segment, and the reconstruction engine directlybuilds the lung-lobe-segment tree. The concept left lung,while being a specialization of the concept lung, has beenrepresented at a refined detail level, although not explicitlystored in the Network table. Initially the reconstruction en-gine looks for the first parent (compositionally represented)the partonomy of which has been explicitly expressed (lung,in this case). Then the system takes into account the conceptlobe as the first level part of the concept lung, attemptingto associate each specific lobe (superior, inferior, and middlelaterality from the Composition table) to left lung, applying
inheritance. The operator VerifyFacts (see the Appendix),which implements a consistency rule based on heuristicknowledge between the two matched concepts, excludes thegeneration of possible unreal concepts. Heuristic knowledgehas been stored in a data structure named Facts in whicheach entry indicates an explicit block of inheritance betweentwo concepts (see Table 5). In particular, by finding an entry
… … …CC00rs001 (left lung) CCfe342112 (medial (is-part-of)

segment)CC00rs001 (left lung) CC00asb22 (middle lobe) (is-part-of)… … …

Note. For the entries the inheritance is to be blocked.

in the Facts table that links left lung to middle lobe, inferenceavoids the generation of the concept middle lobe of theleft lung. By moving to segments (apical, lingular, basal,posterior, anterior, medial, etc.) the inference engine repeatsthe control, excluding, for example, medial segment fromsuperior lobe (see the Appendix for the pseudo code).

Each time a property must be subsumed from one conceptto another, the VerifyFacts operator verifies from the Factstable whether a possible block exists. This corresponds toassuming that inference is automatically enabled apart fromexplicit coded blocking.

Based on this information reconstruction procedure, thedeveloped software system provides the ability to issue thefollowing requests:

• All anatomical concepts that contain or exactly matcha user-defined key word;

• All the synonyms given a preferred term;• Progressive detailed representations for a concept;• All anatomical concepts that belong to a specified body

region and/or to a specified body system;• The hierarchical tree (taxonomic or partonomic—


TABLE 5

The Facts Table Used to Store Heuristic Knowledge

ConceptCodeA ConceptCodeB RelationCode

• The network of characteristics related to a concept (non-hierarchical relationships);

• Constrained views: all anatomical objects involved ina certain function, located in a specific body region, orsatisfying other particular conditions.

3. USE OF THE DEVELOPED KNOWLEDGE BASE

The utility of the developed anatomical knowledge basehas been demonstrated by assembling a prototypical client

by a user-specified relationship. The result consists of tree-


side interface. In addition to knowledge-base query, the clientside has been conceived to provide symbolic access to spatialinformation, constituted by enhanced features like seg-mented pixel regions or contours, of a reduced set of theimages of the Visible Human Dataset (VHD) [16]. The ap-pealing idea here is that of enabling the user to interactivelyconstruct anatomical statements to recover both symbolicand spatial content. The latter consists of partially segmentedimages in the abdomen from the VHD contained in a testset provided by Gold Standard Multimedia. The image seg-mentation has been refined to cope with the detail levelexpressed in the AnatomicalKnowledge database. In particu-lar, each labeled pixel has been assigned to the correspondinganatomical entity (concept code) with the highest level ofgranularity. For example, pixels of the region imaging theleft lung will never be assigned to the left lung concept—thatwould cause information loss—but rather will be assignedto a composite concept like anterior basal segment of theinferior lobe of the left lung. Therefore, at run time whensuch pixels will be picked, the corresponding concept and itsproperties can be reconstructed with any (even user defined)granularity level.

Spatial information has been structured in such a way asto take into account: (a) pixel membership, (b) contours ofanatomical structures, and (c) relations between labeled pix-els and AnatomicalKnowledge concept codes. With respectto the first two points, two types of files have been generatedfor each image: a mask file in which each value, correspond-ing to an image pixel, has been assigned by a label to ananatomical structure, and a contour file, containing pixelcoordinate pair lists each corresponding to a structure con-tour (see Fig. 10). The relation between labeled pixels andAnatomicalKnowledge concept codes has been set upthrough a database named ImageMap in which labels areexternally linked to concept codes. The ImageMap databasealso takes into account the fact that an anatomical structurecan appear in multiple images as an image region as wellas a contour.

The overall software system has been conceived as a two-layer framework: the client side that provides visualizationand user-interface facilities and the server side that receives
user input via RMI (remote method invocation) technology,processes user requests, queries the databases through thereconstruction engine, and delivers results to the client side.The client-side application, consisting of a set of panels,provides the user with the ability to visually formulate con-strained queries to the image database ImageMap via differ-ent user interaction modes.
337

Figure 11 shows the operation pipeline for retrieving se-mantic information about an anatomical structure corres-ponding to a pixel picked by the user. First, the server-sideengine, after receiving the picked pixel coordinates and theindex of the currently displayed image, looks in the corres-ponding mask file for the label of the related anatomicalstructure. Then it opens the corresponding contour file to findout the contour (pixel coordinate pair list) of the involvedanatomical structure and from the ImageMap database itobtains the anatomical structure code mapped into the Anato-myKnowledge database. Then, it queries the AnatomyKnowl-edge database to retrieve general information about the in-volved anatomical concept. All these results are thendelivered to the client side where the contour is depictedas superimposed on the image and concept information isvisualized via the interface.

Some different modes are foreseen for querying theknowledge base. A query by term represents the standardsearch by key words: the user can recover the list of termsthat exactly match, contain, or are synonyms of the requiredkeyword. As shown in Fig. 12, the result consists of a listof terms that can be alternatively selected to recover thebasic semantic information about the corresponding concept(definition and taxonomy). Semantic queries provide accessto anatomical concepts that fit a semantic constraint defined

like views in which each single item can be picked to showrelevant information.

A visual browsing panel provides the user with the abilityto retrieve image content corresponding to a selected itemin a tree of concepts obtained after a semantic query (seeFig. 13). In this case, the user has selected the item “left

FIG. 9. Relational model of the knowledge base.

p da
FIG. 10. The relational schema of the ImageMa
kidney.” This action corresponds to retrieving all of the VHDimages that contain that anatomical structure. In this casethe result is many-fold: in the left box, a list of the retrievedimages is displayed; in the central box, containing a sagittalimage of the whole body reconstructed from axial originalimages, a band (yellow lines) is displayed, superimposed
on the image, which indicates the location in the dataset ofthe retrieved images. The red line accounts for the currentlydisplayed image; in the right box, a representative imageof the retrieved image set is visualized at a user-definedresolution (50% in the example of Fig. 13) with a superim-posed contour (in green) of the structure for which the userwas looking.
tabase in connection with the visual data sources.

Then by right clicking in the proximity of the structurecontour, the main semantic information is listed in a floatingbox. Alternatively, the user can pick (via the left mousebutton) any other pixel in the image in the right box toretrieve information about the corresponding anatomicalstructure.


Figure 14 shows a Constrained query panel in which theuser can construct an arbitrary (yet, system-driven) query(in this case “select all muscle in the arm”) and visualizethe results. Then the user can click with the mouse close tothe contour of any structure (“pronator teres, right” in thiscase) to retrieve features about the corresponding anatomicalstructure listed in a floating box.

339
tio
FIG. 11. Operation pipeline to retrieve informa
4. DISCUSSION AND CONCLUSION

This paper constitutes a first-stage attempt to consistentlyrepresent anatomy in a knowledge base. The main noveltyof our development consists of the ability to reproduce het-erogeneous views over anatomy by modeling genus andpartitive properties through hierarchical relations and physi-cal, spatial, and functional attributes through nonhierarchicalrelations. In particular, the work has focused attention onthe problem of modeling anatomical concepts according toa multiple-classification paradigm. Toward this aim, bothconstrained polyhierarchies and context-dependent strict hi-
erarchies have been used. In addition, the explicit representa-tion of partonomic relations has allowed accommodation ofsystemic and topographical views in a unique framework.The power of the representation has also been enriched bydefining several levels of granularity for concepts in termsof abstraction and detail levels as discussed above. This hasallowed accommodation of the user need for manipulating
n after the user has picked a pixel into the image.

the same concept at different levels of specification. Atten-tion has been paid to the inheritance feature by implementinga reconstruction engine that uses heuristic knowledge toexplicitly block wrong inheritance.

We stressed, moreover, the role of the development strat-egy we adopted (compositional approach) by describing theadvantages and weaknesses. The compositional strategyused to organize concepts has resulted in the following ad-vantages:

• Definition of the main semantics directly at the concep-tual level of the knowledge base disregarding specific terms;

• Prevention of explicitly representing huge amounts ofredundant information in contrast to an enumerative strategy;

• Ability to store both semantic and nonsemantic attri-butes in the knowledge base.

However, an ad-hoc supervisor to reconstruct compositeconcepts is required. Some 3000 anatomical concepts havebeen inserted in the AnatomicalKnowledge database alongwith 1500 relations.

340C

ER

VE

RI

AN

DPIN

CIR

OL

I

eft box). Then he/she selectedan ting boxes). Moreover, in thelow

FIG. 12. The “term search” panel. The user has been looking for anatomical concepts into the abdomen. The results was a term list (upper litem (celiac trunk) of the list and the definition appeared in the right centered box. Then synonyms and taxonomy have been retrieved (floaer left box additional concept information has been listed which has been automatically retrieved from UMLS server.

calstru

FIG. 13. The “visual browsing” panel. After a semantic hierarchiitem (left kidney) to look for the VHD images where the correspondingthe contour to show main semantic information (floating box).

With regard to anatomy, we oriented our modeling towardbeginners, so we focused our conceptualization on someorgans, muscles, bones, nerves, and vessels. However, weacknowledge that other organs and particular anatomicalentities cannot be so easily and consistently represented.

The initial tests performed have demonstrated the reliabil-ity and consistency of the developed knowledge base, al-though the concept category should be systematically vali-
dated with the aid of anatomists for the knowledge base toconstitute the kernel of a representation for clinical concepts[20]. Articulated concepts such as “fracture of the left femur”can be easily composed by linking the composite anatomicalconcept left femur with the anatomical clinic concept offracture in the representation.
Similar to other recent efforts [21, 22] aimed at integrating

query performed into “semantic search” panel, the user can select ancture is imaged. Then the user has picked (right mouse button) nearby

spatial and symbolic information, we developed a prototypi-cal user interface to the knowledge base which allows one tosemantically navigate an anatomical image set (2D) derivedfrom the Visible Human Dataset organized into a frameworkcomposed by a database (ImageMap) and two file systems.Although this software system is in the first stage of develop-ment and no systematic evaluation has been carried out todate, it has been used to demonstrate the reliability and


consistency of the developed knowledge base. Moreover, ithas the great advantage of providing the ability to access(constrained views of) image content through arbitrary user-generated statements (e.g., retrieve all muscles in the leftarm) for which typical results are shown by Figs. 13 and14. As a limitation, the visual user interface lacks three-dimensional visualization. We recognize this as a funda-

FIG. 14. The “constrained query” panel. The user can build a semantthe right panel side. In this case, the constrained query is: “select all musccontour of structure “pronator teres, right.” In this case, a query is submitted to

mental requirement for such types of applications. Therefore,future developments of the proposed work will include im-provement of the anatomical concept organization and intro-duction of 3D spatial information of the indexed anatomi-cal structures.


ic query by incrementally selecting constraints (query by example) inle in the arm.” Then the user has picked with the mouse close to theretrieve semantic characteristics of the involved structure (floating box).

In conclusion, we believe that the proposed work repre-sents a suitable contribution in the field of medical informa-tics in terms of computer-based knowledge representationand semantic access to images.

43
APPENDIX

(A)

Here we report the definitions of the two major relations (obtained from UMLS server: http://umlsks3.nlm.nih.gov)

::is-a-kind-of: This is the basic hierarchical link in the semantic network. If a concept “is-

a-kind-of” another concept, then the first concept is more specific in meaning than the

second concept.

::is-part-of: Composes, with one or more other physical units, some larger whole. This

includes component of, division of, portion of, fragment of, section of, and layer of.

The definition of the other relations does not differ from standard definition.

(B)

The following pseudo code shows the main classes and operators devoted to information reconstruction:

ConceptClass::GetParents(RelationClass relationObj){

string sqlCode 5 SELECT ParentConceptCode FROM NetworkWHERE ConceptCode 5 this.code AND

RelationCode 5 relationObj.codeUNION SELECT ConceptCode FROM Composition

WHERE CompositeConceptCode 5 this.code ANDSemantics 5 TRUE

result 5 RunQuery(sqlCode)}Given a relation object it builds the SQL query to recover the parents of theConceptClass and issues it.ConceptClass ConceptClass::GetNextParent( ){

if result !5 NULLConceptClass parentConcept;parentConcept.code 5 result.GetNextValue( );parentConcept.GetInfo( );return parentConcept

elsereturn NULL

}After issuing the query GetParents it retrieves the parent object.ConceptClass::GetConceptfromTerm(TermClass termObj)
{
for i 5 1 : termObj.nltemsif ( IsConcept(termObj.ParseTerm(i), this.code ))

GetInfo( );return 0

return 1}

I
344 CERVERI AND PINCIROL
Given a TermClass object find, if existing, the corresponding ConceptClass object.ConceptTreeClass::GetAncestors(ConceptClass conceptObj, RelationClassrelationObj, integer currLevel, integer finalLevel){

ConceptClass conceptObjParent;conceptObj.GetParents(relationObj)

while ( (conceptObjFather5conceptObj.GetNextParent (j)) !5NULL )AddFather(conceptObj, conceptObjFather)if (currLevel ,5 finalLevel)

GetAncestors(conceptObjFather, relationObj, 11 currLevel,finalLevel)

}}It is a recursive method of ConceptTreeClass that attaches to the tree object thetree built with the parents.ConceptTreeArray::GetAscendingTree(TermClass termObj, RelationClassrelationObj, integer finalLevel){

integerlevel 5 0ConceptClass conceptObjif (conceptObj).GetConceptfromTerm(termObj)!50

ENDfor all contexts of relationObj{

ConceptTreeClass conceptTreeConceptTree.GetAncestors(conceptObj, relationObj, level,finalLevel)AddTree(conceptTree)

}return conceptTreeArray

}It reconstructs all the ancestor concept trees (one for each context of the involverelation) for a givenConceptClass object derived by a TermClass object and a RelationClass object.ConceptClass::GetSons(RelationClass relationObj){

string sqlCode;if (relationObj.name55is-a-kind-of)

sqlCode 5 SELECT ConceptCode FROM NetworkWHERE ParentConceptCode 5 this.code AND

RelationCode 5 relationObj.code
UNION SELECT CompositeConceptCode FROM Composition
WHERE ConceptCode 5 this.code AND Semantics 5 TRUE//Run query

this.result 5 RunQuery(sqlCode)

else if (relationObj.name55is-part-of)if (this.isAtomic55TRUE)

345
sqlCode 5 SELECT ConceptCode FROM NetworkWHERE ParentConceptCode 5 this.code AND

RelationCode 5 relationObj.code//Run query

this.result 5 RunQuery(sqlCode)else

ConceptClass atomicConceptObjatomicConceptObj 5 this.GetAtomicConcept( )

sqlCode 5 SELECT ConceptCode FROM NetworkWHERE ParentConceptCode 5 atomicConceptObj.code AND

RelationCode 5 relationObj.code//Run query

this.result 5 RunQuery(sqlCode)}According to the kind of involved relation and the typology of the concept(anatomical/nonanatomical) it builds the corresponding query to retrievesthe children and issues itConceptClass ConceptClass::GetNextSon( ){

if result !5 NULLConceptClass sonConcept;sonConcept.code 5 result.GetNextValue( );sonConcept.GetInfo( );return sonConcept

elsereturn NULL

}After issuing the query GetSons it retrieves the son object.ConceptTreeClass::GetOffspring(ConceptClass conceptObj RelationClassrelationObj, integer currLevel, integer finalLevel)){

ConceptClass conceptObjSon

conceptObj.GetSons(relationObj)while ((conceptObjSon5conceptObj.GetNextSon( ))!5NULL){

if ((conceptObj.isAtomic( )55FALSE) AND (relationObj.name!5is-a-kind-of)){conceptObjSon.GetSons(RelationClass(is-a-kind-of, code))while ((conceptObjSon 5 conceptObj.GetNextSon( ))! 5 NULL)

{if (VerifyFacts(conceptObj, conceptObjSon) 55 FALSE)

{AddSon(conceptObj, conceptObjSon)if (currLevel ,5 finalLevel)GetOffspring(conceptObjSon, relationObj, 11currLevel, finalLevel)

}

LI
346 CERVERI AND PINCIRO
}}else{

if (VerifyFacts(conceptObj, conceptObjSon) 55 FALSE){ AddSon(conceptObj, conceptObjSon)if (currLevel ,5 finalLevel)GetOffspring(conceptObjSon, relationObj, 11currLevel, finalLevel)

}}

}}It is a recursive method of ConceptTreeClass that attaches to the tree object thetree built with the children.ConceptTreeArray::GetDescendingTree(TermClass termObj, RelationClassrelationObj, integer finalLevel){

integer level 5 0ConceptClass conceptRootif (conceptRoot.GetConceptfromTerm(termObj)!50)

ENDfor all contexts of relationObj{

ConceptTreeClass conceptTree;conceptTree.AddNode(conceptRoot)conceptTree.GetOffspring(conceptRoot, relationObj, level,finalLevel)

AddTree(conceptTree)

tre

s o

}return conceptTreeArray;

}It reconstructs all the offspring conceptrelation) for a givenConceptClass object derived by a TermClas

REFERENCES

1. Cimino JJ. Desiderata for controlled medical vocabularies in thetwenty-first century. Methods Inform Med 1998;37:394–403.

2. Cimino JJ. From data to knowledge through concept-oriented ter-minologies: experience with the medical entity dictionary. J Am
Med Inform Assoc 2000;7:288–97.
3. Ingenerf J, Giere W. Concept-oriented standardization and statis-tics-oriented classification: continuing the classification versus no-menclature controversy. Methods Inform Med 1998;37:527–39.

4. Musen MA, KE Wieckert, Miller ET, Campbell KE, Fagan LM.Development of a controlled medical terminology: knowledge ac-quisition and knowledge representation. Methods Inform Med1995;34:85–95.

es (one for each context of the involved

bject and a RelationClass object.

5. Lindberg D, Humphrey B, McCray A. The unified medical lan-guage system. Methods Inform Med 1993;32:281–9.

6. Bernauer J, Schoop D. Formal classification of medical conceptdescription: graph-oriented operators. Methods Inform. Med.1998;37:510–7.

7. Niemann H, Sagerer GF, Schroder S, Kummert F. Ernest: A seman-tic network system for pattern understanding. IEEE Trans PatternAnal Machine Intell 1990;12(9):883–905.

8. Cote RA, Rothwell DJ, Patolay JL, Beckett RS, Brochu L, editors.SNOMED international: the systematized nomenclature of humanand veterinary medicine. Northfield, IL: College of AmericanPathologist, 1993.

9. Schultz EB, Price C, Brown PJB. Symbolic anatomic knowledgerepresentation in read codes version 3: structure and application.J Am Med Inform Assoc 1997;4:38–48.

10. Rector A.. Compositional models of medical concepts: towards


re-usable application independent medical terminologies. In: Bara- 16. Ackerman, MJ, Spitzer, VM, Scherzinger, AL, Whitlock, DG.hona P, Christensen JP, editors. Knowledge and decisions in health
telematics. Lisbon: IOS Press, 1994; 109–14.
11. Rector AL, Gangemi A, Galeazzi E, Glowinski AJ, Rossi MoriA. The GALEN CORE model schemata for anatomy: towards are-usable application-independent model of medical concepts. In:Barahona P, Veloso M, Bryant J, editors. Proceedings of the 12thInternational Congress of the European Federation for MedicalInformatics (MIE 94), Lisbon: IOS Press, 1994, 229–33.

12. Rector AL, Rogers JE, Pole P. The GALEN high level ontologymedical informatics Europe ’96, Brender J. et al.. editors. Lisbon:IOS Press, 1996, 174–8.

13. Rector AL, Bechhofer S, Goble CA, Horrocks I, Nowlan WA,Solomon WD. The GRAIL concept modeling language for medicalterminology. Artif Intell Med 1997;9:139–71.

14. Rosse C, Mejino JL, Modayur BR, Jakobovits R, Hinshaw KP,Brinkley JF. Motivation and organizational principles for anatomi-cal knowledge representation. J Am Med Inform Assoc 1998;5:17–40.

15. Schubert R, Pommert A, Hohne KH. Methods for model-basedknowledge representation in anatomy. In: Annual Symposium onComputer Applications in Medical Care, Proc SCAMC, New Or-leans, LA, 1995, 971–7.

The visible human data set: an image resource for anatomicalvisualization. Medinfo 1995;8:1195–8.

17. Spence AP and Mason EB. Human anatomy and physiology. NewYork: West Publishing Co., 1992.

18. Balboni GC, et al. Anatomia umana. Milan, Italy:Edi-ermes, 1990.

19. Schiemann T, Nuthmann J, Tiede U, Hohne KH. Segmentation ofthe Visible Human for high quality volume based visualization,In Hohne KH, Kikinis R, editors. Visualization in biomedicalcomputing, Proc VBC 96, Lecture Notes in Computer Science1131, Berlin: Springer-Verlag, 1996, 13–22.

20. Campbell KE, Das AK, Musen MA. A logical foundation forrepresentation of clinical data. J Am Med Inform Assoc1994;1(3):218–32.

21. Brinkley JF, Bradley SW, Sundsten JW, Rosse C. The digitalanatomist information system and its use in the generation anddelivery of web-based anatomy atlases. Comput Biomed Res1997;30:472–503.

22. Pommert A, Schubert R, Riemer M, Schiemann T, Tiede U, HohneKH. Symbolic modeling of human anatomy for visualization andsimulation. In: Robb RA, editor. Visualization in biomedical com-puting. Rochester, MN: Proc SPIE 2359, 1994, 412–23.

1-s2 0-S1532046401910305-main

Documents

Transcript of 1-s2 0-S1532046401910305-main