EP1456798A2

EP1456798A2 - Evaluation of images of the brain obtained by means of functional magnetic resonance tomography

Info

Publication number: EP1456798A2
Application number: EP02791617A
Authority: EP
Inventors: Bernd SCHÜRMANN; Gustavo Deco; Martin Stetter; Jan Storck; Silvia Corchs
Original assignee: Siemens AG
Current assignee: Siemens AG
Priority date: 2001-12-20
Filing date: 2002-12-09
Publication date: 2004-09-15
Also published as: US20050119558A1; US7349728B2; WO2003054794A3; WO2003054794A2; CN1620666A; DE10162927A1

Abstract

The invention relates to a method for evaluating an image (fMRI-image) of the brain obtained by means of functional magnetic resonance tomography. According to said method, a neuronal network is used to simulate the activities of the brain. Supposed disorders of the brain are simulated in the neuronal network (disturbed neuronal network). The activities determined in the disturbed neuronal network are compared with the activities observed in the fMRI image. The loss of function of any substructures of the brain can be artificially simulated in the model and its effect on the complex synergy of the areas of the brain can be quantified. The comparison with the fMRI image or fMRI activity pattern relating to the patient enables the cause of the disorders to be localised, thus leading to a successful diagnosis.

Description

description

Evaluation of images of the brain obtained using functional magnetic resonance tomography

Thanks to the rapid development in the field of functional magnetic resonance imaging (functional magnetic resonance imaging, fMRI for short, or functional magnetic resonance tomography, fMRI for short), it is increasingly possible to distribute the brain activity of patients while solving complex perception and planning as well as of motor tasks.

However, the enormous possibilities of this technology still conflict with its current use for diagnosis and diagnosis in neurology and neurosurgery.

The main reason for this discrepancy lies in the complexity of the brain: The human cerebral cortex alone can be divided into around 200 functional units, the brain areas, between which there are around 10,000 closed, dense network paths. The network paths each consist of several synaptic connections, ie nerve strands.

As a result of this complex structure and the distributed-parallel signal processing in the brain, malfunctions can only be clearly seen in very rare cases by a change in activity in the fMRI image - for example in an area.

Rather, cerebral dysfunction usually manifests itself in a change in the interaction of the areas compared to healthy people, which in turn modifies the activity state of the entire brain in a complex manner.

For the doctor, this results in the highly non-trivial problem that has so far been solved with the exception of a few special cases the measured complex abnormalities in the fMRI image to determine the actual cause of the cerebral disorder.

The object of the invention is to make fMRI images more usable for diagnosis.

This object is achieved by the inventions according to the independent claims. Advantageous developments of the inventions are characterized in the subclaims.

According to the invention, a method for evaluating an image (fMRI image) of the brain obtained by means of functional magnetic resonance tomography is specified. A neural network is used to simulate the activities of the brain. Suspected disturbances in the brain are simulated in the neural network (disturbed neural network). The determined activities in the disturbed neural network are compared with the activities observed in the fMRI image. Brain disorders are deduced from the comparison.

The possible causes of malfunction can be divided into several classes:

(a) Failures or partial failures in an area.

(b) Complete failure of one or more connection paths between two or more areas.

(c) Failures at the cellular level, which, for example, change the short-term dynamics of the nerve cell populations, e.g. B. within an area.

Only the first class (a) of causes of malfunction can be identified directly in the fMRI image. However, the effects of these disturbances on other areas are complex due to the dense network and require quantification.

The other two classes of fault causes, (b) and (c), can have a very hidden effect. For example, the dense networking of the brain areas means that the fully constant interruption of the connection between two areas does not only affect the two affected areas. The interruption of the connection changes the overall signal propagation through the brain and thus indirectly causes disturbances in other brain functions that appear to be independent of the areas under consideration.

According to the invention, the doctor is provided with a tool with the aid of which the complex interaction of several brain areas can be simulated by the brain while solving defined tasks. The simulator is based on the simulation of the dynamics of coupled populations of neurons, i.e. on a neural network. The dynamics simulate the time course of the activities of the neurons.

The failure of any substructure in the brain can be artificially simulated in the model and its impact on the complex interplay of the areas of the brain can be quantified. The comparison with the fMRI image or activity pattern measured on the patient leads to the localization of the cause of the fault and thus to a successful diagnosis. A quantitative relationship is thus established between the measured spatial distribution of brain activity on the one hand and the medically relevant functional brain state on the other.

In addition to simulating the brain on the basis of a large number of individual neurons, the neural network can also have a structure that is based on the structure of the brain in its division into areas and their connections. On the one hand, this leads to a reduction in the complexity of the neural network. On the other hand, the structure of the neural network corresponds to that of the brain.

A third-generation neurosimulator (neurocognition) is therefore advantageously used for the quantitative interpretation and thus diagnosis of fMRI images. As a neuro- First generation simulators are called models of networks of neurons on a more or less static basis, the classic neural networks. Second generation neurosimulators are models of the dynamic behavior of the neurons, in particular the pulses they generate. Third-generation neurosimulators are hierarchical models of the organization of neurons in pools and pools in areas. A pool contains thousands of neurons.

The method according to the invention can be integrated as an evaluation tool in the operating software of a computer which controls an fMRI tomograph, can be integrated, or can work in the form of an independent diagnostic support device.

The object is further achieved by a computer program which executes the method according to the invention when it is running on a computer, and by a computer program with program code means to carry out all of the steps according to the invention when the program is executed on a computer.

Furthermore, the object is achieved by a computer program with the program code means mentioned, which are stored on a computer-readable data carrier. Furthermore by a

Computer program product with program code means stored on a machine-readable carrier in order to carry out all the steps according to the invention when the program is executed on a computer. Finally, through a data carrier on which a data structure is stored, which executes the method according to the invention after loading into a main memory of a computer.

The invention is explained in more detail below on the basis of exemplary embodiments which are shown schematically in the figures. are posed. The same reference numbers in the individual figures denote the same elements. In detail shows:

Fig. 1 shows an example of an fMRI image;

Fig. 2 simplifies the essential areas of the visual cortex of the brain;

Fig. 3 an abstract representation of the areas of the brain and their synaptic connections; and

Fig. 4 schematically shows the interaction between an area and an associated inhibitory pool.

In the following, the working method of neurocognitive modeling is described using the example of visual attention phenomena.

First, fMRI images of the activity distribution in the brain must be taken. Using positron emission tomography (PET), a procedure that can show metabolic and blood flow changes in the brain with the help of radioactively labeled substances, it could be shown that the activation of certain brain areas leads to a local increase in the blood flow and oxygen consumption of the nerve cells ( Fox PT, Raichle ME: "Focal physiological uncoupling of cerebral blood flow and oxidative metabolism during somatosensory stimulation in human subjects" Procedures of the National Academy of Science of the USA, 1986, volume 83, pages 1140-1144). The supply of oxygen-rich arterial blood necessary for this is locally increased. Normally, however, the blood flow increases disproportionately to the oxygen consumption, so that an excess of oxygenated hemoglobin results in the venous drainage area.

Functional magnetic resonance imaging (fMRI) takes advantage of the fact that the oxygen content of the blood influences its magnetic properties and thus leads to a different borrowed signaling in magnetic resonance imaging leads. Since the magnetic properties and thus the signaling of the blood change with the content of oxygenated or deoxygenated hemoglobin, blood behaves like a (endogenous) contrast medium in functional magnetic resonance imaging. With a high proportion of deoxygenated hemoglobin, a local magnetic field gradient is induced due to its paramagnetic properties in the vicinity of the vessels, which, with a suitable selection of the MRI measurement sequence (e.g. gradient echo sequence or corresponding echo planar imaging sequence), leads to a local signal reduction leads. If the proportion of oxygenated hemoglobin in the blood increases, the so-called susceptibility effect decreases. This leads to an increase in the measurement signal. This relationship is called BOLD contrast (blood oxygen level dependent contrast). With increasing field strength of the magnetic resonance tomograph, this effect increases [Ogawa S, Lee TM, Nayak AS, Glynn P: "Oxygenation-sensitive contrast in magnetic resonance image of circular brain at high magnetic flields." Magnetic Resonance Medicine, 1990, Volume 14, pages 68-78], so that devices with a magnetic field strength of 1.5 Tesla or more are used for functional MRI (fMRI).

The local changes in the oxygen content in the blood during activation processes can be displayed with high spatial resolution using fMRI and precisely assigned to the individual anatomical structures of the brain.

Using fMRI, it can thus be determined which areas of the brain are active at a given point in time or during an activity and which are not, or to what degree individual areas of the brain are active.

A schematic example of an fMRI recording can be seen in FIG. 1. 1 shows a view of the left hemisphere 10 with indicated brain lobes. Shown is a picture of a patient who was given the task, the location to find a point in an image. When solving this task, hatched brain areas 12 are activated, which are visible on the fMRI image.

The aim of the modeling is a detailed neural network model of the areas of the brain, which reflects the real conditions in the brain during activation processes, especially with regard to visual attention control, and thus provides an explanation of these services through physiological mechanisms.

In classic models of image processing, such as digital image processing, a captured image is analyzed by means of a so-called bottom-up approach in successively higher levels of processing.

In contrast to these classic models of image processing, it has been shown that a so-called top-down approach better reflects the realities of the visual cortex. In the top-down approach, intermediate results at a higher processing level are used by feedback to meaningfully reevaluate lower processing levels. What is important is the moment of feedback between the individual levels. In the model to be described in more detail below, this feedback is realized by the interaction of the individual areas.

The feedback leads to a shift in the balance in the attention competition of the individual neurons or groups of neurons. This leads to an uneven situation

Competition for attention ("biased competition"). Only increased attention to a certain spatial area or a characteristic and an associated neglect of the other characteristics or spatial areas enables a reduction in the amount of data in an image and thus a targeted perception of individual objects. When looking for a feature in an image, e.g. B. the Eiffel Tower of Paris, would be differentiated in classic image processing between two questions:

- Where's the Eiffel Tower? This is a so-called "where" question. It looks for the location of the known feature in picture

(Search).

- The second question is: Which object can be seen in the middle of the picture? So a so-called "what" question, the question of recognizing an object at the specified location (object recognition).

The "where" question is traditionally answered by searching the entire image using a given attention window. The "what" question is answered by comparing the known patterns with the predefined image section or by searching the predefined image section for features on which attention is concentrated in order to recognize the image features.

With the new top-down approach, the entire image is processed in parallel. The searched features emerge in the course of processing, i. H. after a while they stand out, e.g. For example, those "grandmother pools" (see below) that have won the competition between the individual pools or characteristics can become active. The "what" and "where" questions are answered using the same model. Only the so-called input bias (see below) is changed, i. h attention is shifted towards "what" or "where". An expectation is generated by means of the bias.

A so-called third-generation neurosimulator is used to model this top-down approach. Third generation neurosimulators are hierarchical models of the organization of neurons in pools and pools in areas, corresponding to the areas of the brain, as described below using the example of the visual cortex is described. A pool contains thousands of neurons.

2 shows in simplified form the essential areas of the visual cortex of the brain 10. The cerebrum 16 and the cerebellum 18 are shown. The cerebrum 16 contains, among others, the areas VI, V4, PP and IT shown and explained in more detail below in the visual cortex , There are many-stranded synaptic connections 20 between these reales.

The structure of the mathematical model will now be described in detail with reference to FIG. 3, which represents an abstract representation of the conditions in the brain.

The IT area (inferotemporal) is used for image recognition or

Object recognition within an image ("what" question). It stores image patterns that can correspond to representations of objects in the visible world. Two patterns, bricks and honeycombs are shown as examples. A pattern is recognized when a so-called "grandmother neuron" assigned to the pattern becomes maximally active. The ability of the "grandmother neuron" to recognize a certain pattern is acquired through training. In the model at hand, the pattern recognition does not work with "grandmother neurons", but with the smallest unit of the model: the pool. A pattern is therefore recognized by a "grandmother pool" when the corresponding grandmother pool is active to the maximum. Accordingly, the IT area in the present model contains as many pools as there are patterns or objects to be recognized.

The area PP (posterior parietal) serves to localize known patterns ("where" question). The area PP in the present model therefore contains as many pools 24 as there are pixels in the image to be recognized. The concentration of neuronal activity in a small number of neighboring pools in PP corresponds to a localization of the object. The areas VI and V4 are combined in the present model to form the A-real V1-V4, which is also referred to as V4. This area is generally responsible for the extraction of features. It contains approximately 1 million pools 24, one pool for each characteristic. The pools 24 speak to individual features of the

Picture. The characteristics of the image result from a Wavelet transformation of the image (see below). A characteristic is thus defined by a certain size or spatial frequency, a spatial orientation and a certain position in the x and y directions (see below). All captured image data initially go to area V1-V4.

In addition, for each area there is at least one inhibitory pool 22, that is to say a pool which has an inhibitory effect on the activity of other pools. The inhibitory pools are coupled to the stimulable pools 24 with bidirectional connections 26. The inhibitory pools 22 result in competitive interaction or competition between the pools. The competition in V1-V4 is carried out with pools 24, which encode both location and object information. PP abstracts location information and mediates a competition on the spatial level. IT abstracts information from classes of objects and mediates competition at the level of classes of objects.

There are synaptic connections 20 between the areas, through which the pools 24 can be stimulated to activity. The IT area is connected to the V4 area. Area PP is connected to area V4. The synaptic connections 20 between the areas simulated in the model reflect that

"what" - and the "where" path of visual processing again. The "what" path connects the V4 area with the IT area for object recognition. The "where" path connects the area V4 with the area PP for localization. The areas IT and PP are not interconnected. The synaptic connections 20 are always bidirectional, ie the data from V4 are further processed in PP or IT. At the same time, results from PP or IT in V4 are also coupled back in order to steer the competition for attention.

The activities of the neural pools are modeled using the mean field approximation. Many areas of the brain organize groups of neurons with similar properties in columns or field assemblies, such as orientation columns, in the primary visual cortex and in the somatosensory cortex. These groups of neurons, the pools, are composed of a large and homogeneous population of neurons that receive a similar external input, are mutually coupled, and are likely to function together as a unit. These pools can form a more robust processing and coding unit because their current population mean response, in contrast to the time average of a relatively stochastic neuron in a large time window, is better adapted to the analysis of rapid changes in the real world.

The activity of the pools of the neurons is described using the mean field approximation. The pulse activity of a pool is expressed by an ensemble mean x the pulse rate of all neurons in the pool. This average activity x of the pool generally results from the excitation of the pool's neurons by an input pulse current I:

F is a real function. For pulsed neurons of the type "integrate and fire" (integrate-and-fire), which reacts deterministically to the input current I, the following applies in an adiabatic approximation (Usher, M. and Niebur, E .: "Modeling the temporal dynamics of IT neurons in Visual search: A mechanism of top-down selective attention ", Journal of Cognitive Neuroscience, pages 311-327 (1996)):

F (I (t)) = -

T _refraclo - τlog (l - -) τl (t) _{f (2)}

where T _{refractory is} the dead time of a neuron after sending out a

Pulses indicates (about 1 ms) and τ the latency of the membrane of the neuron, ie the time between an external input and the complete polarization of the membrane (Usher, M. and Niebur, E .: "Modeling the temporal dynamics of IT neurons in visual search: A mechanism of top-down selective attention ", Journal of Cognitive Neuroscience, pages 311-327 (1996)). A typical value for τ is 7 ms.

The activity of an isolated pool of neurons can be characterized not only by the average activity x but also by the strength of the input current I flowing between the neurons. This develops over time according to the following equation:

τ ^ -I (t) = -I (t) + qF (I (t)) ^dt ■ (3)

The first term on the right describes the decay of the activity and the second term on the right describes the self-excitation between the neurons within the pool. The second term describes the cooperative, exciting interaction in the pool, q parameterizes the strength of self-excitation. A typical value for q is 0.8.

The immediately recorded images are encoded in a gray value image which is described by an nxn matrix T "" ⁸ . A non-square matrix is also possible. A 64x64 matrix is usually used, ie n = 64. The indices i and j denote the spatial one Position of the pixel. The gray value T ™ ⁸ within each pixel is preferably encoded by 8 bits. Bit value 0 corresponds to black and bit value 255 corresponds to white.

In the first processing step, the constant portion of the

Image subtracted. This probably takes place in the brain in the so-called

LGN (lateral geniculate nucleus) of the thalamus. The nxn image matrix T _y is obtained by subtracting the mean:

1 nny yorig _ ~ 'X " ¹ γ ° rig

U Ü 2 ____ £ __ V

'= 1. (4)

According to the model, features are extracted from the image by the pools in area V4 in such a way that the pools carry out a Gabor wavelet transformation of the image, more precisely that the activity of the pools corresponds to the coefficients of a Gabor wavelet transformation ,

The functions G _kqpl used for the Gabor wavelet transformation are functions of the location x and y or of the discrete _ones

Indices i and j and are defined by

G _kpql (x, y) = a- ^k Ψ _θι (a ^{~ k} (x - 2p) - a- ^k (y - 2q))

(5)

in which

^ψ e, (*> y) = ψ (x cos (lθ ₀ ) + y sin (lθ ₀ ), ~ x sm (lθ _ϋ ) + y cos (W ₀ )). (6)

The basic wavelet (x, y) is defined by the product of an elliptic Gaussian function and a complex plane wave:

(7) The Gabor wavelet functions thus have four degrees of freedom: k, 1, p and q.

k corresponds to the size of the feature, expressed by the octave k, ie the spatial frequency, determined by 2 ^A k times the basic frequency, which is scaled by the parameter a; The value 2 is usually chosen for a.

1 corresponds to the angular orientation, expressed by Θ, = l - Θ _Q. θ is a multiple of the angular increment

Θ ₀ = π / L, i.e. the orientation resolution. Values from 2 to 10 are preferably chosen for L.

p and q determine the spatial position of the center point c of the function in the x and y directions, expressed by

Accordingly, the activity l _{pql of} a pool in area V4, which _responds to the spatial frequency at the octave k, the spatial orientation with the index 1 and an incentive whose center is determined by p and q, is stimulated by

VA, E kpql with:

According to the model, this corresponds to the coefficients of the Gabor wavelet transformation. The respective behavior of the pools is determined by previous training (see below).

We now consider the neurodynamic equations that determine the temporal development of the system. The activity l _{pql of} a pool in area V4 with properties that are described by the parameters k, p, q and 1 described above develops in a continuation of equation (3) due to the inhibitory and excitatory input currents over time

The first two terms on the right were explained above. They represent the natural decay of activity or self-excitement within the pool.

The third term on the right-hand side of equation (10), bE (/ ⁴ ' ⁷ ), describes the above-mentioned inhibitory effect of the inhibitory pool 22, which is described in more detail below. The parameter b on the right side of equation (10) scales the strength of the inhibition. A typical value for b is 0.8.

The fourth term on the right side of the equation (10), ll _py f _> describes the excitation by the recorded image according to the Gabor-Wavelet transformation according to equation (9).

The fifth term on the right-hand side of equation (10), I _p ^V _q ^PP r describes the attention control for a feature with the spatial position corresponding to p and q, ie the emphasis on the "where" question, as explained in more detail below becomes.

The sixth term on the right side of equation (10), l _p ^~ , ' ^τ , describes the attention control in V4 for certain patterns from IT, ie the emphasis on the "what" question, as will be explained in more detail below.

The seventh term on the right side of equation (10), I-, describes a diffuse spontaneous background input. A ty- The typical value for I ₀ is 0.025. v stands for a stochastic noise of the activity. It is assumed to be equally strong for all pools. A typical mean for v is zero, with a Gaussian distribution with a standard deviation of 0.01.

The third term on the right side of equation (10), bFUζ ⁴ ' ¹ ), as mentioned above, describes the inhibitory

Effect of inhibitory pool 22 on area V4. In the following reference is made to FIG. 4. The pools 24 within an area are in competition with one another, which is conveyed by an inhibitory pool 22, which receives the exciting input 27 from all excitable pools 24 and directs a uniform inhibitory feedback 28 to all excitable pools 24. This inhibiting feed-back 28 has a stronger effect on less active pools than on more active ones. This allows more active pools to prevail over less active pools.

In addition, an external input current 30 (bias) is shown in Fig. 4, which can excite one or more pools. The exact function of the bias 30 is described below in connection with equation (15).

Activities E ⁴ 'within the inhibitory pool satisfy the equation:

The first term on the right side of equation (11) again describes the decay of the inhibitory pool 22. The second term describes the input current from V4 into the inhibitory pool 22 belonging to V4 with the index k, scaled by the parameter c. A typical value for c is 0.1. The third term represents a self-locking of the inhibitory pool 22 belonging to V4 with the index k. A typical value for d is 0.1.

Experience has shown that the inhibitory effect within V4 acts solely within a spatial structure of a predetermined size, expressed by the octave k. Within the structure of size k there is competition between the locations p and q and the orientation 1, mediated by the sum _y_ F \ lζ _pql {t)). Each index triplet (p, q, 1) inhibits all other pql

Index triple (p, q, 1). Spatial structures of different sizes k, ie different spatial frequencies k, do not influence each other, since the inhibitory effect in equation (10), -bF (I _k ^v * ^J ) only affects k itself.

The fifth term on the right side of equation (10), jv ⁱ - ^pp _^ , as mentioned, describes the attention control for a feature with the spatial position corresponding to p and q, ie the emphasis on the "where" question. Attention control takes place by feedback of the activity I _p ^{v ~ FP} der

Pools with the indices i and j close to the values p and q from the area PP to all pools with the indices p and q in the area V4. This feedback is modeled by

C = ΣΣ Ab) ⁽ 12 ⁾

where the coefficients W are in turn determined from:

dιst ² ((p, q), (ι, j))

WP _Da ΨJ = Ae ^ -B (13;

with the coupling constant A (typical value 1.5), with the spatial scaling factor S, which defines the range of the spatial influence of a feature (typically S = 2), and with the distance function dist (p, q, i, j), which determines the distance between the location i, j and the location defined by p, q center of the Gabor wavelet function. The Euclidean metric is preferably used here:

(i)) = (p -ϊf + (q - j) ^2. (14)

In addition, there is a negative connection B to the surroundings, which leads to an overemphasis on neighboring features and a devaluation of more distant features. A typical value for B is 0.01.

In effect, the pools with the spatial position corresponding to p and q do not directly stimulate the corresponding pools in V4, but only after a convolution with a Gaussian kernel. In other words, V4 and PP are connected with symmetrical, localized connections that are modeled by Gaussian weights.

The temporal development of activity I ^{pp of} the pools in area PP is given by

τ-ι; ^p I ^ + qF (I ^) -bF (IP ^ P.1 '') ₊ I r ^ PP-VH ₊ L τPP, A + / ₀ + V (15) dt "

In the first, second, sixth and seventh term, the equation corresponds to equation (10), but for the area PP.

The third term on the right again describes the inhibitory effect of the common inhibitory pool I on the area PP. Its activity I ^PP satisfies the equation

τ ~ I ^PPJ = -I ^PPJ + c∑F (i; ^r ) ~ dF (l ^pp ''). (16)

^ ι, J

The structure of this equation corresponds to equation (11) already described. There is only one uniform inhibitory effect for the PP area. The fourth term on the right side of equation (15) again describes the attention-controlling feedback from V4 to PP and is given by

j ^pp -vi γ _{w F} (τv * \ ^γ y Z_ι ^yy pψJ ^r Vk _P qi) '(1; k, p, q, l

where the WP _m Ψ _i J _! were defined above in connection with equation (13). The synaptic connections 20 between V4 and PP are thus symmetrical. In PP, V4 controls attention with regard to certain locations ("where" question).

The fifth term 1 u on the right-hand side of equation (15) is an external top-down bias that draws attention to a specific location (i, j). This is represented by arrow 30 in FIG. 4. If the bias is preset, an object is expected at the preset location. A typical value for this external bias is 0.07 for the expected location and 0 for all other locations.

The sixth term on the right-hand side of equation (10), I _kp ^~ , ^IT , describes - as mentioned - the attention control in V4 for certain patterns from IT, ie the emphasis on the "what" question. Attention control takes place by feedback of an activity i 'of the pools, which stand for pattern c, from the area IT to associated pools in area V4. This feedback is modeled by

The determination of the weights w _ck , the input streams from IT in

V4 and thus the pools belonging to pattern c in area V4 are explained below. I _C ^JT is the activity of a pool, which stands for pattern c, in the IT area. The temporal development of I ¹ follows the differential equation:

_{r +} qF (A _c ^τ ) -bF (I ^JT ' ¹ ) + ll ^τ - ^v * + 1 + / ₀ + v. (19)

In the first, second, sixth and seventh term, the equation corresponds to equations (10) and (15), but for the IT area.

The third term on the right side of equation (19), - bF (I ^{, r} ''), in turn describes the inhibitory effect of the inhibitory pool 22 on pattern c of the IT area. The activity I ^{IT of} the inhibitory pool for the IT area satisfies the equation

_τ ι * ^J = -l ' ^r _{+ c} ∑E (/) - </ E (/ ^{/ r} ' ⁷ ). (20) oi _c

The structure of this equation corresponds to the already described equations (11) and (16). There is only one uniform inhibitory effect for the IT area, which causes the competition for attention between the individual patterns c.

The fourth term on the right side of equation (19),

I _c ^~ , describes. in turn, the attention-controlling feedback from V4 to IT and is given by

where w _ckpq already occurred in equation (18) and are explained in more detail below. The synaptic connections 20 between V4 and IT are thus symmetrical. V4 controls the attention in IT with regard to certain patterns ("what" question). The fifth term on the right side of equation (19), A ' ^A , is again an external top-down bias that draws attention to a particular pattern c. If the bias is preset, a certain pattern c or object c is expected. A typical value for this external bias is 0.07 for the expected pattern and 0 for all other patterns.

The weights w _ck , the synaptic connections between V4 and IT are through Hebbian learning (Hebbian Training) (Deco, G. and Obradovic, D.: "An Information-theoretical Approach to Neurocomputing", Springer Verlag (1996)) with known objects educated. To put it simply, pattern c is presented to the neural network one after the other and the weights w _{ckpql are} varied until the grandmother pools c recognize pattern c in IT, ie show maximum activity. In a first approximation, the weights w _ckpql result from the above-described Gabor wavelet transformation of the pattern c stored in IT.

After the training, simulations can be carried out with the neural network. B. evaluate an fMRI image. The evaluation of an fMRI image is basically an inverse problem: the cause (the activity of certain areas) should be used to determine the cause. Due to the complexity of the networking, the effect cannot be deductively deduced from the cause. It is only possible to reproduce the effects by varying a variety of causes.

This is done by varying the parameters of the described neural network, e.g. B. by switching off individual pools or entire parts of an area. The effect of such assumptions on the neural network is calculated by solving the differential equations given above and compared with the measured fMRI images. The system of the given differential equations is highly parallel. It consists of approximately 1.2 million coupled differential equations. These are solved numerically, preferably by discretization with the aid of the Eu-1 or Runge-Kutta method. The time increment chosen is preferably 1 ms, ie approximately T _refracloιy according to equation (2).

With the help of the described neural network, experimental data (Kaster, S .; De Weerd, P .; Desimone, R. and

Ungerleider, L.: "Mechanisms of directed attention in the human extrastriate cortex as revealed by functional MRI"; Science 282 (1998) 108-111. Kaster, S .; Pinsk, M .; De Weerd, P .; Desimone, R. and Ungerleider, L.: "Increased activity in human visual cortex during directed attention in the absence of visual stimulation"; Neuron 22 (1999) 751-761.). The dynamics of the activity of the pools in V4 with significant changes on the scale below one second could be tracked. Likewise, attention control through expectation and the inhibitory effect of simultaneous or neighboring stimuli.

The medically known phenomenon of "visual neglect" could also be clarified by simulation. "Visual neglect" is the fading out of half of the visual field from the

Attention or perception even though the patient's eyes are undamaged. The suppression of attention is a disturbance in the processing of the visual data recorded by the visual cortex of the brain. In the simulation using the described neurocognitive neural network, it was assumed that half of the area PP responsible for localization was damaged. For example, one half of PP may have an increased level of noise v. By solving the differential equations, it could be shown that under these circumstances attention could no longer be drawn to this half of the visual field. The one for this Half of the pools responsible for the field of vision no longer achieved an increased level of activity.

As a diagnosis for "visual neglect", damage to the Areal PP responsible for local localization is therefore possible.

Numerous modifications and further developments of the exemplary embodiments described can be implemented within the scope of the invention. In particular, the invention can be used with a corresponding

Application to suitable areas of the brain to diagnose all neuropsychiatric phenomena, ie all disorders of the functioning of the brain.

List of literature cited:

1. Fox, P.T. and Raichle, M.E .: "Focal physiological uncoupling of cerebral blood flow and oxidative metabolism through-

5 ing somatosensory stimulation in human subjects "Proceedings of the National Academy of Science of the USA, 1986, volume 83, pages 1140-1144.

2. Ogawa S, Lee TM, Nayak AS, Glynn P: "Oxygenation-sensitive 10 contrast in magnetic resonance image of rodent brain at high magnetic flields." Magnetic Resonance Medicine, 1990, volume 14, pages 68-78.

3. Usher, M. and Niebur, E .: "Modeling the temporal dynamics 15 of IT neurons in isual search: A mechanism of top-down selective attention", Journal of Cognitive Neuroscience, pages 311-327 (1996).

4. Deco, G. and Obradovic, D.: "An Information-theoretical Ap-20 proach to Neurocomputing", Springer Verlag (1996).

5. Kaster, S .; De Weerd, P .; Desimone, R. and Ungerleider, L .: "Mechanisms of directed attention in the human extra cortex as revealed by functional MRI"; Science

25 282 (1998) 108-111.

6. Kaster, S .; Pinsk, M.; De Weerd, P-. ; Desimone, R. and Ungerleider, L .: "Increased activity in human visual cortex during directed attention in the absence of visual

^ 0 stimulation "; Neuron 22 (1999) 751-761.

Claims

claims

1. A method for evaluating an image of the brain (fMRI image) obtained by means of functional magnetic resonance tomography, characterized in that a) that a neural network for simulating the activity of the

Brain is used; b) that a disturbance suspected in the brain is simulated in the neural network (disturbed neural network); c) that the activity determined in the disturbed neural network is compared with the activity of the brain observed in the fMRI image; and e) that the comparison suggests disorders in the brain.

2. The method according to the preceding claim, characterized in that the neural network used to simulate the activity of the brain simulates the activity of neurons in such a way that a plurality of neurons of the neural network are combined to form a pool; and that the activity of the pools is simulated.

3. The method according to any one of the preceding claims, characterized in that the neural network used to simulate the activity of the brain simulates the activity of the brain in such a way that the activity of the brain takes place in functionally differentiated areas (IT, PP, V4, VI).

4. The method according to the two preceding claims, characterized in that the areas are selected such that they each have a plurality of pools.

5. The method according to claim 2, characterized in that the pools compete for attention with each other; and that this competition is mediated by at least one inhibitory pool which has an inhibitory effect on the activity of the pools.

6. The method according to any one of the preceding claims, characterized in that the activity of the brain is analyzed during the vision process; and that the neural network uses a wavelet transform to analyze the captured image.

7. The method according to claim 4, characterized in that the activity of the brain is analyzed during the vision process; and that an area (IT) of the neural network has the function

Identify objects in the field of view, as pools in this area specialize in the identification of specific objects.

8. The method according to claim 7, characterized in that the activity of the brain is analyzed during the visual process; and that an area (PP) of the neural network has the function of identifying the location of recognizable objects in the visual field, in that pools of this area specialize in locating objects at specific locations in the visual field.

9. The method according to claim 7, characterized in that the neural network is designed such that attention to a specific object to be identified or can be increased for a specific object to be located.

10. Arrangement for evaluating an image of the brain obtained by means of functional magnetic resonance tomography (fMRI image) a) with a neural network for simulating the activity of the brain; b) with means for simulating a suspected disturbance in the brain in the neural network (disturbed neural network); c) with means for comparing the determined activity in the disturbed neural network with the activity of the brain observed in the fMRI image; and d) with means for identifying disorders in the brain from the comparison.

11. Device for evaluating an image of the brain (fMRI image) obtained by means of functional magnetic resonance tomography a) with means for generating an fMRI image of a patient's brain; and b) with an arrangement for evaluating the fMRI image according to the preceding claim.