Data augmentation

Abstract

A method for generating data representing the volume of a part of a body, the method comprising generating a point distribution model "PDM" based on an input data set, the input data set comprising data representing at least one surface of the part of the body, the PDM defining a surface model data set based on an average data set and one or more weight feature vector pairs, generating a first surface model data set based on the PDM by modifying at least one weight in the one or more weight feature vector pairs, wherein the first surface model data set is different from the average data set, and generating an output volume data set based on the first surface model data set and a first reference data set, the first reference data set comprising data representing the volume of the corresponding part of the body, and the output volume data set comprising data representing the deformed volume of the corresponding part of the body.

Description

Technical Field

[0001] The present disclosure relates to data augmentation. In particular, the present disclosure relates to the generation of datasets representing body parts. Background Art

[0002] Computational models of body parts are used in many scenarios. For example, in medical training, computational models of organs are useful for studying different medical phenomena. In modern applications, for example, artificial intelligence (AI) systems can be used to identify such phenomena in scans from patients. To do this, the system must learn what different phenomena look like so that they can be accurately identified. In many cases, this is achieved by inputting large training data sets into the AI ​​system. The training data set is a model that represents a real patient, which has been studied by experts in the field. This allows the AI ​​system to know what to look for in the data set and thus learn to recognize interesting features present in other input data sets. This is known as machine learning.

[0003] In order to train computing systems as comprehensively and robustly as possible, large training datasets are required. However, only a limited number of real patient models are available, and this number is often insufficient to robustly train AI systems. Therefore, the number of training datasets used for machine learning applications is often augmented to increase the number of examples and, therefore, the stability of the trained models. Typically, this is achieved by modifying an existing dataset to generate a new and different dataset. This modification is performed by rotating, scaling, translating, and / or deforming the existing dataset. In the case of applied deformations, the existing model is deformed in a random, non-systematic manner.

[0004] Taking this approach leads to many problems as the generated datasets may not be realistic. In reality, the body parts of different patients are not simple transformations of other patients’ body parts. Random deformations do not accurately convey the real inter-patient differences. Such unrealistic datasets may lead to incorrect training of AI systems, resulting in poor performance of these systems in identifying features of interest present in other input datasets.

[0005] Therefore, an object of the present disclosure is to create a more realistic dataset representing body parts. Summary of the invention

[0006] According to one aspect of the present disclosure, a method for generating data representing the volume of a part of a body is provided, the method comprising generating a point distribution model "PDM" based on an input data set, the input data set comprising data representing at least one surface of the part of the body, the PDM defining a surface model data set based on an average data set and one or more weight feature vector pairs; generating a first surface model data set based on the PDM by modifying at least one weight in the one or more weight feature vector pairs, wherein the first surface model data set is different from the average data set; and generating an output volume data set based on the first surface model data set and a first reference data set, the first reference data set comprising data representing the volume of the corresponding part of the body, and the output volume data set comprising data representing the deformed volume of the corresponding part of the body.

[0007] Optionally, the part of the body comprises at least one organ. Optionally, the body is a human body. Optionally, the input dataset is derived from at least one real body. Optionally, the input dataset comprises a plurality of datasets, each dataset representing the same part of a respective body.

[0008] Optionally, generating a first surface model dataset includes modifying at least one weight in one or more weight feature vector pairs so that the first surface model dataset generally corresponds to a second reference dataset, and the method also includes modifying at least one weight in one or more weight feature vector pairs to provide a second surface model dataset different from the first surface model dataset, and determining a mapping between the first and second surface model datasets.

[0009] Optionally, the second reference data set represents a corresponding part of the body and is derived from a real body. Optionally, determining the mapping includes using a deformable image registration algorithm based on the first and second surface model data sets. Optionally, the deformable image registration algorithm only produces positive Jacobians. Optionally, generating the output volume data set includes applying the mapping to the first reference data set.

[0010] Optionally, generating the PDM comprises using principal component analysis. Optionally, the average data set is a mean average of the input data sets. Optionally, the one or more weight feature vector pairs comprise a plurality of weight feature vector pairs. Optionally, the plurality of weight feature vector pairs comprise linear combinations of feature vectors. Optionally, the PDM is defined by the following equation

[0011] x m =x a +α0e0+α1e1+α2e2+...α n e n

[0012] Among them, x m is the surface model dataset, x a is the average data set, αx is the weight of the xth weight feature vector pair, e x is the eigenvector of the xth weighted eigenvector pair, and n is the number of weighted eigenvector pairs.

[0013] Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises: modifying the weight within an interval defined based on a standard deviation of the input data set. Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises randomly modifying the weight. Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises incrementally modifying the weight from an initial weight. Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises modifying all weights in one or more weight eigenvector pairs.

[0014] Optionally, the first reference dataset represents a corresponding part of the body and is derived from a real body. Optionally, the method further comprises using the output volume dataset as part of a training set for a deep learning process of the artificial intelligence system.

[0015] According to another aspect, a method for generating data representing the volume of a part of a body is provided, the method comprising generating a point distribution model "PDM" based on an input data set, the input data set comprising data representing at least one volume of the part of the body, the PDM defining an output volume data set based on an average data set and one or more weight feature vector pairs; and generating an output volume data set based on the PDM by modifying at least one weight in one or more weight feature vector pairs, wherein the output volume data set is different from the average data set.

[0016] Optionally, the part of the body comprises at least one organ. Optionally, the body is a human body. Optionally, the input data set is derived from at least one real body.

[0017] Optionally, the input dataset comprises a plurality of datasets, each dataset representing the same part of the corresponding body.

[0018] Optionally, generating the PDM comprises using principal component analysis. Optionally, the average data set is the mean of the input data sets. Optionally, the one or more weight feature vector pairs comprise a plurality of weight feature vector pairs. Optionally, the plurality of weight feature vector pairs comprise linear combinations of feature vectors. Optionally, the PDM is defined by the following equation

[0019] x m =x a +α0e0+α1e1+α2e2+...α n e n

[0020] Among them, x mis the surface model dataset, x a is the average data set, α x is the weight of the xth weight feature vector pair, e x is the eigenvector of the xth weighted eigenvector pair, and n is the number of weighted eigenvector pairs.

[0021] Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises: modifying the weight within an interval defined based on a standard deviation of the input data set. Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises randomly modifying the weight. Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises incrementally modifying the weight from an initial weight. Optionally, modifying at least one weight in one or more weight eigenvector pairs comprises modifying all weights in one or more weight eigenvector pairs.

[0022] Optionally, the method further comprises using the output volume dataset as part of a training set for a deep learning process for an artificial intelligence system.

[0023] According to another aspect, a computer readable medium is provided, the computer readable medium comprising instructions which when executed on a processor causes the processor to perform the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Exemplary embodiments of the present disclosure will now be described with reference to the accompanying drawings, in which:

[0025] Figure 1 A view showing a computational model of a region of interest;

[0026] Figure 2 A method for generating data representing the volume of a part of a body according to the present disclosure is shown;

[0027] Figure 3 Another method for generating data representing the volume of a part of a body according to the present disclosure is shown; and

[0028] Figure 4 A schematic diagram of a computing environment is shown.

[0029] Like reference numerals refer to like parts throughout the specification and drawings. DETAILED DESCRIPTION

[0030] Figure 1A view of a computational model of a region of interest (ROI) 100 is shown. In this case, the ROI is of a human body and includes multiple organs. As can be seen, the ROI 100 includes two lungs 102, 104 and a heart 106. In other embodiments, the ROI may be of an animal body. In other embodiments, the ROI may include a different number of organs, such as a single organ, two organs, or four or more organs.

[0031] The computational model can be derived from a scan of the patient. For example, a computed tomography (CT) scan, a magnetic resonance imaging (MRI) scan, or other types of scans known in the art can be used to capture the ROI 100 in the patient. The scan can then be segmented to define different features in the ROI. For example, the segmentation process can define the first lung 102, the second lung 104, and the heart 106. The segmentation process can also identify internal features of various organs. The resulting computational model may include data representing the geometry of the ROI 100 in three dimensions, such as surface and volume coordinates of the organs 102-106.

[0032] This computational model is an example of a data set that can be used in a training data set for the machine learning process of an AI system, for example, to identify physical features in a patient ROI. Such a training data set should include as many different models as possible so that the AI ​​system can acquire as many different features as possible and therefore accurately identify these features in future input data sets. However, for various reasons, only a relatively small number of patients are used to derive models for any given ROI.

[0033] The present disclosure proposes a method of generating data representing a ROI that is different from existing datasets and is also realistic (human-like), rather than currently known artificial geometry methods that do not produce the true geometry of the ROI.

[0034] The method includes generating a point distribution model (PDM) for a particular ROI based on an input data set. The point distribution model represents the average geometry of the shape and some statistical patterns of geometric variations inferred from the input set of shapes. The input data set includes a plurality of data sets, each data set representing the ROI for which the PDM is to be generated. In some embodiments, each data set in the input data set can be derived from a scan of a patient, as described above. In other embodiments, each data set in the input data set can be derived artificially. In still other embodiments, the input data set can include a combination of data sets derived from real patient scans and artificially derived data sets.

[0035] Each data set in the input data set includes multiple data points, so that the geometry of the original shape can be fully represented. A data set x including m data points can be described in two dimensions as x=(x1, y1, ..., x m ,y m ), which can also be described in three dimensions as x = (x1, y1, z1, ..., x m ,y m , z m ). Advantageously, each data point represents the same location in the ROI across all of the data sets in the input data set, e.g., the first data point in each data set represents the same portion of the first lung 102. In some embodiments, the input data set includes a set of data points that describe a geometric shape in three dimensions. In some embodiments, this may be surface data representing the ROI, e.g., data points representing the surface of the lungs 102, 104 and heart 106 of the ROI 100 in three dimensions. In some embodiments, the data set may be volumetric data representing the ROI, e.g., data points representing the volume of the organs 102-106 of the ROI 100 in three dimensions. Using surface data simply reduces the storage and processing requirements of the method, thereby providing a more efficient PDM generation. Using volumetric data provides a more comprehensive and accurate ROI modeling, thereby providing a more comprehensive and accurate PDM.

[0036] Principal component analysis (PCA) can be used to compute the normalized eigenvectors and eigenvalues ​​of the covariance matrix across the input dataset, where each eigenvector describes the main mode of variation along the set. Finally, linear combinations of the eigenvectors are used to define the shape / model of the new dataset x m , which is defined by the equation:

[0037] x m =x a +α0e0+α1e1+α2e2+...+α n e n

[0038] where x m is the model dataset, x a is the average shape across the input dataset, e x is the eigenvector describing the xth change pattern, α x is the scaling value or weight of the xth eigenvector, and n+1 is the number of variation patterns.

[0039] In some embodiments, the average is a mean average. In some embodiments, there may be only a single variation mode (i.e., n=0). In other embodiments, there are multiple variation modes (i.e., n>0). In this case, e orepresents the most common changes, and subsequent eigenvectors represent increasingly less common changes until e n Indicates the least common changes.

[0040] If all weights α n are all set to zero, the resulting model dataset x m With the average data set x a By changing the value of α0 to α n , an infinite number of model data sets can be generated, each representing a different shape. Since PDM is based on an input data set that includes real geometry, the captured variations are real, and therefore the model data sets are also real.

[0041] Modifying any weight from zero will produce a model dataset x that is different from the mean dataset m In some embodiments, a single weight is modified. In other embodiments, a subset of multiple weights is modified. In other embodiments, all weights α0 to α n are all modified. The weights can be modified within an interval to ensure that unrealistically large deformations are not generated. The interval can be defined based on the standard deviation of the input data set around the mean surface. For example, each value of α can be limited to ±3 standard deviations. The standard deviation of a given principal component can be defined as the square root of its corresponding eigenvalue. In some embodiments, each weight is modified in a random manner. In other embodiments, each weight is modified systematically. For example, each weight is incrementally modified from an initial weight. For example, if an initial value of α2 is given as 0.34, the value of α2 can be incrementally modified, such as to 0.35, 0.36, etc., to ensure that the model data set will change continuously from the first surface model data set. This modification may also be useful when modeling dynamic phenomena such as the breathing cycle.

[0042] By implementing different modifications of the weights in the PDM, different model data sets can be generated. As described above, since the PDM is based on an input data set that includes real geometry, the captured variations are real and therefore the model data set is also real. In embodiments where an ROI such as ROI 100 includes multiple organs, the described method allows the organs to deform relative to each other in the model data set so that one of the organs does not deform unrealistically relative to the other organs (e.g., producing two lungs of different sizes, which is rare in reality). Known geometry generation methods cannot guarantee this.

[0043] In embodiments where the input dataset comprises volume data, the model dataset x generated from the PDM m is a deformed version of the mean volume in the PDM. In this case, each model dataset x generated from the PDM mAll can be used for training data sets for the machine learning process of the AI ​​system. The machine learning process may be a deep learning process. The AI ​​system may be an image segmentation algorithm. By increasing the training data set, the machine learning process becomes more comprehensive and more accurate, and the AI ​​system is more accurate. In this way, the AI ​​system has improved performance. In the example of an image segmentation algorithm, the algorithm may be able to more accurately segment the image of the ROI input into the algorithm. In other embodiments, the geometry can be used for adaptive scenarios that expect to predict the geometry of future patients. For example, the ROI of a given patient may look different at different times. For example, this may be at different stages of the respiratory cycle, where the lungs will have different shapes and sizes. Another example is that if the treatment is effective or has side effects, the geometry of the ROI may change at different stages of treatment. The method discussed above allows the generation of a geometry corresponding to the desired result after a given treatment stage, which can then be compared with the patient's true geometry to indicate whether the treatment is successful. The method described herein allows the generation of geometry for any application field that requires real deformed data.

[0044] Figure 2 A method 200 is shown for generating data representing a volume of a portion of a body. As described above, the portion of the body may be a ROI, such as ROI 100, including at least one organ, such as organs 102-106. In some embodiments, the body is a human body.

[0045] At step 202, a PDM for a ROI is generated based on an input dataset. The input dataset includes data representing a volume of the ROI and, in some embodiments, is derived from at least one real body, such as a patient. The input dataset may include multiple datasets, each representing the same ROI from a different patient.

[0046] The PDM defines an output data set based on a mean data set and one or more weighted eigenvector pairs. In some embodiments, the PDM is generated using principal component analysis and comprises a linear combination of the mean and eigenvectors of the input data sets.

[0047] At step 204, an output data set is generated based on the PDM. The output data set is generated by modifying at least one weight in the weight feature vector pair so that the output data set is different from the average data set. In some embodiments, the weights are modified within an interval defined by the standard deviation of the input data set. The weights can be modified randomly or incrementally from the initial weights. One, some, or all weights in the linear combination can be modified to produce the output data set.

[0048] At step 206, the output data set is used as part of a training set for machine learning of an AI system. For example, this could be deep learning for an image segmentation algorithm. Multiple output data sets may be generated to make the training set as large as possible, thereby ensuring that the AI ​​system is robustly trained. This results in a system with improved performance.

[0049] In some embodiments, the model dataset generated from the PDM can be exploited to generate additional datasets at a relatively low computational cost. As described below, surface data can only be used at different stages to generate different volume datasets. By using surface data rather than volume data, the computational burden associated with calculating feature vectors can be greatly reduced. In order to generate a given model dataset x from the PDM, m To achieve this, the model dataset is related to one or more reference datasets.

[0050] In some embodiments, the PDM is used to generate a first model data set using the above method. In this case, the input data set only includes surface data representing the ROI, and therefore, the model data set generated from the PDM also includes surface data representing the ROI of the new geometry. In some embodiments, the modification of the above weights is performed so that the first model data set generally corresponds to the reference data set as much as possible. The reference data set is obtained from a real body and represents the ROI corresponding to the input data set and the PDM. In some embodiments, the reference data set includes surface data representing the corresponding ROI, while in other embodiments, the reference data set includes volume data representing the corresponding ROI. Using surface data allows for simpler modification of the PDM to match the reference ROI. The reason is that when the model data set (which is a triangular surface) is adapted to correspond to, for example, volume data and images from CT scans, it is necessary to extract different types of image features that describe the edges between different parts of the ROI. This is called image segmentation and may be difficult for ROIs or ROI parts with fuzzy boundaries. By using surface data, experts, such as doctors, have already segmented the surface / ROI / organ from the image, making the PDM more suitable for the existing surface. Rotation, translation and scaling of the model data set generated by the PDM may also be used to provide better correspondence with the reference data set.

[0051] In order to ensure that the correspondence is as accurate as possible, a similarity metric can be applied. In some embodiments, a dice similarity coefficient (DSC) is used. DSC is a number between 0 and 1 that describes the geometric similarity between geometric objects. If the DSC is not high enough after a given number of iterations, it is considered that the adaptation has failed.

[0052] Once the first surface model data set is generated, the weights may be further modified to provide a second surface model data set that is different from the first model data set. The modification of the weights may be performed in any of the ways described above.

[0053] In this way, two surface model data sets are generated. Then the geometric mapping between the first and second surface model data sets can be determined. The mapping describes the deformation required to change the first surface model data set into the second surface model data set. In some embodiments, based on the first model data set and the second model data set, a deformable image registration (DIR) algorithm is used to determine the mapping. The DIR algorithm generates a three-dimensional deformation field or mapping. In this case, the generated mapping will map the first surface model data set to the second surface model data set. If the first and second surface model data sets are the same, the mapping will be invalid. In some embodiments, the deformable image registration algorithm is configured so that it produces a single-shot mapping. That is, it does not map different points to the same point. Only functions that generate positive Jacobian matrices can ensure this. In this way, the mapping does not produce any folds in the generated surface. In most cases, DIR is driven by minimizing certain error metrics between images based on pixel / voxel values. This is usually achieved using two different image scans, such as a pair of inhalation / expiration images. Since the two images are different, one image can be deformed onto another image. However, in the embodiments disclosed herein, only one image scan is used, and two surface models are used to drive the deformation. Then, the image scan may be deformed due to the mapping between the surfaces.

[0054] Once the mapping is determined, an output dataset can be generated. To this end, the mapping is applied to a reference dataset. In this case, the reference dataset includes volumetric data, so that the output of applying the mapping to the reference dataset is a dataset representing the deformed volume of the ROI of interest. The reference dataset can be the same as that used to determine the first model dataset, or it can be a different reference dataset. In order to ensure that a realistic output dataset is generated, the reference dataset is obtained from a real body and represents an ROI corresponding to the mapped ROI. By associating the model dataset with the real geometry of different stages, realistic deformation can be ensured. For example, if the PDM describing the ROI 100 is mistakenly placed on another part of the body, the shape changes of the lungs 102, 104 and the heart 106 of the ROI 100 will be applied to completely different anatomical regions. This situation can be avoided by associating the PDM with the actual geometry of the corresponding ROI.

[0055] Each output dataset generated using the mapping method can be used in a training dataset to provide a more comprehensive and robust machine learning process for the AI ​​system as well as a more accurate AI system. In the example of an image segmentation algorithm, the algorithm may be able to more accurately segment images of ROIs input into the algorithm. In other embodiments, the geometry generated from the PDM can be used in adaptive schemes that desire to predict future patient geometry, or in any application area where realistic deformation data is required.

[0056] Figure 3 A method 300 is shown for generating data representing a volume of a portion of a body. Similar to method 200, the portion of the body may be a ROI, such as ROI 100, including at least one organ, such as organs 102-106. In some embodiments, the body is a human body.

[0057] At step 302, a PDM for a ROI is generated based on an input data set. The input data set includes data representing one or more surfaces of the ROI, and in some embodiments, is obtained from at least one real body, such as a patient. The input data set may include multiple data sets, each representing the same ROI from a different patient. The PDM defines an output data set based on an average data set and one or more weighted feature vector pairs. In some embodiments, the PDM is generated using principal component analysis and includes a linear combination of the mean and feature vectors of the input data set.

[0058] At step 304, a first model data set is generated based on the PDM. An output data set is generated by modifying at least one weight in the weight feature vector pair so that the output data set is different from the average data set. This can include modifying the weights so that the first model data set generally corresponds to a reference data set. The reference data set represents a corresponding ROI obtained from a real body. In some embodiments, the weights are modified within an interval defined by a standard deviation of the input data set. The weights can be modified randomly or incrementally from the initial weights. One, some, or all of the weights in the linear combination can be modified to produce the first model data set.

[0059] At step 306, weight values ​​corresponding to the first model data set are modified to provide a second model data set that is different from the first model data set. In some embodiments, the weights are modified within an interval defined by the standard deviation of the input data set. The weights may be modified randomly or incrementally from the initial weights. One, some, or all of the weights in the linear combination may be modified to produce the second model data set.

[0060] At step 308, a mapping between the first model data set and the second model data set is generated. In some embodiments, the mapping is generated using a deformable image registration algorithm based on the first model data set and the second model data set. In some embodiments, the deformable image registration algorithm is configured so that it produces an injective mapping. That is, the deformable image registration algorithm only generates positive Jacobian determinants.

[0061] At step 310, an output data set is generated based on the first model data set and the reference data set. Specifically, a mapping determined based on the first model data set is applied to the reference data set. The reference data set represents the volume of the corresponding ROI obtained from the real body. In this way, the output data set includes data representing the deformed volume of the ROI of interest.

[0062] At step 312, the output data set is used as part of a training set for machine learning of an AI system. For example, this could be deep learning for an image segmentation algorithm. Multiple output data sets may be generated to make the training set as large as possible, thereby ensuring that the AI ​​system is robustly trained. This results in a system with improved performance.

[0063] The above method alleviates the problem of training an AI system for a ROI of a human or animal. The disclosed method allows for the generation of multiple geometries that can be used to enhance the training set of an AI system, where currently only a limited number of real patient models are available, and this number is generally insufficient to robustly train the AI ​​system. In addition, in cases where the currently generated data set may not be realistic due to random or unnatural deformations, the disclosed method allows for the generation of synthetic but human-like geometries. By generating such realistic geometries, thereby expanding the training set of the AI ​​system, training can be performed more comprehensively and robustly. In this way, the performance of the trained AI system is improved. In the example of an image segmentation algorithm, the algorithm may be able to more accurately segment images of a ROI input into the algorithm. In other embodiments, the geometries generated from the PDM can be used in adaptive schemes that expect to predict future patient geometries, or in any application area where realistic deformation data is required.

[0064] Figure 4A schematic diagram of a computer readable medium 400 encoded with instructions 410 is shown, which performs the above method when executed on a processor. In this embodiment, the computer readable medium 400 is a data disk. In one embodiment, the data disk can be a magnetic data storage disk. The data disk can be configured to carry instructions 410, which when loaded into a controller such as a processor, performs a method or process according to the above disclosed embodiments. The data disk can be arranged to be connected to or in a reading device 420 and read by the reading device 420 to load the instructions into the controller. One such example of a reading device 420 combined with one (or more) data disks is a hard drive. It should be noted that the computer readable medium can also be other media, such as an optical disk, a digital video disk, a flash memory, or other commonly used storage technologies. In such an embodiment, the data disk can be a type of tangible computer readable medium 400.

[0065] The instructions 410 may also be downloaded to a computer data reading device 430, such as a computer or other device capable of reading computer-encoded data on a computer-readable medium, by including the instructions 410 in a computer-readable signal 440, which is sent to the computer data reading device 430 via a wireless (or wired) interface (e.g., over the Internet) to load the instructions 410 into the controller. In such an embodiment, the computer-readable signal 440 is a type of non-tangible computer-readable medium 400.

[0066] The controller is responsible for the overall operation of the computer data reading device 430 and is preferably implemented by any commercially available CPU ("Central Processing Unit"), DSP ("Digital Signal Processor") or any other electronic programmable logic device. The controller can be implemented using instructions that enable hardware functions, for example, by using executable computer program instructions that can be stored on a computer-readable storage medium (disk, memory, etc.) in a general or special processor to be executed by the processor. The controller can be configured to read instructions from the memory and execute these instructions to control the operation of the computer data reading device 430. The memory can be implemented using any known technology for computer-readable memory such as ROM, RAM, SRAM, DRAM, CMOS, FLASH, DDR, EEPROM memory, flash memory, hard drive, optical memory or any combination thereof.

[0067] According to some embodiments, the computer data reading device 430 may further include a user interface, which may include, for example, a display and a keypad or a touch screen. Other user interface elements known in the art may equally be used for a portion of the user interface.

[0068] References to computer programs, instructions, code, etc. should be understood to include software or firmware for a programmable processor, such as, for example, the programmable content of a hardware device, whether instructions for a processor, or configuration settings for a fixed function device, gate array or programmable logic device, etc.

[0069] Benefiting from the teachings presented in the foregoing description and the related drawings, modifications and other variations of the described embodiments will occur to those skilled in the art. Therefore, it should be understood that the embodiments are not limited to the specific example embodiments described in the present disclosure, and modifications and other variations are intended to be included within the scope of the present disclosure. In addition, although specific terms may be employed herein, they are used only in a general and descriptive sense, and not for the purpose of limitation. Therefore, those skilled in the art will recognize that various variations of the described embodiments will still fall within the scope of the appended claims. As used herein, the terms "include" or "comprise" do not exclude the presence of other elements or steps. In addition, although individual features may be included in different claims, these features may be advantageously combined, and the inclusion of different claims does not mean that the combination of features is not feasible and / or disadvantageous. In addition, singular references do not exclude multiple.

Claims

1. A method for generating data representing a volume of a part of a body, the method comprising: generating a point distribution model PDM based on an input data set, the input data set comprising data representing at least one surface of a part of a body, the PDM defining a surface model data set based on an average data set and one or more weight feature vector pairs, wherein the average data set is a mean of the input data set; generating a first surface model dataset based on the PDM by modifying at least one weight in the one or more weight feature vector pairs, the first surface model dataset corresponding to a second reference dataset, wherein the first surface model dataset is different from the average dataset, the second reference dataset comprising data representing a volume of a portion of a body corresponding to the portion of the body of the input dataset; modifying at least one weight in the one or more weight feature vector pairs to provide a second surface model data set different from the first surface model data set; and determining a mapping between the first surface model data set and the second surface model data set; and The mapping is applied to a first reference data set to generate an output volume data set comprising data representing a volume of a portion of a body corresponding to the portion of the body of the input data set, the output volume data set comprising data representing a deformed volume of the corresponding portion of the body.

2. The method according to claim 1, wherein: The body part includes at least one organ.

3. The method according to claim 1 or 2, wherein: The body is a human body.

4. The method according to claim 1 or 2, wherein: The input data set is derived from at least one real body.

5. The method according to claim 1 or 2, wherein: The input data set includes a plurality of data sets, each data set representing the same part of a corresponding body.

6. The method according to claim 1, wherein: The second reference data set is derived from a real body.

7. The method according to claim 1, wherein: Determining the mapping includes using a deformable image registration algorithm based on the first surface model dataset and the second surface model dataset.

8. The method according to claim 7, wherein: The deformable image registration algorithm produces only positive Jacobians.

9. The method according to claim 1 or 2, wherein: Generating the PDM includes using principal component analysis.

10. The method according to claim 1 or 2, wherein: The one or more weight feature vector pairs include a plurality of weight feature vector pairs.

11. The method according to claim 10, wherein: The plurality of weight feature vector pairs include linear combinations of the feature vectors.

12. The method according to claim 11, wherein: The PDM is defined by the following equation X m =X a +α0e0+α1e1+α2e2+...+α n e n Among them, x m is the surface model dataset, x a is the average data set, α x is the weight of the xth weight feature vector pair, e x is the eigenvector of the xth weighted eigenvector pair, and n is the number of weighted eigenvector pairs.

13. The method according to claim 1 or 2, wherein: Modifying at least one weight in the one or more weight feature vector pairs includes modifying the weight within an interval defined based on a standard deviation of the input data set.

14. The method according to claim 1 or 2, wherein: Modifying at least one weight in the one or more weight-feature vector pairs includes randomly modifying the weight.

15. The method according to claim 1 or 2, wherein: Modifying at least one weight in the one or more weight eigenvector pairs includes incrementally modifying the weight from an initial weight.

16. The method according to claim 1 or 2, wherein: Modifying at least one weight in the one or more weight feature vector pairs includes modifying all weights in the one or more weight feature vector pairs.

17. The method according to claim 1 or 2, wherein: The first reference data set is derived from a real body.

18. The method according to claim 1 or 2, further comprising: The output volume dataset is used as part of a training set for a deep learning process for an artificial intelligence system.

19. A computer readable medium comprising instructions which, when executed on a processor, cause the processor to perform the method according to any one of claims 1 to 18.

Images