Model-based image segmentation

Abstract

A method and system for mapping boundary detection features of at least one source triangle mesh of known topology to a target triangle mesh of arbitrary topology. A region of interest in a volumetric image associated with each triangle of the target triangle mesh is provided to a feature mapping network. The feature mapping network assigns a feature selection vector to each triangle of the target triangle mesh. The associated region of interest and the assigned feature selection vector for each triangle of the target triangle mesh are provided to a boundary detection network. A predicted boundary is obtained from the boundary detection network based on features of the associated region of interest selected by the assigned feature selection vector.

Description

TECHNICAL FIELD

[0001] The present invention relates to the field of model-based image segmentation, and in particular to model-based image segmentation using triangular surface meshes. BACKGROUND

[0002] Model-based image segmentation is used in a range of applications to automatically segment objects from images. For example, model-based image segmentation techniques are used in medical image processing to segment organs or some other body part from volumetric medical images.

[0003] Model-based segmentation techniques using triangular surface meshes have proven to be fast, robust, and accurate. In these techniques, shape priors are encoded in a surface mesh, and the mesh is fitted to the image. Shape priors mean that the object in the image can be segmented even if some parts of the object boundary cannot be detected, and because only image data close to the mesh surface is processed to fit the mesh to the image, the image can be segmented quickly.

[0004] Conventional model-based segmentation techniques use features such as image gradients to detect the boundary of an object in an image. Recently, deep learning-based techniques have been developed to improve boundary detection. Brosch, T. et al., 2018 (“Deep Learning-Based Boundary Detection for Model-Based Segmentation with Application to MR Prostate Segmentation”, MICCAI (2018), pp. 512-522) describes a neural network trained to predict the boundary of each triangle of a mesh. The boundary at each triangle of the mesh is detected based on a combination of triangle-specific features.

[0005] However, because these techniques are trained to detect boundaries by assigning specific features to each triangle, they require surface meshes with a fixed topology (i.e. a fixed number of vertices and triangles). This prevents any changes in the topology, such as adding triangles to the mesh, removing triangles from the mesh, and otherwise rearranging or refining the topology, because features for triangles in the new, modified topology are not defined. As a result, such model-based segmentation techniques cannot be used to segment images of body parts whose shape differs significantly from a prior shape, for example due to anatomical abnormalities such as extra vertebrae, or artificial abnormalities such as screws from a spinal fixation. Some body parts, such as the rectum, have such a large shape flexibility that it is not possible to encode a shape prior that reliably delineates their shape.

[0006] Thus, there is a need for a model-based segmentation technique that can use triangular surface meshes of arbitrary topology. SUMMARY

[0007] The invention is defined by the claims.

[0008] According to examples in accordance with an aspect of the invention, there is provided a computer-implemented method of predicting a boundary of an object in a region of interest, the method comprising mapping boundary detection features of at least one source triangular mesh of known topology to a target triangular mesh of arbitrary topology.

[0009] The method comprises providing a region of interest in a volumetric image associated with each triangle of the target triangular mesh to a feature mapping network; assigning a feature selection vector to each triangle of the target triangular mesh using the feature mapping network; for each triangle of the target triangular mesh, providing the associated region of interest and the assigned feature selection vector to a boundary detection network; and for each triangle of the target triangular mesh, obtaining a predicted boundary from the boundary detection network based on features of the associated region of interest selected by the assigned feature selection vector.

[0010] The proposed concept is based on the proposal that triangles of meshes with different topologies that have a similar local image environment should have the same feature selection vector applied to them when determining the boundary at the triangle. The proposed embodiments use a second network, a feature mapping network, to assign a feature selection vector to each triangle of a mesh of arbitrary topology based on the appearance of the image region surrounding the triangle, rather than using a boundary detection network that requires a fixed mesh topology for triangle-specific feature selection vectors for each triangle.

[0011] The ability to predict boundaries of meshes of arbitrary topology means that the shape of the mesh can be adjusted to better depict structures in the image. A further advantage that the proposed embodiments can provide is that once the boundary detection network and the feature mapping network have been trained, no training data can be required to use these networks for new topologies.

[0012] In some embodiments, the method can further comprise obtaining the target triangular mesh of arbitrary topology by adding triangles to or removing triangles from the triangular mesh of known topology. In this way, boundaries can be predicted for meshes that have had triangles added or removed from them, for example to locally refine the mesh to improve the depiction of a body part in a volumetric medical image, or to add structure to the mesh to include anomalies in the shape of the body part.

[0013] In other embodiments, the method can further include obtaining an arbitrarily topological target triangle mesh by obtaining a segmentation of the volumetric image using a voxel-wise segmentation technique, and generating the target triangle mesh based on the obtained segmentation. In this way, a boundary can be obtained for structures for which no existing triangle surface mesh exists. This allows for delineation of body parts with large shape flexibility, such as the rectum.

[0014] The region of interest associated with each triangle can be oriented according to the normal of the triangle. This enables the feature mapping network to assign the same feature selection vector to triangles for which the associated region of interest looks similar when oriented such that the triangles are at the same angle.

[0015] The feature mapping network can be trained using a first training algorithm configured to receive an array of training inputs and known outputs, where the training inputs include regions of interest associated with triangles of arbitrarily topological meshes, and the known outputs include known boundaries of the regions of interest.

[0016] In some embodiments, the first training algorithm can be further configured to: assign a feature selection vector for each region of interest associated with a triangle of an arbitrarily topological mesh; provide the associated region of interest and the assigned feature selection vector to the boundary detection network for each mesh triangle of the arbitrarily topological mesh; obtain a predicted boundary from the boundary detection network for each triangle of the arbitrarily topological mesh; and train the weights of the feature mapping network based on the predicted boundaries from the boundary detection network and the known boundaries. In this way, the feature mapping network can be trained to learn a correspondence between the regions of interest and the feature selection vectors that best select features for predicting boundaries in the regions of interest.

[0017] In some embodiments, the method further includes providing a set of normalized coordinates for each region of interest to the feature mapping network.

[0018] In some embodiments, the method can further include providing a relative position of each triangle of the target triangle mesh to the feature mapping network.

[0019] The boundary detection network can be trained using a second training algorithm configured to receive an array of training inputs and known outputs, where the training inputs include training images of meshes having a known topology, and the known outputs include known boundaries. The training algorithm trains the boundary detection network to detect boundaries at triangles of the meshes of the known topology.

[0020] According to another aspect of the application, there is provided a computer- implemented method of model-based image segmentation comprising mapping features of at least one source triangular mesh of known topology to a target triangular mesh of arbitrary topology according to any of the previously described methods, and segmenting an object from a volumetric image using the target triangular mesh.

[0021] Such proposed methods can enable segmentation of objects that cannot be segmented in other ways using existing model-based segmentation techniques. For example, model-based segmentation techniques that use triangular surface meshes of fixed topology are not suitable for segmenting body parts that have abnormal anatomical structures and shapes that can vary greatly. By mapping features of meshes of fixed topology to meshes of arbitrary topology, these structures can be segmented.

[0022] According to another aspect of the application, there is provided a computer program comprising code means for implementing any of the previously described methods when said program is run on a processing system.

[0023] According to another aspect of the application, there is provided a processing system adapted to: provide a region of interest in a volumetric image associated with each triangle of a target triangular mesh to a feature mapping network; assign a feature selection vector to each triangle of the target triangular mesh using the feature mapping network; provide the associated region of interest and the assigned feature selection vector to a boundary detection network for each triangle of the target triangular mesh; and obtain a predicted boundary from the boundary detection network for each triangle of the target triangular mesh based on features of the associated region of interest selected by the assigned feature selection vector.

[0024] In some embodiments, the processing system is further adapted to orient the region of interest associated with each triangle according to a normal of the triangle.

[0025] In some embodiments, the processing system is further adapted to provide at least one of the following to the feature mapping network for each triangle of the target triangular mesh: a set of normalized coordinates of the associated region of interest and a relative position of the triangle.

[0026] According to another aspect of the application, there is provided a model-based image segmentation system comprising: a processing system as previously described that is further adapted to segment an object from a volumetric image using the target triangular mesh, and a user interface configured to receive and display a segmented image of the object from the processing system.

[0027] These and other aspects of the application will become apparent in light of the embodiments described below. BRIEF DESCRIPTION OF DRAWINGS

[0028] For a better understanding of the present application, and to show more clearly how it can be carried into effect, reference will now be made, by way of example only, to the accompanying drawings in which:

[0029] Figure 1 is an illustration of an object segmented by two meshes of different topologies.

[0030] Figure 2 is an illustration of triangles of a mesh oriented according to their normal. Figure 1

[0031] Figure 3 is a flowchart of a method of mapping boundary detection features of at least one source triangular mesh of known topology to a target triangular mesh of arbitrary topology according to an embodiment of the application.

[0032] Figure 4 is a flowchart of a method of training a feature mapping network according to an embodiment of the application.

[0033] Figure 5 is a flowchart of a model-based image segmentation method according to an embodiment of the application.

[0034] Figure 6 is a schematic illustration of a processing system according to an embodiment of the application.

[0035] Figure 7 is a schematic illustration of a model-based image segmentation system according to an embodiment of the application. DETAILED DESCRIPTION

[0036] The present application will be described with reference to the accompanying drawings.

[0037] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the system and method, are intended for purposes of illustration only and are not intended to limit the scope of the present application. These and other features, aspects, and advantages of the system and method of the present application will become better understood from the following description, appended claims, and accompanying drawings. It should be understood that the drawings are only schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the drawings for like or similar items.

[0038] ​According to the proposed concept, a method of mapping boundary detection features of at least one source triangle mesh of a known topology to a target triangle mesh of an arbitrary topology is provided. A region of interest in a volume image associated with each triangle of the target triangle mesh is provided to a feature mapping network. The feature mapping network assigns a feature selection vector to each triangle of the target triangle mesh. The associated region of interest and the assigned feature selection vector of each triangle of the target triangle mesh are provided to a boundary detection network. From the boundary detection network, a predicted boundary based on features of the associated region of interest selected by the assigned feature selection vector is obtained.

[0039] The illustrative embodiments can be used, for example, in a model-based image segmentation system, such as in a medical imaging analysis system.

[0040] The embodiments can be based at least in part on the recognition that the same feature selection vector can be used for triangles of meshes of different topologies that have similar local environments, and that a neural network can be trained to learn a correspondence between the local environment of a triangle and a feature selection vector.

[0041] Figure 1 An object 110 segmented by two meshes of different topologies is illustrated. In the top image, the object 110 is segmented by a first mesh 120. A region of interest 130 associated with a triangle 125 of the first mesh 120 encompasses the triangle 125 and a portion of a boundary between the object 110 and an adjacent structure 140. In the bottom image, the object 110 is segmented by a second mesh 150, which has more surface triangles than the mesh 120 and thus a different topology than the mesh 120. A second region of interest 160 associated with a triangle 155 of the second mesh 150 encompasses the triangle 155 and a portion of a boundary between the object 110 and a second adjacent structure 170.

[0042] Figure 2 The regions of interest 130 and 160 are illustrated as oriented according to the normal of the triangles 125 and 155, respectively. From Figure 2 It can be seen that the triangles 125 and 155 have similar associated regions of interest when oriented in the same direction. Thus, the same boundary detection features can be used to predict the boundary at the triangles 125 and 155. However, current boundary detection networks detect the boundary at a triangle based on a combination of features specific to the triangle. This means that a boundary detection network trained using the mesh 120 cannot be used to detect the boundary of a triangle of the mesh 150.

[0043] The present invention recognizes that by training the second neural network to map the boundary detection features of the triangles 125 of the mesh 120 to the triangles 155 of the mesh 150, the boundary detection network trained on the mesh 120 can be used to detect boundaries at the triangles 155.

[0044] Figure 3 A computer-implemented method 300 for mapping boundary detection features of at least one source triangle mesh of a known topology to a target triangle mesh of an arbitrary topology is illustrated according to an embodiment of the present invention. The at least one source triangle mesh can be an average mesh generated from N meshes of the same known topology, wherein each of the N meshes is a (ground truth) segmentation of N patient images.

[0045] The method 300 starts with step 320 in which a region of interest in the volumetric image associated with each triangle of the target triangle mesh is provided to the feature mapping network.

[0046] In step 340, the feature mapping network assigns a feature selection vector to each triangle of the target triangle mesh based on the associated region of interest.

[0047] In step 360, the associated region of interest and the assigned feature selection vector are provided to the boundary detection network for each triangle of the target triangle mesh.

[0048] In step 380, a predicted boundary for each triangle of the target triangle mesh is obtained from the boundary detection network based on the features of the associated region of interest selected by the assigned feature selection vector.

[0049] The target triangle mesh can be obtained by modifying a triangle mesh of the known topology or by generating a new mesh from scratch. For example, triangles can be added to or removed from the shape prior encoded in the MBS model to locally refine the mesh or to add additional structures to the mesh. Alternatively, an initial (coarse) segmentation technique, such as a voxel-wise segmentation technique, can be used to obtain an initial segmentation of the volumetric image, and the target triangle mesh can be obtained by generating a mesh based on the initial segmentation, e.g., by using marching cubes or some other mesh generation technique. Suitable initial (coarse) segmentation techniques will be apparent to the skilled person and can involve the use of a deep learning network.

[0050] The volumetric image can for example be a volumetric medical image. For example, the volumetric image can be a computed tomography (CT) image, a magnetic resonance (MR) image, a nuclear medicine image such as a positron emission tomography (PET) image or a single photon emission computed tomography (SPECT) image, or a volumetric ultrasound image.

[0051] Each triangle of the target triangle mesh is associated with a region of interest in the volumetric image. Each region of interest contains its associated triangle, a portion of the object that the target mesh is to depict, and a portion of the surrounding environment of the object. Each region of interest can be oriented so that the associated triangle is at the same angle. For example, each region of interest can be oriented according to the normal of its associated triangle.

[0052] Each region of interest is provided to a feature mapping network. Each region of interest can be accompanied by additional information. For example, in the case where the object to be depicted is pre-registered to a common coordinate system, such as an atlas-based registration for volumetric medical images, the feature mapping network can be provided with a set of normalized coordinates for each region of interest. The feature mapping network can be provided with the relative position of the triangles of the target triangle mesh associated with each region of interest, such as the spherical coordinates of each triangle or a distance vector from each triangle to the barycenter of the mesh.

[0053] The feature mapping network is an artificial neural network trained to assign a feature selection vector to each region of interest. An artificial neural network is a machine learning algorithm, i.e. a self-training algorithm that processes input data in order to produce or predict output data. The input data of the feature mapping network comprises regions of interest associated with triangles of a mesh of arbitrary topology, and the output data comprises the assigned feature selection vector.

[0054] The structure of an artificial neural network (or simply, neural network) is inspired by the human brain. A neural network is composed of multiple layers, each layer comprising multiple neurons. Each neuron comprises a mathematical operation. In particular, each neuron can comprise a different weighted combination of a single type of transformation (e.g. the same type of transformation, sigmoid, etc., but with different weightings). In the process of processing input data, the mathematical operation of each neuron is performed on the input data to produce a numerical output, and the output of each layer in the neural network is sequentially fed to the next layer. The last layer provides the output.

[0055] Methods of training neural networks are well known. Typically, such methods comprise obtaining a training data set comprising training input data entries and corresponding training output data entries. An initialized machine learning algorithm is applied to each input data entry to generate a predicted output data entry. The error between the predicted output data entry and the corresponding training output data entry is used to modify the machine learning algorithm. This process can be repeated until the error converges, and the predicted output data entry is sufficiently similar to the training output data entry (e.g. ±1%). For example, the weightings of the mathematical operations of each neuron can be modified until the error converges. This is commonly referred to as a supervised learning technique. Known methods of modifying a neural network include gradient descent, backpropagation algorithms, etc.

[0056] The feature mapping network can be trained using training input data entries corresponding to regions of interest associated with triangles of a mesh of arbitrary topology and training output data entries corresponding to known boundaries of the regions of interest. The training data set can thus comprise a set of training images having an adapted mesh of arbitrary topology and annotated to delineate boundaries of objects in the images.

[0057] Figure 4 A computer-implemented method 400 for training a feature mapping network according to an embodiment of the application is illustrated.

[0058] The method 400 starts in step 410 in which the feature mapping network receives from a training data set regions of interest associated with triangles of a mesh of arbitrary topology. The regions of interest can be oriented such that their associated triangles are at the same angle.

[0059] In step 420, for each region of interest, the feature mapping network assigns a feature selection vector to the triangle associated with the region of interest.

[0060] In step 430, the associated region of interest and the assigned feature selection vector for each triangle are provided to a boundary detection network.

[0061] In step 440, a predicted boundary for each triangle is obtained from the boundary detection network. The boundary detection network predicts the boundary at each triangle based on the features of the associated region of interest selected by the assigned feature selection vector.

[0062] In step 450, the weights of the feature mapping network are trained based on the known boundaries from annotations on the images comprising the training data set and the obtained predicted boundaries. This step can comprise calculating a distance between each predicted boundary and its corresponding known boundary and training the weights of the feature mapping network based on the calculated distances.

[0063] These steps can be repeated until the boundaries predicted by the boundary detection network using the feature selection vectors assigned by the feature mapping network are sufficiently similar to their corresponding known boundaries (e.g. until they are within 1% of each other).

[0064] Returning to Figure 3 The feature selection vector assigned to each triangle of the target mesh of triangles by the trained feature mapping network in step 340 is provided to the boundary detection network in step 360 along with the associated region of interest for each triangle.

[0065] The boundary detection network is a second artificial neural network trained to predict boundaries in a volumetric image at a triangle of a mesh of known topology based on features of the image selected by the triangle-specific feature selection vector.

[0066] The boundary detection network can be trained using training input data entries corresponding to training images of meshes having at least one known topology and training output data entries corresponding to the known boundaries in the training images. The training data set can thus comprise a set of training images having adapted meshes of known topologies and annotated to depict the boundaries of objects in the images. The training images can be the same images used to train the feature mapping network, but with meshes of known topologies instead of adapted to their arbitrary topologies. The boundary detection network can use the method described in Brosch, T. et al., 2018.

[0067] In this way, the source mesh of known topology can be considered to be an average mesh generated from all N meshes of the same known topology used to train the boundary detection network, and the features assigned to each triangle of the source mesh are generated by training features across all N meshes. Thus, the source mesh of known topology is a representation of all meshes of the known topology used to train the boundary detection network, and the features assigned to each triangle of the source mesh are features representing the best match across all N meshes for that triangle.

[0068] In step 380, the predicted boundaries for each triangle of the target triangular mesh of arbitrary topology are obtained from the boundary detection network. The boundary detection network is able to select features to predict the boundaries at the triangles of the target mesh of arbitrary topology using the triangle-specific feature selection vectors specific to the meshes of the known topology used to train the boundary detection network, as the feature mapping network has determined which of these feature selection vectors will best select features for predicting the boundaries at each triangle of the target triangular mesh.

[0069] In other words, the feature mapping network maps the triangle-specific feature selection vectors to the target triangular mesh, and the boundary detection network uses the mapped feature selection vectors to select which features to use to determine the boundaries at each triangle of the target triangular mesh. The boundary detection network then predicts the boundaries based on the selected features of the region of interest associated with each triangle of the target mesh.

[0070] Figure 5 Fig. illustrates a computer-implemented method 500 of model-based image segmentation according to an embodiment of the invention.

[0071] The method starts in step 510 in which features of at least one source triangular mesh of known topology are mapped to a target triangular mesh of arbitrary topology according to any of the previously described methods.

[0072] At step 520, the object is segmented from the volumetric image using the target triangle mesh. For example, when the volumetric image is a volumetric medical image, the target triangle mesh can be used to segment an organ or some other body part from the volumetric medical image. Methods for segmenting an object from a volumetric image using a triangle mesh are well known and will be apparent to those skilled in the art.

[0073] Figure 6 A processing system 600 according to an embodiment of the application is illustrated. The processing system 600 comprises a feature mapping network 640 and a boundary detection network 660.

[0074] For each triangle of the target triangle mesh, the associated region of interest 620 of the volumetric image is provided to the feature mapping network 640, which assigns a feature selection vector 662 to each triangle based on the associated region of interest 620 of each triangle.

[0075] The associated region of interest 620 and the assigned feature selection vector 662 for each triangle of the target triangle mesh are provided to the boundary detection network 660. The boundary detection network 660 uses the features 664 of the associated region of interest 620 selected by the assigned feature selection vector 662 to determine a predicted boundary 680 for each triangle of the target triangle mesh.

[0076] In some embodiments, the processing system 600 is further adapted to orient the region of interest 620 associated with each triangle so that each triangle is at the same angle. For example, the processing system can be adapted to orient the region of interest 620 associated with each triangle according to the normal of the triangle.

[0077] In some embodiments, the processing system 600 is further adapted to provide additional information to the feature mapping network 640. For example, the processing system 600 can be adapted to provide a set of normalized coordinates for each region of interest 620 to the feature mapping network 640 and / or the relative position of each triangle of the target triangle mesh to the feature mapping network 640.

[0078] Figure 7 A model-based segmentation system 700 according to an embodiment of the application is illustrated. The model-based segmentation system comprises a processing system 600 as previously described and a user interface 720, the processing system 600 being further adapted to segment an object from a volumetric image using a target triangle mesh, thereby producing a segmented image 710.

[0079] The user interface 720 is configured to receive the segmented image 710 from the processing system 600 and to display the segmented image 710. The user interface can also receive and display the original volumetric image and / or the target triangle mesh from the processing system 600.

[0080] It will be appreciated that the disclosed methods are computer implemented methods. As such, there is further presented the concept of a computer program which comprises code means for performing any previously described method when said program is run on a processing system.

[0081] Those skilled in the art will readily develop processors for performing the methods described herein. Thus, each step of the flowcharts can represent a distinct action taken by a processor and can be performed by a corresponding module of the processing processor.

[0082] As described above, the system utilizes a processor to perform data processing. The processor can be implemented in numerous ways, with software and / or hardware, to perform the various functions described herein. The processor typically employs one or more microprocessors that can be programmed using software (e.g., microcode) to perform the various functions described herein. The processor can be implemented as a combination of dedicated hardware to perform some functions and one or more programmed microprocessors and associated circuitry to perform other functions.

[0083] Examples of circuitry that can be employed in various embodiments of the disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field-programmable gate arrays (FPGAs).

[0084] In various implementations, the processor can be associated with one or more storage media such as volatile and non-volatile computer memory such as RAM, PROM, EPROM, and EEPROM. The storage media can be encoded with one or more programs that, when executed on one or more processors and / or controllers, perform the required functions. Various storage media can be fixed within or portable to the processor or controller.

[0085] Variations to the disclosed embodiments can become apparent to those of ordinary skill in the art from a reading of the drawings, the disclosure, and the appended claims. In the claims, the term comprising does not exclude other elements or steps, and the indefinite articles "a" or "an" do not exclude a plurality. A single processor or other unit can fulfil the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures can not be used to advantage. A computer program can be stored / distributed on a suitable medium, such as an optical storage medium or a solid state storage medium supplied together with or as part of other hardware, but can also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. If the term "comprising" is used in the claims, the term "comprising" is not intended to be construed as a statistical term or a term of limitation. If the term "adapted to" is used in the claims, it should be noted that the term "adapted to" is intended to be equivalent to the term "configured to." Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A computer-implemented method for predicting the boundary of an object (110) in a region of interest (160, 620), the method being adapted to map boundary detection features of at least one source triangular mesh (120) with known topology to a target triangular mesh (150) with arbitrary topology, the computer-implemented method comprising: - The regions of interest (160, 620) in the volumetric image associated with each triangle (155) of the target triangular mesh (150) are provided to the feature mapping network (640), wherein the regions of interest (160) include each triangle (155) of the target triangular mesh (150), a portion of the boundary of the object (110) to be depicted by the target triangular mesh, and a portion of the surrounding environment of the object, and the feature mapping network (640) is configured to assign a feature selection vector (662) to each triangle (155) of the target triangular mesh (150) for selection based on the regions of interest, thereby mapping the boundary detection features of the triangles (125) of the at least one source triangular mesh (120) to the triangles (155) of the target triangular mesh (150). - The feature selection vector (662) is assigned to each triangle (155) of the target triangular mesh (150) using the feature mapping network (640). - For each triangle (155) of the target triangular mesh (150), the associated region of interest (160, 620) and the assigned feature selection vector (662) are provided to a boundary detection network (660), wherein the boundary detection network is configured to detect the predicted boundary of each triangle (155) of the target triangular mesh (150) based on features of the region of interest selected by the assigned feature selection vector (662); and - For each triangle (155) of the target triangular mesh (150), a predicted boundary (680) is obtained from the boundary detection network (660) based on the features (664) of the associated region of interest (160, 620) selected by the assigned feature selection vector (662).

2. The computer-implemented method of claim 1, further comprising: The target triangular mesh (150) of any topology is obtained by adding triangles to or removing triangles from a triangular mesh of a known topology.

3. The computer-implemented method of claim 1, further comprising: The target triangular mesh (150) with arbitrary topology is obtained through the following steps: - The volumetric image is segmented using a voxel-by-voxel segmentation technique; as well as - Generate the target triangular mesh (150) based on the obtained segmentation.

4. The computer-implemented method according to any one of claims 1 to 3, wherein the region of interest (160, 620) associated with each triangle (155) of the target triangular mesh (150) is oriented according to the normal of the triangle of the target triangular mesh (150).

5. The computer-implemented method according to any of claims 1 to 3, wherein the feature mapping network (640) is trained using a first training algorithm configured to receive an array of training inputs and known outputs, wherein the training inputs comprise regions of interest associated with triangles of a mesh of arbitrary topology and the known outputs comprise known boundaries of the regions of interest.

6. The computer-implemented method according to claim 5, wherein the first training algorithm is further configured to: - assign a feature selection vector (662) for each region of interest associated with a triangle of a mesh of arbitrary topology; - provide the boundary detection network (660) with the associated region of interest and the assigned feature selection vector for each triangle of a mesh of arbitrary topology; - obtain a predicted boundary (680) from the boundary detection network for each triangle of a mesh of arbitrary topology; and - train the weights of the feature mapping network (640) based on the predicted boundary (680) from the boundary detection network (660) and the known boundary.

7. The computer-implemented method of any one of claims 1 to 3, further comprising: A set of normalized coordinates for each region of interest (160, 620) is provided to the feature mapping network (640).

8. The computer-implemented method of any one of claims 1 to 3, further comprising: The relative position of each triangle (155) of the target triangular mesh (150) is provided to the feature mapping network (640).

9. The computer-implemented method according to any of claims 1 to 3, wherein the boundary detection network (660) is trained using a second training algorithm configured to receive an array of training inputs and known outputs, wherein the training inputs comprise training images of a mesh with known topology and the known outputs comprise known boundaries.

10. A computer-implemented method of model-based image segmentation, comprising: - mapping features of at least one source triangular mesh (120) of known topology to a target triangular mesh (150) of arbitrary topology according to the method of any of claims 1-9; and - segmenting objects from the volumetric image using the target triangular mesh (150).

11. A computer program product comprising code means for implementing the method according to any of claims 1-10 when said program is run on a processing system.

12. A processing system (600) for predicting boundaries of objects (110) in regions of interest (160, 620) of objects (110), and adapted to: - providing the region of interest (160, 620) in the volumetric image associated with each triangle (155) of a target triangle mesh (150) to a feature mapping network (640), wherein the region of interest (160) comprises each triangle (155) of the target triangle mesh (150), a portion of a boundary of the object (110) to be depicted by the target triangle mesh, and a portion of a surrounding environment of the object, the feature mapping network (640) being configured to assign a feature selection vector (662) to each triangle (155) of the target triangle mesh (150) based on the region of interest for selection, thereby mapping boundary detection features of triangles (125) of at least one source triangle mesh (120) to triangles (155) of the target triangle mesh (150); - assigning the feature selection vector (662) to each triangle (155) of the target triangle mesh (150) using the feature mapping network (640); - for each triangle (155) of the target triangle mesh (150), providing the associated region of interest (160, 620) and the assigned feature selection vector (662) to a boundary detection network (660), wherein the boundary detection network is configured to detect a predicted boundary for each triangle (155) of the target triangle mesh (150) based on features of the region of interest selected by the assigned feature selection vector (662); and - for each triangle (155) of the target triangle mesh (150), obtaining a predicted boundary (680) from the boundary detection network (660) based on features (664) of the associated region of interest (160, 620) selected by the assigned feature selection vector (662).

13. The processing system (600) of claim 12, further adapted to orient the region of interest (160, 620) associated with each triangle (155) of the target triangle mesh (150) according to a normal of the triangle of the target triangle mesh (150).

14. The processing system (600) of claim 12 or 13, further adapted to provide, for each triangle (155) of the target triangle mesh (150), to the feature mapping network (640) at least one of: a set of normalized coordinates for the associated region of interest (160, 620), and a relative position of the triangle (155) of the target triangle mesh (150).

15. A model-based image segmentation system (700), comprising: - the processing system (600) of any one of claims 12 to 14, further adapted to segment an object from the volumetric image using the target triangle mesh (150); and - a user interface (720) configured to receive and display a segmented image (710) of the object from the processing system (600).

Images