CN115359467A - Identification method and device for deformable medical target - Google Patents

Abstract

The invention provides an identification method and device for a deformable medical target, which are used for detecting the deformable medical target in an image under the conditions of partial occlusion, clutter and nonlinear illumination change. An integrated method for deformable medical target detection is disclosed, which combines: a matching metric based on the normalized gradient direction of the model points, a decomposition of the model into parts, and a search method that considers all search results for all parts simultaneously. Although the target model is decomposed into sub-parts, the relevant size of the model used to search at the highest pyramid level is not reduced. Thus, the present invention does not have the speed limitations of the reduced number of pyramid stages that exist in prior art methods. Finally, the identification accuracy and speed of the medical target under the deformation condition are improved, and the accurate needle insertion point positioning of the operation is improved.

Description

Identification method and device for deformable medical target

Technical Field

The invention relates to an identification method for a deformable medical target, and belongs to the technical field of machine vision systems.

Background

Minimally invasive puncture surgery is a significant revolution of traditional open surgery. Has the advantages of high operation efficiency, less postoperative complications, small wound, quick postoperative recovery and the like. At present, the traditional surgical puncture operation punctures according to doctor experience under medical images, and the problem of focus target point positioning deviation exists.

In clinical practice, the surgical puncture positioning depends on electromagnetic positioning, mechanical positioning, ultrasonic positioning, optical positioning and other methods.

The mechanical positioning device can play a role in auxiliary positioning to a certain extent, but due to the existence of the device, the operation space is reduced, the degree of freedom and the flexibility of the device are poor, the implementation of an operation is influenced, and meanwhile, the rigid puncture path positioning neglects the influence of tissues and organs on the deformation of a surgical instrument, so that the operation positioning deviation is easily caused.

The ultrasonic positioning technology makes a great breakthrough in precision and speed, and is low in cost, but the error judgment of the position often comes from the environment, including temperature, humidity, air flow, noise and the like, the environmental factors are not easy to control, and corresponding compensation needs to be performed by additionally using an algorithm.

The magnetic positioning navigation has wide application in the field of minimally invasive medical surgery, the method has high measurement accuracy and good repeatability, but the magnetic positioning navigation is easily influenced by physical characteristics of a patient during surgery, including respiration, heartbeat, action and the like. Meanwhile, the magnetic positioning navigation has high cost, small action range and limited operation range of doctors, and the positioning precision is seriously influenced by the existence of some necessary magnetic devices and magnetic interference materials.

Disclosure of Invention

The invention aims to overcome the defects in the prior art, provides the identification method and the identification device for the deformable medical target, is robust to shielding, clutter and nonlinear contrast change, and improves the identification accuracy and speed of the medical target under the deformation condition.

In order to achieve the purpose, the invention is realized by adopting the following technical scheme:

in a first aspect, the present invention provides a method for identifying a deformable medical target, comprising the steps of:

(a) Acquiring an image of a medical target model;

(b) Transforming an image of a target into a multi-level pyramid representation consistent with the recursive subdivision of a search space, the multi-level pyramid comprising at least an image of the target model;

(c) Generating, for each discretization stage of the search space, at least one pre-computed target model of the medical target, the pre-computed target model comprising a plurality of model points having corresponding direction vectors, the model points and direction vectors of the target being produced by an image processing operation returning the direction vector of each model point;

(d) Subdividing the model points into a plurality of sections, wherein deformed instances of the target model are represented by transforming the sections; each section is composed of a plurality of points, the plurality of sections forming an overlapping set of points;

(e) Acquiring a search image;

(f) Transforming a search image into a multi-level representation consistent with a recursive subdivision of a search space, the multi-level representation including at least the search image;

(g) Performing an image processing operation on each transformed image of the multi-level representation, the image processing operation returning direction vectors for a subset of target model points within the search image, the subset corresponding to a transformation range for which the at least one pre-computed target model should be searched;

(h) Clustering target model points to form a plurality of independent parts, matching and calculating each independent part in a certain range to obtain a local measurement result, and calculating the result sum normalized by the number of the model points in each part to obtain global matching measurement;

(i) Determining a target model pose for which the global matching metric exceeds a selectable threshold and for which the matching metric is locally maximum, and generating from the model pose a list of instances of the at least one pre-computed target model at a coarsest discretization level of the search space;

(j) Obtaining an estimate of the deformation by obtaining a best match for each local portion, calculating a deformation transformation describing the local displacement of the portion;

(k) Tracking, by recursive subdivision of the search space, instances of the at least one pre-computed target model at a coarsest level of discretization of the search space until a finest level of discretization is reached;

(l) Calculating a corresponding deformation transformation at each stage and transferring the deformation transformation to a next stage;

(m) providing the deformation transformation and model pose of the target model instance at the finest discretization level.

Further, for the deformable target model, the spline function is defined by a displacement, searching each set of points independently for local displacement, going along the image all the way to the lowest pyramid level, where the displacement is determined at a resolution even higher than that of the original image, instantiated at a sub-pixel precision position, where part of the displacement is defined by the gradient magnitude, determining the position of the target at a resolution higher than the finest discretization level.

Further, in addition to the user-selectable threshold, only selectable instances of the landmark hypothesis that satisfy the deformation requirement are generated into the list of possible matches on the coarsest discretization level.

Further, in step h, the local metric for each segment must exceed a user-selectable local threshold, otherwise the segment is considered occluded and discarded from further processing.

Further, in step d, subdivision is performed by using k-means clustering or by using normalized segmentation.

Further, in step j, the computed transformation is a perspective transformation.

Further, in step j, the computed transformation is a three-dimensional pose, based on receiving as input metrology information about internal geometric parameters of the imaging device and the model.

Further, in step h, the sum of normalized dot products of the transformed target model portion and the direction vector of the search image is used for the local metric result.

Further, in step h, in order to achieve invariance to contrast inversion, polarity information is discarded from the local measurement results, and the absolute value of the sum of normalized dot products or the sum of absolute values of normalized dot products of the transformed target model portion and the direction vector of the search image is used for the local measurement results.

In a second aspect, the present invention provides an identification apparatus for a deformable medical target, comprising program code means stored on a computer readable medium for performing the method for identifying a deformable medical target of the first aspect when said computer program product is run on a computer.

Compared with the prior art, the invention has the following beneficial effects:

1. the invention provides an identification method for accurately identifying a deformed medical target under optical imaging. The method combines invariance matching metrics to decompose a target model into multiple parts. Although the target model is decomposed into sub-parts, the relative size of the model used to search at the highest pyramid level is not reduced. Thus, the present invention does not have the speed limitations of the reduced number of pyramid levels that exist in prior art methods. Robust to partial occlusion, clutter, and non-linear illumination variations;

2. the method can identify the target even when the target is transformed due to perspective deformation or generalized deformation, and improves the accuracy and robustness of target identification.

3. The method is low in cost and high in precision, and the method for identifying the deformable medical target is used for detecting the deformable medical target in the image under the conditions of partial occlusion, clutter and nonlinear illumination change, and can improve the target detection efficiency and accuracy.

Drawings

FIG. 1 is a flow chart of a preferred embodiment of the present invention showing the steps of the method;

FIG. 2 shows an image of a medical target and a region of interest around the target for model generation;

FIG. 3A illustrates model generation of a target, wherein the target is on a planar surface;

FIG. 3B illustrates model generation of a target, wherein the target is occluded by a calibration plate;

FIG. 4A shows a set of target model points generated by an edge filter;

FIG. 4B illustrates an example subdivision of a target model point into portions, depicting a model center and a translation of the portions relative to the model center;

FIG. 4C illustrates a typical deformation of the target model due to local motion of the part in the vicinity;

FIG. 5 shows a current image containing two instances of deformation of a target and two instances of target detection results found for the model;

FIG. 6 shows a deformation map between a rigid template and a deformed template generated by fitting a deformation function, where an example point of the deformation map is the center of the portion.

Detailed Description

The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.

The present invention provides a method for deformable medical target identification that is robust to occlusions, clutter, and non-linear contrast variations.

Fig. 1 gives an overview of the steps of the method. A method of identification for a deformable medical target is provided, comprising the steps of:

step 1: acquiring an image of a medical target model by a camera;

step 2: transforming the obtained target image into a multi-level pyramid representation consistent with the recursive subdivision of a search space, wherein the multi-level pyramid representation at least comprises the target model;

and 3, step 3: generating at least one pre-computed target model of the medical target according to each discretization level of the search space, the pre-computed target model comprising a plurality of model points having corresponding orientation vectors, the model points and orientation vectors of the target being produced by an image processing operation returning the orientation vector of each model point;

and 4, step 4: generating a subdivision that subdivides the plurality of model points into a plurality of portions, wherein a deformed instance of the target model is represented by transforming the portions; each section is composed of a plurality of points, wherein the subdivision generates an overlapping set of points.

And 5: acquiring a search image through a camera;

step 6: transforming a search image into a multi-level pyramid representation consistent with a recursive subdivision of a search space, the multi-level pyramid representation including at least the search image;

and 7: performing an image processing operation on each transformed image of the multi-level representation, the image processing operation returning direction vectors for a subset of target model points within the search image, the subset corresponding to a transformation range for which the at least one pre-computed target model should be searched;

and 8: clustering target model points, performing matching calculation on each independent local within a certain range to obtain a local measurement result, and obtaining a global matching measurement by calculating the sum of results normalized by the number of model points in each part, wherein for the local measurement, parts of the model are searched within a limited transformation range close to a pre-calculated target model, and the most suitable part of each part is regarded as the contribution of the part to the global matching measurement;

and step 9: determining a target model pose for which the global matching metric exceeds a selectable threshold and for which the matching metric is locally maximum, and generating from these model poses a list of instances of the at least one pre-computed target model at the coarsest discretization level of the search space;

step 10: obtaining an estimate of the deformation by obtaining a best match for each local portion, calculating a deformation transformation describing the local displacement of the portion;

step 11: tracking, by recursive subdivision of the search space, instances of the at least one pre-computed target model at a coarsest level of discretization of the search space until a finest level of discretization is reached;

step 12: calculating a corresponding deformation transformation at each stage and transferring the deformation transformation to a next stage;

step 13: the deformation transformation and model pose of the target model instance are provided at the finest discretization level.

The target model to be identified comprises a plurality of points with corresponding direction vectors, which can be obtained by standard image processing algorithms, for example by edge detection methods or line detection methods. In generating the target model, the set of points is divided into a plurality of portions. These parts are movable relative to their original positions during the search, allowing the model to flexibly change its shape. In a preferred embodiment, each portion of the target model comprises only one model point. In another preferred embodiment, each section comprises several adjacent points, which points are kept rigid with respect to each other.

During the search, the original target model is instantiated, for example, for a generalized affine pose range. At each location, the instance of the model is deformed by transforming each part independently with a close-range transform. For each portion, a matching metric is computed at each transform within this restricted range. In a preferred embodiment, the matching metric is a normalized dot product of the direction vector of the portion and the direction vector of the preprocessed search image. The matching metric for the entire target model is the normalized sum of the best-fit portions at the deformation transformation of the best-fit portions. In a preferred embodiment, those parts with a score on the matching metric below the threshold are assumed to be in an occlusion state, so that they are discarded in further processing. The transformation of the portion with the largest matching metric determines the deformation of this portion with respect to the original position. This displacement is used to calculate a pre-selected deformation model. In a preferred embodiment, the model of the non-linear deformation is a perspective transformation. In another embodiment it is, for example, a spline function or another method known in the art for interpolating or approximating a set of points. Once this transformation function is calculated, the distortion found for the image area can be reversed to produce a modified image.

The method is divided into an offline phase of generating a target model and an online phase of finding the target model in a search image. The input to the model generation is an example image showing the target in an undeformed manner. In fig. 2, an example image 202 of a target shape is shown. The region of interest 201 limits the position of the target in the image. Typically, this area is defined by the user during an offline training phase. If the user of the target recognition system is only interested in subsequently revising the target in the search image, sometimes only a portion of the target needs to be selected as the region of interest 203. Which defines the location and size of the region for a model that must then be unwarped.

The region of interest 201 in the image specifies only the position and size of the target in the image. To determine the metrology pose of the target, the internal geometry of the imaging device must be provided. The internal geometry of the imaging device 300 (see FIG. 3A) is typically described by: its focal length, the location of the principal point in the image, the size of the pixel element in the row and column directions, and a distortion factor that simulates pincushion or barrel distortion caused by the lens. The internal and external parameters of the camera can be determined in advance by methods known in the art (see, e.g., MVTec Software GmbH, HALCON 8.0documentation, solution Guide II-F,3d Machine Vision, 2007).

Once these parameters are determined, the relative pose of the region of interest 203 of the target model 301 in the camera coordinate system is required for relative pose estimation of the target (see fig. 3 and 4). Since no a priori metric information of the imaging target is generally available, it cannot be said that the target is, for example, small and close to the camera or large and far from the camera. Both cases here will result in the same image. A typical way to provide such metrology information is to overlay a planar calibration plate 303, which has been measured, on the medical target and acquire an image 302 (see fig. 3B) showing the calibration plate, in a preferred embodiment. In this schematic, the calibration plate 303 contains defined points of a dark circle. Since the dimensions of the calibration plate and the exact metrology positions of the points are known, the relative pose of the calibration plane can be determined in the camera coordinate system. The calibration plate is then removed from the target and a second image of the target is acquired showing the same location as the calibration plate. Because the pose of the calibration plate and the pose used for target model generation are the same in the world and in the image, the corresponding pose of the target is automatically determined. The area of the calibration plate 303 is used directly in conjunction with the image of the target for model generation.

The medical target identification method transforms a model-generated image into a recursive subdivision containing a smoothed and subsampled version of the original image. In the following expressions, recursive subdivision, multi-level representation and image pyramids have the same meaning when used. In a preferred embodiment, the recursive subdivision is a pyramid of mean images. In another preferred embodiment, a gaussian image pyramid is applied. The same multi-level representation is generated from the region of interest defining the location of the target model. For each multi-level representation, the target model generation extracts edge points from the region of the image. The edge detection results are shown in fig. 4A. Wherein the edge detection not only extracts the position, but also the direction of the strong contrast change. The edge detection used is for example Sobel filtering or Canny edge detection filtering, or any other filtering known in the art that extracts directional feature points from an image. The present invention is not limited to edge features, which one skilled in the art can easily extend to line features or point of interest features. For clarity, we are limited to edge points in further discussion. The small arrow 400 of fig. 4A indicates the location and direction of the edge point. For each model point, the extracted edge points are transformed into a model coordinate frame (represented by circle 401) and saved into memory. Thus, the identification method yields a geometric description of the imaged target.

The model coordinate frame 401 defining the origin of the target model is typically computed by taking the center of gravity of the set of points. The coordinate frame is oriented the same as the target. Thus, the transformation that maps the model coordinate frame to the template image coordinate frame is a simple translation. By applying a generalized affine transformation mapping from the model coordinate frame to the image coordinate frame, different instances of the model can be projected into the image.

To account for successive nonlinear target model deformations, the plurality of edge points are organized into sets of sub-multiple points. By locally transforming the sets of points of the plurality, the spatial relationship of the sets to each other changes, resulting in a non-linear shape change of the entire target. Here, the local transformation applied to each group is a sufficiently small affine transformation, or a subset thereof, such as a rigid transformation or a translation. The model breakdown of the target is shown in fig. 4B. The partially generated input is a set of edge points 400 that were previously generated by feature extraction.

Once the edge points are extracted, part of the generated task is to group these points into a spatially correlated structure 403. Here, the present invention assumes that the structure of the spatial correlation remains the same even after the transformation. One aspect of the invention is to perform clustering in a manual manner. Here, the user will keep similar parts selected as a group. Another embodiment of the present invention performs clustering by an automated method. A straightforward approach is to set a fixed subdivision to the model and attribute points within a subdivision unit to a part. The other methodThe method is to compute a neighborhood graph of model points and to sort a fixed number of the nearest points into one part. It is important to note that the present invention is not limited to the case: i.e. the different sets of said secondary points are separate sets. In a preferred embodiment, a set of secondary points is generated for each point and its nearest neighbors. Regardless of the subdivision method used, the model points are divided into n parts, each part containing k _i And (4) model points. In order to speed up the subsequent calculations, a data structure is used which contains, for each part, the index n of the model point it contains _ij . Here, the index i ranges from 1 to n, and defines which part is selected; j is from 1 to k _i And defining the point of the portion. E.g., each section has the same number of model points, a matrix representation is used, where each row defines a section and each column defines an index into that section.

After such a subdivision is defined, the center 402 of each portion 403 is calculated, for example by taking the center of gravity of the corresponding set of points. The transformation 404 between the center of the part and the origin of the target model 401 is saved in the model. Thus, the relative position of the center of the part is converted into a transformation, such as a euclidean transformation, that changes the coordinate frame of the model into that of the part. These transformations allow the transformation of the position and orientation of the model points from the partial coordinate frame into the coordinate frame of the model and vice versa. Changing the relative transformation 404 between the target model and the part, for example by small movements along the x and y axes or rotations around the center of the part, allows to instantiate a deformed version of the target model. Some example deformations due to small translations in the x and y directions are illustrated in fig. 4C.

One aspect of the present invention is to extend the known methods for detecting medical targets in images in the presence of partial occlusions, clutter, and non-linear illumination variations.

The set of directed points of the target model is compared to the dense gradient direction field of the search image. The gradient direction remains the same even if the non-linear illumination delivered to the gradient magnitude varies considerably. Furthermore, hysteresis thresholds or non-maximum suppression are completely avoided in the search image, so that a true invariance with respect to arbitrary illumination variations is achieved. Partial occlusion, noise, and clutter result in random gradient directions in the search image. These effects lower the maximum value of the score for this metric, but do not change its position. The semantics of the score value are the scores that match the model points.

The idea of an efficient search is that target identification globally instantiates only the generalized affine transformation or a subset thereof. The search implicitly evaluates higher level nonlinear transforms by allowing for local movement of the parts and taking the maximum response as the best fitter. This is illustrated in fig. 5, where a scout image with two instances of deformation target models is shown. An example of a transmission transformation 500 of the model is shown on the left. On the right, a more complex arbitrary deformation 501 is depicted. As shown, the locally adapted portion 403 approximates the target in the search image. Changing the local transformation between the rigid position of the part and the locally adapted position allows representing a wide variety of target model appearances.

An important observation is: by transforming the image into a pyramidal representation, only small distortions need to be compensated at each level. For example, even if the object has a complex deformation at the lowest pyramid level, the appearance at the highest pyramid level does not change much. On the other hand, if the target has a large deformation, it can be compensated at the highest level. In the present invention, the deformation is transmitted along the pyramid in a recursive manner. If all higher level deformations are compensated at the higher pyramid level, only a relatively small change in the appearance of the target occurs at each level.

Thus, by dividing the search metric into global portions s _g And a partial section s _l . For the sake of clarity, we only give the formula for translation, which means that the score is only calculated for each row r and column c. It can be directly extended to be used for generalized affine parameters. As described above, the target model is divided into n parts, each part containing k _i And (4) model points.

The global metric is defined as:

meaning that: it is the combination of the score values of the local matches computed for each part defined by index i.

The local match metric is defined as:

here, the ij pairs define an index indicating which model point in the target is in which portion, where each portion has k _i And (4) model points. r is _ij And c _ij Is the row and column displacement of the corresponding model point in the model coordinate system. Local transformation of T _l To change the shape of the target model. Typically, these are euclidean transforms with a small contribution, e.g. 1 pixel shift in each direction. The superscripts m and s define d to be the directional vector of the target model or the directional vector of the corresponding location in the search image.

At each possible pose location, each part has a score value, as independent, for each part, the metric is evaluated in a range around its original affine location. The maximum score in the local neighborhood is the best fit for that part. The global metric is obtained by summing the results of the local metrics normalized by the number of model points in each section. A variant of the invention is that a threshold value can be set for each part that the part must exceed. Otherwise, the part is considered occluded and is therefore discarded without further processing.

Wherein

Another preferred embodiment is when the dimensions of the parts are different. At this point, one weights the impact of each segment by the number of target model points it contains.

The global score values of a set of generalized affine transformations also allow the determination of the approximate location of the target when the exact deformation is not known. Another variation is: to achieve invariance to contrast inversion, polarity information is discarded from the local measurement results. This is done by using the absolute value or sum of absolute values of the sum of normalized dot products of the direction vectors of the model points and the image points in the local measure.

By obtaining the best match for each part, not only the score values but also an estimate of the deformation is obtained. These are the local transformations T that define the maximum local score _l . After having the local displacement of each part, the corresponding non-linear target model is fitted. Even for locations without model points, a smooth deformation can be calculated. Fig. 6 shows an example variation. The center of portion 402 moves to a nearby location 603. The non-linear transformation is fitted to these points, transforming the original rigid space (schematically depicted as grid 601) into a deformation space 602. This is a well known problem in the art and various solutions have been proposed that are implemented based on functional interpolation and approximation. Here, it is an aspect of the present invention to use only the local displacement of each portion as a function point and to fit, for example, a perspective transformation to each point. For deformable target models, the spline function is defined by the displacement. This spline function is, for example, a B-spline function or a thin plate spline function. The coefficients of these functions are calculated by a direct method. However, if, for example, a thin-plate spline function is used, a very large linear system (liner system) must be inverted to obtain the distortion coefficient. Therefore, in another preferred embodiment, a harmonic interpolation method defined by the deformation of the model points is used. At this point, the displacement of the model point is inserted into two images describing the warping in the row and column directions. Then, the deformation is repaired for the region without the model point by a method called harmonic repair. To smooth the twist, the deformation is madeBack to the original regions of the model points. Therefore, not only the interpolation function but also the approximation function is obtained. The advantages of this method are: the runtime is only linearly related to the size of the target, not to the number of anchor points, for example, to the third power as with thin plate splines.

In general, especially for severe deformations, it is not possible to extract the deformations in one step. When given a deformation map, all model points and corresponding directions are transformed. With this transformed target model, each set of points of the model is now searched again independently for local displacements. This gives a loop of determining small displacements and fitting the model being evaluated until convergence is reached. Convergence is typically checked by checking whether the displacement becomes less than a predetermined threshold. For the defined range of global instances that exceed the threshold and are locally maximal, the target hypotheses with position, score and deformation information are put into a list for further examination at the lower pyramid level. In a preferred embodiment, not only is a threshold set on the global score value, but also the maximum number of hypotheses resulting from the highest pyramid level. At this point, all hypotheses are sorted according to their score values, and only a fixed number of best matching candidates are placed in the list of further processed hypotheses.

Once the exact location and deformation of the target model at a particular pyramid level is determined, the deformation must be transferred along the pyramid to the next pyramid level. It is important to do so that only local distortions of a small search range have to be evaluated at a lower level. In a preferred embodiment, the original affine model from the lower level is transformed into the higher pyramid level by recursive subdivision. The already extracted higher-level deformation is applied to the model, while the now transformed model from the lower level is transformed back to its original pyramid level. The search at this level starts with an instance of the model transformed according to the deformation of the higher pyramid level.

The tracking of hypotheses along the image pyramid proceeds until the lowest pyramid level is reached. At the lowest pyramid level, the displacement is determined at a resolution even higher than the original image. Thus, portions are instantiated at sub-pixel precision locations and corresponding maximum edge amplitudes in the image are determined. At this point, the displacement of the section is no longer defined by the gradient direction, but by the gradient magnitude. According to the above method, the deformation function is fitted with high accuracy using small displacements. Once the object is found at the lowest level, the position, pose, and deformation functions are returned. In addition, the value of the global score function is returned to provide a measure for the user to find the target to a better degree.

The second embodiment:

the present embodiments provide an identification apparatus for a deformable medical target comprising program code means stored on a computer readable medium for performing the method of identifying a deformable medical target of an embodiment one, when said computer program product is run on a computer.

As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and so forth) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.

Claims (10)

1. A method of identifying for a deformable medical target, the method comprising the steps of:

(a) Acquiring an image of a medical target model;

(b) Transforming an image of the target into a multi-level pyramid representation consistent with the recursive subdivision of the search space, the multi-level pyramid comprising at least an image of the target model;

(c) Generating, for each discretization stage of the search space, at least one pre-computed target model of the medical target, the pre-computed target model comprising a plurality of model points with corresponding direction vectors, the model points and direction vectors of the target resulting from image processing operations that return the direction vector of each model point;

(d) Subdividing the model points into a plurality of sections, wherein a deformed instance of the target model is represented by transforming the sections; each section is composed of a plurality of points, the plurality of sections forming an overlapping set of points;

(e) Acquiring a search image;

(f) Transforming a search image into a multi-level representation consistent with a recursive subdivision of a search space, the multi-level representation including at least the search image;

(g) Performing an image processing operation on each transformed image of the multi-level representation, the image processing operation returning direction vectors for a subset of target model points within the search image, the subset corresponding to a transformation range for which the at least one pre-computed target model should be searched;

(h) Clustering target model points to form a plurality of independent parts, matching and calculating each independent part in a certain range to obtain a local measurement result, and calculating the result sum normalized by the number of the model points in each part to obtain global matching measurement;

(i) Determining a target model pose for which the global matching metric exceeds a selectable threshold and for which the matching metric is locally greatest, and generating from the model pose a list of instances of the at least one pre-computed target model at a coarsest discretization level of the search space;

(j) Obtaining an estimate of the deformation by obtaining a best match for each local portion, calculating a deformation transformation describing the local displacement of the portion;

(k) Tracking, by recursive subdivision of the search space, instances of the at least one pre-computed target model at a coarsest level of discretization of the search space until a finest level of discretization is reached;

(l) Calculating a corresponding deformation transformation at each stage and transferring the deformation transformation to a next stage;

(m) providing the deformation transformation and model pose of the target model instance at the finest discretization level.

2. Method for the identification of a deformable medical target according to claim 1, characterized in that, for the deformable target model, the spline function is defined by a displacement, each set of points is searched independently for a local displacement, going along the image up to the lowest pyramid level, at which the displacement is determined at a resolution even higher than that of the original image, instantiated at a sub-pixel precision position, when part of the displacement is defined by the gradient amplitude, determining the position of the target at a resolution higher than the finest discretization level.

3. The method of claim 2, wherein in addition to the user-selectable threshold, only selectable instances of the target hypothesis that satisfy the deformation requirement are generated into the list of possible matches on the coarsest discretized level.

4. The method for identifying a deformable medical target of claim 3, wherein in step h, the local metric result for each segment must exceed a user-selectable local threshold, otherwise the segment is considered occluded and the segment is discarded without further processing.

5. An identification method for a deformable medical target as claimed in claim 4, characterized in that in step d, the subdivision is performed by means of k-means clustering or by means of normalized segmentation.

6. The method for identifying a deformable medical target of claim 4, wherein in step j, the computed transformation is a perspective transformation.

7. An identification method for a deformable medical target as claimed in claim 4, characterized in that in step j, the computed transformation is a three-dimensional pose, according to receiving as input metric information about the internal geometric parameters of the imaging device and the model.

8. An identification method for a deformable medical target according to claim 4, characterized in that in step h, the sum of normalized dot products of the transformed target model part and the direction vector of the search image is used for the local metric result.

9. The method of claim 4, wherein in step h, to achieve invariance to contrast inversion, polarity information is discarded from the local measurement results, and the absolute value of the sum of normalized dot products or the sum of absolute values of normalized dot products of the transformed target model portion and the direction vector of the search image is used for the local measurement results.

10. An identification device for a deformable medical target, characterized in that it comprises program code means stored on a computer readable medium for performing the identification method for a deformable medical target according to claims 1-9 when said computer program product is run on a computer.

Images