a prototype symbolic model of canonical functional ...core.ac.uk/download/pdf/82741215.pdf · a...

Available online at www.sciencedirect.com

www.elsevier.com/locate/yjbin

Journal of Biomedical Informatics 41 (2008) 251–263

A prototype symbolic model of canonical functionalneuroanatomy of the motor system

Ion-Florin Talos a,*,1, Daniel L. Rubin b,1, Michael Halle a, Mark Musen b, Ron Kikinis a

a Department of Radiology, Brigham and Women’s Hospital, Harvard Medical School, 75 Francis St., Boston, MA 02115, USAb Stanford Medical Informatics, Stanford University, Stanford, CA, USA

Received 6 February 2007Available online 22 November 2007

Abstract

Recent advances in bioinformatics have opened entire new avenues for organizing, integrating and retrieving neuroscientific data, in adigital, machine-processable format, which can be at the same time understood by humans, using ontological, symbolic data represen-tations. Declarative information stored in ontological format can be perused and maintained by domain experts, interpreted bymachines, and serve as basis for a multitude of decision support, computerized simulation, data mining, and teaching applications.

We have developed a prototype symbolic model of canonical neuroanatomy of the motor system. Our symbolic model is intended tosupport symbolic lookup, logical inference and mathematical modeling by integrating descriptive, qualitative and quantitative functionalneuroanatomical knowledge. Furthermore, we show how our approach can be extended to modeling impaired brain connectivity in dis-ease states, such as common movement disorders.

In developing our ontology, we adopted a disciplined modeling approach, relying on a set of declared principles, a high-level schema,Aristotelian definitions, and a frame-based authoring system. These features, along with the use of the Unified Medical Language System(UMLS) vocabulary, enable the alignment of our functional ontology with an existing comprehensive ontology of human anatomy, andthus allow for combining the structural and functional views of neuroanatomy for clinical decision support and neuroanatomy teachingapplications.

Although the scope of our current prototype ontology is limited to a particular functional system in the brain, it may be possible toadapt this approach for modeling other brain functional systems as well.� 2007 Elsevier Inc. All rights reserved.

Keywords: Functional neuroanatomy; Ontology; Neural network; Motor system; Basal ganglia; Disease model; Parkinson’s disease; Chorea; Hemibal-lism; Neural circuit

1. Introduction

The brain is arguably the most complex organ of thehuman body, and our understanding of its structure andfunction is fragmentary. The past few decades have seenan enormous accumulation of neuroscientific data, makingit impossible for any one individual to comprehend andassimilate more than a fraction of the available data. Thisfact becomes evident while performing a simple search ofthe literature databases, such as Medline. For instance, a

1532-0464/$ - see front matter � 2007 Elsevier Inc. All rights reserved.

doi:10.1016/j.jbi.2007.11.003

* Corresponding author. Fax: +1 617 582 6033.E-mail address: [email protected] (I.-F. Talos).

1 These authors contributed equally to this work.

recent search in the Medline database performed by theauthors, entering the term ‘‘brain’’, yielded 1,003,745entries (http://www.ncbi.nlm.nih.gov/entrez, accessed on01/25/2007).

Recent advances in bioinformatics have opened entirenew avenues for organizing, integrating and retrieving neu-roscientific data, in a digital format, which can be at thesame time understood by both humans and machines,using ontological, symbolic data representations [1]. Inaddition to providing an understanding of the physicalorganization of the nervous system, neuroanatomy mayalso serve as a common frame of reference for organizingall types of neuroscience data [2].

http://www.ncbi.nlm.nih.gov/entrez

mailto:[email protected]

252 I.-F. Talos et al. / Journal of Biomedical Informatics 41 (2008) 251–263

In a general sense, there are two complementary viewsof neuroanatomy: (a) a structural view, concerned withshape, dimensions, spatial location and relationships, andembryologic origin of neural structures, and (b) a func-

tional view, dealing with functional (physiologic) relation-ships between entities assembled into neural functionalsystems (connections between these entities via neuralpathways, physiologic actions—e.g. excitation or inhibi-tion—of one entity on another exerted via neural path-ways); these entities often do not share a commonembryologic origin and may be spatially remote. The twoperspectives are by no means mutually exclusive. On thecontrary, they must be viewed as complementary, as bothare essential for problem solving, in the basic as well asclinical neuroscience domain.

The Foundational Model of Anatomy (FMA), devel-oped by Rosse and colleagues, is an excellent example ofmodeling the structural view of a biomedical domain [3].The FMA is a comprehensive domain ontology describingthe concepts and relationships that pertain to the structuralorganization of the human body, including the nervous sys-tem, from the molecular to the macroscopic level. TheFMA has been successfully used for developing knowl-edge-based applications that rely on inference to supportclinical decision-making, such as reasoning about conse-quences of penetrating chest injuries [4–8].

Structural information contained in the FMA enablesqueries such as: ‘‘which structures are adjacent to or contin-

uous with other structures?’’, or ‘‘what are the parts of a par-

ticular anatomic structure?’’. However, structuralinformation is insufficient to support logical inferences onthe functional consequences of anatomic lesions.

The entities in the FMA are grouped according toembryologic origin and spatial adjacency criteria. Func-tional systems in the brain, on the other hand, include mul-tiple, often spatially remote, and embryologically unrelatedstructures. For instance, from a structural point of view,the striatum and globus pallidus are part of the telenceph-alon, the ventral anterior nucleus of thalamus and the sub-thalamic nucleus are part of the diencephalon, whilesubstantia nigra is part of the midbrain. From a functionalperspective, all these structures are part of a single subcor-tical neural network, controlling the initiation of voluntarymovement (Fig. 1).

The purpose of the present study was to develop a pro-totype symbolic model of canonical functional neuroanat-omy of the motor system in an ontology, representingnormal and disease states, based on a set of declared prin-ciples, Aristotelian definitions, and a high-level schema. Wechose the motor system because it displays little anatomicvariability and its function is better characterized than thatof other, more complex functional systems in the brain.Furthermore, the motor system is involved in a host ofpathologic processes with high impact on public health,such as movement disorders.

Our current prototype ontology, based on the principlesof functional connectionism, is limited to a single func-

tional system. However, since functional connectionismapplies to the entire brain, it may be possible to extendthe modeling approach we describe in this paper to otherbrain functional systems as well. As we will show in the fol-lowing sections, the ontological modeling approach we areproposing can be employed to describe normal, as well aspathologic states.

Symbolic models of functional neuroanatomy, alone orin combination with MRI-based digital brain atlases, couldopen the way for developing knowledge-based applicationsfor clinical decision support (e.g. surgical planning applica-tions capable to identify potential targets for functional ste-reotactic surgery), and computer applications forneuroanatomy teaching.

2. Materials and methods

First, we extracted the relevant functional neuroana-tomical information from authoritative neuroscience text-books [9–11]. In addition, we collected information onmovement disorders that result from impaired connectiv-ity between key anatomical components of the motorsystem.

Next, we created an ontology of functional anatomy ofthe motor system, based on the anatomic knowledge weextracted in the previous step. The ontology was createdusing a disciplined modeling approach, inspired by thatadopted by the developers of the Foundational Model ofAnatomy [3]. Our ontology was implemented using theProtege suite of tools (http://protege.stanford.edu), aframe-based ontology editor.

In this section, we provide a brief overview of the func-tional anatomy of the motor system. We then describe thetheoretical framework and the principles we employed fordeveloping our prototype functional ontology.

2.1. Functional anatomy of the motor system

Older classifications used to make a distinction betweenthe ‘‘pyramidal’’ and ‘‘extrapyramidal’’ motor system.Modern neuroscience has outgrown this rather simplisticview, and it is currently well established that the concertedaction of both these strongly interconnected groups of neu-ral circuits is required for normal motor function.

2.1.1. The ‘‘Pyramidal’’ systemThis functional system is responsible for voluntary

movement. From a functional perspective, it is composedof two major types of neurons: upper motor neurons andlower motor neurons.

The upper motor neurons (UMN) are represented bythe giant pyramidal neurons (Betz’s cells), located in theinternal pyramidal layer (layer V) of the primary motorcortex (precentral gyrus, motor strip).

The lower motor neurons (LMN) are located in thebrainstem (motor nuclei of cranial nerves), and in the ante-rior horn of the spinal cord gray matter.

http://protege.stanford.edu

Fig. 1. Functional connections (red arrows) between subcortical nuclei, relevant to motor function (motor initiation network). The different nuclei arerepresented as part of the FMA class to which they belong (yellow). (For interpretation of the references to colors in this figure legend, the reader isreferred to the web version of this paper.)

I.-F. Talos et al. / Journal of Biomedical Informatics 41 (2008) 251–263 253

The precentral gyrus is located in the posterior frontallobe. The central sulcus separates it from the postcentralgyrus (site of the primary somatosensory cortex). Medially,it is contiguous with the paracentral lobule, while inferiorlyit is separated from the superior temporal gyrus by the lat-eral fissure (Sylvius). In the primary motor cortex, the dif-ferent regions of the body are represented in asomatotopical fashion, with the foot and leg area locatedon the medial aspect of the cerebral hemisphere, followedby the trunk, upper limb and face areas, from medial to lat-eral (‘‘motor homunculus’’). The surface of the cortical rep-resentation is not proportional with the size of therespective body region, but rather with the complexity ofmovements performed by a particular part.

The axons of the upper motor neurons form the cortico-spinal (pyramidal) tract, which descends through the cor-ona radiata and converge in the posterior limb of theinternal capsule. After passing the internal capsule, theseaxons continue their descent through the ventral brain-stem. At the brainstem level, some of these axons cross-over to the contralateral side and synapse with lower motorneurons located in the motor nuclei of the cranial nerves.From a clinical perspective, it is important to point out thatthe dorsal part of the facial nerve nucleus receives bothcontralateral and ipsilateral cortical input.

In the ventral medulla oblongata, the corticospinalfibers converge into two compact tracts, prominently visi-ble at the surface (pyramids). At the junction betweenmedulla and spinal cord, most (80–90%) of these axonscross-over to the contralateral side (pyramidal decussa-tion), and continue their descent in the lateral spinal cord(lateral corticospinal tract). The uncrossed axons continuetheir descent in the anterior spinal cord, as the anterior cor-ticospinal tract. This smaller contingent of fibers cross-overat the spinal cord level, shortly before reaching their targetlower motor neurons. Most corticospinal axons synapsewith their target lower motor neurons via spinal interneu-rons, while a smaller fiber contingent synapses directly onlower motor neurons.

2.1.2. The ‘‘Extrapyramidal’’ system (motor initiation

system)

The motor initiation system consists of a family of par-allel circuits linking subcortical structures with the motorcortex. Its principal components are the basal ganglia (stri-atum, globus pallidus), the subthalamic nucleus, the ventralanterior nucleus of thalamus, the substantia nigra and themotor cortex.

Input from the motor cortex (glutaminergic) reaches thebasal ganglia via the striatum (caudate nucleus and puta-


men). The basal ganglia process this information and pro-ject back to the motor cortex, via the internal pallidal seg-ment and ventral anterior thalamic nucleus. The onlyoutput from the basal ganglia is inhibitory, and it origi-nates in the GABA-ergic medium-spiny neurons of theinternal pallidal segment.

The state of activity in the basal ganglia is regulated viatwo dopaminergic projection systems, originating in thesubstantia nigra pars compacta: the direct and the indirectprojection systems (a.k.a. direct and indirect pathways).The facilitating effects on movement of the direct projec-tion system are mediated by D1-dopamine receptors. Theindirect projection system has an inhibitory effect on move-ment, mediated by D2-dopamine receptors (Fig. 2).

2.2. Diseases of the ‘‘Extrapyramidal’’ system (motorinitiation system)

Lesions of the basal ganglia frequently result in move-ment disorders:

Parkinson’s disease is a classical example of hypokinetic

movement disorder. Clinically, this condition is character-ized by impaired initiation of movement (akinesia), reduced

Fig. 2. Functional organization of the motor initiation neural network. GPi sethalamus, in turn, sends excitatory projections to the cortex. The subthalamiarrows) facilitates movement by directly inhibiting the GPi, and thus disinhibarrows) inhibits movement, by inhibiting the GPe. This results in a decreased inan increased inhibition on the thalamocortical projections (black arrow). Thespinal projection fibers (pathways) originating in the motor cortex. (For interprto the web version of this paper.)

velocity and amplitude of movement (bradykinesia), restingtremor and increased muscle tone (rigidity).

The prevalence of idiopathic Parkinson’s disease is esti-mated at 128–168 cases per 100,000 [12]. There is a dra-matic increase in Parkinson’s cases with increasing age.About 1% of the population aged 50–64 years suffers fromParkinson’s disease. This rate increases to 14.9% in the 65–74 years age group and 52.4% of individuals over 85 yearsold [13].

Pathophysiologically, degeneration of dopamine pro-ducing cells in the substantia nigra pars compacta leadsto a decrease in the activity of the direct basal gangliapathway relative to the indirect pathway activity. This,in turn, results in an increased inhibitory output fromthe internal pallidal segment (globus pallidus pars interna,GPi).

The therapeutic approaches to Parkinson’s diseaseinclude pharmacologic agents (dopamine agonists), andstereotactic functional surgery, such as ablation of theinternal pallidal segment (GPi).

Hyperkinetic movement disorders are characterized byexcessive, uncontrollable motor activity resulting in abnor-mal, involuntary movements of the extremities, head and

nds inhibitory projections to the thalamus (ventral anterior nucleus). Thec nucleus sends excitatory projections to GPi. The direct pathway (greeniting the thalamus (ventral anterior nucleus). The indirect pathway (red

hibition of the GPi and of the subthalamic nucleus, which, in turn results inyellow arrows represent cortico-striatal, cortico-subthalamic and cortico-

etation of the references to colors in this figure legend, the reader is referred


trunk, and decreased muscle tone (hypotonia). Dependingon the anatomic structure affected, the involuntary move-ments can take the form of writhing movements of thearms and hands (athetosis), brief, non-rhythmic move-ments spreading from one muscle group to the next (cho-

rea), violent, large amplitude movements of the proximallimbs (ballism).

Huntington’s disease is a typical example of hyperkineticmovement disorder. It is an inherited, autosomal-dominantdisorder with complete penetrance. Its prevalence in theUnited States is estimated at about five cases per 100,000.Huntington’s disease is a slowly progressive condition,leading to death of the affected individual within 15–20years from symptom onset [14].

The responsible gene is located on chromosome 4.This gene encodes a large protein, huntingtin. The func-tion of huntingtin has yet to be characterized. It is theo-rized, however that huntingtin plays a role in triggeringapoptosis (programmed cell death) in certain neuronpopulations in the central nervous system, including thebasal ganglia.

Fig. 3. Pathologic alterations of basal ganglia neural circuits. The affected pagray matter nuclei appear is gray. See also Fig. 2 for reference.

As opposed to Parkinson’s disease, the inhibitory inter-nal pallidal output is decreased in hyperkinetic disorders,such as chorea and hemiballism.

Disease-induced alterations of the basal ganglia neuralcircuits in Parkinson’s disease, chorea and hemiballismare presented in Fig. 3.

2.3. Ontological modeling

In developing our ontology, we adopted a disciplinedmodeling approach, as described by Rosse and colleaguesin their seminal work on the Foundational Model of Anat-omy (FMA) [3]. This development was accomplished in afour step process: (1) establishing the appropriate theoret-ical framework and identification of the biological con-cepts, attributes and relationships that will form thebuilding blocks of the symbolic representation; (2) defininga rational modeling approach, the elements of the symbolicmodel structure (high-level schema), and the set of proper-ties and modeling rules to be employed; (3) identification ofan appropriate software authoring system, that will not

thways are represented with interrupted arrows, the names of the affected


only allow, but also enforce the modeling rules; (4) evalua-tion of the symbolic model. Our ontology was implementedusing the Protege suite of tools (http://protege.stanford.edu).

2.3.1. Theoretical frameworkThe functional organization principles of ‘‘cellular con-

nectionism’’ (Wernicke, Sherrington, Cajal) provide thebiological foundation for our symbolic model:

(1) Neuron doctrine: the elementary signaling unit in thenervous system is the neuron. Each neuron is a dis-tinctive cell with distinctive processes (multiple den-drites, one axon).

(2) Dynamic polarization principle: the signal (actionpotential) flow in the neuron is unidirectional, fromthe dendrites to the cell body to the axon.

(3) Connectional specificity principle: each neuron is con-nected with certain other neurons (target cells), butnot with others. The connections between neuronsprovide the physical basis for signal processing, i.e.for brain function. Neurons are arranged in specificfunctional groups or neural networks (e.g. primarymotor and somatosensory cortex, subcortical nuclei).Each neural network is concerned with specific ele-

mentary signal processing operations, as part of a spe-cific neurologic function. These groups of neurons arelinked via serial and parallel connections (neuralpathways). Specific brain functions are divisible intoelementary signal processing operations, performedby specialized neural networks. Consequently, abrain functional system (e.g. motor system) can beviewed as a collection of specialized neural networks.

(4) The signals are transmitted from one neuron to theother via synapses. This process is mediated by chem-ical neurotransmitters. Neurotransmitter release maylead to depolarization (excitation) or hyperpolarization

(inhibition) of the postsynaptic membrane. Forinstance, glutamate is the most common excitatoryneurotransmitter in the nervous system, while gaba-amino-butyric acid (GABA) is the most commoninhibitory neurotransmitter.

(5) The action potential is the universal mechanism forsignal transmission in the nervous system. The speci-ficity of a given signal (e.g. motor, sensory) is deter-mined by the neural network(s) it travels through.

(6) Somatotopic representation: individual parts of thebody are represented at specific sites in the motorand somatosensory systems.

In summary, elementary signal processing operations arethe building blocks of brain function, and neural networks

provide the anatomical basis for these operations, i.e. theyrepresent the anatomo-functional units of the nervous sys-tem. The defining feature of a network is the link (connec-tion) between its nodes. As detailed in the following

sections, the neural network concept plays a central rolein our symbolic representation.

2.3.2. Disciplined modeling

One important requirement for ontological models isthat they rigorously and consistently conform to thedomain they are designed to model [3,15–19]. A commonpitfall when developing symbolic models of anatomy ismixing spatial and functional information in the samehierarchy [3,15,20]. In order to avoid this problem, thescope of our ontology is strictly limited to representingfunctional anatomy, specifically the functional anatomyof motor neural networks (abstraction level principle) [3].Consequently, our design plan conforms to a functionalcontext (unified context principle) [3], and the entities ofthe symbolic model, along with their attributes and rela-tionships are defined in respect to their functional role,and not to spatial relationships or embryologic origin cri-teria (definition principle, relationship constraint principle)[3].

Defining the dominant concept (dominant concept princi-

ple) and the organizational unit (organizational unit princi-ple) [3] are crucial steps in ontology design. The dominantconcept serves as reference for defining all classes (con-cepts) of the ontology.

As discussed in the previous section, the neural networkis the anatomo-functional unit of the nervous system.Hence, in order to represent brain function, we identifythe ‘‘neural network’’ concept as the dominant concept ofour ontology. We define this concept as follows:

Neural network is a biological entity consisting of neu-rons and their processes, connected by synapses in a spe-cific, genetically determined pattern, which performsspecific elementary signal processing operations, and con-stitutes the functional organization unit of the nervoussystem.

Microscopic neuronal networks are grouped togetherinto functionally specialized, gray matter structures (sub-cortical nuclei, cortical areas). Gray matter structures areconnected in a specific, genetically determined pattern,via long axonal processes (neural pathways, fiber tracts),into functionally specialized macroscopic neural networks

(e.g. networks of subcortical nuclei).The organizational unit of our symbolic model is the

neuron, since all elements (classes) of our symbolic rep-resentation can be derived from neurons or parts ofneurons. At its current stage of development, ourontology represents anatomic structures at macroscopicscale.

In accordance with the content constraint principle [3],the largest entity that can be modeled is a functional system(collection of macroscopic neural networks performing ele-mentary signal processing operations as part of a specificbrain function, e.g. motor function), and the smallest entitythat could be represented is a biological molecule con-cerned with synaptic transmission in the nervous system(neurotransmitter).




2.3.3. High-level schema

The general structure (high-level schema) of our ontol-ogy consists of the following elements: anatomo-functional

abstraction, neural network taxonomy, and neural network

component taxonomies (nodes taxonomy, connectionstaxonomy).

2.3.3.1. Anatomo-functional abstraction. The anatomo-functional abstraction describes the properties and rela-tionships between the entities of the functional ontology.

There are two types of relationships that need to be rep-resented in an ontology of functional anatomy: meronymic

(partitive) and functional (physiologic) relationships.(1) Meronymic (‘‘is part of’’, ‘‘has parts’’) relationships.

Each neural network is part of a functional system (e.g.motor system, sensory systems), and has parts, i.e. networknodes (subcortical gray matter nuclei/cortical areas), andnetwork connections (white matter fiber tracts, neuralpathways), respectively.

Examples:

‘‘The motor initiation neural network ispart of the motor system.’’‘‘The motor system has parts: the voluntary

motor neural network and motor feedback

neural networks.’’

‘‘Globus pallidus is part of the motor initia-tion network, and thus it is part of the motorsystem.’’

White matter structures (white matter fiber tracts, neuralpathways) are also parts of a neural network.

(2) Functional (physiologic) relationships. Neural net-works provide the anatomic basis for neurologic function.Typically, the physiologic influence of one network nodeon another is conveyed via neural pathways (neural net-work connections). A pathway can only have one origin(cortical area/subcortical nucleus) and one target (corticalarea/subcortical nucleus), and it conveys one type of phys-iologic effect to the target (excitation, or inhibition). Thesignal propagation is unidirectional, from the origin tothe target (dynamic polarization).

Example:

‘‘The internal pallidal segment exerts an

inhibitory influence on the ventral anterior thalamicnucleus; this inhibitory influence is conveyed via the pal-lido-thalamic pathway.’’

Since neural pathways convey the physiologic action ofthe origin node to the target node, they can be representedas verbs in the syntactic structure of the ontology, while theorigin node can be represented as subject, and the targetnode as object (Fig. 4).

2.3.3.2. Neural network taxonomy and neural networkelement taxonomies. Since neural network elements cannotbe represented in the form of ‘‘is a’’—subclasses of the gen-

eric neural network class, we accomplished the goal of cre-ating a comprehensive class-subsumption hierarchy in anoperational manner.

First, we created a template taxonomy of neural net-works, with ‘‘neural network’’ as its root class (Fig. 5).

Specific neural networks belonging to particular func-tional systems (e.g. motor neural network, somatosensoryneural network) are represented as subclasses of the rootclass. We further elaborated the ‘‘motor neural network’’class, by adding the following subclasses: ‘‘voluntary move-ment neural network’’, and ‘‘motor feedback neural net-work’’. The latter subclass also has two subclasses:‘‘motor initiation neural network’’ and ‘‘motor modulationneural network’’. The right and left motor initiation neuralnetworks are represented as instances of the ‘‘motor initia-tion neural network’’ subclass.

Second, using the template neural network taxonomy,we created separate taxonomies for each network elementtype (nodes and connections, respectively).

The classes in the nodes taxonomy were given the fol-lowing properties: ‘‘is part of’’ (neural network), ‘‘output’’,‘‘input’’. The ‘‘input/output’’ slots can only be filled withinstances of classes from the network connectionstaxonomy.

One important feature of neural networks participatingin the same brain function is the fact that they share someof their nodes. For instance, the motor cortex is at the sametime part of the voluntary motor network and of the motorinitiation network. The basal ganglia send output back tothe motor cortex, where the initial motor command origi-nates (see Fig. 1). We accounted for this feature by creatinga ‘‘shared node’’ subclass of the ‘‘motor neural networknode’’ class.

The classes in the connections taxonomy were given thefollowing properties: ‘‘origin’’, ‘‘target’’, and ‘‘physiologic

effect’’ (excitation/inhibition). The ‘‘origin’’ and ‘‘target’’slots can only be filled with class instances from the nodestaxonomy.

In the final step, we combined the two network elementtaxonomies into a comprehensive functional ontology. Inthis design, the classes of the nodes taxonomy fill slots ofclasses in the connections taxonomy (Figs. 6 and 7).

2.3.3.3. Representing somatotopy and pathway cross-over.

Specific sites of the motor system control the motor activityof different body regions (head, neck, trunk, limbs). This iscalled somatotopic representation. For example, the corti-cal area located on the medial aspect of the precentralgyrus controls voluntary movement of the contralateralleg, while voluntary movement of the face muscles is con-trolled by an area located in the inferior portion of the pre-central gyrus. In order to represent somatotopy in ourontology, we added instances to the ‘‘neural networknode’’ and ‘‘neural network connection’’ classes, accordingto the body regions they represent. For example, we addedinstances to the ‘‘motor cortex’’ class corresponding to leg,trunk, upper limb, lower limb, etc.

Fig. 4. Basic syntactic and ontological representation of functional relationships in the brain.

Fig. 5. Template neural network taxonomy. Subclasses at the same level of the hierarchy are represented using the same color. (For interpretation of thereferences to colors in this figure legend, the reader is referred to the web version of this paper.)


Due to the fact that the corticospinal tract crosses overto the contralateral side, voluntary movement of the twobody halves is controlled by the contralateral motor cortex.Pathway cross-over is represented in our ontology in thefollowing manner:

origin node (right side) fi pathway

(instance of neural network connec-

tion) fi target node (left side).

3. Results

Using the approach presented in the previous sections,we developed of a prototype ontology describing the basicfunctional organization of the motor system. There is a

one-to-one mapping from each instance in the ontologyand a gray matter structure in the brain involved withmotor function (e.g. right and left primary motor cortex,right and left putamen, globus pallidus pars interna andexterna, etc.). For example, the internal pallidal segmentis represented as an instance of the NeuralNetworkNodeclass. Furthermore, the connections (neural pathways)between gray matter structures, each having the appropri-ate attributes to specify their functional action on the targetnetwork node (gray matter structure), i.e. excitatory orinhibitory influence, are also represented in our ontology,as instances of the NeuralNetworkConnection class. Theontological representation of the different neural networksthat compose the motor system is accomplished in an oper-ational manner, i.e. the network is reconstructed from itselements: class instances of the neural network nodes tax-

Fig. 6. Symbolic representation of the ‘‘neural network’’ concept. As in the previous figure, the subclasses at the same level of the hierarchy arerepresented with the same color. This representation is accomplished in an operational manner, i.e. the network is ‘‘reconstructed’’ from its elements: classinstances of the neural network nodes taxonomy (dark blue) fill ‘‘origin’’ and ‘‘target’’ slots (gray) of classes in the neural network connections taxonomy.(For interpretation of the references to colors in this figure legend, the reader is referred to the web version of this paper.)

Fig. 7. Functional ontology outline in Protege. Left: class browser showing the neural network connection taxonomy. The VAN_thalamo-corticalpathway (connection) class, representing the pathway connecting the ventral anterior thalamic nucleus (VAN) with the motor cortex is highlighted.Middle: instance browser, showing the instances of the thalamo-cortical pathway. Right: instance editor, showing the properties of the left VAN_thalamo-cortical pathway. The origin and target slots are filled with instances from the neural nework nodes taxonomy (see also Fig. 6).


onomy fill ‘‘target’’ slots of classes in the neural networkconnections taxonomy.

Our modeling approach is suitable not only for repre-senting neural connectivity in the normal brain. It can beeasily extended to represent impaired brain connectivityas well.

We used the prototype ontology described above to cre-ate a symbolic, computable model of neural connectivityand function of the motor initiation network in the normal

brain (Fig. 8B). While this ontology-based symbolic modelhas a very similar appearance to the original graphical rep-resentation from which it was derived (Figs. 1 and 8A), theentire model is a computable representation. Each nodeand connection in the diagram represents an object in theontology. For example, the arc connecting thalamus andcortex is an instance in the ontology in which the thalamusis linked with the motor cortex via an excitatory connec-tion. This instance contains information about the struc-

Fig. 8. (A) Diagrammatic representation of the key nuclei and their connections for the motor initiation neural network. (B) Representation of the nodesand connections in a simplified ontology-based model of this network in Protege. While the model can be viewed in an intuitive graphical form, eachcomponent of this diagram is an ontological object, and the entire network can be analyzed using computer programs.


ture from which it originates and the target structure towhich it projects, as well as the type of physiologic influ-ence it exerts on the target. Likewise, other connectionobjects are instances of an ‘‘inhibitory connection’’ class,meaning that they inhibit nodes to which they connect.

The entire model is computable, and by traversing alllinks, it is possible to compute the net excitation or inhibi-tion of every anatomic structure.

To demonstrate the extensibility of our approach forrepresenting pathologic conditions, we created a secondmodel representing impaired neural connectivity in Parkin-son’s disease, by modifying the normal model. We replacedtwo arcs in the normal model with arcs that are of type‘‘impaired excitatory neural connection’’. These arcs repre-sent neural connections that produce less excitatory outputthan normal. Accordingly, the ontological representationof the Parkinson’s disease model permits us to assess theconsequences of the impaired connectivity—that there willbe increased inhibitory activity in the internal pallidal seg-ment (GPi), and consequently increased inhibition of themotor cortex (Fig. 9). In a similar manner, we could modelthe consequences of impaired connectivity in hyperkinetic

movement disorders, such as Chorea and Hemiballism byreplacing links with the appropriate functional typesaccording to the pathology of the disorder.

The net inhibitory effect of the internal pallidal segmenton the motor cortex, via the anterior thalamic nucleus, isdetermined by the ratio between activation levels of thedirect vs. indirect pathway (ADP:AIDP) [21]. Voluntarymovement can only be initiated when the activation levelof the direct pathway is greater than that of the indirectpathway. In hypokinetic movement disorders, such as Par-kinson’s disease, the ADP:AIDP ratio is lower than normal,whereas in hyperkinetic movement disorders (Chorea,Hemiballism), this ratio is higher than normal.

By attributing arbitrary strength values to the differentnetwork connections (positive for excitatory connections,negative for inhibitory connections), it is possible to createa computer reasoning application that computes the netexcitation levels of the motor cortex under normal andpathologic circumstances, because our ontology-basedmodel of neural connectivity is machine-processable andnet excitation of all nodes can thus be computed (Figs. 8and 9). For example, in the Parkinson’s disease model,

Fig. 9. A simplified model of the neural connections and functional consequences in Parkinson’s disease. (A) Diagrammatic representation. (B)Ontological representation, produced by modifying the model of the normal state (Fig. 8B), by replacing two normal connections with functionallyimpaired connections.


the ontological representation of the functional aspects ofneural network connections would permit a computer rea-soning service to evaluate the net activation in the differentnodes of the MotorInitiationNeuralNetwork andconclude that there is net inhibition of the PrimaryMo-

torCortex node. Accordingly, the value of creating ourontological representation of the functional organizationof the motor system is to make the anatomic and func-tional aspects of neural structures accessible to intelligentcomputer reasoning services.

Such reasoning services, combined with patient-specificimaging-based brain atlases, may be used in creating deci-sion support applications to help surgical planning andpersonalized patient care. Image-based, geometric modelsof brain anatomy provide spatially accurate, implicit repre-sentations of brain structure. However, they lack explicitknowledge about their contents, such as the functional roleof the anatomic structures they represent or the functionalconsequences of pathologic changes. By combining theexplicit functional knowledge provided by ontology-basedmodels with image-based geometric models of brain anat-

omy, it may be possible to develop surgical planning appli-cations designed to predict consequences of injuries tobrain structures resulting from particular surgicalapproaches, or to support identification of appropriate tar-gets for stereotactic functional surgery in movement disor-ders, that can be highlighted on patient-specific imagedatasets.

4. Discussion and conclusion

In this paper, we introduced an ontological modelingapproach of functional neuroanatomy, and presented aprototype ontology of canonical functional anatomy ofthe motor system. Our ontology-based model is intendedto support symbolic lookup, logical inference and mathe-matical modeling by integrating descriptive, qualitativeand quantitative functional neuroanatomical knowledge.We have shown that our approach permits us to generatesymbolic models of impaired brain connectivity underpathologic conditions, such as movement disorders. Ourmethods also provide a computational framework in which


to create applications that can reason about the functionalconsequences of brain injuries.

Our functional ontology shares several important fea-tures with the Foundational Model of Anatomy (FMA)[3]: common vocabulary, common modeling principles,and a common modeling platform (Protege). Aside fromthe limited coverage of our ontology compared to theFMA, one essential difference is the fact that these ontolo-gies provide two different, but complementary views ofneuroanatomy. While the FMA describes the spatial orga-

nization of the nervous system, our ontology describes itsfunctional organization. The structural knowledge of theFMA enables automated identification of anatomic struc-tures that may be affected by certain injuries with a givenspatial distribution. However, structural knowledge is notsufficient to provide an understanding of how the neuralstructures work, or to predict functional consequences ofinjuries.

Recently, parameterized models have been designed forspecific applications, such as simulating the effects of deepbrain stimulation (DBS) on the activity of basal ganglia[22,23]. Such parameterized models incorporate actual clin-ical and experimental observations, and use systems of par-tial differential equations to describe temporal variations ofphysiologic signals in a quantitative manner. These param-eterized models are highly specialized applications devel-oped with the purpose of solving a particular clinicalproblem. They lack explicit declarative anatomic informa-tion about the components of which they are comprised.

To our knowledge, no ontology of functional neuro-anatomy has been developed to date. As opposed toparametrized models, domain ontologies, are repositoriesof coherent, explicit knowledge, stored in a format under-standable by both humans and computers. They are notintended as end-user applications or designed for solvinga specific problem. They represent generalizable and reus-able sources of knowledge about a particular domain.Ontologies of functional neuroanatomy enable qualitativereasoning about functional consequences of abnormalitiesand can serve as basis for a multitude of specific applica-tions, including parameterized modeling of neurologicfunction. Furthermore, they are extensible, and can bemapped to structural ontologies, such as the FMA, as wellas to medical images.

One limitation of our ontology is that it currently coversonly large macroscopic neural components. We are cur-rently working on extending this representation to moregranular levels of anatomic and functional detail.

Another limitation of our prototype ontology is the factthat it currently covers the narrow domain of functionalorganization and abnormal neural connectivity of themotor system. While our ontology is potentially useful toenable different types of intelligent applications, it maynot be able to tackle a broader range of reasoning applica-tions beyond the scope of our focused domain. However,since the principles of functional connectionism, that layat the foundation of our ontological representation, apply

to the entire brain, it may be possible to extend the scope ofour ontological representation to incorporate other func-tional systems as well, with the final goal of creating adomain ontology of functional neuroanatomy.

In summary, in this preliminary report, we have shownthat functional neuroanatomical knowledge about themotor system can be represented using an ontology, whichcan be exploited by computer reasoning applications. Thedeclarative knowledge encoded in our ontology can be per-used and maintained by domain experts, interpreted bymachines, and serve as basis for a multitude of decisionsupport, computerized modeling, and teachingapplications.

Acknowledgment

This work was partially supported by the following NIHGrants: P41RR013218, U41RR019703, U54HG004028,and LM 007885.

References

[1] Blake JA, Bult CJ. Beyond the data deluge: data integration and bio-ontologies. J Biomed Inform 2006;39(3):314–20.

[2] Brinkley JF, Rosse C. Imaging and the Human Brain Project: areview. Methods Inf Med 2002;41(4):245–60.

[3] Rosse C, Mejino Jr JL. A reference ontology for biomedicalinformatics: the Foundational Model of Anatomy. J Biomed Inform2003;36(6):478–500.

[4] Rubin DL, Bashir Y, Grossman D, Dev P, et al. Using an ontologyof human anatomy to inform reasoning with geometric models. StudHealth Technol Inform 2005;111:429–35.

[5] Rubin DL, Bashir Y, Grossman D, Dev P, et al. Linking ontologieswith three-dimensional models of anatomy to predict the effects ofpenetrating injuries. Conf Proc IEEE Eng Med Biol Soc2004;5:3128–31.

[6] Rubin DL, Dameron O, Bashir Y, Grossman D, et al. Usingontologies linked with geometric models to reason about penetratinginjuries. Artif Intell Med 2006;37(3):167–76.

[7] Rubin DL, Dameron O, Musen MA. Use of description logicclassification to reason about consequences of penetrating injuries.AMIA Annu Symp Proc 2005:649–53.

[8] Rubin DL, Grossman D, Neal M, Cook DL, et al. AMIA AnnuSymp Proc 2006:664–8.

[9] Kandel E, Schwartz J, Jessell T. Principles of neural science. 4thed. McGraw-Hill; 2000. 1414.

[10] Nieuwenhuys R, Voogd J, Van Huizen C. The human nervoussystem. Springer; 1980. 253.

[11] Purves D, Augustine G, Fitzpatrick D, Hall W. et al. Neuroscience.3rd ed.: Sinauer Associates; 2004. p. 773.

[12] Schrag A, Ben-Shlomo Y, Quinn NP. Cross sectional prevalencesurvey of idiopathic Parkinson’s disease and Parkinsonism inLondon. BMJ 2000;321(7252):21–2.

[13] Bennett DA, Beckett LA, Murray AM, Shannon KM, et al.Prevalence of parkinsonian signs and associated mortality in acommunity population of older people. N Engl J Med1996;334(2):71–6.

[14] Anderson KE. Huntington’s disease and related disorders. PsychiatrClin North Am 2005;28(1):275–90.

[15] Noy NF, Musen MA, Mejino JL, Rosse C. Pushing the envelope:challenges in a frame-based representation of human anatomy. DataKnowledge Eng 2004;48:335–59.

[16] Bard J. Ontologies: formalising biological knowledge for bioinfor-matics. Bioessays 2003;25(5):501–6.


[17] Burgun A. Desiderata for domain reference ontologies in biomedi-cine. J Biomed Inform 2005.

[18] Smith B. From concepts to clinical reality: an essay on thebenchmarking of biomedical terminologies. J Biomed Inform2005.

[19] Smith B, Ceusters W, Klagges B, Kohler J, et al. Relations inbiomedical ontologies. Genome Biol 2005;6(5):R46.

[20] Bard JB. Anatomics: the intersection of anatomy and bioinformatics.J Anat 2005;206(1):1–16.

[21] Wichmann T, DeLong MR. Functional and pathophysiologicalmodels of the basal ganglia. Curr Opin Neurobiol 1996;6(6):751–8.

[22] Rubin JE, Terman D. High frequency stimulation of the subthalamicnucleus eliminates pathological thalamic rhythmicity in a computa-tional model. J Comput Neurosci 2004;16(3):211–35.

[23] Miocinovic S, Parent M, Butson CR, Hahn PJ, et al. Computationalanalysis of subthalamic nucleus and lenticular fasciculus activationduring therapeutic deep brain stimulation. J Neurophysiol2006;96(3):1569–80.

a prototype symbolic model of canonical functional ...core.ac.uk/download/pdf/82741215.pdf · a...

Documents