METHOD AND ARRANGEMENT FOR ENHANCING IMAGE QUALITY IN A COMPUTED TOMOGRAPHY IMAGING SYSTEM

TECHNICAL FIELD

The present invention relates to a method and an arrangement in a computed tomography imaging system and, particularly, to an arrangement allowing for enhancing image quality of a computed tomography image as well as a method for such enhancement. The invention further relates to a computer-readable medium containing computer program for enhancing image quality of a computed tomography image.

BACKGROUND OF THE INVENTION

Computed Tomography (CT) is a diagnostic procedure that uses special x-ray equipment to obtain cross-sectional pictures of a body. The CT computer displays these pictures as detailed images of organs, bones and other tissues. CT may e.g. be used to detect or confirm the presence of a tumor inside the body of a person. During a CT scan, the person lies very still on a table, which slowly passes through the center of a large x- ray machine. Alternatively, the x-ray machine rotates continuously around the body. This is called spiral CT.

The CT system converts x-ray attenuation measurements into "CT numbers" or Hounsfield Units (HU) that indicate the density of a particular area (pixel or voxel). A detailed description of how a CT system generates image data is not included here. This is, however, a well-known technique which e.g. may be found in US 5 594 767.

One of the main difficulties in the interpretation of CT images is the presence of noise. In order to achieve a good image quality, it is necessary to expose the examined patient to high radiation doses and a major hinder for the use of lower radiation doses is the accompanied increase in noise. To reduce the effect of noise, several image processing filters have been proposed. One prior art approach is found in US 6 963 670, which discloses a method for CT dose reduction that uses segmentation-based filtering on a shrunken version of the image.

Because the measured CT values indicate tissue density, different organs appear in different HU ranges. Low density regions such as lung tissue are indicated by values between -1000 and -500 HU for example. Regions containing much water appear around the origin of the Hounsfield scale and high density regions such as bone tissue show values up to 2000 HU. This direct relation between the measured density and the corresponding tissue type may be exploited to adjust image processing schemes to various tissue types.

Known methods for tissue dependent processing are e.g. shown in US 5,594,767, which describes a method for enhancing image sharpness, where the amount of sharpening, by unsharp masking, of a CT image is controlled by the Hounsfield value. In US 5,761 ,333 a method is disclosed where the grey values of the grey-white matter region are stretched with an amount depending on the Hounsfield value in order to increase the contrast of the image. These methods, however, do no aim at enhancing image quality by reducing noise in the image.

In US 6,819,735 a method is presented that uses organ-specific, based on the HU value, transfer functions to provide processing adjusted to organs. According to this method, the HU values are calculated from a previous recording of a layer of the body by a CT device and a first image is created in which the frequency distribution of the HU values is reproduced. A second image is created on the basis of the calculated HU values. The images are filtered and then mixed together.

Another prior art approach is shown in US 6,463,167, which discloses a method of enhancing a medical image. According to this method the image is split into different regions based on both the HU value and other tissue characteristics such as the local gradient. Subsequently these regions are processed differently and the results combined to a result image.

There is a need for a method and an arrangement, which provide an enhanced image quality by reducing noise in the images and which utilize the advantages of using a tissue dependent image processing.

SUMMARY OF THE INVENTION

Accordingly, it is an objective with the present invention to provide an improved method of enhancing image quality of a computed tomography image.

This objective is achieved through a method in accordance with the characterizing portion of claim 1.

Another objective with the present invention is to provide an improved arrangement of enhancing image quality of a computed tomography image.

This other objective is achieved through an arrangement in accordance with the characterized portion of claim 13.

A further objective with the present invention is to provide an improved computer- readable medium comprising computer program for enhancing image quality of a computed tomography image.

This further objective is achieved through a computer-readable medium according to the characterising portion of claim 25.

Further embodiments are listed in the dependent claims.

Thanks to the provision of a special method that utilizes the relationship between measured density and corresponding tissue type to control the setting of a particular noise reduction scheme based on local signal model approximations, the image quality of the scanned CT images is enhanced.

Still other objects and features of the present invention will become apparent from the following detailed description considered in conjunction with the accompanying drawings. It is to be understood, however, that the drawings are designed solely for purposes of illustration and not as a definition of the limits of the invention, for which reference should be made to the appended claims. It should be further understood that the drawings are not necessarily drawn to scale and that, unless otherwise indicated,

they are merely intended to conceptually illustrate the structures and procedures described herein.

BRIEF DESCRIPTION OF THE DRAWINGS

In the drawings, wherein like reference characters denote similar elements throughout the several views:

Figure 1 is an exemplary schematic CT slice through a human thorax;

Figure 2 is a flowchart of the inventive filtering method;

Figure 3 shows a few robust error norms in comparison to a least squares norm;

Figure 4 shows the influence functions for the example error norms of figure 3.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

A cross-sectional image of a human thorax acquired by computed tomography (CT) indicating the different visible types of tissue, is shown in figure 1. The indicated anatomical structures in the image 10 contain lungs 11 , a heart 12, several ribs 13, muscle 14, an aorta 15, a spinal cord 16 enclosed by a backbone 17 and, a vertebral body 18. The CT system converts x-ray attenuation measurements into "CT numbers" or Hounsfield Units (HU) that indicate the density of a particular area. Different tissue types, such as lungs 11 , liver or bone have different characteristics and the tissue density varies from -1000 HU for e.g. lungs (low density regions) to 2000 Hu for e.g. bones (high density regions). The present invention provides a method to adaptively reduce the noise in CT data. Towards this end the data is locally approximated by robust estimation of a local signal model where both the model and the estimation scheme are depended on the Hounsfield value of the central data pixel/voxel. As different tissue types such as lung, liver or bone have different characteristics the dependence on the Hounsfield value makes it possible to adjust the filtering to the tissue type.

The operation replaces the signal value at each point (i,j,k) by its approximation by a local signal model:

s(i,j,k) → m(i,j,k) , (1 )

where m is the model evaluated at the point of interest. The estimation of the signal model is done in a local neighbourhood, e.g. 9x9 pixels, using a robust estimator. Typically the model consists of a low order polynomial and the estimation uses a robust estimator based on influence functions, also called M-Estimator. The main point of the given method is that the parameters controlling the estimation are made dependent on the Hounsfield value of the target value. Most importantly are the parameters controlling the influence function dependent on the HU which means that the estimation norm is changed between different regions.

It may be advantageous to low-pass filter the data and use the thus smoothed Hounsfield values to control the processing. This ensures a smooth transition between different tissue areas and results in more pleasing edges.

In a preferred embodiment of the present invention, the procedure of enhancing image quality of a computed tomography image, shown in figure 2, is as follows:

- generating image data (step 21 ) by x-raying a body with a CT scanner. The generated image may be a two-dimensional image made up of pixels or a three-dimensional image made up of voxels; receiving CT image data as Hounsfield values (step 22), wherein each pixel or voxel of the image has a HU-value. The CT image data may be received directly from the CT scanner or received from e.g. a disk on which the generated image data has been stored;

- for each of said Hounsfield values, i.e. for each pixel/voxel, selecting a signal model based on the Hounsfield value, the signal model being described in more detail below; and also based on the Hounsfield value in each pixel/voxel, selecting a robust estimation technique for estimating the selected signal model, the robust estimation being described in more detail below;

- estimating the locally selected signal model by using the selected robust estimator (step 23), whereby the parameters controlling the model estimation are calculated based on the original HU-values;

optionally, pre-filtering the received image data (step 24) before initializing the robust estimation (step 25). Example pre-filters are a median filter and a low- pass filter.

- optionally the image data (HU-values) is low-pass filtered (step 26) and the parameters controlling the model estimation are computed based on the low- pass filtered version of the original HU-values (step 27).

- determining an approximation value for each pixel/voxel based on the received Hounsfield value using the estimated signal model, i.e. the signal model is evaluated at the origin of local coordinate system; replacing each received Hounsfield value with said determined approximation value;

- generating a result image using said determined approximation values for each pixel. The result image thus having a reduced amount of noise compared to the original CT-image; and,

- storing, displaying or transferring over a network, the generated result image (step 29).

Signal model

In its general form a signal model approximates the signal by a number of basis functions:

Here the neighbourhood is denoted as a vector obtained by stacking all pixel on top of each other. The matrix B contains the basis functions as row vectors stacked in the same way as the signal. Common basis functions are polynomials, for example does the second order polynomial basis in 2D become:

^ _1,2,3,4,5,6 = \\-,χ,y,χ - y,0.5 - x ² ,0.5 - y ² ) (3)

Other choices are for example trigonometric functions or polar harmonics. As stated above the result of the described filtering operation is given by the evaluation of the

estimated signal model at the point of interest here denoted as the origin in the local coordinate system:

m = ∑c,b,(x) with Jc = O (4)

Note that for the polynomial model this evaluation reduces to the zero order coefficient m = c ₀ . However the estimation of this coefficient changes with the order of the model used. To understand the used robust estimation better it is intuitive to recall the standard least squares approximation.

Least squares estimation finds that coefficient vector that minimises the difference between the signal and the model:

I n? Sc -? (5)

The solution can be found using standard techniques, e.g. by the pseudo-inverse, c = B ⁺ s . If B has full column rank, which is the case for sensible basis functions, the pseudo inverse is readily obtained from B ⁺ = (B ^T B) ^~l B ^T , otherwise it can be computed via the singular value decomposition of B.

In image processing it is often desirable to give points further away less influence on the least squares estimation. This can be incorporated by the use of a weighted norm:

min|i?c -5| (6)

Here W denotes a weight function, a typical example is a Gaussian. Writing the weight as a diagonal matrix the solution can then be found from:

c = (B ⁷ WB) ⁺ B ^τ Ws (7)

Hence the coefficients can be found by correlation of the signal with a filter kernel that is given by the rows of (B ⁷ WB) ⁺ B ^τ W .

Robust estimator

The main drawback with least squares estimation is that it degrades rapidly in the presence of outliers. This stems from the fact that each data point has an influence on the result that is dependent on the square of the difference between model and this point. For outliers this difference can become very large and thus a single outlier can have a very strong impact on the result. When modelling CT data locally, boundaries between different anatomical structures as well as bad data points due to noise are often encountered, both will create outliers with respect to the local model.

One group of robust estimation techniques is based on random sampling. In order to estimate the coefficients as many data points as sought coefficients are needed. The considered neighbourhood should however contain many more points. This implies that it is possible to randomly select a set of points, compute the model from that and then check how close this model is to all the other points. By repeatedly doing this one does eventually find a good model. In a final step it is then possible to classify each point as inlier or outlier with respect to the chosen model and compute a refined model from the inliers only. However these sampling techniques are often computationally expensive.

As a faster alternative to sampling techniques also robust estimators based on influence functions are used. As mentioned above does least squares estimation suffer from giving outliers too much weight. This can be avoided by using a different error norm that reduces the influence of large deviations. Figure 3 shows several such robust error norms in comparison to the least squares norm which is denoted with 31. The shown robust error norms in figure 3 are: the error norm according to Huber 32, the error norm according to Tukey 33, the Fair error norm and the Gaussian error norm 34. The estimation then becomes

min p (Bc -s) (8)

With the error norm p, choosing p(x) = x ² leads to the least squares estimation. The influence each data point has on the solution is determined by the derivative ψ of the error norm, ψ is also called influence function. The following table lists a few commonly used error norms and associated influence functions:

Figure 4 shows the influence functions for the example error norms from figure 3. The influence functions shown in figure 4 are: the influence function for the least square error norm 41 , the influence function for the Huber error norm 42, the influence function for the Tukey error norm 43, the influence function for the Fair error norm 44 and the influence function for the Gaussian error norm 45. It can be seen that the influence for the least squares norm increases for large values while that of the robust estimators do not. The robust estimators can be converted into an iteratively re-weighted least squares problem, where the weights are used to reduce the influence of the outliers. Initialization of the iterative estimation can for example be based on the result of a least squares estimation.

Start by formulating the error sum as a weighted least squares sum

∑P(*,) = ∑w, -*, ² . (9)

minimisation is performed by setting the derivative to zero:

∑y (x ₁ ) = 2∑w ₁ - x ₁ ≡ 0 (10)

Thus the weights can be obtained from the influence function:

w, = ψCO (1 1 )

For example does a Gaussian error norm result in a Gaussian weight function:

p(x) = l -exp - (12)

For a simple constant signal model the estimation reduces to an iterated weighted average:

w w{x,y,z;ι,j,k)

with d initialised either as the original data d ⁰ = D or as the result from a pre-filtering operation d ⁰ = f(D) . A typical choice of the weight function is a Gaussian:

(14)

Where the standard deviations and neighbourhood size are functions of the Hounsfield value:

O _x = r(HU _{1 J>k} )...σ _d = f«(HU _1J>k ) X = g ^* {HU _UJtk )..Z = g ^z (HU _l>J>k ) (15)

This may be implemented by using look-up tables. If only one iteration is chosen this simply becomes a weighted average with the weight function being controlled by the Hounsfield value. If furthermore the spatial size is the same in all dimensions and constant, X=Y=Z=const, and the spatial standard deviations are the same in all dimensions and constant, σ _x = σ _y = σ _z = const , and the standard deviation in the data component is constant, σ _d = const , the operation reduces to a Bilateral filter.

In order to speed up the computation or to adapt to varying size of different organs it is sometimes desirable to apply the described method on several spatial scales. Furthermore can the filtering be used on several frequency bands simultaneously to allow for an increased adaptability to structures of varying frequency content.

To facilitate understanding, many aspects of the invention are described in terms of sequences of actions to be performed by, for example, elements of a programmable computer system. It will be recognized that the various actions could be performed by specialized circuits (e.g. discrete logic gates interconnected to perform a specialized function or application-specific integrated circuits), by program instructions executed by one or more processors, or a combination of both.

Moreover, the invention can additionally be considered to be embodied entirely within any form of computer-readable storage medium having stored therein an appropriate set of instructions for use by or in connection with an instruction-execution system, apparatus or device, such as computer-based system, processor-containing system, or other system that can fetch instructions from a medium and execute the instructions. As used here, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction-execution system, apparatus or device. The computer-readable medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic,

infrared, or semiconductor system, apparatus, device or propagation medium. More specific examples (a non-exhaustive list) of the computer-readable medium include an electrical connection having one or more wires, a portable computer diskette, a random access memory (RAM), a read only memory (ROM), an erasable programmable read only memory (EPROM or Flash memory), an optical fibre, and a portable compact disc read only memory (CD-ROM).

Thus, a computer-readable medium containing computer program according to a preferred embodiment of the present invention of enhancing image quality of a computed tomography image, is provided wherein the computer program performs the steps of: receiving computed tomography image data as Hounsfield values; for each of said received Hounsfield values, selecting a signal model based on said Hounsfield value; based on said Hounsfield value, selecting an estimation technique for estimating said selected signal model; determining an approximation value based on said received Hounsfield value using said estimated signal model; replacing each received Hounsfield value with said determined approximation value; and, generating a result image with said determined approximation values.

Thus, while there have been shown and described and pointed out fundamental novel features of the invention as applied to a preferred embodiment thereof, it will be understood that various omissions and substitutions and changes in the form and details of the devices illustrated, and in their operation, may be made by those skilled in the art without departing from the spirit of the invention. For example, it is expressly intended that all combinations of those elements and/or method steps which perform substantially the same function in substantially the same way to achieve the same results are within the scope of the invention. Moreover, it should be recognized that structures and/or elements and/or method steps shown and/or described in connection with any disclosed form or embodiment of the invention may be incorporated in any other disclosed or described or suggested form or embodiment as a general matter of design choice. It is

the intention, therefore, to be limited only as indicated by the scope of the claims appended hereto.

Expressions such as "including", "comprising", "containing", "incorporating", "consisting of, "have", "is" used to describe and claim the present invention are intended to be construed in a non-exclusive manner, namely allowing for items, components or elements not explicitly described also to be present. Reference to the singular is also to be construed to relate to the plural and vice versa.

Numerals included within parentheses in the accompanying claims are intended to assist understanding of the claims and should not be construed in any way to limit subject matter claimed by these claims.