DETECTION OF COGNITIVE IMPAIRMENT IN HUMAN BRAINS FROM IMAGES

Abstract

A computer implemented method by which digital images of the human brain can be used to diagnose or to predict cognitive impairment, such as Alzheimer's disease and other forms of cognitive impairment such as so-called prodromal Alzheimer's disease. Methods of classifying or stratifying cohorts of human subjects such as for the purpose of clinical trials and/or to assess the impact of therapies are included. In some embodiments the images comprise T1 weighted MRI images.

Claims

1-117. (canceled)

118. A computer implemented method of predicting or diagnosing cognitive impairment in a human subject based on images of the subject's brain, the method comprising: determining, based on processing by the computer of digital image data obtained from the images, image metrics comprising at least one of: a complexity metric of an image region corresponding to the right middle temporal gyms; an image texture metric of a image region corresponding to the right rostral middle frontal; an image texture metric in the image region corresponding to the right supramarginal; and an image intensity metric in the image region corresponding to the right temporal pole; the method further comprising: determining an indicator based on said image metrics according to a predetermined method; and predicting or diagnosing cognitive impairment state in the subject based on the indicator.

119. The method of claim 118 wherein predicting or diagnosing cognitive impairment state comprises distinguishing between: (a) Alzheimer's disease; and (b) non-Alzheimer's disease.

120. The computer implemented method of claim 118 wherein the image metrics comprise the complexity metric of the image region corresponding to the right middle temporal gyms; and the image texture metric of the image region corresponding to the right rostral middle frontal.

121. The computer implemented method of claim 120 wherein the image metrics comprise the image texture metric in the image region corresponding to the right supramarginal.

122. The computer implemented method of claim 120 wherein the image metrics comprise the image intensity metric in the image region corresponding to the right temporal pole.

123. The computer implemented method of claim 118 wherein the image metrics comprise the complexity metric of the image region corresponding to the right middle temporal gyms; and the image texture metric in the image region corresponding to the right supramarginal.

124. The computer implemented method of claim 123 wherein the image metrics comprise the image intensity metric in the image region corresponding to the right temporal pole.

125. The computer implemented method of claim 118 wherein the image metrics comprise the image texture metric of the image region corresponding to the right rostral middle frontal; and the image intensity metric in the image region corresponding to the right temporal pole.

126. The computer implemented method of claim 118 wherein the image metrics comprise the image texture metric of the image region corresponding to the right rostral middle frontal; and the image texture metric in the image region corresponding to the right supramarginal.

127. The computer implemented method of claim 126 wherein the image metrics comprise the image intensity metric in the image region corresponding to the right temporal pole.

128. The computer implemented method of claim 118 wherein the image metrics comprise the image texture metric in the image region corresponding to the right supramarginal and the image intensity metric in the image region corresponding to the right temporal pole.

129. The computer implemented method of claim 118 wherein the image metrics comprise the complexity metric of the image region corresponding to the right middle temporal gyms; and the image metrics comprise the image intensity metric in the image region corresponding to the right temporal pole.

130. The computer implemented method of claim 118 wherein the complexity metric of the right middle temporal gyms comprises a measure of fractal dimension such as a minimum fractal dimension, optionally wherein the image data in the region corresponding to the right inferior lateral ventricle is modified using an HLH filter.

131. The computer implemented method of claim 118 wherein the image texture metric in the right supramarginal comprises a measure of correlation, such as the GLCM correlation.

132. The computer implemented method of claim 118 wherein the image texture metric the right rostral middle frontal comprises a measure of correlation, such as the GLCM correlation.

133. The computer implemented method of claim 118 wherein the image intensity metric in the right temporal pole comprises a measure of central tendency such as the mean.

134. The method of claim 118 wherein the predetermined method comprises computing a weighted sum of the image metrics.

135. The method of claim 118 comprising obtaining reference data configured to indicate a cognitive impairment state using reference indicators determined according to the predetermined method, and comparing the indicators for the subject to the reference indicators to perform said predicting or diagnosing of the cognitive impairment state in the subject.

136. (canceled)

137. The computer implemented method of claim 118 comprising operating a processor to automatically segment the images to provide digital image data corresponding to ROIs in each of the image regions, and determining the image metrics by operating the processor to perform, on the digital data, image processing steps configured to provide said image metrics.

138. A computer program product or computer apparatus configured to perform the method of claim 118, and to provide an output indicating said prediction or diagnosis.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Embodiments of the disclosure will now be described in detail, by way of example only, with reference to the accompanying drawings, in which:

[0109] FIG. 1 shows an apparatus configured to predict or diagnose cognitive impairment;

[0110] FIG. 2 shows a plot of the ROC analysis corresponding to Table 1 and Table 3;

[0111] FIG. 3 shows a plot of the ROC analysis corresponding to Table 2 and Table 4;

[0112] FIG. 4 shows the results of an ROC analysis for the comparative tests set out in Table 7;

[0113] FIG. 5 shows the results of an ROC analysis for the comparative tests set out in Table 8;

[0114] FIG. 6 shows a set of overlays of particular image features in sub-regions of the cortex;

[0115] FIG. 7 shows further overlays obtained from particular image features in sub regions of the cortex;

[0116] FIG. 8 shows further overlays obtained from particular image features in sub regions of the cortex;

[0117] FIG. 9 shows the results of an ROC analysis for the feature set defined in Table 10 an indicates the results of a comparative analysis;

[0118] FIG. 9 shows the results of an ROC analysis for the feature set defined in Table 11 and indicates the results of a comparative analysis;

[0119] FIG. 10 shows the results of an ROC analysis for the feature set defined in Table 12 and indicates the results of a comparative analysis;

[0120] FIG. 11 shows the results of an ROC analysis for the feature set defined in Table 11 and indicates the results of a comparative analysis;

[0121] FIG. 12 shows the results of an ROC analysis for the feature set defined in Table 12 and indicates the results of a comparative analysis;

[0122] FIG. 13 shows the results of an ROC analysis for the feature set defined in Table 13 and indicates the results of a comparative analysis; and

[0123] FIG. 14 shows the results of an ROC analysis for the feature set defined in Table 14 and indicates the results of a comparative analysis.

DETAILED DESCRIPTION

[0124] A computer implemented method of predicting or diagnosing cognitive impairment in a human subject will now be described with reference to the apparatus shown in FIG. 1.

[0125] In overview, this apparatus segments the images to identify in those images, image regions corresponding to selected anatomical features in the brain. It then determines an image metric or metrics for each anatomical features (segmented image region). Each of these image metrics provides a quantitative indication of structure in that image region, which are combined according to a predetermined method to determine an indicator. The apparatus then obtains reference data against which the indicator is compared to predict or diagnose cognitive impairment in the subject.

[0126] The apparatus illustrated in FIG. 1 comprises a subject image data obtainer 20, a reference data store 26, and a controller 10. It may also be coupled to a source of image data 28.

[0127] The subject image data obtainer 20 may comprise a data interface 22 for communicating data with the controller 10 and/or with the source of image data 28. It may also comprise volatile and/or non-volatile data storage 24, such as a cache, connected to the data interface 22. The subject image data obtainer 20 may be connected (e.g. via the interface 22) to communicate with one or both of: [0128] an imaging system such as a CT scanner, or an MRI scanner or other imaging system capable of acquiring appropriate images of the human brain; or [0129] a store of appropriate images of the human brain.

[0130] Other sources of image data (such as network connections to local and/or wide area networks may also be used). Images of subjects can thus be provided to the controller 10 from a variety of sources.

[0131] The controller 10 may comprise a general purpose processor or similar processing logic, which is configured to segment images of the human brain according to a brain atlas model such as that defined in the freesurfer utilities which are available from https://surfer.nmr.mgh.harvard.edu/fswiki/CorticalParcellation. It will be appreciated in the context of the present disclosure that such utilities may assign neuroanatomical labels to each location on a cortical surface model based on probabilistic information estimated from a manually labelled training set (e.g. that which is made using FreeSurfer). Such procedures may incorporate both geometric information derived from the cortical model, and neuroanatomical convention, as found in the training set. The atlases used may comprise: [0132] gyral based atlases such as the ‘Desikan-Killiany’ cortical atlas described in “An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest” Neuroimage 31 (2006) 968-980, by RS Desikan et al. [0133] atlases based on parcellation schemes that divide the cortex into gyral and sulcal regions such as the ‘Destrieux’ cortical atlas described in “Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature” NeuroImage 53, Issue 1, 15 Oct. 2010, Pages 1-15, Destrieux et al. [0134] the DKT40 atlas https://mindboggle.info/data.html or other similar atlases. Whatever atlas is used, the controller may be configured to identify, in image data having appropriate contrast and resolution, some or all of the following structures: [0135] Left Cerebral White Matter [0136] Left Inferior Lateral Ventricle [0137] Left Cerebral Cortex [0138] Right Cerebral Cortex [0139] Left Hippocampus [0140] Left Amygdala [0141] Left choroid plexus [0142] Brain Stem [0143] White matter (WM) hypointensities, which may comprise white matter lesions [0144] Left Pallidum [0145] Corpus Callosum Posterior [0146] Left Inferior Lateral Ventricle [0147] Right Hippocampus

[0148] The controller may also be configured to segment, from the images regions of interest (ROI) corresponding to the structures which it identifies. This too may be done using approaches such as those defined in the FreeSurfer package or other equivalent packages.

[0149] The controller also comprises image processing functionality arranged to determine some or all of the features defined below. For example, the controller may be configured to determine for a given ROI: [0150] First order statistics such as measures of central tendency, measures of spread, skewness and kurtosis; [0151] Morphological features; [0152] Grey Level Size Zone Matrix features; [0153] Grey Level Co-occurrence Matrix (GLCM) features; [0154] Neighbourhood grey tone difference based features; [0155] Grey Level Run Length Matrix (GLRLM) features; and [0156] Fractal dimension features.

[0157] Examples of such features and the way in which they may be determined by the controller are explained below. It will be appreciated in the context of the present disclosure that other equivalent and/or comparable image features may be used. In addition, the formulae given for the features listed are intended only to serve as an illustrative description of the way in which they can be provided. Typically, such metrics operate on digital data which may be encoded as a discrete grey level in each voxel of an ROI. It will be appreciated in the context of the present disclosure that the notations set out below, using indices and summations are intended to be implemented by the controller stepping through the intensity values (e.g. digital grey level data) stored in the pixels of the images (or ROI's of those images). The indices mentioned in the mathematical formulae below may thus connote the indices of stepwise computational methods, such as may be implemented using a counter incremented to step through the relevant set of digital data. Other methods, such as vector methods, may also be used.

Morphological Features

[0158] Morphological features, such as the surface area, A, and volume, V, of a region identified in an image may be determined based on voxel representations of that volume. Mesh based representations of the outer surface of such a volume may also be used to determine the surface area and volume of the region, for example a marching cubes algorithm may be used.

[0159] One morphological feature which may be used is the degree to which a region is spherical. One measure of this feature is sphericity, which may be defined by F.sub.morph.sphericity

[00005]$F_{\mathrm{morph}.\mathrm{sphericity}}=\frac{{\left(36\pi V^2\right)}^{1/3}}{A}$

wherein V is the volume of the image region concerned, and A is its surface area. Other measures include: [0160] compactness, which may be based on

[00006]$F_{\mathrm{morph}.\mathrm{comp}\text{.1}}=\frac{V}{{\pi }^{1/2}{A}^{3/2}}$

(so called compactness 1), or

[00007]$F_{\mathrm{morph}.\mathrm{comp}\text{.2}}=36\pi \frac{V^2}{A^3}$

(so called compactness 2) [0161] asphericity, which may be based on the ratio A.sup.3/V.sup.2 and spherical disproportion, which may be based on the ratio A/V.sup.2/3.

Grey Level Size Zone Matrix Features

[0162] The grey level size zone matrix (GLSZM) counts the number of groups (or zones) of linked voxels. Voxels are linked if the neighbouring voxel has an identical discretised grey level. Whether a voxel classifies as a neighbour depends on its connectedness. In a 3D approach to texture analysis all of the 26 neighbouring voxels in a 3D rectilinear grid are considered. In the 2D the neighbouring 8 voxels in the same 2D image are considered.

[0163] Let M be the Ng×Nz grey level size zone matrix, where Ng is the number of discretised grey levels present in the ROI intensity mask and Nz the maximum zone size of any group of linked voxels. Element s.sub.ij of M is then the number of zones with discretised grey level i and size j. Furthermore, let N.sub.v be the number of voxels in the intensity mask and

N.sub.s=Σ.sub.i=1.sup.N.sup.gN.sub.j=1.sup.N.sup.zs.sub.ij

be the total number of zones. Marginal sums can likewise be defined. The number of zones with discretised grey level i, regardless of size is s.sub.i.

s.sub.i.=Σ.sub.j=1.sup.N.sup.zs.sub.ij

[0164] Likewise, the number of zones with size j, regardless of grey level is s.Math.j

s.sub..j=Σ.sub.i=1.sup.N.sup.gs.sub.ij

[0165] A grey level non-uniformity measure can then be defined which assesses the distribution of zone counts over the grey values. The feature value is low when zone counts are equally distributed along grey levels. One example of a grey level size zone matrix based grey level non-uniformity measure, denoted F.sub.szm.glnu, comprises:

[00008]$F_{\mathrm{szm}.\mathrm{glnu}}=\frac{1}{N_s}\underset{i=1}{\overset{N_g}{.Math.}}{s}_{i.}^j$

[0166] The controller may be configured to determine an image metric which gives emphasis to the prevalence of large zones. One example of such a measure, which may be based on the grey level size zone matrix and indicate the presence of large zones, is the large zone emphasis of the grey level size zone matrix. This may be denoted F.sub.szm.lze and comprises:

[00009]$F_{\mathrm{szm}.\mathrm{lze}}=\frac{1}{N_s}\underset{j=1}{\overset{N_z}{.Math.}}{j}^2{s}_{.j}$

[0167] The controller may be configured to determine an image metric indicating the variance in zone counts over the different zone sizes (the GLSZM variance)

[00010]$F_{\mathrm{szm}.\mathrm{zs}.\mathrm{var}}=\underset{i=1}{\overset{N_g}{.Math.}}\underset{j=1}{\overset{N_z}{.Math.}}{\left(j-\mu \right)}^2{p}_{\mathrm{ij}}$

In which p.sub.ij=s.sub.ij/N.sub.s and μ=Σ.sub.i=1.sup.N.sup.gΣ.sub.j=1.sup.N.sup.zjp.sub.ij

Grey Level Co-Occurrence Matrix (GLCM) Features

[0168] The grey level co-occurrence matrix (GLCM) is a matrix that expresses how combinations of discretised intensities (grey levels) of neighbouring pixels, or voxels in a 3D volume, are distributed along one of the image directions. Generally, the neighbourhood for GLCM is a 26-connected neighbourhood in 3D and a 8-connected neighbourhood in 2D. Thus, in 3D there are 13 unique direction vectors within the neighbourhood for Chebyshev distance δ=1, i.e. (0, 0, 1), (0, 1, 0), (1, 0, 0), (0, 1, 1), (0, 1, −1), (1, 0, 1), (1, 0, −1), (1, 1, 0), (1, −1, 0), (1, 1, 1), (1, 1, −1), (1, −1, 1) and (1, −1, −1), whereas in 2D the direction vectors are (1, 0, 0), (1, 1, 0), (0, 1, 0) and (−1, 1, 0).

[0169] A GLCM is calculated for each direction vector, as follows. M.sub.m is the N.sub.g×N.sub.g grey level co-occurrence matrix, with N.sub.g the number of discretised grey levels present in the region of interest (ROI) intensity mask, and m the particular direction vector.

[0170] Element (i, j) of the GLCM contains the frequency at which combinations of discretised grey levels i and j occur in neighbouring voxels along direction m.sub.+=m and along direction m.sub.−=−m. Then, M.sub.m=M.sub.m++M.sub.m−=M.sub.m++M.sup.T.sub.m+. As a consequence the GLCM matrix M.sub.m is symmetric.

[0171] A probability distribution for grey level co-occurrences, P.sub.m, is obtainable by normalising M.sub.m by the sum of its elements. Each element p.sub.ij of Pm is then the joint probability of grey levels i and j occurring in neighbouring voxels along direction m.

[0172] The row marginal probability p.sub.i is

p.sub.i.=Σ.sub.j=1.sup.N.sup.gp.sub.ij

and the column marginal probability is

p.sub..j=Σ.sub.i=1.sup.N.sup.gp.sub.ij

as P.sub.m is by definition symmetric, p.sub.i.=p.sub..j.

[0173] Measures of correlation of the grey level co-occurrence matrix GLCM can be defined such as

[00011]$F_{\mathrm{cm}.\mathrm{corr}}=\frac{1}{{\sigma }_{i.}{\sigma }_{.j}}\left(-{\mu }_{i.}{\mu }_{.j}+\underset{i=1}{\overset{N_g}{.Math.}}\underset{j=1}{\overset{N_s}{.Math.}}{\mathrm{ijp}}_{\mathrm{ij}}\right)$

in which μ.sub.i. and σ.sub.i. are the mean and standard deviation of row marginal probability p.sub.i., respectively. Likewise, μ.sub..j and σ.sub..j are the mean and standard deviation of the column marginal probability p.sub..j, respectively. An informational measure of correlation of the grey level co-occurrence matrix GLCM, F.sub.cm.info.corr.2:

[00012]$F_{\mathrm{cm}.\mathrm{info}.\mathrm{corr}\text{.2}}=\sqrt{1-\exp \left(-2\left({\mathrm{HXY}}_2-\mathrm{HXY}\right)\right)}$$\mathrm{wherein},$$\mathrm{HXY}=-{.Math.}_{i=1}^{N_g}{.Math.}_{j=1}^{N_g}{p}_{\mathrm{ij}}{\log }_2{p}_{\mathrm{ij}},$${\mathrm{HXY}}_2=-\underset{i=1}{\overset{N_g}{.Math.}}\underset{i=1}{\overset{N_s}{.Math.}}{p}_{i.}{p}_{.j}{\log }_2\left(p_{i.}{p}_{.j}\right)$

[0174] p.sub.ij is the joint probability of grey levels i and j occurring in neighbouring voxels along a direction in which the GLCM is defined;

[0175] p.sub.i. is the row marginal probability of the GLCM, and

[0176] p.sub..j is the column marginal probability of the GLCM.

[0177] Other measures of correlation of the GLCM may be used.

[0178] The controller may be configured to determine the above and other measures of correlation. In addition, the controller may be configured to determine an auto correlation of the GLCM based on:

[00013]$F_{\mathrm{cm}.\mathrm{auto}.\mathrm{corr}}=\underset{i=1}{\overset{N_g}{.Math.}}\underset{j=1}{\overset{N_g}{.Math.}}{\mathrm{ijp}}_{\mathrm{ij}}$

[0179] The GLCM difference variance may be computed as the difference variance for the diagonal probabilities thus:

[00014]$F_{\mathrm{cm}.\mathrm{diff}.\mathrm{var}}=\underset{k=0}{\overset{N_g-1}{.Math.}}{\left(k-\mu \right)}^2{p}_{i-j.k}$

[0180] The sum average for the GLCM is:

[00015]$F_{\mathrm{cm}.\mathrm{sum}.\mathrm{avg}}=\underset{k=2}{\overset{2N_g}{.Math.}}{\mathrm{kp}}_{i+j.k}$

[0181] The sum variance for the GLCM is defined as:

[00016]$F_{\mathrm{cm}.\mathrm{sum}.\mathrm{var}}=\underset{k=2}{\overset{2N_g}{.Math.}}{\left(k-\mu \right)}^2{p}_{i+j.k}$

Where μ is equal to the value of the sum average for the GLCM.

[0182] The controller may be configured to determine an entropy measure of the GLCM, such as a sum entropy:

[00017]$F_{\mathrm{cm}.\mathrm{sum}.\mathrm{entr}}=-\underset{k=2}{\overset{2N_g}{.Math.}}{p_i}_{+j.k}{\log }_2{p}_{i+j.k}$

[0183] It can thus be seen that a variety of metrics of image texture may be obtained from the GLCM. Another example is the inverse variance:

[00018]$F_{\mathrm{cm}.\mathrm{inv}.\mathrm{var}}=2\underset{i=1}{\overset{N_g}{.Math.}}\underset{j>1}{\overset{N_g}{.Math.}}\frac{p_{\mathrm{ij}}}{{\left(i-j\right)}^2}$

[0184] The controller may also be configured to determine other measures derived from the GLCM—such as cluster based measures of texture. One such measure is the cluster tendency

[00019]$F_{\mathrm{cm}.\mathrm{clust}.\mathrm{tend}}=\underset{i=1}{\overset{N_g}{.Math.}}\underset{j=1}{\overset{N_g}{.Math.}}{\left(i+j-{\mu }_{i.}-\mu {.}_j\right)}^2{p}_{\mathrm{ij}}$

in which μ.sub.i is the mean row marginal probability, and μ.sub..j is the mean column marginal probability. As noted above, these parameters may be computed as

μ.sub.i.=Σ.sub.i=1.sup.N.sup.gi p.sub.i. and μ.sub..j=Σ.sub.j=1.sup.N.sup.gj p.sub..j

[0185] Another such cluster based measure is the so called GLCM cluster shade, which may be computed as

[00020]$F_{\mathrm{cm}.\mathrm{clust}.\mathrm{shade}}=\underset{i=1}{\overset{N_g}{.Math.}}\underset{j=1}{\overset{N_g}{.Math.}}{\left(i+j-{\mu }_{i.}-{\mu }_{.j}\right)}^3{p}_{\mathrm{ij}}$

[0186] Another such cluster based measure is the so called GLCM cluster prominence, which may be computed as

[00021]$F_{\mathrm{cm}.\mathrm{clust}.\mathrm{prom}}=\underset{i=1}{\overset{N_g}{.Math.}}\underset{j=1}{\overset{N_g}{.Math.}}{\left(i+j-{\mu }_{i.}-{\mu }_{.j}\right)}^4{p}_{\mathrm{ij}}$

Neighbourhood Grey Tone Difference Based Features

[0187] The controller may also be configured to determine neighbourhood grey tone difference matrix (NGTDM). This contains the sum of grey level differences of pixels/voxels with discretised grey level i and the average discretised grey level of neighbouring pixels/voxels within a Chebyshev distance δ.

[0188] The average grey level within a neighbourhood centred at (k.sub.x, k.sub.y, k.sub.z), but excluding (k.sub.x, k.sub.y, k.sub.z) itself is:

[00022]${\stackrel{\_}{X}}_k=\frac{1}{W}\underset{m_{x}m-\delta }{\overset{\delta }{.Math.}}\underset{m_{y}m-\delta }{\overset{\delta }{.Math.}}\underset{m_{x}m-\delta }{\overset{\delta }{.Math.}}{X}_d\left(k_x+m_x,k_y+m_y,k_z+m_z\right)$$\left(m_x,m_y,m_z\right)\ne \left(0,0,0\right)$

where X.sub.d,k is the discretised grey level of a voxel at position k.sub.x, k.sub.y, k.sub.z), for a 3D neighbourhood W=(2δ+1)3−1 is the size of the neighbourhood. For a 2D neighbourhood W=(2δ+1)2−1, and averages are not calculated between different slices.

[0189] Neighbourhood grey tone difference s.sub.i for discretised grey level i is:

[00023]$s_i=\underset{i=1}{\overset{N_v}{.Math.}}.Math.i-{\stackrel{\_}{X}}_k.Math.\left[X_d\left(k\right)=i\mathrm{and}{W}_k\ne 0\right]$

where, W.sub.k is neighbourhood size for the voxel (k.sub.x, k.sub.y, k.sub.z) and N.sub.v is equal to the number of voxels in the neighbourhood that are part of the ROI mask

[00024]$W_k=\underset{m,m-\delta }{\overset{\delta }{.Math.}}\underset{m_{8}m-\delta }{\overset{\delta }{.Math.}}\underset{m,m-\delta }{\overset{\delta }{.Math.}}.Math.m\ne 0\mathrm{and}k+m\mathrm{in}\mathrm{ROI}.Math.$

where [ . . . ] is an Iverson bracket, which is 1 if the conditions within it are true and zero otherwise.

[0190] In NGTDM grey level probabilities p.sub.i are defined, thus p.sub.i=n.sub.i/N.sub.v,c. N.sub.v,c is total number of voxels that have at least one neighbour. If all voxels have at least one neighbour N.sub.v,c=N.sub.v

[0191] The controller may be configured to determine a texture strength based on the NGTDM. One example of such a texture strength comprises:

[00025]$F_{\mathrm{ngi}.Math.\mathrm{strength}}=\frac{{.Math.}_{i_1=1}^{N_g}{.Math.}_{i_2=1}^{N_g}\left(p_{i_1}+p_{i_2}\right){\left(i_1-i_2\right)}^2}{{.Math.}_{i=1}^{N_g}{s}_i},$$p_{i_1}\ne 0\mathrm{and}{p}_{i_2}\ne 0$

Where N.sub.g is the number of discretised grey levels in the ROI intensity mask.

[0192] The controller may be configured to determine a contrast based on the NGTDM, which may be based on depends on the dynamic range of the grey levels and the spatial frequency of intensity changes in said grey levels. One example of such a contrast comprises:

[00026]$F_{\mathrm{ngt}.Math.\mathrm{contrast}}=\left(\frac{1}{N_{g,p}\left(N_{g,p}-1\right)}\underset{i_1=1}{\overset{N_g}{.Math.}}\underset{i_2=1}{\overset{N_g}{.Math.}}{p}_{i_1}{p_{i_2}\left(i_1-i_2\right)}^2\right)$$\left(\frac{1}{N_{v,c}}\underset{i=1}{\overset{N_g}{.Math.}}{s}_i\right)$

[0193] Grey level probabilities p.sub.i1 and p.sub.i2 are copies of p.sub.i with different iterators, i.e. p.sub.i1=p.sub.i2 for i.sub.1=i.sub.2. The first term considers the grey level dynamic range, whereas the second term is a measure for intensity changes within the volume. If N.sub.g,p=1, F.sub.ngt.contrast=0.

[0194] The controller may also be configured to determine a measure of busyness, for example based on the prevalence of large changes in grey levels between neighbouring voxels. One such metric may be defined

[00027]$F_{\mathrm{ngt}.Math.\mathrm{busyness}}=\frac{{.Math.}_{i=1}^{N_g}{p}_i{s}_i}{{.Math.}_{i_1=1}^{N_g}{.Math.}_{i_2=1}^{N_g}.Math.i_1{p}_{i_1}-i_2{p}_{i_2}.Math.},$$p_{i_1}\ne 0\mathrm{and}{p}_{i_2}\ne 0$

[0195] The controller may also be configured to determine a measure of complexity such as an image metric indicating the prevalence of complex textures in which rapid changes in grey levels are common. One example of such a metric is texture complexity, or NGTDM complexity, which may be defined:

[00028]$F_{\mathrm{ntg}.Math.\mathrm{complexity}}=\frac{1}{N_{v,c}}\underset{i_1=1}{\overset{N_g}{.Math.}}\underset{i_2=1}{\overset{N_g}{.Math.}}.Math.i_1-i_2.Math.\frac{p_{i_1}{s}_{i_1}+p_{i_2}{s}_{i_2}}{p_{i_1}+p_{i_2}},$$p_{i_1}\ne 0\mathrm{and}{p}_{i_2}\ne 0$

[0196] The controller may be configured to provide a metric of coarseness based on the fact that grey level differences in coarse textures are generally small due to large-scale patterns. Such measures may be determined by summing differences to give an indication of the level of the spatial rate of change in intensity (coarseness). One measure of NGTDM coarseness may be defined as:

[00029]$F_{\mathrm{ngt}.Math.\mathrm{coarseness}}=\frac{1}{{.Math.}_{i=1}^{N_g}{p}_i{s}_i}$

Where N.sub.g, s.sub.i and p.sub.i are defined as above.

Grey Level Run Length Matrix (GLRLM) Features

[0197] Another way to define texture features is to use the grey level run length matrix (GLRLM). A run length is defined as the length of a consecutive sequence of pixels or voxels with the same grey level along direction m. The GLRLM then contains the occurrences of runs with length j for a discretised grey level i.

[0198] M.sub.m is an Ng×Nr grey level run length matrix, where Ng is the number of discretised grey levels present in the ROI intensity mask and Nr is the maximum possible run length along direction m. Matrix element r.sub.ij of the GLRLM is the occurrence of grey level i with run length j. If N.sub.v is the total number of voxels in the ROI, and Ns is the sum over all elements in M.sub.m

[0199] In the context of GLRLM measures, r.sub.i. is the marginal sum of the runs over run lengths j for grey value i.

r.sub.i.=Σ.sub.j=1.sup.N.sup.rr.sub.ij.

[0200] Similarly, in GLRLM measures r.sub..j is the marginal sum of the runs over the grey values i for run length j.

r.sub..j=Σ.sub.i=1.sup.N.sup.gr.sub.ij.

[0201] The controller may be configured to determine grey level non-uniformity, e.g. based on r.sub.i. the marginal sum of the runs over run lengths j for grey value i, viz.

[00030]$F_{\mathrm{rlm}.Math.\mathrm{glnu}}=\frac{1}{N_s}\underset{i=1}{\overset{N_g}{.Math.}}{r}_i^2$

where N.sub.s is the sum over all elements in M.sub.m in the GLRLM.

[0202] The controller may also be configured to determine run length non-uniformity based on r.sub..j the marginal sum of the runs over the grey values i for run length j. For example, this may comprise an image metric whose value is low when runs are equally distributed along run lengths, e.g.

[00031]$F_{\mathrm{rlm}.Math.\mathrm{rlnu}}=\frac{1}{N_s}\underset{j=1}{\overset{N_g}{.Math.}}{r}_j^2$

[0203] The controller may also be configured to determine measures of run percentage such as the fraction of the number of realised runs and the maximum number of potential runs. Strongly linear or highly uniform ROI volumes may produce low run percentages. One example of run percentage may be defined:

[00032]$F_{\mathrm{rlm}.Math.r.Math.\mathrm{perc}}=\frac{N_s}{N_v}.$

[0204] Any feature of any one of the examples disclosed herein may be combined with any selected features of any of the other examples described herein. For example, features of methods may be implemented in suitably configured hardware, and the configuration of the specific hardware described herein may be employed in methods implemented using other hardware.

Fractal Dimension Features

[0205] A fractal dimension of a region may be based on a statistical index of complexity comparing how detail in a pattern, such as the boundary of an image region changes with the scale at which it is measured. It may also be based on the space-filling capacity of such a pattern. Fractal dimensions considers the space filling properties of the images, and may be defined based on methods such as those defined in “A multifractal approach to space-filling recovery for PET quantification.” Med Phys. 2014 November; 41(11):112505. doi: 10.1118/1.4898122. Willaime J M, Aboagye E O, Tsoumpas C, Turkheimer F E.

Spatial Filtering

[0206] In addition to the image metrics defined above, the controller may also be configured to apply filters such as high pass and/or low pass filters to the image data (e.g. prior to calculation of the image metrics). These filters may be wavelet based. It will be appreciated in the context of the present disclosure that three-dimensional wavelets can be constructed as separable products of 1-D wavelets by successively applying a 1-D analyzing wavelet in three spatial directions (x, y, z). The volume F (x, y, z) is firstly filtered along the x-dimension, resulting in a low-pass image L(x, y, z) and a high-pass image H(x, y, z). Both L and H are then filtered along the y-dimension, resulting in four decomposed sub-volumes: LL, LH, HL and HH. Then each of these four subvolumes are filtered along the z-dimension, resulting in eight sub-volumes: LLL, LLH, LHL, LHH, HLL, HLH, HHL and HHH.

[0207] In 1D dimension, the continuous wavelet transform is defined as the convolution of x(t) with a wavelet function, W(t), shifted in time by a translation parameter b and a dilation parameter a:

[00033]$X_W\left(a,b\right)=\frac{1}{\sqrt{a}}\underset{-\infty }{\overset{+\infty }{\int }}W\frac{t-b}{a}x\left(t\right)\mathrm{dt}$

[0208] The discrete form of the wavelet transform is based upon the discretization of parameters (a, b) on the time-scale plane corresponding to a discrete set of continuous basis functions. This can be achieved defining:

[00034]$W_{j,k}\left(t\right)=\frac{1}{\sqrt{a_j}}W\frac{\left(t-b_k\right)}{a_j}$

[0209] For a.sub.j=a.sub.0.sup.j and b.sub.k=kb.sub.0 a.sub.0.sup.j where j, k∈Z, a.sub.0>1, b.sub.0≠0 where j controls the dilation and k controls the translation. Two popular choices for the discrete wavelet parameters a.sub.0 and b.sub.0 are 2 and 1 respectively, a configuration that is known as the dyadic grid arrangement resulting in:

[00035]$W_{j,k}\left(t\right)=a_0^{-\frac{j}{2}}W\left(a_0^{-j}t-k{b}_0\right)=2^{-\frac{j}{2}}W\left(2^{-j}t-k\right)$

[0210] Wavelet analysis is simply the process of decomposing a signal into shifted and scaled versions of a mother (initial) wavelet. An important property of wavelet analysis is perfect reconstruction, which is the process of reassembling a decomposed signal or image into its original form without loss of information. For decomposition and reconstruction the scaling function Φ.sub.jk(t) and the wavelet W.sub.jk(t) are used in the form:

[00036]${\Phi }_{jk}\left(t\right)=2^{-\frac{j}{2}}{\Phi }_0\left(2^{-j}t-k\right)$$W_{jk}\left(t\right)=2^{-\frac{j}{2}}{\Psi }_0\left(2^{-j}t-k\right)$

where m stands for dilation or compression and k is the translation index. Every basis function W is orthogonal to every basis function Φ.

[0211] The one-dimensional wavelet transform of a discrete-time signal x(n) (n=0, 1, . . . , N) is performed by convolving signal x(n) with both a half-band low-pass filter L and high-pass filter H and downsampling by two.

[00037]$c\left(n\right)=\underset{n-0}{\overset{L-1}{.Math.}}{h}_0\left(k\right)x\left(n-k\right)$$d\left(n\right)=\underset{n-0}{\overset{L-1}{.Math.}}{h}_1\left(k\right)x\left(n-k\right)$

where c(n) represent the approximation coefficients forn=0, 1, 2 . . . ,N−1 and d(n) are the detail coefficients, h0 and h1, are coefficients of the discrete-time filters L and H respectively where c(n) represent the approximation coefficients for n=0, 1, 2 . . . , N−1 and d(n) are the detail coefficients, h.sub.0 and h.sub.1, are coefficients of the discrete-time filters L and H respectively.

{h.sub.0(n)}.sub.n−0.sup.L−1=(h.sub.0(0),h.sub.0(1), . . . ,h.sub.0(L−1))

{h.sub.1(n)}.sub.n−0.sup.L−1=(h.sub.1(0),h.sub.1(1), . . . ,h.sub.1(L−1))

resulting in the separable, sub-band process.

[0212] It will be appreciated in the context of the present disclosure that whilst wavelet based filters may have particular advantages, other types of spatial filters may be used.

IMPLEMENTATION

[0213] By applying the image processing steps outlined above, the controller is configured first to segment the image data for the subject to identify regions of the image data (anatomical neuroanatomical regions of interest, ROIs). It will be appreciated in the context of a 3D image (such as one made up of a set of slices) that an ROI may comprise a 3D volume, e.g. a cluster of voxels spanning more than one slice of a volumetric image.

[0214] For each ROI, the controller determines a selected one or more of the features mentioned above, e.g. [0215] First order statistics such as measures of central tendency, measures of spread, skewness and kurtosis; [0216] Morphological features; [0217] Grey Level Size Zone Matrix features; [0218] Grey Level Co-occurrence Matrix (GLCM) features; [0219] Neighbourhood grey tone difference based features; [0220] Grey Level Run Length Matrix (GLRLM) features; and [0221] Fractal dimension features.

[0222] The controller may store a list of weightings which define the weighting to be given to that feature (image metric) in combining the feature score from each region to provide an indicator. Two examples of such lists of weights have been defined for predicting or diagnosing cognitive impairment in human subjects. The first of these is defined in Table 1, below. The second is defined in Table 2. These two lists have been found to have excellent prediction accuracy, but prediction accuracy which is sufficient to make reliable diagnosis may be provided by other embodiments such as those described and claimed elsewhere herein.

TABLE-US-00001 TABLE 1 Region Feature Weight a Left Cerebral median, and ROI may be 0.3119 White Matter filtered, e.g. using an LHH filter b Left Cerebral Cortex Compactness 1 −0.4431 c Left Cerebral Cortex Mode 1.2220 d Left Inferior A correlation measure of the −1.4074 Lateral Ventricle GLCM (e.g. information based correlation) e Left Hippocampus Compactness 2 −1.0632 f Left Hippocampus NGTDB Based measure of 1.7741 texture strength, and ROI may be filtered, e.g. using an HLL filter g Left Amygdala Compactness 1 −2.5940 h Left Amygdala GLRLM run length non −0.0486 uniformity, and ROI may be filtered, e.g. using an HLL filter i Left Amygdala GLRLM grey level non −0.8351 uniformity, and ROI may be filtered, e.g. using an HHH filter j Right Cerebral Cortex mode 1.70381

[0223] It may be advantageous to age normalise image data before applying the feature analysis defined in Table 1.

TABLE-US-00002 TABLE 2 Region Feature Weight A Left Cerebral Cortex Shape to volume ratio 0.7590 B Left Cerebral Cortex Sphericity −0.4395 C Left Cerebral Cortex Maximum fractal dimension, and −0.4646 ROI may be filtered, e.g. using an HLH filter. D Left Inferior Standard deviation, and ROI −0.0156 Lateral Ventricle may be filtered, e.g. using an LHL filter E Left Inferior A correlation measure of the −1.2023 Lateral Ventricle GLCM (e.g. information based correlation), and ROI may be filtered, e.g. using an LHL filter F Left Pallidum GLSZM large zone emphasis −0.6226 G Brain Stem Maximum fractal dimension, and 0.1561 ROI may be filtered, e.g. using an HHL filter. H Left Hippocampus Minimum, and ROI may be −0.5028 filtered, e.g. using an LLL filter. I Left Hippocampus GLRLM run length non uniformity −1.4729 J Left Amygdala GLSZM grey level non uniformity −1.4162 K Left Amygdala correlation measure of 0.6181 the GLCM (e.g. information based correlation), and ROI may be filtered, e.g. using an HHH filter L Left choroid plexus Maximum fractal dimension −0.0894 M Right Cerebral Cortex Minimum, and ROI may be −0.9432 filtered, e.g. using an LLL filter N Right Cerebral Cortex Maximum, and ROI may be −0.5396 filtered, e.g. using an HLH filter O Right Hippocampus Compactness −1.5872 P WM hypointensities GLCM correlation, and ROI −0.3731 may be filtered, e.g. using an HLL filter Q Corpus Callosum GLCM Maximum probability, −0.2500 Posterior and ROI may be filtered, e.g. using an LHL filter. R Corpus Callosum GLCM Entropy, and ROI 0.5234 Posterior may be filtered, e.g. using an LHL filter

[0224] The controller may apply either (a) the combination of weightings listed in Table 1, or (b) the combination of weightings listed in Table 2 to scale each feature's value in the corresponding ROI before combining the scaled feature values (e.g. by summing them) to provide the indicator. The controller then compares this indicator value with reference data thereby to determine a cognitive impairment diagnosis for the indicator.

[0225] The reference data may be predetermined using the same set of regions, features, and weights as is used to determine the indicator, but based on image data from subjects having a known cognitive impairment diagnosis. The reference data may thus comprise reference value (or range of values) associated with each population on image data from subjects having a known cognitive impairment diagnosis (e.g. control (no impairment), MCI, AD, etc.). The reference data may also be correlated with cognitive testing scores, which may enable a cognitive testing score to be estimated or predicted based on the image data.

[0226] The indicators defined by the combinations of regions, features and weights defined in Table 1 are listed in Table 3. The rows shown in Table 3 indicate the outcome of an ROC analysis.

[0227] In Table 3 the first of the two columns under the heading APV1 (CNvsAD) indicates the application to a trial population of Alzheimer's disease (AD) subjects and control subjects (CN), and the accuracy of the method defined in table 1 in discriminating AD from CN. The second of the two columns under the heading APV1 (CNvsMCI) indicates the application to a trial population of mild cognitive impairment (MCI) subjects and control subjects (CN), and the accuracy of the method defined in Table 1 in discriminating MCI from CN. The columns under the heading Vol Hippocampus indicate the accuracy of discrimination based on the so-called “gold standard” measure provided using the volume of the hippocampus. Compared to the gold standard, our method shows higher AUC, specificity, sensitivity, accuracy, negative and positive predictive values, likelihood ratios and diagnostic odds ratios in both the discriminations between CN/MCI and CN/AD. Table 4 provides the same ROC analysis for indicators defined by the combinations of regions, features and weights defined in Table 2.

[0228] Table 5 shows the same ROC analysis for indicators defined by the combinations of regions, features and weights defined in Table 1 to a different cohort. Table 6 shows the same ROC analysis for indicators defined by the combinations of regions, features and weights defined in Table 2 to a different cohort. The data used to establish tables 5 and 6 provided by OASIS, namely OASIS-3: Principal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P50AG00561, P30NS09857781, P01AG026276, P01AG003991, R01AG043434, UL1TR000448, R01EB009352. Any AV-45 doses were provided by Avid Radiopharmaceuticals, a wholly owned subsidiary of Eli Lilly.

[0229] It will be appreciated in the context of the present disclosure that the weights listed herein need not be given the accuracy quoted here. It is believed that weightings which provide comparable relative contributions of at least two of the more heavily weighted contributions may provide reliable prediction/diagnosis.

TABLE-US-00003 TABLE 3 ApV1 Vol Hippocampus CNvsAD CNvsMCI CNvsAD CNvsMCI AUC 0.9578 0.8477 0.8600 0.7144 Threshold −2.1161 −2.4288 0.3793 0.4399 Specificity 0.9027 0.8240 0.8564 0.7222 Sensitivity 0.9214 0.7281 0.7382 0.6313 Accuracy 0.9115 0.7759 0.8009 0.6766 NPV 0.9285 0.7510 0.7872 0.6610 PPV 0.8934 0.8061 0.8197 0.6954 LR+ 9.4779 4.1387 5.1437 2.2728 LR− 0.0869 0.3299 0.2444 0.5104 YI 0.8242 0.5521 0.6505 0.3535 DOR 108.9524 12.5441 18.0225 4.4525

TABLE-US-00004 TABLE 4 ApV2 Vol Hippocampus CNvsAD CNvsMCI CNvsAD CNvsMCI AUC 0.9642 0.7596 0.8792 0.7146 Threshold −4.8105 −5.2179 0.3573 0.4399 Specificity 0.9537 0.8518 0.8981 0.7175 Sensitivity 0.9162 0.6036 0.8315 0.6313 Accuracy 0.9361 0.7274 0.8669 0.6743 NPV 0.9279 0.6814 0.8584 0.6595 PPV 0.9459 0.8036 0.8777 0.6919 LR+ 19.7905 4.0748 8.1679 2.2355 LR− 0.0878 0.4652 0.1875 0.5137 YI 0.8699 0.4555 0.7296 0.3489 DOR 225.3125 8.7587 43.5397 4.3514

TABLE-US-00005 TABLE 5 ApV1-OAS Volume hip CNvsAD CNvsUND CNvsAD CNvsUND AUC 0.7409 0.7050 0.7774 0.7328 Threshold −2.0260 −2.1677 0.4137 0.5011 Specificity 0.7467 0.7012 0.6667 0.7308 Sensitivity 0.6666 0.6923 0.8766 0.6688 Accuracy 0.7326 0.7000 0.8396 0.6778 NPV 0.9126 0.9310 0.5366 0.2714 PPV 0.3606 0.2812 0.9247 0.9364 LR+ 2.2402 2.2792 7.1037 2.2065 LR− 0.3798 0.4314 0.3802 0.4532 YI 0.4134 0.3936 0.5433 0.3996 DOR 5.8974 5.2826 14.2098 5.4818

TABLE-US-00006 TABLE 6 ApV1-OAS Volume hip CNvsAD CNvsMCI CNvsAD CNvsMCI AUC 0.7884 0.6726 0.7357 0.6655 Threshold −4.9773 −5.5365 0.4054 0.5461 Specificity 0.7878 0.6484 0.6000 0.7429 Sensitivity 0.6857 0.6857 0.8606 0.5636 Accuracy 0.7700 0.6550 0.8150 0.5950 NPV 0.9219 0.9067 0.4773 0.2653 PPV 0.4067 0.2926 0.9103 0.9118 LR+ 2.5068 2.0633 4.3042 1.7023 LR− 0.3093 0.5126 0.2323 0.5874 YI 0.4735 0.3341 0.4606 0.3065 DOR 8.1038 4.0250 9.2604 3.7259

[0230] It can be seen that in all cases, and in both cohorts of subjects, the ability of the method described herein to distinguish control subjects from those with MCI or AD matches or exceeds measurements based on the volume of the hippocampus.

The Use of Reduced Feature Sets

[0231] It is believed that the weightings having the highest absolute values may contribute most strongly to the predictive power of the indicators defined by the lists of tuples defined in Table 1 and Table 2. Therefore not all of the features and regions need be used in order to predict cognitive impairment. Particularly advantageous combinations of features and regions include those described and claimed elsewhere herein.

[0232] To determine the extent to which reduced feature sets may be useful in predicting cognitive impairment two comparative studies were conducted. In the first comparative study the predictive power of the full feature sets listed above in Table 1 was compared with the predictive power of incomplete versions of that same set. In summary—it was determined that the proposal set out in the preceding paragraph is correct, and that methods of predicting or diagnosing cognitive impairment based only on the more strongly weighted image metrics and image regions provide effective predictive power.

[0233] In Tables 7 and 8 listed below the tuples of feature and region listed in Tables 1 and 2 are referred to using row labels. This labelling relates to the image regions and image metrics identified in those rows of Table 1 and Table 2 being used to determine an indicator for comparison with reference data, thereby to predict or diagnose cognitive impairment in the subject in the manner described and claimed herein.

[0234] In the first study, the indicated combinations of features from the rows listed in Table 1 were selected, and tested using an ROC analysis. The complete ROC data is shown in FIG. 4, but summarised below using the AUC (area under curve) obtained from the ROC analysis of each feature set.

TABLE-US-00007 TABLE 7 Rows of ROC- Combination Table 1 AUC Ftot All 0.9578 Ftest4 c, f, g, j 0.9518 P1 c, f, g 0.9207 P2 c, f, j 0.8633 P3 c, g, j 0.9184 P4 f, g, j 0.9296 P5 c, f 0.8635 P6 c, g 0.9145 P7 c, j 0.7197 P8 f, g 0.8563 P9 f, j 0.867 P10 g, j 0.9181

[0235] The performance achieved using only four of the heavily weighted feature-region tuples (Ftest4) is closely comparable with that achieved by using the entire feature set listed in Table 1. Where only three features are used, effective performance is still achieved and the best performance is given by permutation 4 (P4), which involves he Left Hippocampus, Left Amygdala and Right Cerebral Cortex—see rows f, g and j of Table 1.

[0236] Where only two of the feature-region tuples were used, the best performance was provided by the permutation labelled P6, which involves Left Cerebral Cortex and Left Amygdala—see rows c and g of Table 1. The worst performance was given by the combination of Left and Right Cerebral Cortex alone, labelled P7. Even this however provided a reasonable degree of class separation, so a measurable predictive effect is present even in this less preferred embodiment.

[0237] It can thus be seen that whilst there are some particularly advantageous combinations, selection of any two of the four heavily weighted feature-region tuples listed above provides effective prediction or diagnosis of cognitive impairment. The full table with complete ROC data is set out in FIG. 4.

[0238] In the second study, combinations of features from the rows listed in Table 2 were selected and tested using an ROC analysis. The complete ROC data is shown in FIG. 5, but is summarised below using the AUC (area under curve) obtained from the ROC analysis of each feature set.

TABLE-US-00008 TABLE 8 Rows of ROC- Combination Table 2 AUC Ftot All 0.9642 Ftest4 E, I, J, O 0.9691 P1 E, I, J 0.9545 P2 E, I, O 0.9679 P3 E, J, O 0.9584 P4 I, J, O 0.9597 P5 E, I, 0.9587 P6 E, J 0.9327 P7 E, O 0.9396 P8 I, J 0.9222 P9 I, O 0.9662 P10 J, O 0.9487

[0239] As with the reduced feature sets explained with reference to Table 7, Table 8 also shows that the performance achieved using only four of the heavily weighted feature-region tuples from Table 2 (Ftest4) is closely comparable with that achieved by using the entire feature set of Table 2.

[0240] Where only three features are used, effective performance is still achieved and the best performance is given by permutation 4 (P4), which involves the Left Hippocampus, Left Amygdala and Right Hippocampus—see rows I, J and O of Table 2.

[0241] Where only two of the feature-region tuples were used, the best performance was provided by the permutation labelled P9, which involves Left Hippocampus and Right Hippocampus—see rows I and O of Table 2. The performance of all of the reduced sets using only two feature-region tuples of Table 2 provided AUC performance of well above 0.9

[0242] It can thus be seen that whilst there are some particularly advantageous combinations, selection of any two of the four heavily weighted feature-region tuples listed above provides very effective prediction or diagnosis of cognitive impairment. The full table with complete ROC data is set out in FIG. 5.

The Use of Sub-Regions of the Cortex

[0243] In addition, or as an alternative to, the use of the image regions outlined above embodiments of the present disclosure comprise methods which identify regions of interest ROIs comprising sub-regions of the cortex (which may be referred to as sub cortical regions). Such methods of predicting or diagnosing cognitive impairment may find particular application in providing a visual guide to the stage of cognitive impairment and/or the progress of disease states such as Alzheimer's disease (AD).

[0244] The apparatus which performs these methods may be identical to that outlined above, other than in that it may comprise a display means for providing an overlay of the results of the methods disclosed herein on an image of a subject's brain and/or on a standard brain such as an anatomical atlas.

[0245] FIG. 6, FIG. 7 and FIG. 8 illustrate the type of visual output which may be obtained from such methods FIG. 7 relates to the data obtained from mild cognitive impairment (MCI) and alzheimer's patients (AD) in age normalised image data. FIG. 8 relates to the data obtained from mild cognitive impairment (MCI) and alzheimer's patients (AD) in non-age normalised image data. Such visual indications may be determined by calculating the feature/region tuples defined in any one of Tables 9 to 13 below, and displaying the results as an overlay on an anatomical/structural image of the human brain to assist a clinician in diagnosing/stratifying patients.

[0246] The controller may also be configured to identify, in image data having appropriate contrast and resolution, certain sub-regions of the cortex. The sub-regions of the cortex may comprise: [0247] Banks of the Superior Temporal Sulcus [0248] Caudal Middlefrontal [0249] Entorhinal [0250] Frontal pole [0251] Fusiform [0252] Inferior parietal [0253] Inferior temporal [0254] Insula [0255] Isthmus cingulate [0256] Lateral Occipital [0257] Lateral Orbitofrontal [0258] lingual [0259] Middle temporal [0260] Parahippocampal [0261] Pars Orbitalis [0262] Pars Triangularis [0263] Pericalcarine [0264] Precentral [0265] Precuneus [0266] Rostral anterior cingulate [0267] Rostral middlefrontal [0268] Superior parietal [0269] Superior temporal [0270] Supramarginal [0271] Temporal pole [0272] Transverse Temporal

[0273] In the embodiments in which these sub regions of the cortex are used, the controller is able to identify both the left and right brain locations of these regions. The image metrics employed in each of these regions may comprise any one or more of those outlined above. The image data may be age normalised as described above.

[0274] In particular embodiments, the controller is configured to identify: the left entorhinal, left fusiform and the right temporal pole, and the transverse temporal pole. In each of these regions the controller then determines one or more of the following image metrics: [0275] a metric of image texture; [0276] a metric of image intensity; and [0277] a morphological metric of the image region.

[0278] Different implementations can be used—and some of the metrics which have been found to be of particularly strong predictive power are outlined in Table 9, Table 10, Table 11, and Table 12, below. It can be seen from these different tables of data that, in addition to the consistent usefulness of these four regions, the insula was also found to play a significant part in some approaches.

[0279] Table 9, below, sets out one set of tuples of sub-cortical regions, and the image metrics for each region. These have been applied to age normalised data and found to have a particularly strong predictive effect. The weightings listed can be used to combine the image metrics from the identified regions, for example in a linear sum. Other predetermined methods may also be used. The resultant indicator may be used in the prediction and/or diagnosis of cognitive impairment as described above.

[0280] In addition to the quantitative predictive/diagnostic effectiveness of these methods, it has also been found that overlays of the relevant image metrics on images of the human brain provide a useful marker of the extent of cognitive impairment. Examples of such overlays are illustrated in FIG. 6. The display of visual indicators of the metrics of sub-regions of the cortex as defined herein may therefore provide a useful diagnostic aide, to assist clinicians in making an assessment of cognitive impairment in a subject.

TABLE-US-00009 TABLE 9 Region Feature Weight L Entorhinal shape and size surface to 3.295269 spherical disproportion L Entorhinal GLSZM small zone high grey level 1.987232 emphasis with HHL filter L Entorhinal GLSZM small zone high grey level 1.365442 emphasis with LHL filter L Entorhinal GLCM difference variance 1.319203 with LHH filter L Entorhinal GLRLM short run emphasis 0.379947 L Entorhinal GLSZM small zone high grey level 0.053123 emphasis with HLL filter L Fusiform GLRLM short run length low grey 3.272863 level emphasis with LHL filter L Fusiform First Order Statistics 1.28828 mode with LHL filter L Fusiform first order statistics 5.42E−06 minimum with LHL filter R Insula GLRLM short run emphasis 2.922947 with LLL filter R Temporalpole First Order Statistics −1.70099 median with LLH filter R Temporalpole First Order Statistics −0.58523 mean with LLH filter R Transverse Mean Absolute Deviation −3.11412 Temporal with LHH filter R Transverse mean fractal dimension −2.88058 Temporal with HLL filter R Transverse First Order Statistics 1.712749 Temporal mode with HHH filter R Transverse Mean Absolute Deviation −0.18015 Temporal with HHH filter R Transverse First Order Statistics 6.43E−06 Temporal minimum with HHH filter

[0281] An extract of the ROC analysis of using this approach to distinguish between age normalised images of control subjects and subjects having mild cognitive impairment (MCI) is illustrated in FIG. 9.

[0282] It is strongly indicated by the data set out above in Table 7 and Table 8, that when combining image metrics from different regions in this way, the predictive power of the method may remain even when using reduced feature sets. This hypothesis was tested by making comparative assessments of the predictive power of the method defined in Table 9, above. In this comparative test, the predictive power of the full feature set listed in Table 9 was compared against the power of a reduced feature set in which only 5 tuples were selected. The tuples chosen for the comparative test were: [0283] 1. a morphological feature of the left entorhinal (e.g. spherical distortion) [0284] 2. a texture feature of the left fusiform (e.g. GLRLM short run low grey level emphasis) [0285] 3. an image intensity of the right temporal pole (median intensity) [0286] 4. an image intensity of the right transverse temporal cortex—(e.g. mean absolute deviation, which may comprise the mean distance of all intensity values from the Mean Value of the image array

[00038]$\mathrm{Mean}\mathrm{Abs}\mathrm{Dev}=\frac{1}{N_P}{.Math.}_{i=1}^{Np}.Math.X\left(i\right)-\overline{X}.Math.;\mathrm{and}$ [0287] 5. a texture feature of the right insula (e.g. GLRLM short run emphasis)

[0288] It can be seen from FIG. 9 that the performance of the reduced feature set (labelled F5 in FIG. 9) which uses just five (5) of the 17 tuples listed in Table 9 provides a predictive/diagnostic effect nearly equivalent to that provided by the full feature set. This is consistent with the observations set out above—namely that a reduced feature set using the more strongly weighted feature/region tuples provides effective performance.

[0289] It can be seen from an inspection of Table 9, above, that whilst certain texture, intensity, and morphology features have been selected in each of these regions other such features also provide significant contributions to the predictive power of the approach. It is therefore believed that the combination of the particular image regions, and texture/morphology/intensity based image metrics derived from those regions provides the predictive/diagnostic effect of the present disclosure. The precise detail of the metrics themselves is a lesser consideration than the underlying anatomical/physiological effect which they reveal. Other image metrics of texture, morphology and so forth may be used.

[0290] A further predictor was derived using control patients and patients with Alzheimer's Disease. The data used to establish this model was age normalised in the manner described above. The feature/region tuples used in this predictor are set out below in Table 10

TABLE-US-00010 TABLE 10 Regions long name Weights L Banks of the GLSZM small zone low grey 0.170415 Superior level with LLL filter Temporal Sulcus L Caudal grey level size zone matrix 0.144365 Middlefrontal small zone emphasis with LHL filter L Enthorinal shape and size surface to 3.049285 volume ratio L Enthorinal GLSZM zone percentage with 0.956581 LLL filter L Enthorinal NGTDM busyness with LLH filter −0.77649 L Enthorinal GLSZM small zone high grey level 0.229988 emphasis with HLL filter L Enthorinal GLCM Correlation with −0.01224 HHH filter L Enthorinal shape and size surface −0.00135 to sphericity L Frontalpole First Order Statistics mean 0.153934 with HHL filter L Fusiform GLSZM grey level variance 0.46578 L Fusiform mean fractal dimension −0.41496 with HHL filter L Fusiform NGTDM complexity with LLL filter 0.178951 L Fusiform GLCM cluster prominence 0.022571 with LLH filter L Inferiorparietal GLRLM run length nonuniformity −2.95444 L Inferiorparietal GLCM information based measure of −0.0713 correlation 1 with HLL filter L Insula First Order Statistics mode with −2.74864 HLL filter L Insula variance of the fractal dimension −0.63036 with a HHL filter L Insula first order statistics minimum −0.00016 with HLL filter L Isthmus maximum fractal dimension −0.48581 cingulate with HHH filter L Isthmus Grey level co-occurrence matrix −0.30605 cingulate joint entropy with a HLH filter L Isthmus First Order Statistics mode −0.03194 cingulate with LLH filter L Isthmus first order statistics minimum −1.24E−05 cingulate with LLH filter L Lateral fractal dimension standard −2.1886 Occipital deviation with LHL filter L lingual grey level size zone matrix 0.314208 size zone non uniformity with LLL filter L lingual GLCM cluster shade with HLL filter 0.089456 L lingual Grey level co-occurrence matrix −0.01972 maximum probability with LLH filter L Parahippocampal shape and size, compactness 1 −0.94001 L Parahippocampal NGTDM busyness with LHL filter −0.61542 L Parahippocampal minimum fractal dimension 0.400015 with LLL filter L Parahippocampal NGTDM strength with LLL filter 0.091937 L ParsOrbitalis First Order Statistics mean −0.49261 with HLL filter L Parstriangularis First Order Statistics mean 1.283595 with HLH filter L Precentral GLCM difference entropy 0.617321 L Precentral GLSZM small zone high grey level 0.053608 emphasis with LLL filter L Precuneus GLCM information based measure −0.6645 of correlation with LHL filter L Rostral GLSZM zone size variance 0.964688 middlefrontal L Rostral NGTDM strength with HHH filter 0.598154 middlefrontal L Superiorparietal GLSZM grey level variance 0.08388 with HLL filter L Superiortemporal GLSZM zone low grey level 1.521496 emphasis with LLH filter L Superiortemporal First Order Statistics maximum −1.35223 with LLL filter L Superiortemporal Grey level co-occurrence matrix 1.115139 sum average with HHH filter L Superiortemporal First Order Statistics skewness 0.494638 with HLL filter L Supramarginal GLSZM zone high grey level −2.20863 emphasis with HLL filter L Supramarginal First Order Statistics median 0.211957 with HLH filter L Temporalpole GLCM cluster prominence with 0.264933 LHL filter L Temporalpole GLCM sum entropy with LLL filter −0.17607 R Banks of the First Order Statistics median −1.48796 Superior with LLL filter Temporal Sulcus R Banks of the Grey level co-occurrence matrix −1.15302 Superior joint entropy with a LLH filter Temporal Sulcus R Banks of the Grey level co-occurrence matrix −0.70663 Superior joint entropy with a LHL filter Temporal Sulcus R Banks of the maximum fractal dimension −0.61167 Superior with HHL filter Temporal Sulcus R Banks of the GLCM difference entropy 0.199848 Superior with LLL filter Temporal Sulcus R Banks of the GLSZM zone percentage with 0.17245 Superior LLL filter Temporal Sulcus R Banks of the GLCM information based measure of 0.032353 Superior correlation with HLL filter Temporal Sulcus R Banks of the First Order Statistics maximum −0.02894 Superior with LLL filter Temporal Sulcus R Caudal GLSZM zone size variance 1.102021 Middlefrontal with LLL filter R Enthorinal shape and size, −1.10584 compactness 2 R Enthorinal GLSZM zone percentage with 1.012529 LLL filter R Enthorinal minimum fractal dimension 0.64708 with LHL filter R Enthorinal GLRLM long run emphasis −0.51931 R Enthorinal grey level size zone matrix small 0.470111 zone emphasis with LLH filter R Enthorinal GLCM information based measure 0.312182 of correlation with LLH filter R Enthorinal shape and size surface to 0.105788 volume ratio R Fusiform maximum fractal dimension −1.8349 with LHL filter R Inferiorparietal GLSZM large zone high grey level −1.28437 emphasis with LLL filter R Inferiorparietal GLCM information based measure of −1.20131 correlation 1 with HHL filter R Inferiortemporal GLRLM high grey level long run −0.52235 emphasis with HHL filter R Insula First Order Statistics mode 2.524577 with LLL filter R Insula First Order Statistics mean 0.507135 with LHH filter R Insula first order statistics minimum 0.000107 with HLL filter R LateralOccipital minimum fractal dimension 1.137846 with LLL filter R Middletemporal grey level size zone matrix size zone −2.25961 non uniformity with HHL filter R Pars mean fractal dimension 3.920725 Triangularis with LLL filter R Pars lacunarity of fractal dimension −0.4736 Triangularis with LLL filter R Precentral First Order Statistics median −0.56497 with LHL filter R Precuneus mean fractal dimension with 1.061564 LLL filter R Supramarginal GLSZM small zone high grey level 0.113365 emphasis with LHL filter R Temporalpole minimum fractal dimension 1.981675 with LHH filter R Transverse First order statistics mean absolute −0.83475 Temporal deviance with HLL filter R Transverse GLRLM short run high grey level −0.48063 Temporal emphasis with HHH filter R Transverse First Order Statistics maximum −0.47761 Temporal with LHL filter

[0291] Significantly, it can be seen that a morphological feature of the entorhinal cortex is again very heavily weighted in this data. In this instance the parameter is the surface to volume ratio, but it will be appreciated in the context of the present disclosure that such parameters are not unrelated to the sphericity of a region, or indeed the spherical disproportion. Embodiments of the present disclosure may thus comprise methods of predicting or diagnosing cognitive impairment based on morphological features of the entorhinal cortex.

[0292] Other features which are heavily weighted in Table 10 are also heavily weighted in Table 9. For example the insula, the fusiform, and the temporal pole amongst other regions are heavily weighted in both Table 9 and Table 10.

[0293] To test the hypothesis that the most strongly weighted feature/region tuples can provide an effective predictor without use of the full feature set, a set of comparative tests were performed. In these comparative tests, the full 80 tuple feature set defined in Table 10 was compared against the reduced feature sets shown in FIG. 10.

[0294] In FIG. 10, F.sub.tot indicates the full 80 tuple feature set. The following feature sets are also defined: [0295] F35: 35 features correspondent to the 35 highest weighted features per region [0296] F18: the 18 highest weighted among the 35 [0297] F10: the 10 highest weighted among the 18

[0298] An ROC analysis was performed to compare the predictive/diagnostic power of the full feature set against the reduced sets. It can be seen that even the 10 tuple feature set, F10, provides an effective predictor with an AUC of 0.73. Again, the entorhinal, temporal pole, and insula provide a significant contribution to the effectiveness of this predictor.

[0299] Table 11, below shows the feature/region tuples developed for a non-age normalised data set comprising control patients and patients with mild cognitive impairment (MCI)

TABLE-US-00011 TABLE 11 Regions long name Weights L Caudal NGTDM complexity with −1.31758 Middltefrontal LHL filter L Caudal GLCM difference variance 0.116458 Middltefrontal L Enthorinal GLCM Correlation with 1.540307 HLL filter L Enthorinal maximum fractal dimension −0.95594 with LHL filter L Enthorinal GLSZM small zone emphasis 0.950351 L Enthorinal minimum fractal dimension 0.695691 with HHL filter L Enthorinal First Order Statistics mean −0.60892 with LHL filter L Enthorinal minimum fractal dimension 0.176888 L Enthorinal GLCM inverse variance −0.1582 L Enthorinal maximum fractal dimension −0.13014 with HHH filter L Inferiortemporal First Order Statistics RMS 0.242942 with HLL filter L Isthmus First Order Statistics mean −0.66677 cingulate with HLL filter L LateralOccipital NGTDM complexity with HHL filter 2.161983 L Middletemporal GLCM Contrast 0.343334 L Middletemporal GLSZM large zone low grey level −0.24811 emphasis with HHH filter L Parahippocampal Grey level co-occurrence matrix 0.322795 sum variance with LHH filter L Parahippocampal Grey level co-occurence matrix 5.30E−10 cluster tendency with LHH filter L ParsOrbitalis mean fractal dimension 1.38805 L Superiortemporal maximum fractal dimension −1.10252 with HHL filter L Superiortemporal mean fractal dimension 0.150033 L Superiortemporal NGTDM complexity 0.022989 R Banks of the mean fractal dimension with −2.40888 Superior LHL filter Temporal Sulcus R Banks of the Grey level co-occurrence matrix 0.742583 Superior sum variance with HHL filter Temporal Sulcus R Banks of the First Order Statistics −0.60401 Superior median with LLH filter Temporal Sulcus R Banks of the Grey level co-occurence matrix 2.78E−10 Superior cluster tendency with HHL filter Temporal Sulcus R Enthorinal shape and size, compactness 2 −1.15717 R Enthorinal Grey level co-occurrence 0.318659 matrix sum variance R Enthorinal First Order Statistics mean −0.19651 with LLH filter R Enthorinal Grey level co-occurence 0.000878 matrix cluster tendency R Inferiorparietal minimum fractal dimension 0.90721 with HHL filter R Insula minimum fractal dimension 0.584533 with HLL filter R Insula First Order Statistics mean 0.129373 with HHL filter R Middletemporal GLSZM small zone high grey level 0.745496 emphasis with HHL filter R Middletemporal minimum fractal dimension 0.165184 with HLL filter R Pars Orbitalis NGTDM strength with LLL filter 0.538842 R Pars Orbitalis NGTDM busyness −0.35922 R Rostral First Order Statistics 1.051034 Middlefrontal median with HHL filter R Rostral GLSZM small zone low grey 0.323497 Middlefrontal level with LLL filter R Temporalpole GLSZM small zone low grey 0.709318 level with LLL filter R Temporalpole GLSZM large zone high grey −0.23521 level emphasis with HHH filter R Temporalpole GLSZM large zone emphasis −0.20974 with HLH filter

[0300] It can be seen that this analysis gives rise to 41 feature/region tuples in 17 regions of the brain. A comparative test was performed in which the performance of a predictor based on 17 tuples (the highest weighted feature for each region) was tested against the full feature set and a further reduced feature set was also used in which only the 8 highest weighted tuples from amongst that 17 were used. These are labelled F.sub.tot, F17, and F8 respectively in FIG. 11. It can be seen from the ROC analysis presented in FIG. 11 that both reduced feature sets provide effective predictive/diagnostic power.

[0301] It can also be seen from a comparison of the performance of the predictive models defined in Tables 9 and 11 that although age normalisation may have some advantages it is by no means essential. An effective prediction can be made without it. In addition the morphology of the entorhinal cortex again plays a significant role—in this instance the compactness.

[0302] Table 12 below shows the feature/region tuples for a predictor developed from a data set comprising control patients and patients with Alzheimer's Disease. The image data was not age normalised.

TABLE-US-00012 Regions Features Weights L Banks of the GLCM information based 0.888641 Superior measure of correlation Temporal Sulcus with LHL filter L Banks of the mean fractal dimension 0.870555 Superior with HHL filter Temporal Sulcus L Banks of the First order statistics 0.740675 Superior mean absolute deviance Temporal Sulcus with HLL filter L Enthorinal neighbourhood grey tone 2.559715 difference matrix (NGTDM) contrast, with LLL filter L Enthorinal Grey level run length 1.624852 matrix long run length low grey level emphasis L Enthorinal GLCM difference variance 1.557292 L Enthorinal GLCM information based 1.076048 measure of correlation with HLH filter L Enthorinal GLRLM short run high 0.916666 grey level emphasis with HHH filter L Enthorinal GLSZM small zone high 0.727289 grey level emphasis L Enthorinal GLSZM grey level variance −0.34558 L Enthorinal GLCM information based 0.127109 measure of correlation L Enthorinal GLCM information based 0.053807 measure of correlation with HLL filter L Enthorinal GLSZM small zone high 0.005602 grey level emphasis with HHH filter L Frontalpole GLSZM small zone high 0.536633 grey level emphasis with HLL filter L Inferioritemporal First Order Statistics 3.526849 RMS with HLL filter L Inferioritemporal First Order Statistics 0.178556 maximum with HLL filter L Inferioritemporal First Order Statistics −0.0899 median with HLL filter L Inferiorparietal GLRLM run length −0.73427 nonuniformity with LHH filter L Inferiorparietal First Order Statistics −0.53161 Energy with HHH filter L Inferiorparietal GLSZM small zone high 0.48923 grey level emphasis with LHH filter L Insula GLRLM short run high 0.995209 grey level emphasis with HHH filter L Isthmus First Order Statistics −1.65017 cingulate median with LLH filter L Lateral First Order Statistics 0.225136 Orbitofrontal maximum with LLH filter L Middletemporal maximum fractal dimension −2.19676 with HHL filter L Middletemporal GLRLM run percentage 1.499056 with HLL filter L Parahippocampal First Order Statistics 0.620384 skewness with LHL filter L Parahippocampal GLCM information based 0.595259 measure of correlation with LLL filter L Parahippocampal First Order Statistics 0.348604 maximum with HLL filter L Parahippocampal GLCM Correlation 0.134036 with LHH filter L Parahippocampal GLSZM grey level variance −0.01686 with HLL filter L ParsOrbitalis mean fractal dimension 0.588439 with LHL filter L Precentral maximum fractal dimension 1.278405 with HHH filter L Precentral GLCM grey level variance 0.117173 L Precuneus maximum fractal dimension −1.56941 with LHL filter L Rostral grey level size zone 1.777615 middlefrontal matrix small zone emphasis L Rostral GLRLM long run high grey −0.07205 middlefrontal level emphasis with HHL filter L Rostral minimum fractal dimension 0.064977 middlefrontal L First Order Statistics 0.47461 Rostralanteriorcingulate maximum with HHL filter L Superiorparietal grey level co-occurrence 2.218497 matrix autocorrelation with HHL filter L Supramarginal GLRLM short run high 0.040094 grey level emphasis with LHH filter L Temporalpole First Order Statistics 0.867909 RMS with LLH filter L Temporalpole Grey level co-occurrence −0.61069 matrix sum average with HLL filter R Banks of the maximum fractal dimension −0.48825 Superior with LLL filter Temporal Sulcus R Banks of the GLSZM small zone low 0.242826 Superior grey level emphasis Temporal Sulcus R Enthorinal GLCM information based 1.475704 measure of correlation with LLH filter R Enthorinal grey level size zone 1.042112 matrix small zone emphasis with LLL filter R Enthorinal maximum fractal dimension −0.81562 with LHL filter R Enthorinal GLCM difference variance 0.124307 R Enthorinal GLCM Correlation 0.109949 with HHH filter R Enthorinal Grey level co-occurrence −0.04415 matrix inverse different moment normalised R Frontalpole First order statistics 1.197703 entropy with HHL filter R Frontalpole GLSZM small zone low 0.939372 grey level with HLH filter R Frontalpole GLSZM zone low grey level 0.727302 emphasis with HLH filter R Frontalpole GLRLM short run length 0.109486 low grey level emphasis with LHH filter R Fusiform NGTDM busyness with −1.23079 HLL filter R Fusiform First Order Statistics 0.278442 median with HHH filter R Fusiform GLSZM grey level variance −0.03732 R Inferiorparietal GLSZM small zone high 0.668561 grey level emphasis with HLH filter R Inferiortemporal GLSZM small zone emphasis 0.423731 with HHH filter R Isthmus First Order Statistics −0.85146 cingulate mode with HLL filter R Isthmus GLCM Autocorrelation −0.79826 cingulate with HHH filter R Isthmus first order statistics −4.12E−06 cingulate minimum with HLL filter R Lateral GLSZM zone size variance −1.56146 Orbitofrontal R Middletemporal GLSZM small zone high 0.215538 grey level emphasis with HHL filter R Parahippocampal maximum fractal dimension 0.741955 with HHL filter R Parahippocampal GLCM information based 0.65008 measure of correlation with HLH filter R Parahippocampal Grey level co-occurrence −0.6041 matrix inverse different moment normalised with HLH filter R Parahippocampal GLCM difference variance −0.16653 with HLL filter R Parahippocampal maximum fractal dimension −0.07323 R Pars Orbitalis Grey level co-occurrence 0.187851 matrix sum variance with LHL filter R Pars Orbitalis GLCM cluster prominence 0.167818 R Pars Orbitalis Grey level co-occurrence 0.002109 matrix cluster tendency with LHL filter R Pericalcarine NGTDM contrast −1.33727 R Precentral GLRLM long run high −0.2258 grey level emphasis with HHL filter R Rostral fractal dimension 1.905555 anterior cingulate standard deviation R Superior First Order Statistics −1.99215 parietal minimum with HLL filter R Transverse Fractal dimension −0.90794 Temporal lacunarity with LLH filter R Transverse First order 0.634112 Temporal statistics range

[0303] Table 12 comprises 78 tuples in 34 different sub-regions of the cortex. As with the data listed in Table 11 in the comparative test, the highest weighted tuple in each region was selected an used to provide a reduced feature set, labelled F34 in FIG. 12. A further reduced feature set was also constructed by using the most heavily weighted 14 tuples from amongst these 34, labelled F14 in FIG. 12. Again, it can be seen that the reduced feature sets both provide effective diagnostic/predictive power. Again the entorhinal cortex plays a significant role, although in this predictor the NGTDM contrast, a texture feature of the entorhinal is significant.

[0304] The consistent finding of the predictors investigated herein is that it is possible to define texture and morphology based models of certain brain regions which can be used to predict the presence of cognitive impairment and/or Alzheimer's disease. A wide variety of models/predictors are possible. Without wishing to be bound by theory, it is believed that those which take account of at least the texture and/or morphology of the entorhinal cortex, the fusiform gyrus, the temporal pole, and the transverse temporal cortex may be the most effective predictors. The insula may also make a significant contribution.

[0305] In further investigation to establish the robustness of the methods described and claimed herein a further dataset was investigated in which the control subjects comprised both healthy controls and subjects with Parkinson's and frontotemporal disease. The other subjects, referred to herein as group ADrp, had prodromal Alzheimer's disease and Alzheimer's disease. Predictors were developed that, in a first stage classification were able to distinguish group ADrp from the control subjects. A further predictor was also developed that is able to distinguish within group ADrp. These two predictors differ from those described above, not only in that the control subjects comprise subjects with Parkinson's and frontotemporal disease, but also in that the predictor uses both cortical regions and sub-regions of the cortex.

[0306] FIG. 13, and Table 13 below set out a predictor to distinguish between (a) control subjects comprising both healthy controls and subjects with Parkinson's and frontotemporal disease, and (b) a group ADrp having both prodromal Alzheimer's disease and Alzheimer's disease.

TABLE-US-00013 TABLE 13 Region Feature Filter Ftot Left Lateral GLCM Correlation LHL 0.081215 Ventricle Right Hippocampus NGTDM Coarseness HLL 0.159293 Left entorhinal GLCM Sum Variance 0.024377 Left entorhinal GLCM Cluster Tendency 4.55E−12 Left lateraloccipital NGTDM Strength HLL 0.038394 Left paracentral NGTDM Strength 0.083421 Left supramarginal First Order Statistics - median HLL −0.1052 Left supramarginal NGTDM Strength LLH 0.022109 Right corpuscallosum First Order Statistics - mean HLL −0.21223 Right corpuscallosum GLCM Cluster prominence HLL 0.085458 Right fusiform GLCM Cluster prominence 0.103157 Right middletemporal Minimum fractal dimension HLH −0.25151 Right postcentral GLCM Cluster prominence 0.02635 Right postcentral Minimum fractal dimension HHH −0.12378 Right GLCM correlation LHL 0.376363 rostralmiddlefrontal Right superiorfrontal GLRLM run length HHH 0.091141 non uniformity Right superiorfrontal Minimum fractal dimension HHH −0.17817 Right supramarginal GLCM correlation LHL 0.757222 Right supramarginal GLCM correlation HLL 0.024897 Right temporalpole First Order Statistics - mean LLH −0.69006

[0307] This predictor based on the above features and weights set out in Table 13 was found to provide an AUC of 0.986 in an ROC analysis, and a specificity of 0.9831. It is thus shown to have very reliable predictive power and specificity—even where the control group comprise subjects having other cognitive impairments. A comparative test was performed to demonstrate the possibility to employ reduced feature sets and to demonstrate that the predictive power is preserved when selected ones of the image regions and image metrics are used in the predictor and other are disregarded. As illustrated in FIG. 13, a reduced feature set was constructed using at least two regions selected from the following four: [0308] Right Middle Temporal [0309] Right Rostral Middle Frontal [0310] Right Supramarginal [0311] Right Temporal Pole

[0312] It was found that all of the predictors generated using these regions reliably distinguished the group ARDP subjects from the control group—even where the control subjects comprised both healthy controls and subjects with Parkinson's and frontotemporal disease. The use of spatial filters such as those indicated in column 13 of table 13 is optional.

[0313] It can be seen that when reduced to three image region-feature tuples, the best performance was given by the predictor labelled permutation 2 (Ftest3-p2) in FIG. 13. This involves Right Middle Temporal, Right Rostral Middle Frontal and Right Temporal Pole. When reduced to 2 Features, best performance given by permutation 7 (Ftest2-p7), which involves Right Middle Temporal and Right Temporal Pole.

[0314] In the Right Middle Temporal, this embodiment may employ a measure of complexity such as fractal dimension (e.g. minimum fractal dimension). In the right temporal pole a measure of central tendency of pixel intensity may be used, such as the mean intensity.

[0315] FIG. 14, and Table 14 below set out the results obtained from this approach when the predictor is developed to distinguish between patients having prodromal Alzheimer's disease and those having Alzheimer's disease

TABLE-US-00014 TABLE 14 Region Feature Filter Weight Left Cerebral First order statistics range HLL 0.01755 Cortex Left Cerebral GLSZM grey level LLL −0.0804 Cortex non uniformity Left Inferior mean fractal dimension HLL −0.0762 Lateral Ventricle Fourth Ventricle GLSZM zone low grey LLL −0.0021 level emphasis Brain Stem NGTDM Coarseness 0.11031 Brain Stem NGTDM Coarseness LLL 0.11706 Left Amygdala NGTDM Coarseness LLL 0.11844 Left choroid FOS minimum −0.1448 plexus Right Cerebral GLCM cluster shade LHH −0.0568 White Matter Right Cerebral FOS Range LLH 0.08329 Cortex Right Inferior GLRLM run length non HLL 0.24727 Lateral Ventricle unformity Right Inferior NGTDM Busyness HLH 0.00999 Lateral Ventricle Right Cerebellar GLCM sum average LLH 0.124 cortex Right Hippocampus shape to volume ratio 0.13459 Right Ventral DC NGTDM Coarseness LLL 0.06172 Right Vessel minimum fractal dimension LHH −0.0011 Right banks of the GLCM cluster shade −0.0215 superior temporal sulcus R fusiform Kurtosis HHH 0.02041

[0316] This predictor based on the above features and weights set out in Table 14 was found to provide an AUC of 0.8121 in an ROC analysis, and a specificity of 0.65. It is thus shown to have reasonable predictive power and specificity—even where distinguishing between different Alzheimer's disease states. A comparative test was performed to demonstrate the possibility to employ reduced feature sets and to demonstrate that the predictive power is preserved when selected ones of the image regions and image metrics are used in the predictor and other are disregarded. As illustrated in FIG. 14, a reduced feature set was constructed using at least two regions selected from the following four: [0317] Left choroid Plexus [0318] Right Inferior Lateral Ventricle [0319] Right Cerebellar cortex [0320] Right Hippocampus

[0321] Surprisingly in this predictor higher accuracy may be achieved by the reduced feature sets. When reduced to these 4 regions & features, the accuracy of the classification goes to 0.72 (compared to 0.68 given by the complete set of features & regions). When reduced to 3 Features, best performance is given by permutation 4 (labelled Ftest3-p4 in FIG. 14), which involves Left choroid Plexus, Right Cerebellar cortex and the Right Hippocampus.

[0322] When reduced to 2 Features, the higher accuracy is obtained with permutation 7 (Ftest2-p7), which involves Left choroid Plexus and Right Hippocampus. The feature used in the left choroid plexus may comprise a measure of intensity such as the minimum intensity. The feature used in the right hippocampus may comprise a morphological measure, such as the shape to volume ratio. This permutation gives the highest accuracy.

[0323] Embodiments of the disclosure may employ these methods sequentially—in a first classification step, a predictor such as any one of those described above with reference to FIG. 13 and/or Table 13 may be used to diagnose or to predict a cognitive impairment state corresponding to Alzheimer's disease. Then, in the event that this first method step does indicate a cognitive impairment state corresponding to Alzheimer's disease, a predictor such as any one of those described above with reference to FIG. 14 and/or Table 14 may be applied to diagnose or to distinguish between Alzheimer's disease and prodromal Alzheimer's disease in the patient. It will be appreciated in the context of the present disclosure that the prodromal stage of Alzheimer's disease may also referred to as mild cognitive impairment (MCI) due to Alzheimer's disease, and this is the stage where there are obvious symptoms of brain dysfunction. Accordingly, in the present disclosure MCI may comprise the prodromal stage of Alzheimer's disease.

[0324] These and other methods of the disclosure may be used to stratify patients for clinical trials and/or to assess the effectiveness of treatments or therapies applied to said patients. It will be appreciated in the context of the present disclosure that whilst particular metrics of image structure in the image regions identified herein have been found to have particular differentiating power, other metrics—e.g. such as metrics of structure, shape, complexity and texture may also be used.

[0325] The reference data store used to store data against which test values are compared may comprise volatile and/or non-volatile data storage for storing the above described reference data. The reference data may comprise data calculated by applying the image metrics described herein to a set of reference images of the human brain (a training data set). This training data set may comprise a large number of images of different subjects for which the cognitive impairment diagnosis of each subject is known—e.g. having been verified by other means.

[0326] It will be appreciated from the discussion above that the embodiments shown in the Figures are merely exemplary, and include features which may be generalised, removed or replaced as described herein and as set out in the claims. With reference to the drawings in general, it will be appreciated that schematic functional block diagrams are used to indicate functionality of systems and apparatus described herein. It will be appreciated however that the functionality need not be divided in this way, and should not be taken to imply any particular structure of hardware other than that described and claimed below. The function of one or more of the elements shown in the drawings may be further subdivided, and/or distributed throughout apparatus of the disclosure. In some embodiments the function of one or more elements shown in the drawings may be integrated into a single functional unit.

[0327] In some examples the functionality of the controller may be provided by a general purpose processor, which may be configured to perform a method according to any one of those described herein. In some examples the controller may comprise digital logic, such as field programmable gate arrays, FPGA, application specific integrated circuits, ASIC, a digital signal processor, DSP, or by any other appropriate hardware. In some examples, one or more memory elements can store data and/or program instructions used to implement the operations described herein. Embodiments of the disclosure provide tangible, non-transitory storage media comprising program instructions operable to program a processor to perform any one or more of the methods described and/or claimed herein and/or to provide data processing apparatus as described and/or claimed herein. The controller may comprise an analogue control circuit which provides at least a part of this control functionality. An embodiment provides an analogue control circuit configured to perform any one or more of the methods described herein.

[0328] The embodiments described herein need not perform any diagnosis in order to provide a technical advantage. In particular, there is no need for any comparison with standard data because patients can simply be stratified according to the indicators obtained from the combinations of features (image metrics), image regions, and weights described herein. Such indicators may be used to stratify subjects, for example to identify cohorts for clinical trials Accordingly, methods of the present disclosure comprise computer implemented methods of processing images, such as T1 weighted MRI images to determine any one or more of the indicators based on image regions and image metrics described herein. Such methods may further comprise conducting clinical trials and/or processing clinical measurement data obtained from the subjects to investigate the efficacy of therapies, such as drug treatments. It will thus be appreciated that the methods and apparatus described herein offer a new method of physiological measurement.

[0329] The above embodiments are to be understood as illustrative examples. Further embodiments are envisaged. It is to be understood that any feature described in relation to any one embodiment may be used alone, or in combination with other features described, and may also be used in combination with one or more features of any other of the embodiments, or any combination of any other of the embodiments. Furthermore, equivalents and modifications not described above may also be employed without departing from the scope of the invention, which is defined in the accompanying claims.