WO2000036563A1 - Method for contouring body structures in ct data sets - Google Patents

Abstract

According to the inventive method for contouring body structures in CT data sets, interpolation points are sought first, and the interpolation points sought are then used to establish a first model contour. The interpolation point which is least compatible with the model contour is eliminated and a new model contour is established using the remaining interpolation points until the deviation of the least compatible interpolation point falls below a threshold value or the number of remaining interpolation points falls below a stop value.

Description

 Process for contouring body structures in CT data sets

The invention relates to a method for contouring body structures in CT data records, in which a model contour is adapted to the actual contour of the body structure.

Contouring risk organs, such as the eyes, is particularly important when planning tumor radiation. This creates the prerequisite that the radiation can be selected optimally without straining the risk organs more than is absolutely necessary.

Until now, the risk organs had to be contoured manually using graphic tools. For the eyes alone, this takes about 6 minutes per patient. to estimate.

It is an object of the present invention to enable automatic or semi-automatic contouring of body structures in CT data records with sufficient or controllable reliability.

As a solution, the invention proposes a method for contouring body structures in CT data records, in which a model contour is adapted to the actual contour of the body structures, with Next, support points are sought, a first model contour is determined on the basis of the support points sought, and the support point that is most incompatible with the model contour is removed, and a new model contour is determined on the basis of the remaining support points until the deviation of the worst-compatible support point becomes a threshold value or the number of remaining ones Support points falls below a demolition value.

The method provided in this way proves to be extremely robust with regard to faulty support points, since not all the support points determined but only support points within a certain error hand width are used to determine the model contour.

In practice, it has been shown that, for example, in most cases with one eye with a suitable choice of the threshold value, more than 30 support points passed the elimination routine. Even with a percentage of faulty support points of 35%, these can be recognized and removed. The usual proportion of faulty support points is between 6% and 12%.

It goes without saying that instead of a CT data record, other data records containing image information can also be used, in which structures are to be represented by a model contour.

The method makes it possible to reduce a quantity of data, which is present as a set of pixels, which represents the contour of a body structure, to a model contour comprising only a few parameters. For example, a certain number of support points of an eye contour can be reduced to the center position and the semiaxes of an ellipse. As can be seen immediately, the number of parameters required for the respective model contour (e.g. center position and semi-axis) determines the abort value, from which the method is aborted as unsuccessful. For safety reasons, this cancellation value can also be chosen higher.

The accuracy of the method can be increased in that the CT data record is pre-processed by a filter before the search for a support point.

In particular, it is possible to use a function dependent on the Hesse matrix as a filter on the CT data record. Such a filter can be selected in such a way that it sharpens the CT data record along the intrinsic direction of the larger intrinsic value of the Hesse matrix and smoothes it perpendicularly along the intrinsic direction of the intrinsically smaller intrinsic value. The filter thus highlights line-like structures. For example, the filter can include the two independent SO (2) invariants of the Hesse matrix Tr // and Tr H ² . These SO (2) invariants are invariant to spatial rotations, so that they can be used regardless of a specific orientation. This ensures that the structures highlighted by the filter are independent of the position of the body in the CT data set. With these functions, it is not necessary to calculate the eigenvectors and eigenvalues of the Hesse matrix, so that the filter can be applied to the CT data record relatively quickly. The search for a support point is made easier if it starts from a location inside the body structure to be contoured. At such a starting point, the edge of the body structure can usually be detected as the first structure seen from the inside. Individual structures or shades in between can be eliminated by suitable evaluation methods, such as the contouring method according to the invention. On the other hand, such a choice of a location that can be used as a starting point is also advantageous for image evaluation, regardless of the features of the contouring method.

The location can be selected by the user, in particular in the case of semi-automatic methods, by marking a corresponding location inside the body structure to be contoured. On the other hand, it is also conceivable that such a location is determined by suitable additional method steps.

For this purpose, a function F can be adapted to the filtered Hounsfield values, for example in the vicinity of each point, for example in an environment of 7 x 7 pixels, of the CT data set, which function has the shape of a curved comb, the curvature of the comb line should be constant. Any function that is axisymmetric and falls from zero can serve as the comb line function G in this case. The ridge line function G can, for example, a bell curve, such as. B. be a Gaussian function. In this respect, the function F to be adapted can be adjusted in its radius of curvature and the center of the curvature. They can also be constant Fit parameters can be added. In this way, a comb line can be determined for each environment of each point, which represents a segment of a circle with a center. A frequency distribution of the center points can be determined by calculating the center point coordinates for each point of the CT data set. In order to avoid confusion with other structures of the CT data set, only those center points are included in the distribution at which the respective radius of curvature approximately corresponds to the radius of the body structures, for example between a minimum and a maximum eye radius. The distribution of the center coordinates obtained in this way contains two dominant peaks which correspond, for example, to the centers of the two eyes. In this respect, the distribution only needs to be smoothed, the background removed and the position of the peaks determined in order to determine a location in the body structure.

To search for support points, local maxima of the CT data set can be determined from a location selected in the interior of the body structure, along which the difference between a previous minimum and the respective maximum lies between a lower support point threshold and an upper support point threshold. Here, the term one-dimensional path denotes a path starting from the selected location through the two-dimensional image of the CT data record of any type, which has no crossings with itself. The method is advantageous when lines are used, so that grinding cuts with the contour are avoided.

The difference between minimum and maximum proves to be a relatively easy to implement and yet reliable criterion for the search for a support point. The lower support point threshold value and the upper support point threshold value can be selected relatively uncritically as process parameters, so that such a search for a support point is also advantageous independently of the other features of the contouring process. In particular in interaction with the contouring method according to the invention, this search for a support point proves to be uncritical in the selection of the process parameters, since faulty support points can be eliminated relatively reliably.

Instead of a difference between minimum and maximum, an increase between minimum and maximum can also serve as a criterion for the search for a support point. The slope should also lie between two threshold values and be used as a criterion. The advantages described above also result here. An absolute threshold value can also be specified, which must be exceeded, the next maximum then being processed as the relevant variable.

In order to avoid that the contours of the body structures are too small and that damage is unintentionally caused as a result, after the above-mentioned rough search for a support point, for example, a local maximum and a closest local minimum of the CT data set, starting from the body structure along a one-dimensional path, are selected. In such an arrangement, the interpolation point can be defined by a suitable or by an interpolation between the maximum and the nearest outer minimum that can be determined via a parameter.

It goes without saying that not only a CT data record of Hounsfield values can be used in the applications described above for the CT data record, but that other image data or CT data or image data pretreated by a filter can also be processed here.

A reliable search for a support point can be ensured, for example, by using a curve for this, which in a range between the local maximum and the closest local minimum between a 3rd degree polynomial, which contains the CT data set in a selected interval approximated and whose quadratic approximation lies in this interval.

Here, the interval can, for example, make up half of the range between the successive minima framing the local maximum.

The curve is preferably an interpolation between the 3rd degree polynomial and its quadratic approximation around the maximum of the 3rd degree polynomial. This curve combines the advantages of a support point determination by means of 2nd and 3rd degree approximations to the CT data set, but does not have the disadvantages of these approximations. The position of the support point is always defined, as when using a 2nd degree polynomial approximation, which does not necessarily have to be the case with a 3rd degree polynomial approximation. In addition, this curve takes into account a skew of the CT data set around the local maximum, similar to a 3rd degree polynomial approximation. This could not be achieved by a 2nd degree approximation.

The use of such a curve can also advantageously be used independently of the other features of the contouring method for a base point search. In particular, such a support search can also be combined with methods other than the inventive method for determining a model contour.

Before the first model contour is ascertained, substructures of the body structure can be ascertained and corresponding support points marked or selected. In this way, these support points do not adversely affect the subsequent model contour determination. For example, the cornea (cornea in front of the lens) or the optic nerve can be segmented in advance in one eye. For the determination of the substructure, the above-described methods can be used, for example, as described using the contouring of the overall body structure. It goes without saying that the method according to the invention can in principle be used in any dimensions, in particular 2 and 3 dimensions. Here, the 3-dimensional application is also advantageous regardless of the other features of the contouring method, since starting from a location inside the body structures, grinding along the contour can be avoided with each cut.

Further advantages, goals and properties of the present invention are illustrated by the description of the attached drawing, in which an inventive contouring method is shown by way of example.

The drawing shows:

1 shows a schematic representation of the process flow in a contouring method according to the invention,

2 shows a schematic illustration of a position parameter interval determination in the method according to FIG. 1,

3 is a schematic diagram for the rough determination of the base position in the search for a base,

Fig. 4 is an enlarged view of the sketch of Figure 3 for

Presentation of a fine adjustment when searching for a support point, 5 shows an exemplary representation of a CT data record via a one-dimensional path,

Fig. 6 found support points and determined model contour for a first pair of lower and upper support point threshold and

7 found support points and determined model contour for a second pair of lower and upper support point threshold values.

The contouring method shown in the figures is used to adapt an ellipse shape to an actual contour of an eye. For this purpose, as shown in FIG. 1, the CT data are first subjected to pre-processing by a filter. In a second step, support points are first searched for on the respective eye edge as part of a rough search. Following this, the lens is searched for on the basis of the support points found and the cornea is re-segmented. Before a first model contour is determined on the basis of the support points sought in order to eliminate defective support points, the support points are fine-tuned. Following this, as already indicated above, a model contour is determined, defective support points being eliminated. After a final ellipse adjustment, an ellipse obtained in this way as a model contour is output as a contour. In the filter used, a function is created from the two independent SO (2) invariants of the Hesse matrix Tr H and Tr H ²

h '(x, y) = h (x, y) - Tτ (H (x, y) ² ) - Tτ ² H (x, y) sign (TrH (x, y)); 0 <α <l

applied to the Hounsfield values of the two-dimensional CT data. This is shown by way of example for a section along a one-dimensional path through a CT data record in FIG. 5. Here, h denotes the Hounsfield values, while h 'represents the values after the filtering. The filter highlights line-like structures in the two-dimensional data set, but leaves individual pixels that stand out from the environment unchanged. In this way, line-like contours present in the CT data record are sharpened. As can be seen directly from FIG. 5, increases are thus exaggerated. Smoothing along the line-like structures cannot be represented in such a section. The filter can be iterated.

In the next process step, support points are roughly determined. Here, a location inside the eye is first selected. This is done manually, so that the method described here is a semi-automatic contouring method. It has been shown that the present contouring method is relatively insensitive to a choice of location in this regard, so that the location should only be chosen within an ellipse with half the half-axis as large as that of the eye in order to achieve sufficiently good results. In a further embodiment of this method, the location inside the eye or the locations inside the eyes could be determined automatically. For this purpose, a function F is adapted to the filtered Hounsfield values K, which has the shape of a curved comb, the curvature being adapted to the surroundings, for example the surroundings of 7 x 7 pixels, of each point (xu, yu) of the CT data record the ridge line is constant. The function F has the form:

where a, c, x _z , y _z, and r are parameters that are adjusted. Here (x _z , y _z ) are the coordinates of the center of curvature and r the radius of curvature. The ridge line function G here has the following properties: G (t) = G (-t), G '(t)> 0, for t <0. In the present exemplary embodiment, a Gaussian bell curve fixed to the maximum value 1 is selected. For each environment of each point, the ridge line represents a segment of a circle, the center of which is at (x _z , y _z ).

A frequency distribution of the center points can be determined by calculating the center point coordinates (x _z , y _z ) for all points (XT _J , y ^,) of the CT data set. In order to obtain only ridge lines belonging to the eyes, only centers of curvature (x _z (x _Us yu), y _z (XT, yu)) are included in the distribution, whose radius of curvature r lies between a selected minimum and a selected maximum. The distribution of the center coordinates determined in this way has two dominant peaks which correspond to the centers of the two eyes. The distribution is then smoothed, the background removed and the position of the peaks determined. These peaks are chosen as starting values or as the starting point for the threshold search.

Starting from the starting point, one-dimensional cuts are made radially outward through the CT data record. Such cuts are exemplified in FIG. 5 and in FIGS. 3 and 4. Next, starting from this location, which is denoted by 0, local maxima X _! determined, in which the difference Δh between the previous minimum XQ and the maximum x lies between a lower support point threshold value and an upper support point threshold value. This is shown in Figure 3.

After support points have been found in this way, the partial structures, namely the cornea of the eye, are determined. This is done by making one-dimensional cuts radially outward from the starting point in the eye through the CT data record, not only the local maximum lying at the starting point but also a subsequent local maximum which meets the same condition being selected . However, these cuts are only selected in a forward-looking angle segment of 120 °, in which a lens is to be expected. This is possible because the position of a head in the CT data record is known. On the other hand, it is conceivable to use further method steps, similar to that in the determination of the Eye centers used to determine the position of this angular segment.

Following this, starting from the clockwise side of the angle segment, the last support point S _{0 is} selected, from which there are two alternatives S, and S, - seen counterclockwise. These two alternatives are each connected to S ₀ via a line and the smallest Hounsfield values are determined along these lines. The support point at which the smallest Hounsfield value h 'is greater is selected as a component of the contour to be selected. Starting from the support point selected in this way, the closest support point pair is selected iteratively and a support point is selected from these in the same way as a component of the contour.

This process is repeated until the entire angular segment of 120 ° has been run through or until there is only one alternative.

In this way, a line-like contour is determined that continues as continuously as possible from the already determined eye edge, which is characterized by only one alternative present. If the alternative determined in this way is the external one, then this alternative is the cornea. The inside alternative is the lens.

If this alternative is the inner one, then the outer alternative belongs to the lid or to another structure that does not belong to the eye.

In this way, there is a set of interpolation point values, which include the edge of the eye and the cornea and which are fine-tuned in a next process step. For this purpose, a polynomial of the 3rd degree is formed around the local maximum x, in an interval which makes up half of the range between the previous minimum Q and the subsequent minimum x ₂

P3 (x) = ax ³ + bx ² + ex + d

or in the coordinate system of the local maximum x _M = ± x,

P3 (xx _M ) = ₃ (xx _M ) ³ + ₂ (xx _M ) ² + ₀

with the quadratic approximation

P2 (xx _M ) = a ₂ (xx _M ) ² + a ₀

adjusted (see Figure 4). The intersection points of this polynomial approximation P3 are now used to calculate the corresponding support points a constant function C = h '(x ₀ ) + η as a function of the coefficient α _{3 of} the cubic term of the polynomial approximation P3. From these intersections, the innermost on the outside of the eye is selected. Its position is then developed in the coefficient α ₃ Taylor. To determine the interpolation position I _St , the Taylor polynomial of the 1st degree in α ₃

ΔAη 1 ΔÄη k ¹ ^ ⁾ - ^X M = α

² «?

used. Here, the tuples [1 = 1 _St (η), h '= h' (x ₀ ) + r / Δh] with variable parameters η provide a curve K which runs between the 3rd degree polynomial approximation P3 and its quadratic approximation P2.

A support point can thus be determined as a function of the parameter η. The η parameter can be used to set how closely the outer contour should sit on the eye. With η = l the outer contour sits on the ring of the local maxima, while with η = 0 it lies much further outside in the vicinity of the next outer minimum (see Figure 4). In practice, this can be used to adjust a support point by 2 to 3 mm, whereby the treating physician can specify the degree of adjustment via η. Doctors usually choose a value of η - 0.5.

In a further method step, an ellipse is then adapted to the remaining or determined support points. Following this, the deviation of the worst support point from the ellipse is determined and so long the worst interpolation point is removed and then a new ellipse is adjusted until the error of the worst interpolation point falls below a certain value determined by the resolution limit of the CT data.

The process terminates at the latest when there are only 4 support points left. However, the contour muzzle is considered to have failed if less than a critical number remain from an initial number of reference points. For example, a critical number of 8 can be selected for an initial 36 support points. However, this never occurred during test runs. In the vast majority of cases, more than 88% of the support points go through the elimination routine. The routine for faulty support points is robust and effective. It can recognize and remove even with a percentage of faulty support points of 35%. The usual proportion of faulty support points is between 6% and 12%. It goes without saying that a different number of support points can also be selected. The process is largely independent of the exact number of support points.

The result of the elimination routine is shown by way of example in FIGS. 6b and 7b, while FIGS. 6a and 7a show the support points used as black crosses and the eliminated support points as white crosses. FIGS. 6a and 6b show a procedure in which the threshold values for determining the local maxima x and minima Xo and x _{2 are set} relatively well. FIGS. 7a and 7b, on the other hand, show an exemplary embodiment in which the node threshold values are not set as well. As a result of this, the support points deviate greatly from one another, that is, a relatively large number of faulty support points have been determined. However, as can be seen from FIG. 7b, the method according to the invention provides excellent contouring of the eye even in such a case.

It goes without saying that the method sequence shown in the exemplary embodiment can be subjected to minor deviations. For example, a first, relatively coarse model can be determined from the first segmentation of the edge of the eye before the cornea and lens are segmented. It is also possible to make a fine adjustment after the first rough search for a support point before segmenting the cornea and lens.

Claims

Claims: 1. A method for contouring body structures in CT data records, in which a model contour is adapted to the actual contour of the body structure, characterized in that - firstly, support points are sought, that a first model contour is determined on the basis of the searched support points, that the with the Model contour most incompatible Support point removed and a new model contour is determined on the basis of the remaining support points until the deviation of the worst-tolerated support point falls below a threshold value or the number of remaining support points falls below a termination value. 2. The method according to claim 1, characterized in that the CT data record is pre-processed by a filter before the search for a support point. 3. Contouring method according to claim 1 or 2, characterized in that a location in the interior of the body structure is first selected to search for support points. 4. The contouring method according to claim 3, characterized in that local maxima of the CT data record are determined for starting point searches from this location along one-dimensional paths in which the difference between a previous minimum and the respective maximum lies between a lower support threshold and an upper support threshold. 5. Contouring method according to one of claims 1 to 4, characterized in that a local maximum and a closest, starting from the inside of the body structure along a one-dimensional path minimun of the CT data carrier are selected to search for support points. 6. Contouring method according to one of claims 1 to 5, characterized in that a local maximum and a closest local minimum of the CT data set, starting from the body structure along a one-dimensional path, are selected for the search for a support point. 7. Contouring method according to claim 6, characterized in that a curve is used for the search for a reference point, which is in a range between the local maximum and the nearest local minimum between a 3rd degree polynomial which contains the CT data set in a selected interval approximated, and its quadratic approximation lies in this interval. Contouring method according to claim 7, characterized in that the curve is an interpolation between the 3rd degree polynomial and its quadratic approximation.