AUTOMATED 3D ANALYSIS OF LARGEBRAIN MRI DATABASES
ALAN C. EVANS
In recent years, the study of gross neuroanatomy and itsrelationship tobehavior and brain functionhas been reener-gized by the advent of imaging techniques and the powerfulcomputational tools with which to analyze high-resolutionthree-dimensional (3D) brain images (10,11,22,52,53).However, such high technology tools often demand thatscientific questions be restated and made more amenable toquantitative analysis. Questions such as ‘‘How much nor-mal variation is there in the size, shape, or location of anindividual brain structure?’’ or ‘‘To what extent does func-tional architecture of the cortex depend on the anatomicboundaries betweenanatomic regions?’’carry with them theassumption that the borders of individual structures can bespecified accurately in any brain. In the past, basic questionsof functional neuroanatomy were difficult to address in asystematic way in the living brain. We have learned muchfrom anecdotal reporting of individual patients with variousforms of brain lesion or from direct cortical stimulationduring brain surgery, but the generalization of individualobservation to the wider population has been confoundedby the normal variation in brain structure itself. There isthen a fundamental interest in understanding the nature of anatomic variability in the population, both for its relation-ship to functional variability and for the potential of usingstructuralabnormality asameasureof development, normalaging, and disease. For instance, in some degenerative dis-eases like Huntington’s disease and Alzheimer’s disease, thesulcibecomemoreopenandtheventriclesbecomeenlarged.Magnetic resonance imaging (MRI)-based measurements of these changes can lead to early diagnosis and treatment, but we need to understand the variation among normal brainsfirst. Although the study of postmortem neuroanatomy is along-established science, the ability to accumulate the num-bersofbrainsnecessarytomakestatisticallymeaningfulcon-clusions about cerebral anatomy is a relatively recent phe-
Alan C. Evans:
Department of Neurology, McGill University, MontrealNeurological Institute, Montreal, Quebec, Canada.
nomenon. It is still difficult to identify reliably in any singlebraintheanatomiclandmarks,boundaries,andotherdelim-iting features necessary for any subsequent analysis. Thus, we face a new problem posed by this newfound technology and its inflexible demand that anatomic questions be posedin numerical rather than descriptive terms. The tools existto image large numbers of brains noninvasively with MRI,but we are still struggling with how to extract the anatomicmeasurements necessary to answer the questions posedabove. It is relatively easy to identify the precentral gyrus,but few researchers attempt to define its ‘‘top’’ and ‘‘bot-tom.’’ Where does the inferior frontal sulcus end? Tradi-tional brain atlases identify brain regions only by pointingto the middle of the region or surface feature, leaving theinterfaces between regions unspecified. Neuroanatomistsdebate the exact boundary of even relatively simple struc-tures such as the thalamus or caudate nucleus. With thiscontext, new initiatives at various laboratories are attempt-ing to standardize and codify the partitioning of the humanbrain into named regions, not without controversy. Tradi-tional neuroanatomists debate among themselves about what parcellation scheme and nomenclature to use. Com-puter scientists argue among themselves about whether touse hierarchical, relational, object-oriented, or some otherformofdatabasestructuretoorganizethebrainparcellation.Both groups tend to misunderstand the importance of theother’s concerns. Neurobiologists or physicians are not usedto thinking in terms of inclusive sets where, for instance,every structure at one level is wholly included within ahigher level organization, where all 3D pixels, i.e., voxels, within the brain space
be labeled as one of the struc-tures in the partitioning scheme, or that the cerebrospinalfluid (CSF) ventricular spaces may be declared as being out-sideof‘‘brain.’’Computerscientiststend toignorethereali-ties that many cortical sulcal features do not exist in every brain, and may be fragmented orhave multiple occurrences.Some sophisticated analytic approaches forquantifying ana-tomic variability assume that a particular landmark can beperfectly identified in any brain when the reality is thaterrors of 5 to 10 mm typically occur, an error that is about