- Home
- Documents
*Multivariate statistical methods for the analysis of ...xrm.phys. Multivariate statistical methods*

If you can't read please download the document

View

216Download

2

Embed Size (px)

Multivariate statistical methods for the analysis ofmicroscope image series: applications in materials science

N. BONNETINSERM Unit 314 (IFR 53) and University of Reims (LERI), 21, rue Clement Ader, BP 138,51685 REIMS Cedex, France

Key words. Automatic classification, dimensionality reduction, electronmicroscopy, image series, microanalysis, multivariate image segmentation,multivariate statistical analysis, neural networks, watersheds.

Summary

Multivariate data sets are now produced in several types ofmicroscopy. Multivariate statistical methods are necessaryin order to extract the useful information contained in such(image or spectrum) series. In this review, linear andnonlinear multivariate methods are described and illu-strated with examples related both to the segmentation ofmicroanalytical maps and to the study of variability in theimages of unit cells in high-resolution transmission electronmicroscopy. Concerning linear multivariate statistical ana-lysis, emphasis is put on the need to go beyond the classicalorthogonal decomposition already routinely performedthrough principal components analysis or correspondenceanalysis. It is shown that oblique analysis is often necessarywhen quantitative results are expected. Concerning non-linear multivariate analysis, several methods are firstdescribed for performing the mapping of data from ahigh-dimensional space to a space of lower dimensionality.Then, automatic classification methods are described. Thesemethods, which range from classical methods (hard andfuzzy C-means) to neural networks through clusteringmethods which do not make assumptions concerning theshape of classes, can be used for multivariate imagesegmentation and image classification and averaging.

Introduction

Physical and chemical sensors provide data concerning thephysical and chemical state and composition of specimens.At the microscopic level, even at very high resolution,several different signals (or different attributes of the samesignal) can be recorded simultaneously, thus providingmultivariate data sets. Such multivariate data sets arerecorded more and more often, because of the intuitive ideathat multivariate measurements can provide informationwell beyond the limits achievable with individual measure-ments. However, this intuitive idea becomes a reality only in

the cases where data analysis methods are available forextracting useful compact information from the enormousamount of data recorded. From this point of view, one couldsay that, although several groups of methods are already inuse, there is still a lot to do before the full recordedinformation can be optimally extracted.

Since we are dealing with multivariate data sets, thegeneric term for these data analysis methods (which oftenrely on some kind of statistics) should be multivariatestatistical analysis. However, from the point of view ofterminology, the situation is slightly confused, becausemultivariate statistical analysis (MSA) is the name generallyused for representing a specific group of methods dealingwith the analysis of data sets by linear methods. However,linear methods (such as principal components analysis(PCA), correspondence analysis (CA), KarhunenLoeveAnalysis (KLA), . . .) represent only a small part of allavailable methods; nonlinear methods are also verypromising, and will probably constitute the largest part offurther developments.

In this paper, I will attempt to cover nonlinear methodsas well as linear ones. On the other hand, analysis of datasets (which, as we will see, can be considered as themapping onto a subspace) is not the only piece ofinformation extraction; another aspect consists of theinterpretation of the projections. Although several facetscan again be recognized, I will concentrate on the questionof data (or object) classification. Obviously, this aspect islargely connected to the topics of pattern recognition and,in some way, of artificial intelligence. Supervised classifica-tion methods, which require a preliminary training phase(the classifier is trained with examples constituting thetraining set), can be used. Alternatively, unsupervisedclassification methods can be attempted, where the dataare gathered into several classes without the help of anexpert, on the basis of their information content(signature) only.

Journal of Microscopy, Vol. 190, Pts 1/2, April/May 1998, pp. 218.Received 10 March 1997; accepted 19 September 1997

2 q 1998 The Royal Microscopical Society

It should be stressed that some of the concepts andmethods described in this paper (linear MSA, classificationof images in the representation space) were introduced inelectron microscopy by researchers working in the field ofthree-dimensional reconstruction of macromolecules (vanHeel & Frank, 1981; Frank & van Heel, 1982; van Heel,1984, 1989; Frank, 1990; Borland & van Heel, 1990).Although these methodological contributions and the greatsuccess they have had in recent years in elucidating the 3Dstructure of important biological macromolecules are notdescribed in this paper, their importance in the diffusion ofrelated (although different) techniques for materials scienceapplications should be recognized.

The outline of the paper is the following. In the nextsection, I give some examples of multivariate physical datasets and I introduce those that I will use for illustration inthe rest of the paper. The following section is devoted tolinear multivariate statistical analysis. Since these methodsare described in many textbooks (Lebart et al., 1984) andare already in use in several laboratories, only a briefdescription of them will be given. Emphasis will be put onthe extension of orthogonal MSA to oblique MSA. Thesection thereafter will be devoted to nonlinear mapping, anextension of linear MSA. Several approaches will bediscussed and illustrated, ranging from the minimizationof a criterion (cost function) to neural networks approaches.The last section will be devoted to automatic classification. Iwill concentrate on unsupervised classification, which doesnot mean that supervised classification does not deserveattention. After briefly describing some classical statisticalclassification techniques (which make assumptions con-cerning the shape of clusters in the parameter space), I willput emphasis on new methods which do not makeassumptions concerning the shapes of classes.

Some examples of multivariate data sets

Multivariate data sets produced in the domain of physicalsciences (as well as in other scientific domains) are verydiverse in nature. They can be simple data, series ofspectra, series of two- or three-dimensional images,spectrum-images, etc.

Examples of simple data are:X the results of measurements (concentrations of differentelements for instance) made at different positions on aspecimen. Examples are described in Quintana (1991) andQuintana & Bonnet (1994a,b).X different preparation conditions related to some char-acteristics of the specimens obtained (see for instance thepaper by Simeonova et al. (1996), which concerns theconditions of preparation of high-temperature supercon-ducting thin films).

Examples of multivariate spectra are:X sets of spectra recorded as a function of time (time-

resolved spectroscopy) (Ellis et al., 1985). Examples ofmultivariate statistical analysis of such data sets can befound in Bonnet et al. (1991) and Jbara et al. (1995).X sets of spectra recorded as a function of position, throughan interface for instance (Tence et al., 1995). Examples ofmultivariate statistical analysis of such data sets can befound in Gatts et al. (1995), Mullejans & Bruley (1995),Brun et al. (1996) and Titchmarsh & Dumbill (1996).

Multivariate two-dimensional (2D) image sets include(Bonnet, 1995a):X sets of different elemental (or chemical) maps of aspecimen recorded in different microanalytical modes(Auger, EELS, X-ray emission, X-ray fluorescence, X-raydifferential absorption). Examples of the processing of thistype of data can be found in Bonnet et al. (1992), Prutton etal. (1990, 1996), Cazaux (1993), Quintana & Bonnet(1994a,b), Colliex et al. (1994) and Trebbia et al. (1995). Asan illustration of this type of data, I have selected a series of14 X-ray fluorescence maps of a specimen of granite(courtesy of K. Janssens and collaborators, Department ofChemistry, University of Antwerp: Wekemans et al., 1997).The series is displayed in Fig. 1.X sets of images of unit cells recorded by high-resolutiontransmission electron microscopy (HRTEM) of interfacesbetween two crystals. Such data sets have been analysed(with the purpose of visualizing the gradual change ofcomposition across the interface) by pattern recognitiontechniques (Ourmazd et al., 1990; De Jong & Van Dyck,1990; Kisielowski et al., 1995). Their analysis by multi-variate statistical methods has also begun (Rouviere &Bonnet, 1993; Aebersold et al., 1996). I will also show somepossibilities of this technique. Such a data set is displayed inFig. 2.

Multivariate three-dimensional (3D) images can also beobtained with some microanalytical techniques (secondaryion mass spectroscopy (SIMS) (Van Espen et al., 1992) orfluorescence confocal laser microscopy, for instance).Techniques working with 2D images can be extended to3D images relatively easily, thanks to the increasingcapabilities of computers.

Four-dimensional (4D) image data sets are, for instance,3D images recorded as a function of time, a mode whichbegins to be feasible in fluorescence (confocal or not)videomicroscopy.

Spectrum-images (or a variant of them: image-spectra)will be the multivariate data sets of choice at the beginningof the 21st century (Jeanguillaume & Colliex, 1989).Combining spatial and full spectral information, they willmix the advantages of spectroscopy and microscopy.Although several acquisition procedu