AU2003255745B2 - Method and apparatus for quantifying tissue histology - Google Patents

Description

WO 2004/010862 PCTIGB2003/003367 1 Method and Apparatus for guantifying tissue histologv This invention relates to a method and apparatus for quantifying tissue histology. In particular the invention relates to methods using an analysis of the spectra of remitted light to establish information on the properties of the tissue. The invention is also applicable to methods and apparatus which rely upon a spectral analysis of light remitted, emitted, and/or transmitted from any material or object under test where they have parameters exhibiting wavelength specific optical effects.

There exists the need for a system, which can recover histological parameters from biological tissue in a way which is invariant to the intensity of the incident illumination and scene geometry. It is an objective of the present invention to provide such a technique. Such a system would be of value in systems where the topology of the tissue or image surface in not known a priori. It would also be of value in.a system where the intensity of the illuminating light cannot be assumed constant. Potential applications include but are not limited to imaging and analysis of the tissue of the gastrointestinal track with an endoscope and imaging and analysis of skin over areas where there is a significant change in curvature, such as the face.

A system is currently in existence, which is able to assist clinicians in their diagnosis of melanoma. The technique is based on a patent, international patent application publication number W098/22023. This system is based on the discovery that when the range of colouration of normal human skin is plotted in a standard RGB colour space, it lies on a well-defined surface.

Furthermore, if an abnormality such as dermal melanin is present, the colouration of the skin changes in such a way as to move points away from the surface which describes healthy skin. By incorporating a calibration calculation which allows variation of dermal thickness to be taken into account, the technique is able to detect abnormalities and thus assist clinicians in their diagnosis of melanoma.

CONFIRMATION COPY WO 2004/010862 PCTIGB2003/003367 2 The fundamental principle behind this system is that it is possible to construct a mathematical function that relates image values, measured using a digital camera, to appropriate histological parameters. Using this functional relation, it is possible to obtain the value of each parameter at every point across a given image. A parametric map can then be produced which gives a grey-scale representation of the parameter value across the whole image.

Although this system has been proved to be clinically effective, it requires exact calibration of the illuminating light source and does not take into account any variation in surface geometry. Thus the technique is limited to problems where a probe can be placed in contact with the region of interest.

This ensures that the incident light is controlled and calibrated and that it's angle of incidence remains constant.

The proposed invention relates to a method for imaging tissue in such a way to give quantitative spectral data independently of the surface geometry or the intensity of the illuminating light. This will allow a non-contact form of imaging and analysis which will be applicable to many different applications. The method may be used with the technique described in W098/22023 and subsequent related patents but is not exclusive to it.

The method concentrates upon the analysis of light remitted by the tissue, ie the illuminating light which penetrates the tissue to some depth and is reflected (or scattered or/and absorbed) to different degrees at different depths due to different parameters of the tissue. Effects due to surface reflection are to be eliminated from the analysis.

Substantial work has been carried out to develop image analysis algorithms which are able to identify different objects irrespective of the illuminating light. Many of the techniques developed are based around the linear model WO 2004/010862 PCT/GB2003/003367 3 of surface reflectance as proposed in L. Maloney and B. Wandell, "Color constancy: a method for recovering surface spectral reflectance", J. Opt.

Soc. Am. A 3, 29-33 (1986). This approach is based on the idea that the surface reflectance of any object within an imaged scene can be expressed as a weighted sum of basis spectral reflectance functions: S(X) ZcyS,(X) (1) j=l and that the illuminating light can similarly be expressed as a weighted sum of basis lights. It has been shown that only a small number of basis functions are required to obtain accurate approximations to the surface reflectances of many naturally occurring objects and also the spectral variation of natural daylight.

With this technique it is possible to recover the vector of weighting constants (7j from a vector of image values and thus specify the spectral reflectance of the imaged object at every pixel. Every potential imaged object/object characteristic will have a unique spectral reflectance. Thus, if the spectral reflectance can be determined using a linear model, then the parameter vector can be specified. With this approach it should be possible to recover a parameter vector from the vector of image values at each pixel.

Unfortunately the method is only able to recover the weighting constants oj to within a multiplicative scaling factor and thus cannot be used to specify the exact spectral reflectance and therefore the exact parameter vector.

An approach to geometry-insensitive segmentation of images has been developed in G. Healey, "Using colour for geometry-insensitive segmentation," J. Opt. Soc. Am. A 6, 920-937 (1989), and is based on the idea of normalised colour. With this approach image values are first divided by an estimate of normalised colour. This estimate is based on approximating the incoming signal from colour pixel values by using a finite-dimensional linear approximation to represent the colour signal.

WO 2001010$62 PIG2013.6 4 US~ UinPg these norinalised. values., different metal and dielectric materias. -can bic identified across an imaged scene in which the geometry Varies conlsiderably.

V) A similar technique has been appidoevlaebn injure tI Afromowitz, G. S. -vani Liew anid D. M. [[ejinbacli, "'Clinical evaluation of burn injuries using ant optical reflectaitce technique.", LREE trans. B iorned.

CI ltag. BME-34, [14-127 (1987), and A. Af omowitz, E. B_ Callis D. M.

fleimbach, L. A. Desoto and M. I I. Norton, "Mutitspectrat imaging of bumn wovd i ew clinical intstrument for evalucating burn depth", MEEE Iran- Biomed. Eng. 35, 8.42-8-49 (1988)). In this case, KGB inilage valuies w ere.

normalised -by dividing themi by the, response oan1kfle.Po th normalised values it was possible to assess the extent of burn damage across 4 given area. of imaged skin- T66re exists a- need. for a non-invasive tedtitiqtue for anialysing an object or.

material (Which may be comp-lex, tor example. niulti-comiponent and/or.

multi-layer ahd which. may be- solid. gaseous, liquid, etc) which- does* not require. ciibration to take into account cliangi~g illumination cdndition's.

Summary of the Invention According to a first aspect of the invention there is provided, an apparatus for Sdetermining the distribution of chromiophores and/or the thickness of structural t 5 layers in a sample of epithelial tissue, the apparatus comprising: n means for illuminating a sample of epithelial tissue with polarised light m having wavelengths falling within a first, second, and third predetermined O wavebands; a polarising filter positioned so as to filter light remitted from a sample of epithelial tissue said .polarising filter being such to filter out light polarised in the manner generated by said means for illuminating a sample of epithelial tissue with polarised light; an image generator operable to detect filtered remitted light from a sample of epithelial tissue and generate image data indicative of the intensity of filtered remitted light received by said image generator having wavelengths falling withii said first, second and third predetermined wavebands; ratio determination means operable to process image data generated by said image generator to determine for positions within images represented by said image data, a first ratio corresponding to the ratio of light received by said image generator having wavelengths within said second waveband relative to light having wavelengths within said first waveband, and a second ratio corresponding to the ratio of light received by said image generator having wavelengths within said third waveband relative to light having wavelengths within said first waveband; concentration determination means operable to determine for positions within an image represented by image data generated by said image generator the concentrations of chromophores and/or the thickness of structural layers of said epithelial tissue at said positions in a sample of epithelial tissue represented by said image data utilising said first and said second ratios determined for said positions by said ratio determination means; and output means operable to output data representing determined concentrations of chromophores and/or thickness of structural layers for points a sample of epithelial tissue as determined by said concentration determination means.

According to a second aspect of the invention there is provided a method for determining the distribution of chromophores and/or thickness of structural layers in a sample of epithelial tissue, the method comprising: illuminating a sample of epithelial tissue with polarised light having wavelengths falling within a first, second, and third predetermined wavebands; filtering light remitted from a sample of epithelial tissue so as to filter out light polarised in the manner corresponding to the polarised light utilised to illuminate the sample of epithelial tissue; generating image data indicative of the intensity of filtered remitted light having wavelengths falling within said first, second and third predetermined wavebands; processing generated image data to. determine for positions within images represented by said image data, a first ratio corresponding to the ratio of filtered remitted light having wavelengths within said second waveband relative to filtered remitted light having wavelengths within said first waveband, and a second ratio corresponding to the ratio of filtered remitted light having wavelengths within said third waveband relative to filtered remitted light having wavelengths within said first waveband; determining for positions within an image represented by generated image data the concentrations of chromophores and/or thickness of structural layers of said epithelial tissue at said positions in the sample of epithelial tissue represented kr) by said image data utilising said first and said second ratios determined for said .n .positions; and outputting data representing said determined concentrations of :chromophores and/or thickness of epithelial tissue for points in said sample of epithelial tissue.

,:Thus the invention lies in the appreciation that by skilful selection of the wavebands of light remitted by a biological component, usually human or animal tissue, the ratio between two such wavebands can be used to create a useful parametric image of the biological component. The wavebands may:be calculated using a biological or mathematical model to monitor the relationship between a particular image ratio and the parameter to create a function which then can be used for monitoring that parameter in the biological component. As an alternative to creating a function the measured waveband ratios can be compared with the predictions of a model either mathematical or experimentially measured.

The method is such that the effects of reflection rather than remittance will be ignored. Although this would appear to limit the application to components which do not have a specular component of reflection, such as many organic .objects, image processing algorithms have been developed to allow removal of this component of reflection giving greater applicability of the technique. For example, it has been shown that it is possible to )WO 2041010862. PCTIGB2003boo3147? c-I 8 0remo ve tile highl ights from coimplex image s containing- inhomogeneous *dielectrics. It is' also possible to remove the suirface, componuent of reflection usifig. -a polarising filter. Onethis component has bee6n removed from~ image data, it will be po6ssible; to use the techtnique ,s described here.

lieferably the effects of 'Surface reflectioa are elimirnate by providing.a par ofcrossolaxised -linear p~lariig'flesTefrt of these is 'placedn.front of thte'source. of illumination and, the' second in front of the image -capture system.: There are howver- other miethods which. will 'be* Apparent e, 10 the skilled reader which could be -used to eliminate, surface reflection.

effects.

The* body comnponent may be any biological componenit but: is mnost usefully, -anima tissue.

Each wavband- referred to may co pie a sngle wavqlt but -in prac .tice will preferably. comprise -a -band of wavelenqgths, dietectable by the The light emitted by the. light source may be a white lighit or light of -a plurality of- wavelengths, some of which aire outside the predeterminted Wavebands, and filters may be used to limit the light received by the photorece~ptor to the'desited wavebanlds. Such filters tuay be placed be-tween the lighit source and the tissule, between the tissue and. the photoucceptor or -at both posi"tions. Alternatively. white light may 'be emnitted by -the ligt source and. -received by the phtotoreceptor' with. the analysis meant es tablishingte amouxnt of light within the desire6d waveband.

*To understand Why this. Process eliminates any variation in illumination 30' initensity. and sutrface geo-metry it is nece ssary to consider thedeioai reflec.tion maodel. This was First proposed by S. -Shlafer -in 'Usirk& coloure to *separate reflection conipoflnents" Color-. R.Appl. 4; 210_2I8'(l(1985).aPd W6 200416 2 PCtICGR20O3/00U37 states~ that light- remitted from an object is the sum of two components, the "body" component and the "Surface" coniponenit. The body component reesto physiical processes occurring after. penetration of liH& into the material find the surface term. to reflections which take place6 at. the'surface of the object. The body component. Is a fucinof the spectral charateristics of th bet hra h urface component depends only on the 9bjec geometry and the incideat lighit. The theory states further that each compoinent. can b~e considered-the product of a geomri~aal termi and a *wav,6eengtht dependent term., 9 i The pro~posed nvninis 4sed where an optical system -makes, spectral measure-s of. tis-sue. One embodimeont of such ani optical- system ussa' colour digital camera as the ph6toreceptor, although -a mnonochrome, -digital camera arranged to take- seuentiit image wit difeetclurdgt ts L sources could Also be used-. Both the-se embodiments mray have millions of iaepixels -or very few piXels,. or even just one pixel in. the case of the moinochrome. -systemn.. The o .ptical. sytem may work in* the visualt spectrum, or over an. extended spectrum to include. non visible' wavelengths., These non visible wavelengths may include infra-fed light This tnfra-red light may itnclude wavelengths in the 600am to 800tim band.

it the case. of a conventional colour digital cameka, thesystem measures light throutgh a number of opticAl filters. Imagt values fo.r a spe'cific image location, corresponding to the nth fitter, are given by' i" aKC *where

K,

3 anrd KS are the geometric terms -of the -body and surface.

r- compoient respw(,ively. and Cb and C, are colour tertas. By using the s'ystem of polarising filters described above -it is posbe oeimnt surface rflectio n_ Image values- are then -given as.-a simplpodctfa geometric'

U

WOi 20041i0862 PCU1G42003.1003367 term anid a colour, or waeeghdpendent term. The illurminaing light is now written as where e 0 is a wavelength independent scaling factor determined by thie initensity of the light source but which' does not -change, or Phnages in- a known Iuanuer, will wavelength. This allows the..dichtomatic reftection model to be written as =e jE 0 *where ce The function R(A) defines the spectral. res ponse of the nth *filter and the remitted spe ctrum of the Illuminated tissue. It -is 1. essential that. both E 0 and are known for the giveji i'maging.

ytm.Tusthe inivention. is prferentially utilisedt in to syts where tis.Ue of Lu~ncest is illum6inated withi light of knoaw s pectial Chaf acteris tics.

Iftheoptial ystmtreords a M-dimensjonal vector ofiaevles at.

20 eacr-pixel then -it spsilodfn a N-dimiensional vector of imag .ratios, -which is obtained by defining appropriate ratios of- imtage values An example of smch. a vector is r

(Z

Athe constantt s eed nyo position within an im age'all compoet.

of the ratio vector r will,.6be independent of -the conlstant e and 'thus inidependent -of the: illumianation -intonsity -and geomatrical. -factors -III the' imaged scene- WO 20041010942 'ICT/GRflo03fOO3347.

Tile invention is applicable. to.,problemus in whti all [,istological vadition can be described by K parame 4ters. The cocp fapraee eti introduced and defined-as

-J

pP where the space. P defines all possible Parameter variation and thus Variation in tissuie histology. Using, the curretit invention it is possible to rOiecover a* parameter vector froM a vector of -image ratios. To achieve this it is necessary to have some tecbnique far pricting a vector of image. ratios from a given parameter VectOr. This can beahee i oeeprimental -technique or with an appropriate mathematical model of light 'traspor within'the tissue of interesCT Techniques -such as Monte Carlo -modelling or.

the Kubelka-Munk approxkimation have :been developed for. this puIrpose.

With .suchi a model -it is possible to' predict a reiiiittance. spectrum whirh Corresponds to a uniqte point in parameter space, that is a uniique isbe histology. With a knowledge 'of the spectral response of the illuminatiig!light source And the spectral re~aeof the filters, used in the ae acquisition system., it is possiblo to pvedict a Vector of -Image values -for a given point in param~eter space,- This can be expressed as Af te I1 where the space I definecs all possible Me asurements made by the optical systemn-- Using an appropriate de-finition 'Of image ratios su-ante. one.

give abveit S possible to obanavector of imnAge, ratios. Thi s can b IPCTIGB2hO3O4) 33 6 7 12 where the space P, deftin es all possible image ratios that can be Obtained from the space of imlage me asuremenits- A function f can, now be define __-which maps fromb points in. pAramneter space to points in. the space of image ratios. To. imple.ment :this function it is first nrecessary to conite thle sectal eflcta ce f te aterial of .interest fo tegi n set of vvarameter values, (or point in 'pararaeter space. Usintg this. spectral rfetne ln With. the spectial responses each Of the filters R"A a vector of iniage values can be calculated. pinal[oithsaecr frtosane obtoined. This threie:::Sta~iapping can written. as to I dottemapnfrmpaXiameters space to the space of 'image ratios..

Provded hata reittace Pectrui can be defined for any. possible paaee obnton then this mapping is defined for the whole or *parameter space. Th6 propoe ietionj delswith te l nverse. of this uncti: u- defikned as, Which denotes. the- mapping from the space of imuage ratios back to parameter space-. A -key Part. of -the invenition is to establish whiether. a Suitable functiong can be defined which will allow aymeasured ratio to be mapped back tQ the appropriate pafameter combijitation.- Such a pxappinig 25 must be li-I_.-That is, for evey poiit int the Space Of image ratios there Must be a corresponding- unique point io parameter Space. If thsis not -the case, .ambiguity -will arise as it couild be possible to -recover muore that one set of 1 aaee aus rmagn vector of image ratios_ To. esgtablish 'this ~condition,itit its Ct lkecessary to de-al with the function fihc utb confsidred. a v'ector Valued func-tionl Of a- vector variable, thtatis, o a, In In t\ Cc, WO 2004/010862 FCTIGB2O3(00o;367 13 r f To establish whether this function is t-4L the determinadt of the Jacobian minatrix, corresponding to this mapping, can be analysed. This is defined as af, 4 tf 4 ara r, ar, a <)P2 ;Pr, apt aP2 PK t'J,02 ai~r) k' t aP a, aV If th:edetermiiant of this matrix is nodzero at a point in parameter space then there exists a neighbourhood around this point where the function fcan be approximated linearly Tihis means that any: points within this. regioa will map under a mapping to a unique point.in parameter space. If, when using a system to image a given tissue, it can be established thit the Jacobian is aonzero across the whole of paraeter space then the fnctionf will be 1. -everywhere.

:4~3, Once this cotidition has been established it is necessary to find either an apiroximation or an exact analytic expression for the function g which will enable image ratios to be.mapped to specific parameters. Although in some cases it may be possible to obtain.an analytic function, in most cases it will be necessary to construct a piecewise continuous approximation. This can be achieved by discretising parameter space in suitably small intervals and generating the corresponding. image ratio values for every point within a discretised space. Some formr of multidimnsional interpolation technique, such as a cubic spline, is then used to construct a continuous piecewise approximation to the function g. This then allows processing of pixels from an imaged tissue to give the.corresponding set of parameter .values. Any nuqiber of.pixels may :4e propessed in this way to produce a anumber of *WO 204100"g2 PCTI/fllOO3/6Q34 7 14 0Paxrametric maps, which give. quantitative information Ott the parameters of" iterest-across the whok6 of :the imaged scene.

Sui maps are o)f oIjACUeSe Value to clinijcianJs and other persons interested 51 tile Composition of 8pocific tissues.

The imptematatin of the OrOposed invention proceeds along the Thlodwing L For the tissue to be imaged. itify all parameters whose variation could cause a ehatgr- in spectral re'mittan0ce Wieft ittuininated With light- Have, 'by_ some meants a ni6thod for predicting the spectral.

remuitancae of a- given4 tissue -for any coibina tiogj of -the ideintified tssuecarmtes 3. Establish the spectral respOLnses'of each c-hannuel-of the givent imaging.

system and from this. define. an..ap propriate setoimag rais 4. Check that the mapping- from the space of pamterP otesaeo un~iag tratios is. 1-4 over: the range of all paramtrviain 5. If this conditon hlds obttn' sortiefnct it either exato approxiwate, Which Maps poli1ts -in the space -of -image ratios -to: thle corresponding poat in. parameter space, Usig this fun4ctioR images catt thRb, proces .sed to give qoantitative information -on the uaderlyiag,-tissue histology.

-The waveblands hazving Image ratioswihmpL oaprmtro omoetVary dependling -upon thle pa tcua copoent, and the particular pAaaMeter to be aaaty-se& TYpically the method anid appAratas axe used- to anialyse. all theprtutr to characteri-ze a parfticWlAr compoent,' With the light source and.

photoreceptor emiitting -and -receivn fo ach p0arameter, a. pair of wavebatds, dhosef stuch .that teriobetween the aiourt Of igt rmte by te cmponnt f ech wvebnd (c te iage ratio for that pair of wavecbands) is a i-I function- of the particular* paralireter. In-practice, the tidiiuum. nlumber' of wavebands to -be moit~red 'Will be -equal to n+t, Where u equals. the numberu of parametaers.

.1 It has been found for slcint that three parameters characterize the ti.s'sue 1 Stkaiiely. skin~ thicknes3, melailin Oncentratio-n and blood concentration and .25 melanin. and blood. concenjtratio .n may be analysed effetively using thie methods and apparattis of the. .inventi on.

The requiired predeterminied. %avebands may be found using- the method.

described above iteratively.

'ef ,inbdj m e o th pr s n i v e t o may In vo lv e th e calcu latio nt o f an errr, ndiative Of. the accuracy Of pamer re~covery. -Obtained usinug said 0"aPin~g nction. The error may be calCulated as folliows: a)Icakculat the ero so~a~dwith Image acquisition for eacht vector Of each" Image ratio; b) from te image ratio, vectrer,.cacltth mxiu Possible error int e-ach componenat of the picameter Vco~rs h woeof paatrneter-space; and t) se thle vector of parameoter errors at each poin ihn aaee *Space to measure the.acc-uracyfpamerrcory Alterirativ d-y co ectionS ma~y be mtade 'or error by star mahmail errcorrection algtoritlns. the. choi-c _9f -Whichwilbapretoth sicle drse of.'the spe-cificatiOn..

TePresent invention mnay be incorPorated 'itto ftafiy differenlt filte Pr-operty calculation schlemes. For erample,inash e Ligageec glgqnt6M, the method mlay be setoideixtify a pralty fcniaefle pretret.The ikiethod of the pr-esent invelition. is then applied rePeOatedl Y to ftind an opt i 41 al fite "parauaeter set using h caniae Alternatively! the n%._ethod-may. be employed 'in a schernie using a gradient descent Y agiti 1 nSuch a scemne, the fttod o h tdapcto e Apeettivention. is. ,empl16yed. to identf Y Ais ca date set. of fite -6arao fietere T h i s e j t h n s e d t o m a k a s e l c t o n o.f a n t h e s t O f f i l e Pfpeie, ndth pocssreeaedas often as necebssary- to arrive xt o masluin Of course, any -suitabfe- optin jvij 0 algoritlim can be.

usdto comlpute an -optimal solution or a solution Which ha11 fiin accuracyac ti etemethod steps 1) to 4) are -carrie-d out te -tnag ratios may .,or may Rot be changed. That is -to say- that, for eahepetto h oeta Wavebands and~ the image, ratios may' be, -changed, or only' thie potential waveband ma-Y be. changed- Althougtis invention -is applicable with particuljar advantage to"t~ o inv si e a al zin o tisu -typically-animal anad prfeail unan tissue illit be apprec'ated that' thec method -and apparatus cou6ld also )be usW t monitor parabaeters ofa. mateial Where the parametelrs cha4acfteizig the m te ia h a e a v le gti s e c fic o p ic l ro p e rttie s a n d W h e re t is possible to conrol the sVpctral characrersticofh iluiato lit Will be appreciatedl that the filters of the appa.atus may be i-mpltented n-o pticat.,electgi.a~ltoi -Os~fta EMfc lRie~f Descriptjon of theDawib Metods -adatpts -according to the-various aspects- of tae ivention.

Wilt -DOW be described, -by way of example onily ihrfrec.t arcconAPa4Yiag d IraWingg, -in -Which:igure 1 -illustrats a function f whch maps. an imgrtiitoa aeil Parafli&!r

P;

Figue 2Ilustrates a comb ination of errors* in parameter space, mapped to lfflageo ratio vdctor space; Fiur 3 illtrates a model of the layered srcueo kn WO 2004/010862 PCT/GB2003/003367 18 Figures 4a and 4b show how the remitted spectrum (intensity vs wavelength) varies for different melanin and blood levels respectively; Figure 5 illustrates a set of filters suitable for analysing blood and melanin levels in skin; Figure 6 is a flow diagram illustrating a method of defining a set of filter properties for use in analysing the properties of an object or material; Figure 7 is a flow diagram illustrating a method of defining a set of suitable wavebands for use in analyzing the parameters of tissue; Figure 8 is a schematic view of a method and apparatus for analyzing facial skin; and, Figures 9a 9b and 9c are respectively, a colour image of a human face taken with a standard digital camera and parametric maps, showing a greyscale representation of then quantitative measurements of melanin and blood derived using the method and apparatus in accordance with the second and third aspect of the invention..

The proof of the theory behind the selection of appropriate wavebands and image ratios for a given parameter will now be described with reference to figures 1 ,2 and 6 In a typical analysis system, light remitted from an object can be measured using a digital camera with a small number of optical filters representing a number of wavebands. Image values brightness or "intensity" for each image location y) for a given filter (the nth filter) are given by: in(x,y) K= b+ KgC- 2- K JE(s(A)S(A)R() d K, J d (2) where Kb and Ks are the geometric terms of the body and surface components respectively and C. and are colour terms. The first integral in equation is the product of three terms: E(A) is the illuminating light, S(A) is the spectral remittance from the body of the imaged object, and WO 2004/010862 PCT/GB2003/003367 19 is the spectral response of the nth optical filter. In the second integral there are only two terms as there is no wavelength dependence on the surface component of reflection. The dichromatic reflection model is very important for 3-D scene analysis as it allows for both colour and geometrical analysis of objects within a scene.

A key issue is to show that the technique proposed here is valid for problems where the intensity of the illuminating light is unknown (whilst assuming that the spectral definition of the illuminating light is known).

For this purpose the incident light is written as: E(A) COEO(A) (3) where so is a wavelength independent scaling factor. Equation now becomes V) f d (4) where E o0Kb. A digital camera records an N-dimensional vector of image values at each location If a mapping, which is independent of the constant E, can be established between the vector of image values and the vector of parameters, then it will be possible to recover scene parameters from image data in a way that does not depend on illumination intensity or scene geometry.

We now introduce the concept of an image ratio, obtained by dividing one image value, calculated from equation by another. For a given image vector, the nth image ratio is given as: n m. Simple consideration of equation shows that any ratio defined in this way will be invariant to a change in the parameter S. Thus any method for the recovery of parameter values from image ratios will be independent of scene geometry and illumination intensity.

WO 2004/010862 PCTIGB2003/003367 The objective here is to extract quantitative parameters upon which the object colouration depends, not to find statistical similarities. Moreover, the specific filters are chosen to maximise the distance in the image ratio space between vectors corresponding to similar parameter values, as this minimises the error on the parameter value recovered from the colour image.

The technique described is generally applicable to scenes in which a small number of parameters are required to describe all possible objects/object characteristics. In the formulation, the parameters will be considered to vary continuously. Thus, the technique will be particularly applicable to problems where object characteristics need to be measured across an image.

For example a medical imaging system may be required to analyse a particular tissue. The underlying structure of the tissue will not vary, only specific characteristics such as thickness of the different layers (including zero thickness) or the concentration of a particular chemical constituents (including zero concentration). In this situation a small parameter vector can describe all possible variations in the characteristics of the imaged scene. For K scene parameters the parameter vector is defined as:

K

p=Zp. P P (6) k= and the space P defines all potential object characteristics. Ultimately, a mapping from image ratios back to the parameter vector is required, but first the forward problem of obtaining image ratios for a given parameter vector is considered. A reflectance spectrum, corresponding to a given point within parameter space, can be described by the vector in M dimensional wavelength space:

.M

A, AEA(7) m=where the space A defines all possible spectral reflectance functions. The where the space A defines all possible spectral reflectance functions. The WO 2004/010862 PCT/GB2003/003367 21 mapping a, defined as a P A (8) is introduced to denote the mapping from parameter space to wavelength space. This mapping gives the spectral reflectance of the object specified by the vector p. Such a mapping can be achieved either by a spectroscopic measurement, or by using a mathematical model which takes as input the parameters and produces a corresponding reflectance spectrum. Models of light propagation in different media, such as the Monte Carlo method or the Kubelka Munk approximation, can be used for this purpose. It must be possible to perform this mapping across the whole of parameter space, thus defining every possible spectral reflectance function.

A digital camera with N optical filters records an N-dimensional image vector at each pixel. The image vector is given as: i= i ieI (9) n=l where I describes the space of all possible image values. The process of image acquisition can be considered as the projection of points in wavelength space to points in filter space. This projection is performed by the mapping function: b20: A I. Equation performs this mapping b in continuous form. In discrete form, the response of the nrth optical filter, is given as:

M

A(11) m=l where Eo(A)S(A) and the positive weights at each wavelength are given by R m thus defining each filter response function. A digital camera effectively performs the mapping b, projecting points from a large WO 2004/010862 PCTIGB2003/003367 22 dimensional space (wavelength space) to a small dimensional space (filter space). With such a mapping there will be a substantial loss of information.

However, even with this loss of information, it should be possible to define the mapping in such a way that accurate information regarding the original parameter values can still be recovered from image data. Conditions for this mapping will be discussed in the following later.

Most current image acquisition systems use an RGB system of filters.

Although this defines a potential mapping b, it may not be the best mapping with which to recover parameter values from image data. However, it is known to select specific filters to obtain better clarity of data than that possible with an RGB system (although mainly for visualization or image segmentation, not for quantification). Also, in spectrometry, particular spectral wavelengths are selected using statistical methods to improve quantification of components in mixtures. It will therefore be appreciated that an objectively defined set of optical filters is able to perform the task of recovery of parameters, which describe the variation in human skin, better than a standard RGB system.

Once the vector of image values has been obtained, a vector of image ratios can be calculated using equation The vector of image ratios is given as: iI(12) .n=1 where I describes the space of all possible image ratios. The mapping from filter space to the space of image ratios is performed after image acquisition and will be referred to as mapping c, defined as: c: I. (13) There are many ways to define the image ratios and thus the mapping c. For example, pairs of image values could be used to define image ratios as: WO 2004/010862 PCT/GB2003/003367 23 (14 12n 2 or a single image value could be taken as the denominator with which to calculate image ratios from the remaining image values, for example: r- 2,3 N. At most the dimensionality, N, of the new space will be one less that that of the original filter space, N. This would correspond to the definition given in equation Alternatively, if the image ratios were defined as given in equation then the dimensionality of the new space will be half that of the original filter space. The aim is to recover a K-dimensional parameter vector from N image ratios. Thus there must be at least as many image ratios as parameters, that is, N K.

The function f defined as: f=aoboc f: P---41 (16) represents the three stage mapping from parameter space to wavelength space, to image space, and finally to the space of image ratios. For a given set of optical filters, it will be possible to perform this mapping across the whole of parameter space, provided that it is possible to obtain a spectrum for any given parameter vector. The inverse of function f is defined as: (17) and maps from the space of image ratios directly to parameter space. If it is possible to define an appropriate f 1, it will be possible to recover parameter values from image data in a way that is independent of illumination intensity and scene geometry. The ultimate aim is to find the optimum f -1 which maximises the accuracy of parameter recovery. Before a detailed discussion of this mapping is presented, it is important to emphasise that the form of the function f will depend on the mappings a, b WO 2004/010862 PCTIGB2003/003367 24 and c. Although mapping a is fixed for a given problem, mapping b will vary with the choice of optical filters and mapping c will vary depending on how the image ratios are defined.

Any mapping function which is to map from the space of image ratios (Ispace) to parameter space (P-space) must be 1 to 1. That is, for a given point in P-space, there must be a corresponding unique point in I-space and vice-versa. If this is not the case, ambiguity will arise as it could be possible to recover more that one set of parameter values from a given vector of image ratios. Once this condition has been established, it is necessary to consider the error associated with parameter recovery as, using a digital camera, it will only be possible to obtain image values to within a given uncertainty. This will introduce an uncertainty into the final recovery of the parameter vector. There could also be an error associated with the prediction/measurement of a spectrum from the parameter vector. For simplicity the analysis presented here will be restricted to problems in which the error associated with the spectral measurement can be neglected.

Initially, the problem where one parameter is sufficient to describe all variation in an imaged scene will be analysed. The methodology will then be extended to problems where the number of parameters is greater than one.

Consider the case where one image ratio (two image values) is used to recover a single parameter value. Figure 1 illustrates a function f which gives the image ratio as a function of the parameter p. It is clear that in order to satisfy the 1 to 1 condition, the curve must not have any turning points: that is, it must increase or decrease monotonically in the appropriate range of p. Mathematically this is expressed as: >0o vpE P. (18) dp WO 2004/010862 PCT/GB2003/003367 Measurement of an image ratio value i 0 corresponding to a parameter value po, is now considered. Associated with acquisition of each image value is an uncertainty due to camera error. It is straightforward to show, using standard error analysis, that the error associated with an image ratio i, which has been calculated from the two image values i7 and i 2 is given as: AAi i+i 2 (19) where Ai is the camera uncertainty. This error has been shown on the ordinate of the graph in Figure 1. If the derivate of f is non-zero in some neighbourhood of po then it is possible to approximate this function linearly. Assuming the error Ai to lie within this neighbourhood, the corresponding error in the parameter value is given as: Ap Thus, it is possible to obtain a value for the error Ap, associated with parameter recovery, at any point in P-space. An optimisation criterion can then be defined based on some measure of this error. For most applications it will be necessary to minimise the error equally across the whole of Pspace, For others it may be that high accuracy parameter recovery is required within a certain range of parameter values. For example, in a medical image application, imaged tissue could be deemed pathological once a characterising parameter changes beyond a threshold level. This would need to be accounted for with some form of weighting in the optimisation criterion. It is interesting to note that in order to minimse Ap, it is necessary to maximise the magnitude of the derivative given in equation This will ensure that any search, carried out to find an optimum f, will tend to move towards regions of search space where the 1 to 1 condition is satisfied.

In theory it is possible to recover the parameter using more than one image ratio. In this case it will be necessary to calculate the error associated with WO 2004/010862 PCT/GB2003/003367 26 parameter recovery for each of the image ratios and select the one, at each point in P-space, which has the smallest associated error It may be that the optimisation procedure gives a single image ratio which performs better than any other across the whole of P-space. In this situation there is no benefit to using more that one image ratio.

The analysis is now extended to the general problem where the recovery of a K-dimensional parameter vector is required from an N dimensional vector of image ratios. Initially the analysis will be restricted to the case where N K and will then be extended to include situations where N K.

As discussed earlier, if N K, then it is not possible to recover Kdimensional data from an N dimensional measurement.

The mapping function f, defined as: i=f(p) (21) must now be considered a vector valued function of a vector variable. In the following analysis specific results from differential geometry will be used. For further details the reader is directed to for example M. M.

Lipschutz, Differential geometry (McGraw-Hill Book Company, New York, 1969). To establish whether the function fprovides a 1 to 1 relationship, it is first necessary to consider the behaviour of the determinant of the Jacobian matrix, simply referred to as the Jacobian. This is defined as: 0f 9f, 8fp OP1 09O2 OP de (22) af, 9fn Afn pi 19p2 01)p The Jacobian can be considered the multidimensional equivalent of the one dimensional derivative given in equation The inverse function theorem states that, if the Jacobian is non-zero at a point po in P-space, then there WO 2004/010862 PCT/GB2003/003367 27 exists a neighbourhood around po where the function f can be approximated linearly as f(p) f(po) df(po)(p Po) (23) where df is the differential of f and is given as: f f Sf df -dp, dpj 2 f-dpk. (24) 59p ap2 9pk It follows that in this neighbourhood the function f provides a 1 to 1 relationship. Thus, if it is possible to establish that the Jacobian is strictly positive or strictly negative throughout the whole of P-space, the function f will be 1 to 1 everywhere. Once this condition has been established, it is necessary to consider how the error associated with image acquisition maps under f to give the corresponding error in parameter recovery. The error associated with each image ratio is calculated using equation The combination of errors maps out a hypervolume in I-space, centred on the point io. This has been illustrated in Figure 2 for the case of a 2D P-space, where an ellipse is obtained (or a circle if the errors are equal). An ellipsoid is obtained in 3D space and a hyperellipse in higher dimensions.

Although the following analysis will be based on a 2D P-space, the arguments are equally valid in higher dimensions.

The ellipse in I-space represents all possible image ratio vectors which could correspond to a camera measurement i= io. It is assumed that the region of error lies within the neighbourhood of i =i 0 where the mapping function f can be approximated linearly. Thus, under the mapping the ellipse in I-space maps directly to another ellipse in P-space. This new ellipse is centred on the point p po and represents all possible parameter vectors which could be recovered from the vector of image ratios i*= 0 The error associated with parameter recovery is obtained by considering the worst case scenario: that is the point within the ellipse in P-space which is WO 2004/010862 PCT/GB2003/003367 28 at the maximum distance from the point p po. This maximum distance must be calculated separately for each component, Pk, of the parameter vector to obtain the error associated with recovery of each individual component. To calculate these errors it is necessary to consider how the ellipse is transformed under the mapping f 1 which is linear provided the Jacobian is non-zero.

Under a linear mapping the ellipse will be translated, scaled and rotated.

The translation associated with the linear mapping defines the point p po which is mapped to from the point i=i 0 The two other transformations, scaling and rotation, are best understood by considering how a vector di= df in I-space, maps under f' to give a corresponding vector dp in Pspace. The vector dp can be calculated from the inverse form of equation (24) which, in matrix form, is given as: dp J- d where J denotes the Jacobian matrix. Note that J- 1 exists only if the Jacobian in non-zero. This must be the case if the 1 to 1 condition is to be satisfied.

The vectors A and B correspond to the major and minor axes of the ellipse in I-space and are given as: A (26) A= B= C26) Under the mapping f-1 these vectors map to the vectors A' and B' which correspond to the major and minor axes of the ellipse in parameter space.

Solving equation (25) for each of these vectors gives: A'B= (27) Ap )ApB WO 2004/010862 PCT/GB2003/003367 29 where Ap' and Ap' are the components of the vector A' in the direction of pi and P2 respectively. Similarly, Ap and Ap' are the components of the vector B' in the direction of pi and p2 respectively. To calculate the error in each component of the parameter vector it is necessary to consider the worst case scenario. It can be seen from Figure 2 that this corresponds to taking the maximum of ApA and Apf as the error in p, and taking the maximum Ap and Apf as the error in P2. This error can be specified by a vector Ap and can be calculated for any given point in parameter space.

With this measure of the accuracy of parameter recovery across the whole of parameter space, it is possible to define an optimisation criterion. This could simply be based on a sum of the errors at every point in P-space or could be chosen to favour accuracy of recovery of a subset of the original parameters. Once this optimisation criterion has been defined, a search can be used to find the optimum mapping function f. It is important to note that, although the above discussion is based on a 2D parameter space, the methodology is equally applicable to any K-dimensional parameter space.

An algorithm for the implementation of the proposed methodology is given as follows: 1. Establish a suitable search space from a parameterisation of mappings b and c.

2. For a given mapping function f calculate the vector of image ratios for each point within a discretised parameter space.

3. For each point, check that the Jacobian is either strictly positive or strictly negative across the whole of parameter space. If this condition is held then compute the inverse of the Jacobian matrix. If not then return to step 1 and define a new mapping function f 4. Using equation calculate the error associated with image acquisition for each vector of image ratios.

5. From the image ratio vector error calculate the maximum possible error in each component, pk, of the parameter vector across the whole WO 2004/010862 PCT/GB2003/003367 of parameter space.

6. Use the vector of parameter errors at each point within parameter space to measure the accuracy of parameter recovery.

7. Repeat steps 2-6 with some optimisation technique which enables an optimum mapping function f to be determined.

It is fairly straightforward to extend this methodology to the case in which N> K: that is, where there are more image ratios than parameter values.

Initially every possible K-dimensional subspace of image ratios will need to be defined from the original N-dimensional space of image ratios. It will then be necessary to go through the above procedure for each potential subspace and obtain the vector of parameter errors at each point within parameter space. To achieve the maximum possible accuracy the best Ap must be selected at every location within parameter space. Thus every point in P-space will be linked to a specific image ratio combination. It will then be necessary to link every region of the original N -dimensional space of image ratios to the particular subspace of image ratios which should be used for parameter recovery. It is important to note that it is necessary to recover the whole parameter vector at each point i, within a particular K-dimensional subspace of image ratios. It is not possible to attempt to improve the accuracy of the system by recovering different components of the parameter vector from different K-dimensional subspaces of image ratios. This is mathematically invalid.

The mapping function f is a composite function of three separate mappings.

Although the first mapping a, from parameter space to wavelength space, is fixed for a given problem, mappings b and c can vary depending on the choice of optical filters and definition of image ratios. Thus, to define an appropriate search space it is necessary to parameterise mappings b and c.

Mapping b, which represents image acquisition, is defined by the positive NxM matrix given in equation Typically this matrix will contain many elements and an appropriate parameterisation should be based WO 2004/010862 PCTIGB2003/003367 31 on typical filter response functions. For example, the position of the central wavelength and a measure of width could be used to define a Gaussian shape.

Parameterisation of the mapping function c will be fairly straightforward as there are only a limited number of ways of combining image values to produce independent image ratios. In some applications the form of this mapping may be fixed apriori. Thus, it will not increase the overall dimensionality of the search space.

An optimisation method should search the whole space of possible mappings using the optimisation criterion outlined in the previous section.

One technique which is ideally suited to this type of search is a genetic algorithm, GA, {see T. Back and H. P. Schwefel, "An overview of evolutionary algorithms for parameter optimisation," Evolutionary Computation 1, 1-23 (1993)1 as it is straightforward to define a fitness function which measures the accuracy of parameter recovery. Genetic algorithms have been shown to work well on a wide range of problems with objective functions that do not possess "nice" properties such as continuity, differentiability or satisfaction of the Lipschitz Condition {see L. Davis, The handbook of genetic algorithms (Van Nostrand Reingold, New York, 1991), and D. Goldberg, Genetic algorithms in search, optimization and machine learning (Addison-Wesley, London, 1989)1.

The above techniques will now be further exemplified by considering their application to the analysis of a body component, in this case a normal skin composition. Firstly, the prediction of spectral reflectance is considered.

In order to perform mapping a it is necessary to have either a mathematical model which can predict spectral reflectance for a given set of parameter values or some technique for measurement of the appropriate spectrum. For this application we use the mathematical model developed by Cotton and WO 2004/010862 PCT/GB2003/003367 32 Claridge {see S. D. Cotton and E Claridge, "Developing a predictive model of human skin colouring," Proc. of SPIE Med. Imag. 2708, 814-825 (1996)}. With this model it is possible to predict the spectral reflectance for a given set of parameters. An outline of the model is now given.

Skin can be considered to be the four-layer structure depicted in Figure 3.

A negligible amount of light is reflected from the surface of the skin, thus the surface term in equation can be neglected. Although not absorbing any radiation, the stratum corneum scatters the incoming light in all directions. Light which penetrates this layer can thus be considered diffuse.

In the epidermis light is absorbed by the pigment melanin. The absorption at each wavelength can be calculated using the Lambert-Beer law and will depend on the product of the melanin extinction coefficient and the pigment concentration. After passing through the epidermis the light is both scattered and absorbed by the papillary dermis. The absorption results from the presence of blood and scattering from the underlying collagen structure.

The simple Kubelka- Munk light theory Kubelka and F Munk, "Ein Beitrag zur Optik der Farbanstriche", Tech. Opt" 11, 593-611 (1931)} can be used to model the interaction of light with the papillary dermis as the necessary condition of diffuse incident illumination is satisfied. Any light which passes through the papillary dermis into the recticular dermis can be neglected as no significant backscattering occurs in this layer.

Using this two-layer light transport model it is possible to obtain the remitted spectra for given concentrations of melanin and blood. A more detailed description of this model can be found in S. D. Cotton and E Claridge, "Developing a predictive model of human skin colouring", Proc.

of SPIE Med. Imag. 2708, 814-825 (1996).

For a given papillary dermal thickness, changes in melanin and blood characterise all histological variation and thus define a 2-D parameter space for healthy skin. To carry out the optimisation procedure described above it is necessary to discretise parameter space. This is done at equal intervals to WO 2004/010862 PCT/GB2003/003367 33 define 10x10 points, each of which corresponds to a spectrum generated by the mathematical model. For simplicity, concentration values will be denoted by a number between 1 and 10. Figures 4a and 4b show how the remitted spectrum changes as melanin and blood are varied respectively.

With a change in melanin concentration, the intensity of the whole spectrum is seen to decrease, with a more pronounced change in the blue region. As the blood concentration is decreased the most significant reduction in intensity is observed in the green region, the resulting shape reflecting the two absorption maxima of oxyhaemoglobin, a blood born pigment.

To define a suitable search space it is necessary to parameterise the mappings b and c. A parameterised form of b is chosen to define a typical interference filter. This is modelled as a square profile with Gaussian decay at each side. Two parameters are required to specify this shape: the central wavelength and a full width half maximum (FWHM). Optimisation is carried out for three such filters, defining a 6-D search space. With three filters giving three image values, i 1 i 2 and i 3 the only possible definition of image ratios, if we assume i\ i 3 is equivalent to i 3 ii, is given as: il (28 73 '3 In this instance the mapping c does not increase the dimensionality of the search space.

The optimisation procedure was implemented following the algorithm given above. Initially the vector of image ratios was calculated for every point within the discretised parameter space. This was done using the mathematical model to perform mapping a, the parameterised form of matrix to perform mapping b and the equations (28) to perform mapping c. The derivative of each image ratio, with respect to each parameter, was obtained at each point within discretised parameter space using three-point finite difference approximations. The Jacobian matrix WO 2004/010862 PCTIGB2003/003367 34 was then constructed at every point within parameter space, and providing its determinant was non-zero everywhere, the inverse calculated. If this condition was violated then a new mapping f was defined. The errors associated with image acquisition were then calculated using equation (19).

The absolute value of the error in each image value will vary depending on the camera gain setting. Although this constant will not affect the mapping f, it must be estimated in order to calculate the effective camera error. For this application it was taken to be 0.78% of the maximum value of all the image values across parameter space. This corresponds to a camera which has been set to give a maximum reading for the largest spectral reflectance and a camera error of two grey scale levels in an 8-bit representation.

Using the procedure outlined above the error associated with parameter recovery in both melanin and blood was obtained for each point within the discretised parameter space. In order to find an optimum f, it is necessary to minimise the errors in recovery of both melanin and blood across the whole of parameter space. Thus the fitness function for the GA was taken to be the sum of the errors in both melanin and blood. This procedure was implemented in matlab

T

M using a standard GA to search the space of available mappings.

The boundaries of the search space were chosen such that the central wavelength was constrained to lie in the visible region (400nm-700nm) and such that the widths of the filters were allowed to vary from a FWIIM of to 200nm. Although it is now possible to engineer almost any shape of interference filter, this corresponds to an economically viable range of such filters.

Although it was originally assumed that an image ratio defined as il I i 3 would be equivalent to i 3 il, the results of the GA search showed that this was not the case. The search was intitialised for a random seed and, although the same central wavelengths were always obtained, different WO 2004/010862 PCTIGB2003/003367 filters were selected corresponding to i 3 defined in equation Further investigation showed that these local maxima in the search space corresponded to differing distributions of errors both, across parameter space and between the two parameters. This is because the fitness function, or measure of accuracy, was defined as the sum the errors across parameter space for both melanin and blood. Thus, a loss of accuracy in one parameter could be compensated for with an increase in the other. It may be that, with a more exact specification of the error distribution in the fitness function, it would be possible to obtain the same results for every GA search.

Figure 5 shows a filter combination which gave a similar error in the recovery of both melanin and blood. The image ratios were calculated by dividing the filter centred at X 473nm and 4 560nm by the response of the filter centred at 700nm. To understand why these specific filters were selected it is necessary to analyse the spectral curves shown in Figure 4. The filters centred at X 473nm and X 560nnmm correspond to spectral locations where there is a large change in intensity with the parameters melanin and blood respectively. A third filter was then required in a region of the spectrum in which the remitted light which was either significantly less or significantly more than that of the other two filters.

The filter centred at 2 700nm was chosen as it always gave the largest response at any point within parameter space. This ensured that the derivatives of each image ratio decreased monotonically across the whole of parameter space. The Jacobian, calculated from these derivatives, was strictly positive across the whole of parameter space. It is interesting to note that some alternative filter combinations gave Jacobians which were strictly negative across parameter space, corresponding to alternative local maxima in the search space. If two filters are chosen, to define an image ratio, which vary similarly across parameter space, there will be minimal change in that image ratio and thus it will be of limited value for parameter WO 2004/010862 PCTIGB2003/003367 36 measurement.

It has been demonstrated that, using an objectively defined set of optical filters, it is possible to recover scene parameters from image data in a way which is insensitive to geometry and incident illumination. In the example problem, discussed above, the error associated with this parameter recovery was found to be relatively small. The invariance of this mapping means that the technique will be particularly applicable to medical image applications where there is significant curvature of the surface of the imaged tissue, such as near a joint. It also means that the method can be used for whole body imaging. It will also be unnecessary to calibrate the camera to determine the intensity of the incident light. This could help to significantly increase the speed of image acquisition and later processing.

The methodology set out here has been developed for a measurement task, where the scene parameters are known to vary continuously. The technique can be also be applied to problems of recognition, where it is necessary to differentiate discrete objects based on some measure of their spectral reflectance. This approach has been discussed in the article G. Healey, "Using colour for geometry-insensitive segmentation," J. Opt. Soc. Am. A 6, 920-937 (1989) who used the idea of normalised colour to identify different regions of normalised colour space corresponding to different metal and dielectric materials. This enabled geometry-insensitive segmentation of an imaged object comprised of a number of different materials.

It will be appreciated that in order to implement the proposed methodology, a look-up table should be established between all possible image ratios and scene parameters. Although this may be time consuming, it is only necessary to carry out this procedure once. Once established, this look-up table will ensure no significant processing after image acquisition, making this technique particularly suitable to real-time applications.

WO 2004/010862 PCTIGB2003/003367 37 Figure 6 is a flow diagram showing the key steps in the method described above.

The method and apparatus for analysing at least one parameter of a body component, in this case animal tissue in the specific form of facial skin is illustrated in figure 8. A light source 100 provides illumination to the tissue and remitted light is received at photoreceptor 200 which in this case is a digital camera. Two cross polarised linear polarising filters 300 are used to eliminate the effects of surface reflection from the skin. One filter 300 is placed between the light source 100 and the skin and the other filter 300 is placed between the skin and the digital camera 200.

In this case the digital camera is provided with Red, Green and Blue filters so that light in those wavebands is received by the camera. These wavebands are used to provide image ratios of which the concentration of melanin and the concentration of blood are one to one function.

The procedure outlined in figure 7 was applied to image data in the following way.

1. Two parameters: the concentration of melanin and blood were identified as sufficient to describe all histological variation of healthy tissue.

2. A Kubelka-Munk model of light transport was used to predict the remitted spectrum of tissue for any given combination of melanin and blood concentration.

3. The spectral responses of each of the RGB channels of the colour camera were established and image ratios defined as a. Ratio 1= Green Red Ratio 2 Blue/ Red WO 2004/010862 PCTIGB2003/003367 38 4. The mapping from the 2-D space of parameter variation to the 2-D space of image ratios was checked to ensure that is was 1-1 across the whole range of appropriate parameter variation.

A piecewise continuous approximation was constructed to define a function relating image ratios to histological parameters.

6. Images were acquired using a system of crossed polarising filters, as described above. The experimental set up has been illustrated in figure 8.

7. The function described in step 5 was then used to process the image data.

8. Parametric maps were then produced of melanin and blood across the imaged tissue.

In one experiment this method was applied to an image obtained using a JAI CV-M7CL+ camera imaging facial skin. Parametric maps, showing a grey-scale representation of then quantitative measurements of melanin and blood derived using this technique, are shown in figure 9b and 9c.

It should be noted that in 9b illustrating the concentration of haemoglobin concentration across the image, spot S is identified but mole M is not identified. However in 9c illustrating the concentration of melanin across the image, spot S is not identified while mole M is identified. This illustrates simply how useful a tool this can be for a clinician.

A second specific embodiment involves the analysis of images of the human gastrointestinal track obtained using an endoscope. The endoscope system can take two alternative embodiments. In one case the endoscope is equipped with a conventional colour camera and white light source WO 2004/010862 PCTIGB2003/003367 39 equipped with cross polarizing filters'. In a second case the endoscope is equipped with a monochrome camera and a light source equipped with cross polarizing filters', with the light source that changes colour sequentially between red, green and blue, and these changes are synchronised with the camera to produce a sequence of red, green and blue images.

The procedure outlined in figure 7 is applied to this problem, using data from an endoscope equipped with a conventional colour camera, as follows: 1. Two parameters blood concentration and tissue thickness are identified as sufficient to describe all histological variation.

2. A Monte Carlo model of light transport is used to predict the remitted spectrum of the given tissue for any possible combination of blood concentration and tissue thickness.

3. For an endoscope and camera system, the spectral responses of each of the RGB channels is established and image ratios defined as a. Ratio 1 Green Red Ratio 2= Blue/ Red 4. The mapping from the 2-D space of parameter variation to the 2-D space of image ratios is checked to ensure that it is 1-1 across the whole range of appropriate parameter variation.

A piecewise continuous approximation is constructed to define a function relating image ratios to histological parameters.

6. Images are acquired the endoscope with a system of crossed polarising filters.

7. The function described in step 5 is then used to process the image data.

8. Parametric maps are then produced to display variation in blood and tissue thickness across the given image.

one filter being placed between the source of illumination and the component, and the other filter placed between the component and the photoreceptor or photoreceptors with the filters being set at 90 degrees to one another.

WO 2004/010862 PCTIGB2003/003367 The procedure can be modified to analyse additional histological parameters with the addition of additional wavebands as described in the equations shown above. These additional wavebands may be obtained by a monochrome camera and light source with cross polarising filters taking a series of images of the subject illuminated by a sequence of coloured lights of known spectral characteristics. The spectral characteristics of one or more of colours may lie outside the visible spectrum.

Claims (29)

1. Apparatus for determining the distribution of In chromophores and/or the thickness of structural layers n 5 in a sample of epithelial tissue, the apparatus Scomprising: C-i means for illuminating a sample of epithelial tissue with polarised light having wavelengths falling within a first, second, and third predetermined wavebands; a polarising filter positioned so as to filter light remitted from a sample of epithelial tissue said polarising filter being such to filter out light polarised in the manner generated by said means for illuminating a sample of epithelial tissue with polarised light; an image generator operable to detect filtered remitted light from a sample of epithelial tissue and generate image data indicative of the intensity of filtered remitted light received by said image generator having wavelengths falling within said first, second and third predetermined wavebands; ratio determination means operable to process image data generated by said image generator to determine for positions within images represented by c1 42 5505253 O said image data, a first ratio corresponding to the In ratio of light received by said image generator having wavelengths within said second waveband relative to light having wavelengths within said first waveband, 'r 5 and a second ratio corresponding to the ratio of light Mc, received by said image generator having wavelengths C' within said third waveband relative to light having wavelengths within said first waveband; concentration determination means operable to determine for positions within an image represented by image data generated by said image generator the concentrations of chromophores and/or the thickness of structural layers of said epithelial tissue at said positions in a sample of epithelial tissue represented by said image data utilising said first and said second ratios determined for said positions by said ratio determination means; and output means operable to output data representing determined concentrations of chromophores and/or thickness of structural layers for points a sample of epithelial tissue as determined by said concentration determination means.

2. Apparatus in accordance with claim 1 wherein said image generator comprises a digital camera. C- 43 5505253 U

3. Apparatus in accordance with claim 1 or 2 wherein said first waveband comprises a waveband corresponding In to red light. Vt (N M

4. Apparatus in accordance with claim 1, 2 or 3 CI wherein said second waveband comprises a waveband corresponding to green light. -0

5. Apparatus in accordance with any preceding claim wherein said third waveband comprises a waveband corresponding to blue light.

6. Apparatus in accordance with claim 1 or 2 wherein said first waveband comprises a waveband centred on a wavelength of 700nm

7. Apparatus in accordance with claim 1, 2 or 6 wherein said second waveband comprises a waveband centred on a wavelength of 560nm.

8. Apparatus in accordance with claim 1, 2, 6 or 7 wherein said third waveband comprises a waveband centred on a wavelength of 473nm. S44 5505253 U

9. Apparatus in accordance with claim 1 or claim 2 In wherein one of said first, second or third wavebands comprises infra red light. In

10. Apparatus in accordance with any preceding claim M wherein said concentration determination means (1 comprises a look up table associating pairs of first and second ratios generated by said ratio determination means with items of data identifying concentrations of blood and melanin which when illuminated with polarised light are liable to remit cross polarised light having wavelengths falling within said first, second and third wavebands at said first and second ratios.

11. Apparatus in accordance with claim 10 wherein said pairs of first and second ratios and said concentrations of blood and melanin comprise ratios and concentrations determined by analysing samples of epithelial tissue.

12. Apparatus in accordance with claim 10 wherein said pairs of first and second ratios and said concentrations of blood and melanin comprise ratios and concentrations determined utilising a mathematical 5505253 U Smodel of the expected remittance of illuminated light In by samples of epithelial tissue having differing concentrations of blood and melanin. In t 5

13. Apparatus in accordance with any of claims 1-9 wherein said concentration determination means Ci comprises means operable to determine values representative of concentrations of blood and melanin by applying a predetermined mathematical function to 0 first and second ratios for a position as determined by said ratio determination means.

14. Apparatus in accordance with any of claims 1-9 wherein said concentration determination means comprises a look up table associating pairs of first and second ratios generated by said ratio determination means with items of data identifying concentrations of blood and tissue thickness which when illuminated with polarised light are liable to remit cross polarised light having wavelengths falling within said first, second and third wavebands at said first and second ratios. Apparatus in accordance with claim 14 wherein said pairs of first and second ratios and said 46 5505253 O Sconcentrations of blood and tissue thickness comprise In ratios and concentrations determined by analysing samples of epithelial tissue.

In n 5

16. Apparatus in accordance with claim 14 wherein M said pairs of first and second ratios and said concentrations of blood and melanin comprise ratios and concentrations determined utilising a mathematical model of the expected remittance of illuminated light by samples of epithelial tissue having differing concentrations of blood and tissue thickness.

17. Apparatus in accordance with any of claims 1-9 wherein said concentration determination means comprises means operable to determine values representative of concentrations of blood and tissue thickness by applying a predetermined mathematical function to first and second ratios for a position as determined by said ratio determination means.

18. Apparatus in accordance with any preceding claim wherein said means for illuminating a sample of epithelial tissue with polarised light comprises: a light source and a polarising filter operable to polarise light generated by said light source. c- 47 5505253 U

19. Apparatus in accordance with claim 18 wherein said light source is operable to illuminate a sample In of epithelial tissue sequentially with light having n 5 wavelengths falling within different ones of said c, first, second, and third predetermined wavebands.

A method for determining the distribution of chromophores and/or thickness of structural layers in a sample of epithelial tissue, the method comprising: illuminating a sample of epithelial tissue with polarised light having wavelengths falling within a first, second, and third predetermined wavebands; filtering light remitted from a sample of epithelial tissue so as to filter out light polarised in the manner corresponding to the polarised light utilised to illuminate the sample of epithelial tissue; generating image data indicative of the intensity of filtered remitted light having wavelengths falling within said first, second and third predetermined wavebands; processing generated image data to determine for positions within images represented by said image data, a first ratio corresponding to the ratio of S48 5505253 U Sfiltered remitted light having wavelengths within said In second waveband relative to filtered remitted light having wavelengths within said first waveband, and a In second ratio corresponding to the ratio of filtered remitted light having wavelengths within said third Swaveband relative to filtered remitted light having 1 wavelengths within said first waveband; determining for positions within an image represented by generated image data the concentrations 0 of chromophores and/or thickness of structural layers of said epithelial tissue at said positions in the sample of epithelial tissue represented by said image data utilising said first and said second ratios determined for said positions; and outputting data representing said determined concentrations of chromophores and/or thickness of epithelial tissue for points in said sample of epithelial tissue.

21. A method in accordance with claim 20 wherein said first waveband comprises a waveband corresponding to red light. c( 49 5505253 U

22. A method in accordance with claim 20 or 21 In wherein said second waveband comprises a waveband corresponding to green light. In i 5

23. A method in accordance with any of claims 20, 21 Sor 22 wherein said third waveband comprises a waveband C-i corresponding to blue light.

24. A method in accordance with claim 20 wherein one 0 of said first, second or third wavebands comprises infra red light.

A method in accordance with claim 20 wherein said first waveband comprises a waveband centred on a wavelength of 700nm.

26. A method in accordance with claim 20 or wherein said second waveband comprises a waveband centred on a wavelength of 560nm.

27. A method in accordance with any of claims 20, or 26 wherein said third waveband comprises a waveband centred on a wavelength of 473nm. ri 50 5505253 O

28. A method in accordance with any of claims 20-27 wherein said determining for positions within an image represented by generated image data the concentrations In of chromophores and/or thickness of structural layers tf 5 of said epithelial tissue comprises determining for c, positions within an image represented by generated (1 image data the concentrations of blood and melanin in said epithelial tissue at said positions in the sample of epithelial tissue represented by said image data utilising said first and said second ratios determined for said positions.

29. A method in accordance with any of claims 20-27 wherein said determining for positions within an image represented by generated image data the concentrations of chromophores and/or thickness of structural layers of said epithelial tissue comprises determining for positions within an image represented by generated image data the concentrations of blood in said epithelial and tissue thickness of said epithelial tissue at said positions in the sample of epithelial tissue represented by said image data utilising said first and said second ratios determined for said positions. c\ 51 5505253 O A method in accordance with any of claims 20-29 wherein said illuminating a sample of epithelial tissue with polarised light having wavelengths falling In within a first, second, and third predetermined k~ 5 wavebands comprises sequentially illuminating said (N M n sample of epithelial tissue with polarised light C-i having wavelengths falling within different ones of said first, second and third predetermined wavebands.