WO2008065590A1

WO2008065590A1 - Improved segmentation

Info

Publication number: WO2008065590A1
Application number: PCT/IB2007/054751
Authority: WO
Inventors: Cristian Lorenz; Hans Barschdorf; Jens Von Berg; Thomas Blaffert; Sebastian P. M Dries; Sven Kabus
Original assignee: Koninklijke Philips Electronics N.V; Philips Intellectual Property & Standards Gmbh
Priority date: 2006-11-28
Filing date: 2007-11-22
Publication date: 2008-06-05

Abstract

A computer program product is described to adapt a shape model to a patient data set, the data set representing at least a region of anatomy of the patient, and in which the shape model is pre-adapted to a data set representing substantially the same anatomy of the same patient. This provides a surprisingly better segmentation of the data in the data set. In addition, A computer program product is described for the comparison of quantitative patient data and arranged to access quantitative patient data contained in a patient data record, characterized in that the quantitative patient data is in the form of an adapted shape model, adapted from a pre-adapted shape model. This extends the improvement derived by use of pre-adapted shape models and extends them to the use of pre-adapted shape models for data comparison.

Description

Improved segmentation

FIELD OF THE INVENTION

The invention relates to a computer program product to adapt a shape model to a patient data set, the data set representing at least a region of anatomy of the patient.

A shape model can be described as a data structure which represents a true anatomical shape in the body of a patient. It is a mathematical model which can be made to encode the physical, geometrical shape of an anatomical object, or a collection of anatomical objects.

BACKGROUND OF THE INVENTION A specific embodiment of a shape model is described in 'A Comprehensive

Shape Model of the Heart' by Cristian Lorenz and Jens von Berg, Medical Image Analysis 10 (2006); 657-670, available online at www.sciencedirect.com since 18 May 2006. This document discloses the setting up of a shape model from mean, i.e. average, patient data to produce a model based on a list of nodes and vertices in a model coordinate system which describe the shape of an average human heart. The nodes and vertices define a collection of connected triangles which form a virtual mesh taking the shape, in virtual space, of the mean heart. The model can be linked to a patient image data set representing or including the real heart of the patient and the model can be caused to adapt to the real heart in the data set using a method of alignment based on the maximization of two terms. The mesh as disclosed above is constructed from triangles, but it is found that other geometrical shapes can be used to form a geometrical mesh.

Alternatively, shape models may be constructed from any other mathematical means which allow description of the geometrical shape of an anatomical object or objects and examples include for example models based on geometrical functions and models based on series expansions. Regardless of the mathematical manner in which the model is set up, it can be used to describe anatomical objects such as the heart, the brain, the liver, the internal structure of the heart including the heart chambers, and also collections of such objects.

It is found that the shape model, when linked to a patient data set containing an example of the anatomical object which is the subject of the model and subsequently adapted according to the known process for that model, aligns itself to the shape of the real object as represented in the data set and thereby forms an acceptable segmentation of the real anatomical object.

SUMMARY OF THE INVENTION

It is an object of the invention to produce an improved segmentation using the shape model.

This is achieved according to the invention in which the shape model is pre- adapted to a data set representing substantially the same anatomy of the same patient. The shape model, as a mathematical construct to be applied to a data set, can be usefully applied to any data set representing or containing an example of the anatomical structure represented by the shape model and can be adapted to fit that object using the known method. However, in the same way that the mean shape model as initially applied to each data set is simply a mathematical representation of an average form of the anatomical object, the resultant adapted model is also merely a mathematical representation of the object, this time the object as represented in the data set to which the mean model has already been applied. The mathematical representation of the adapted model is in the same mathematical form as the mathematical representation of the mean shape model and can therefore also be applied to the contents of a data set. In the case of the embodiment described above, the mathematical model is in the form of a list of points describing a series of nodes and vertices which together make up the triangles of the model mesh. Although it might be expected that the mean model for an anatomical object, as representative of an average human anatomy of that object, would provide the best and most consistent segmentation result when adapted to a data set comprising an example of that object, surprisingly, it is found that when the shape model is already adapted to data representing substantially the same anatomy of the patient, an improved and more accurately segmented shape model results from the adaptation process. The known model, as described in the prior art, is a mean model and as such describes the average geometrical shape of the anatomy which constitutes the subject of the model. The adaptation of the mean model to data representing the patient introduces a congruency into the model which when applied to a data set for further adaptation produces a better and more consistent segmentation.

This pre-adaptation can be achieved by instigating an adaptation procedure using the known mean model according to the known method to produce the resulting adapted model. This resulting adapted model now becomes the pre-adapted model for the next adaptation procedure for the particular anatomy under consideration and for that particular patient. The pre-adapted model can be stored electronically prior to use, although this is not necessary for the invention to work. However, the most common use envisioned is that a patient will present for an image acquisition and will be imaged according to the known procedures of the imaging modality, a mean model will be adapted to the resultant image data to produce an adapted model and therefore to produce also medical results, as is known in the art, and the model will be stored for a period of time until the next similar or equivalent imaging session for the same patient at which a new similar data set is produced. The adapted model will then be retrieved from data storage, according to known data handling methods, and applied as the new 'mean model' to the newly acquired data set for adaptation. This will result in an improved segmentation of the data in the data set.

It is possible, however, that upon adaptation of a mean model to a data set to produce a resulting adapted data set, the user will immediately wish to reapply the newly adapted model, i.e. the pre-adapted model, to the same data set to produce an improved segmentation.

As is known from the disclosure above, in the adaptation process the vertex positions of the triangular surface mesh are the parameters which are varied and it is the vertices V₁ fully describe an individual surface. Mesh deformation is performed by minimizing the energy term

E = E ext + Cl E _lnt

The external energy E_ext drives the mesh towards the surface points obtained in a surface detection step. The internal energy E_mt restricts the flexibility by penalizing differences from the shape model. The parameter α weights the influence of both terms. A fixed number n of such minimization steps is performed on the mesh. The different components of the deformation algorithm are described below.

As is known, surface detection is carried out for each triangle barycenter X₁ .

Within a sampling grid of points Ck, defined in a local co-ordinate system, that point C₁ is chosen that maximizes the objective function

C₁ = argmax ^ ^(X₁ +M₁Cj -BIcJI² I M₁ is a rotational matrix that rotates the z-axis of the local co-ordinate system to the triangle surface normal Xi₁ and

C

is the new surface point for X₁. The parameter δ controls the trade-off between feature strength and distance. The sampling grid

c_k = (0, 0, kε), k = -l, , 1)

was used, that results in (2/ + 1) equidistant sampling points along the triangle surface normal.

The feature function,

was used that projects the image gradient V/(jc) onto the face normal Ti₁ and damps its value so that surface points with image gradients stronger than g_max do not give higher response. σ = {1,-1} determines the gradient orientation and the restriction to a dedicated intensity range may make the feature function more specific and thus makes adaptation less vulnerable to adjacent false attractors.

The external energy,

E_ext = £w_!(e_wM_! c_! )²,w_! = max{θ,F_!(x_i + M₁C₁ ) -δ H c₁ ||²}

drives each triangle barycenter X₁ towards the detected surface point X₁ e_Vl is the unit vector in the direction of the image gradient at the surface point

JC, . Since only the projection onto e_Vl is penalized, this allows the triangle center to locally slide along an iso-contour. This method is superior to direct attraction by the candidate in the case of intermediate false attractions.

The internal energy,

E_INT = ∑ ∑ ((V_j -Vj -JR(V_j -Vj)²,

J keN(j)

preserves shape similarity of all mesh vertices V₁ to the model vertices V₁ . N(j) is the set of neighbors of vertex/ The neighboring vertices are those connected by a single triangle edge. The scaling factor s and the rotational matrix R are determined by a closed- form point-based registration method based on a singular value decomposition prior to calculation of the internal energy. As only interdependences between neighbor vertices exist and the energy terms are of a quadratic form, the conjugate gradient method could be used for minimization of the energy term with a sparsely filled matrix. The labels assigned to each face of the multi- surface model may be used to parameterize interfaces between different anatomical entities specifically.

In addition, a shape model, in being representative of the shape of an anatomical object, also contains a representation of the constituent structures within that anatomical object and as such contains information relating to size, structure, relation in space etc. By analogy, an adapted model contains information pertaining to the specific structure of the relevant organ within an individual patient. The adapted shape model therefore serves as a useful source of information for the patient record and also serves as a useful store of information for data analysis on the patient. The information contained in the adapted shape models can be accessed by a suitably arranged data analysis program to allow comparison of information in the shape models. As an example, change in left ventricular volume over a period in time can be calculated by comparing left ventricular volume from two adapted shape models, both adapted to the patient heart.

The invention therefore also relates to a computer program product for the comparison of quantitative patient data and arranged to access quantitative patient data contained in a patient data record, characterized in that the quantitative patient data is in the form of an adapted shape model, adapted from a pre-adapted shape model. This has the further advantage that the comparison of shape model data can be improved by use of adapted shape models which were themselves adapted from pre-adapted shape models.

The invention also relates to a patient data record characterized in that it comprises at least one stored pre-adapted shape model. This has the advantage that once calculated from a suitable data set in combination with a mean model, the pre-adapted shape model can be stored in the patient data record for access at a suitable future point in time, for example, after the next patient imaging session when the use of the pre-adapted shape model will be required for adaptation in the new data set.

BRIEF DESCRIPTION OF THE DRAWING

These and other aspects of the invention will be further elucidated and described with reference to the following drawings.

Fig. 1 shows a flow diagram showing how the invention can be used in practice.

Fig. 2 shows a further advantageous embodiment of the invention. Fig. 3 shows the use of a data comparator to compare results derived from pre- adapted shape models.

DETAILED EMBODIMENTS

Figure 1 describes an embodiment in which the invention can be used. A patient 101 presents for imaging at a first imaging session 102 as a result of which a first patient data set 103 is produced. A mean shape model 104, conveniently stored in computer storage 105 is linked to the first patient data set 103 and following the known adaptation process 106 results in the shape model adapted to the data set 107. The adapted shape model 108 is a mathematical model of the same form as the original mean model 104. This first adapted shape model 108 can be stored in storage 109 which may or may not be the same as storage 105. Upon completion of the next imaging session 110 for the same patient 101, which may occur some time later, and which produces a second patient data set 111, the first adapted shape model 108 can be linked to the second patient data set 111 and following adaptation process 112 will produce a shape model adapted to the data set 113. This further adapted shape model 114 can be stored in storage 115 which may or may not be the same as storage 105 or 109. The advantages of using a pre-adapted shape model in the known adaptation process are so great that as an alternative, shown in Figure 2, the patient 201 presents for imaging session 202 resulting in data set 203 to which the mean model 204 is applied, the model normally being available from some sort of data storage means 205. The adaptation process 206 results in a shape model adapted to the data set 207 and an adapted shape model 208 which is now immediately re-linked to the original data set 209, equal to data set 203, to undergo a new adaptation process 210 to produce a new data set plus adapted model combination 211. This time, the adaptation process 210 produces an improved output shape model 212. Figure 3 shows the use of a data comparator to compare results derived from pre-adapted shape models, in which a mean model 301 is linked to data sets 302 and 303 which both undergo the known adaption process 304 to produce respective pre-adapted shape models 305 and 306. The pre-adapted shape models 305 and 306 are then linked to data sets 307 and 308 which then undergo the adaptation process 304. Data set 307 may or may not be equal to data set 303. Data set 308, may or may not be equal to data set 302. Following adaption process 304 the resulting shape models 309 and 310, both the result of having used pre-adapted shape models 305 and 306, are supplied as input to a data analysis package 311 and used for comparison 312 of the results and data information held within the respective shape models. It can therefore be seen that the use of pre-adapted shape models as input to the shape model adaptation process produces an improved shape model adaptation process.

Claims

CLAIMS:

1. A computer program product to adapt a shape model to a patient data set (111), the data set representing at least a region of anatomy of the patient, characterized in that, the shape model is pre-adapted (108) to a data set (103) representing substantially the same anatomy of the same patient.

2. A computer program as claimed in claim 1 characterized in that the computer program is arranged to adapt the shape model by minimizing the addition of two mathematical terms, the first term being a term which drives the shape model towards a plurality of points on the surface of the structure represented in the data set and to which the shape model is to be adapted, the second term being a term which minimizes the difference in shape between the unadapted shape model and the adapted shape model.

3. A computer program as claimed in claim 1 characterized in that the pre- adapted shape model is stored prior to adaptation to the data set.

4. A computer program product (311) for the comparison (312) of quantitative patient data and arranged to access quantitative patient data contained in a patient data record, characterized in that the quantitative patient data is in the form of an adapted shape model (309, 310), adapted from a pre-adapted shape model (305, 306).

5. A patient data record characterized in that it comprises at least one stored pre- adapted shape model.