Determining image similarity by analyzing registration

Abstract

Disclosed are computer-implemented methods comprising determining whether two medical image acquisitions are of the same patient. In a first aspect, this is done by analyzing the registration of the two images to each other. The registration can be a direct registration between the two images, or an indirect registration (e.g. via an atlas to which each image is registered). In other aspects, a machine learning algorithm is trained based on image registrations to determine whether two medical image acquisitions are of the same patient. The disclosed methods are aimed at enabling medical images taken of the same patient to be grouped together without the need to provide or otherwise process data regarding the identity of the patient.

Description

TECHNICAL FIELD

[0001] The present invention relates to a computer-implemented method, a corresponding computer program, a computer-readable storage medium storing such a program and a computer executing the program, as well as a system comprising an electronic data storage and the above-mentioned computer, wherein the computer-implemented method is for determining whether a plurality of medical image recordings are of the same patient and for training a learning algorithm to determine whether a plurality of medical image recordings are of the same patient. BACKGROUND

[0002] When comparing medical image data sets recorded for the same patient to each other, it is often faced with the problem that the image content of the sets to be compared can not be comparable, e.g. not fully comparable, e.g. due to deviations between the data sets caused by changes in imaging conditions or anatomical structures. It is known to compare image features which are characteristic for an individual patient and which are present in the data sets to be compared to each other in order to determine, e.g., whether these data sets are from the same patient. However, when applying such a comparison method, changes in imaging conditions or anatomical structures can lead to a result that the data sets are from different patients, thus this can lead to a false result, whereas the opposite is true.

[0003] Therefore, it can be an object of the present invention to provide, e.g., a more reliable method of comparing medical images which, e.g., takes into account changes in imaging conditions or anatomical structures between the time points at which the data sets to be compared are generated.

[0004] US 2013 / 0251099 A1 discloses a method of verifying the identity of a patient, wherein an image of an individual who is to receive radiation treatment is obtained and compared to a reference image of a patient for whom radiation treatment is expected. Based on the comparison of the image of the individual to the reference image of the patient, a negative confirmation that the individual is the expected patient can be made on the basis of a threshold for defining a similarity resulting from the comparison.

[0005] US 2018 / 0300540 A1 discloses techniques for identifying individuals in digital images. A digital image or multiple digital images containing a scene of one or more persons are captured. A single digital image can be used as input to a single machine learning model. In some embodiments, a single machine learning model can be trained to perform a non-face feature recognition task and a face-related recognition task. An output can be generated on the single machine learning model based on the input. The output can include first data indicative of a non-face feature of a given person of the one or more persons and second data indicative of a location of a face of the given person relative to the non-face feature in the digital image. In various embodiments, the given person can be identified based at least in part on the output.

[0006] Various aspects, examples, and exemplary steps of the present invention and embodiments thereof are disclosed in the following. Different exemplary features of the present invention can be combined according to the present invention as far as technically advantageous and feasible. SUMMARY

[0007] BRIEF DESCRIPTION OF THE INVENTION

[0008] In the following, a brief description of specific features of the present invention is given, which is not to be understood as limiting the present invention to the features or combinations of features described in this section.

[0009] The method of the present disclosure comprises determining whether two medical image acquisitions are of the same patient. In a first aspect, this is done by analyzing the registration of the two images to each other. The registration can be a direct registration between the two images, or an indirect registration, e.g. via an atlas to which each image is registered. In other aspects, a machine learning algorithm is trained based on image registrations to determine whether two image acquisitions are of the same patient. The disclosed method serves to be able to group medical images taken of the same patient together without the need to provide or otherwise process data about the identity of the patient. SUMMARY

[0011] In this section, a description of general features of the present invention is given, e.g. by referring to feasible embodiments of the present invention.

[0012] Generally, the present invention achieves the above-mentioned objects by providing, in a first aspect, a computer-implemented method (e.g. a medical method) for determining whether a plurality (e.g. two) or exactly two medical image acquisitions are of the same patient. The method according to the first aspect comprises, on at least one processor of at least one computer, executing the following exemplary steps performed by the at least one processor.

[0013] In a (e.g. first) exemplary step of the method according to the first aspect, first patient image data is acquired, the first patient image data describing a first medical image of a first anatomical body part of a first patient.

[0014] In a (e.g. second) exemplary step of the method according to the first aspect, second patient image data is acquired, the second patient image data describing a second medical image of a second anatomical body part of a second patient, the second anatomical body part corresponding to the first anatomical body part. For example, the first anatomical body part and the second anatomical body part are identical in terms of their anatomical function. Alternatively, the first anatomical body part and the second anatomical body part are different adjacent parts of a larger anatomical structure comprising the first anatomical body part and the second anatomical body part, for example.

[0015] The first and second patient image data are generated by applying an imaging modality, such as a tomographic imaging modality, e.g. magnetic resonance imaging or computed tomography or ultrasonic tomography, or a non-tomographic imaging modality, like radiography, to the first and second anatomical body part, respectively. Thus, depending on the applied imaging modality, the first and second patient image data can be three-dimensional or two-dimensional image data. The first and second patient image data have been generated using the same imaging modality or different imaging modalities.

[0016] In an (e.g. third) exemplary step of the method according to the first aspect, registration data are determined based on the first and second patient image data, wherein the registration data describe a registration between the first and second medical images. In the context of the present disclosure, a registration is a positional relationship between images, e.g. between image components. The registration is embodied, e.g., by a transformation, in particular an elastic transformation or elastic deformation, such as a transformation matrix. For example, the registration defines a positional transformation between the positional reference systems for defining positions in the images and / or between positions of certain image components in the images. In an example, the registration is a direct registration between the first and second medical images. In another example, the registration is an indirect registration, which can be defined as a concatenation of multiple, e.g. two, or exactly two, registrations. For example, both the first and second medical images are registered into an atlas defined by atlas data (as described below with reference to the examples of the method according to the first aspect), and then the registration between the first and second medical images is defined as a concatenation, e.g. multiplication, addition or convolution, of the registration between the first medical image and the atlas and the inverse of the registration between the second medical image and the atlas.

[0017] In an (e.g. fourth) exemplary step of the method according to the first aspect, patient image comparison data are determined based on the registration data, wherein the patient image comparison data describe a measure of similarity between the first and second medical images. For example, the measure of similarity is defined from or derived from the registration, or, if the registration is defined as a concatenation of two or more registrations, the measure of similarity is defined as, e.g., a similarity of the registration.

[0018] In an (e.g. fifth) exemplary step, similarity threshold data are acquired, which describe a threshold value for the measure of similarity. The threshold value can be defined as a single numerical value or as multiple numerical values, e.g. as multiple numerical values in a matrix of the same dimensionality, if the measure of similarity is defined as a matrix.

[0019] In an (e.g. sixth) exemplary step of the method according to the first aspect, the patient similarity data is determined based on the patient image comparison data, wherein the patient similarity data describes that the first medical image and the second medical image were taken of the same patient if the measure of similarity reaches a predetermined relation with respect to (e.g. exceeds or falls below) a threshold value.

[0020] Using the registration as a means for comparing the first medical image and the second medical image is related to the technical effect of producing a correct comparison result even in case of deviations in the imaging conditions in one aspect, in case of using registrations for different imaging conditions and / or different imaging modalities, or in another aspect in case of anatomical deviations of any anatomical state present in the images using a model (e.g. a biomechanical model described below) describing possible anatomical changes in the individual patient between the images to be compared over time. For example, the registration is determined for the anatomical deviation that is most similar to the anatomical deviation that can be reproduced using the model.

[0021] In an (e.g. seventh) exemplary step, atlas data is acquired, the atlas data describing an image-based model of the first and second anatomical body part. For example, determining the registration data then comprises determining first registration data based on the first patient image data and the atlas data, the first registration data describing a first registration between the first medical image and the image-based model, and determining second registration data based on the second patient image data and the atlas data, wherein the second registration data describes a second registration between the second medical image and the image-based model. Determining the patient image comparison data then comprises determining registration comparison data based on the first registration data and the second registration data, wherein the registration comparison data describes a measure of similarity between the first registration and the second registration. The patient image comparison data is then further determined based on the registration comparison data. In this example, the registration between the first medical image and the second medical image is defined as or can be considered or interpreted as a concatenation of the inverse of the first registration and the second registration.

[0022] In an example of the method according to the first aspect, determining the registration comparison data comprises acquiring variability model data describing a variability model of the first and second anatomical body part. For example, the variability model is a finite element model or a coupled spring model of the first and second anatomical body part. The variability describes the deformability of the first and second anatomical body part in the example. For example, the variability is defined as the degrees of freedom to change the boundary conditions of the finite element model or coupled spring model, for example by moving nodes, altering or adding or removing mass points, or altering or adding or removing forces, respectively. For example, the variability model contains statistical information about a range of values within which the degrees of freedom can change due to changes of the anatomy over time, for example due to body motion or natural growth, atrophy, wear and tear, disease, swelling, or physiological cycles, depending on the influence of the state of hydration. By comparing the difference between the first and second image to the degrees of freedom of the variability model, it can be inferred whether the model can reproduce the respective anatomical situation, and thus whether the difference between the images is due to anatomical factors or external factors, for example due to respiratory motion or due to tumor resection. For example, determining the registration comparison data then comprises determining registration difference data describing a difference between the first registration and the second registration. For example, the difference is determined by determining the inverse of the second registration and concatenating it with the first registration. For example, the first registration and the second registration are both defined as displacement vector fields, and the concatenation results in a displacement vector field whose difference to the variability transformation is determined. In a particular example, the displacement vector field does not include displacement vectors longer than a predetermined value. The variability transformation describes a configuration of the variability model, for example an anatomical or physiological change. For example, the configuration is defined by a space, for example a geometry, of the variability model.

[0023] In an example of the method according to the first aspect, determining the registration comparison data comprises determining registration difference data describing a difference between the first registration and the second registration. For example, the difference is determined by determining the inverse of the second registration and concatenating it with the first registration, wherein the first registration and the second registration are both defined as displacement vector fields. For example, the concatenation results in a displacement vector field which does not include displacement vectors longer than a predetermined value.

[0024] In an example of the method according to the first aspect, determining the registration comparison data comprises determining registration difference data describing a difference between the first registration and the second registration. For example, the first registration and the second registration are both defined as displacement vector fields. For example, the difference between the first registration and the second registration is determined by determining the inverse of the second registration and concatenating it with the first registration, and subtracting a rigid transformation from the concatenation. For example, the rigid transformation describes a rotation and / or a translation of at least a portion of the first medical image relative to at least a portion of the second medical image. For example, the rigid transformation describes essentially only a rotation and a translation without elastic deformation.

[0025] In an example of the method according to the first aspect, the first anatomical body part is an anatomically invariant or an anatomically variant anatomical body part. Within the framework of the present disclosure, an anatomically variant part of an anatomical body part is understood to be a part that can change its shape within the patient’s body (like soft tissue). Within the framework of the present disclosure, an anatomically invariant part of an anatomical body part is understood to be a part whose shape is substantially invariant, i.e. a rigid part (like bone tissue or cartilage). However, both anatomically invariant parts and anatomically variant parts can change their position, e.g. relative to other parts of the patient’s body (e.g. by moving an arm or due to respiratory motion).

[0026] In a second aspect, the present invention relates to a computer-implemented method (e.g. a medical method) of training a learning algorithm to determine whether a plurality (e.g. two) or exactly two medical image acquisitions are of the same patient. The method according to the second aspect comprises the following exemplary steps performed by at least one processor of at least one computer (e.g. the at least one computer is part of a cloud-based data processing system) performing the following exemplary steps.

[0027] In an (e.g. first) exemplary step of the method according to the second aspect, training patient image data is acquired, the training patient image data describing a plurality of medical patient training images showing an anatomical body part. The training patient image data is generated by applying an imaging modality, such as a tomographic modality (e.g. magnetic resonance imaging or computed tomography or ultrasonic tomography) or a non-tomographic modality (like radiography) to the anatomical body part. Thus, depending on the applied imaging modality, the training patient image data can be three-dimensional or two-dimensional image data. For example, a plurality (but not necessarily all) of the medical patient training images are of the same patient.

[0028] In an (e.g. second) exemplary step of the method according to the second aspect, patient identifier data is acquired, the patient identifier data describing a patient identifier for each medical patient training image, the patient identifier identifying the patient of which the respective medical patient image is of.

[0029] In an (e.g. third) exemplary step of the method according to the second aspect, training registration data is determined based on the training patient image data, wherein the training registration data describes registrations between pairs of images from the plurality of medical patient training images, i.e. each registration is established between two or exactly two medical patient training images. For example, the registrations are established directly between the plurality of medical patient training images.

[0030] In an (e.g. fourth) exemplary step according to the method according to the second aspect, patient association data is determined for the images belonging to the plurality of medical training images (e.g. for each pair of medical training images), wherein the patient association data is determined based on the training patient image data and the patient identifier data and comprises patient association information describing whether the image acquisitions belonging to each pair of medical images are of the same patient or not. The patient association information can be human readable and / or machine readable. In essence, the association of each training image with a specific patient as defined by the patient identifier data is used as an indicator for later training of a learning algorithm to determine whether, for example, two image acquisitions are of the same patient or not.

[0031] In an (e.g. fifth) exemplary step according to the method according to the second aspect, image similarity data is determined, the image similarity data describing model parameters of a learning algorithm establishing a relationship between the registrations and the patient association information, wherein the image similarity data is determined by inputting the training registration data and the patient association data into an establishing relationship function. Thereby, the learning algorithm is trained to learn the above described relationship (e.g. the registration and the association between images taken of the same patient (or different patients)).

[0032] In an example of the method according to the second aspect, atlas data is acquired, the atlas data describing an image-based model of the first and second anatomical body parts. Further, first registration data is determined based on the first image of each pair of medical patient training images, the first registration data describing a first registration between the first image and the image-based model, and second registration data is determined based on the second image of each pair of medical patient training images, the second registration data describing a second registration between the second image and the image-based model. Then, the learning algorithm is trained to determine a statement indicating whether the first image and the second image are taken of the same patient based on the first registration data and the second registration data, wherein the statement can be associated with a certain confidence level. In this example, the registration between the first image and the second image of each pair is defined as a concatenation of the inverse of the first registration and the second registration.

[0033] In an example of the method according to the second aspect, the anatomical body parts are anatomically invariant or anatomically variant, and registration analysis data is determined for each of the first registration and the second registration based on the registration data, wherein the registration analysis data describes statistical values for characterizing a set of registration vectors for at least one anatomically invariant or anatomically variant part of the first and second anatomical body parts. Then, the learning algorithm is trained to determine a statement indicating whether the first image and the second image are taken of the same patient based on the first registration data and the second registration data, wherein the statement can be associated with a certain confidence level.

[0034] In an example of the method according to the second aspect, atlas data is acquired, the atlas data describing an image-based model of the first and second anatomical body part. Further, first registration data is determined based on the first image of each pair of medical patient training images, the first registration data describing a first registration between the first image and the image-based model, and second registration data is determined based on the second image of each pair of medical patient training images, the second registration data describing a second registration between the second image and the image-based model. Then, a learning algorithm is trained to determine, based on the first registration data and the second registration data, a statement indicating whether the first image and the second image were taken of the same patient, wherein the statement can be associated with a certain level of confidence. In a variation of this example, the learning algorithm is trained such that it determines that the first image and the second image were taken of the same patient if the product results in at least substantially one.

[0035] In a third aspect, the present invention relates to a computer-implemented method (e.g. a medical method) determining whether a plurality (e.g. two) or exactly two medical image acquisitions were taken of the same patient. The method according to the second aspect comprises the following exemplary steps performed by at least one processor of at least one computer (e.g. the at least one computer is part of a cloud-based data processing system).

[0036] In an (e.g. first) exemplary step of the method according to the third aspect, specific registration data is acquired, the specific registration data describing specific registrations between a plurality of medical images of an anatomical body part. Each registration is defined between two or exactly two medical images.

[0037] In an (e.g. second) exemplary step of the method according to the third aspect, specific image similarity data is determined describing a relationship between the patient association data and the specific registration data, the patient association data describing whether images belonging to the plurality of medical images (e.g. each pair of images between which a registration is defined) were taken of the same patient. The specific image similarity data is determined by inputting the specific registration data into a function establishing the relationship, the function being part of a learning algorithm trained by performing the method according to the second aspect. This allows grouping of images into sets of images taken for the same patient according to the determined similarity between the images. Using the registrations as a means for correctly grouping medical images is associated with the technical effect of resulting in a correct grouping result, even in case of deviations in the anatomy or imaging conditions between the images to be grouped together.

[0038] For example, the learning algorithm used in the methods according to the second and third aspects comprises or consists of a machine learning algorithm. For example, the learning algorithm comprises or consists of a convolutional neural network. For example, the model parameters define learnable parameters (e.g. weights) of the learning algorithm used in the methods according to the second and third aspects.

[0039] In an example of the methods according to the first, second and third aspects, the learning algorithm can be a random forest algorithm. With reference to Antonio Criminisi, Jamie Shotton, E. Konukoglu, Decision Forests: A Unified Framework for Classification, Regression, Density Estimation, Manifold

[0040] A random forest is an ensemble learning method for classification or regression that works by building a large number of decision trees at training time and outputting the class of the mode of the classification (classification) or the average prediction of the individual trees (regression).

[0041] The basic building block of a random forest is a single decision tree. A decision tree is a set of questions organized in a hierarchical manner and represented graphically as a tree. A decision tree estimates an unknown attribute (a “label”) of an object by asking successive questions about known attributes (so-called “features”) of the object. What question to ask next depends on the answer to the previous question, and this relationship is represented graphically as a path through the tree that an object follows. Then, a decision is made based on the terminal node (so-called “leaf node”) on the path. Each question corresponds to an internal node (so-called “split node”) of the tree.

[0042] Each split node has a so-called test function associated with it. The test function at split node j is denoted by a function with binary output:

[0043]

[0044] where 0 and 1 can be interpreted as “false” and “true”, respectively, denotes a parameter of the test function at the j-th split node.

[0045] v is thus a current object (“data point”) represented by a vector where the components x i represent certain attributes (features) of the data point, all components x i form a feature space

[0046] In the simplest form, the test function is a linear model that selects one feature axis in the feature space and classifies each data point according to whether the value of the respective feature is below or above a learnable threshold. Other more complex non-linear test functions are also possible.

[0047] To train a decision tree, a set of training data points is used, whose features as well as the desired labels are known. The purpose of the training is to automatically learn suitable test functions at all split nodes that are best suited to determine the label from the features of the data points. Later, such a trained decision tree can be evaluated for new data points with unknown labels by sending the data points through the tree trained based on their features.

[0048] To understand the training process, it is beneficial to represent subsets of the training points as associated with different tree branches. For example, S1 represents the subset of training points that reach node 1 (nodes are numbered in breadth-first order from the root F starting with 0), while S2 represents the subset of training points that reach node 2. S1L and S1R represent the subsets of training points that go to the left and right child nodes of node 1, respectively.

[0049] Training is responsible for selecting the type and parameters of the test function h(v, θj) associated with each split node (indexed by j) by optimizing a selected objective function defined on the available training set. j ) of the test function h(v, θj) associated with each split node (indexed by j) by optimizing a selected objective function defined on the available training set.

[0050] The optimization of the split function is done in a greedy manner. At each node j, a function that “best” splits Sjinto SjL and SjR is learned from the subset of the training set Sj. j j This problem is formulated as the maximization of the objective function at this node:

[0051]

[0052] where

[0053]

[0054]

[0055]

[0056] As mentioned before, the notation SjL and SjR is used to represent the subsets of training points that go to the left and right child nodes of node j, respectively. j ​​​​The set of training points before and after the split is denoted. The objective function is in abstract form here. The precise definition of the objective function and the meaning of "optimal" depend on the task at hand (e.g. whether there is supervision, continuous or discrete output). For example, for binary classification, the term "optimal" can be defined as the split of the training subset S j such that the resulting child nodes are as pure as possible, i.e. contain only training points of a single class. In this case, the objective function can be defined as information gain, for example.

[0057] During training, the structure (shape) of the tree also needs to be optimized. Training starts at the root node (j = 0), where the optimal split parameters are found, as described before. Consequently, two child nodes are constructed, each receiving a different disjoint subset of the training set. This process is then applied to all newly constructed nodes, and the training phase continues. The structure of the tree depends on how and when it is decided to stop growing individual branches of the tree. Different stopping criteria can be applied. For example, it is common to stop the tree when a maximum number of levels D is reached. Alternatively, a maximum value of the minimum value of the information gain can be imposed, in other words, the tree is stopped when the sought attributes of the training points within a leaf node are similar to each other. The growth of the tree can also be stopped when the node contains too few training points. In terms of generalization, it has been shown that it is beneficial to avoid growing the full tree.

[0058] During training, randomness is injected into the tree: when training at the j-th node, instead of optimizing over the entire parameter space of the test function, only a small random subset of parameter values is available. Thus, under the randomness model, the tree is trained by optimizing each split node j as follows:

[0059]

[0060] This random setup results in that multiple decision trees can be trained in parallel afterwards, each exploiting a different set of attributes from the data points.

[0061] At the end of the training phase, one obtains: (i) the (greedy) optimal weak learner associated with each node; (ii) the learned tree structure; and (iii) the different set of training points at each leaf.

[0062] After training, each leaf node remains associated with a subset of the (labeled) training data. During testing, a previously unseen point is threaded through the tree until it reaches a leaf. Since the split nodes act on features, it is possible that the input test point ends up in a leaf associated with training points that are similar to itself. It is therefore reasonable to assume that the relevant label must also be similar to the labels of the training points in that leaf. This justifies the use of the label statistics collected in that leaf to predict the label associated with the input test point.

[0063] In the most general sense, leaf statistics can be collected using the posterior distribution:

[0064] p(c|v) and p(y|v),

[0065] where c and y represent discrete or continuous labels, respectively. v is the data point being tested in the tree, and the conditional notation indicates that the distribution depends on the fact that the particular leaf node was reached by the test point. Different leaf predictors can be used. For example, in the discrete case, the maximum a posteriori (MAP) estimate can be obtained as c* = argmax c p(c|).

[0066] Based on the above principles of construction of a decision tree, one can now proceed to a decision forest, also known as a random forest.

[0067] A random decision forest is a collection of randomly trained decision trees. A key aspect of the forest model is the fact that the constituent trees of the forest model all differ randomly from each other. This leads to decorrelation between the individual tree predictions, which in turn leads to improved universality and robustness.

[0068] In a forest with T trees, the variable is used to index each constituent tree. All trees are trained independently (and possibly in parallel). During testing, each test point v is simultaneously pushed through all trees (starting at the root) until it reaches the respective leaf. Tree testing can also be done in parallel, thus achieving high computational efficiency on modern parallel CPU or GPU hardware. All tree predictions can be combined into a single forest prediction by a simple averaging operation. For example, in classification:

[0069]

[0070] where p t (|) denotes the posterior distribution obtained by the t-th tree. Alternatively, the tree outputs can also be multiplied (although the trees are not statistically independent):

[0071]

[0072] The partition function Z is used to ensure probability normalization.

[0073] In an example of the method according to the first, second and third aspect, the learning algorithm can be a convolutional neural network. In the following, reference is made to Figure 1 An explanation of a convolutional neural network as an example of a machine learning algorithm to be used with the disclosed invention is provided below.

[0074] Convolutional networks (also known as convolutional neural networks, or CNNs) are an example of neural networks used to process data with a known grid-like topology. Examples include time series data (which can be viewed as a one-dimensional grid sampled at regular time intervals) and image data (which can be viewed as a two- or three-dimensional grid of pixels). The name "convolutional neural network" indicates that the network employs convolutional mathematics. Convolution is a linear operation. Convolutional networks are simple neural networks that use convolution in place of general matrix multiplication in at least one of their layers. There are multiple variants of the convolution function that are widely used in practice for neural networks. Generally, the operations used in convolutional neural networks do not exactly correspond to the definition of convolution used in other fields, such as engineering or pure mathematics.

[0075] The primary building block of a convolutional neural network is an artificial neuron. Figure 1 is an example of a single neuron being depicted. The node in the middle represents the neuron, which takes all inputs (x1,..., x n ) and multiplies each input by its specific weight (w1,..., w n ). The importance of an input depends on its weight value. These computed values are summed up, called the weighted sum, which will be inserted into an activation function. The weighted sum z is defined as:

[0076]

[0077] The bias b is a value independent of the inputs, which corrects the threshold's boundary. The resulting value is processed by the activation function, which decides whether to pass the input to the next neuron or not.

[0078] CNNs typically take 1st or 3rd order tensors as their input, for example, an image with H rows, W columns, and 1 or 3 channels (R, G, B color channels). However, CNNs can process higher order tensor inputs in a similar fashion. The input then sequentially goes through a series of processes. One process step is commonly referred to as a layer, which can be a convolutional layer, a pooling layer, a normalization layer, a fully connected layer, a loss layer, etc. Detailed information of these layers is described in the following sections.

[0079]

[0080] Equation 5 above illustrates how a CNN runs layer by layer in forward propagation. The input is x 1 , which is typically an image (1st or 3rd order tensor). The parameters involved in the first layer process are collectively referred to as the tensor w i . The output of the first layer is x 2 , which also serves as the input for the second layer process. This process continues until all layers in the CNN are processed, which outputs x LHowever, an additional layer is added for backpropagation, which is a method for learning good parameter values in CNN. Assume that the current problem is an image classification problem of class C. A common strategy is to output x L as a C-dimensional vector, whose i-th entry encodes the prediction (x 1 is the posterior probability from the i-th class). To make x L a probability mass function, the processing in the (L - 1)-th layer can be set to the softmax transformation of x L-1 . In other applications, the output x L can have other forms and interpretations. The last layer is the loss layer. Assume that t is the corresponding target value (truth) of the input x 1 , then a cost or loss function can be used to measure the difference between the CNN prediction x L and the target t. It should be noted that some layers may have no parameters, that is, for some i, w i may be empty.

[0081] In the example of CNN, the rectified linear unit (ReLu) is used as the activation function of the convolutional layer, while the softmax activation function provides information to give a classification output. The following sections will illustrate the purpose of the most important layers.

[0082] The input image is fed into the feature learning part of the layer including convolution and ReLu, followed by the layer including pooling, and then the further paired repetition of the layer of convolution and ReLu and the layer of pooling. The output of the feature learning part is fed into the classification part, which includes the layers for flattening, fully connecting, and max softening.

[0083] In the convolutional layer, multiple convolutional kernels are usually used. Assume that D convolutional kernels are used, and the spatial span of each convolutional kernel is HxW, then all convolutional kernels are represented as f. f is a 4th-order tensor in l <D l and 0 ≤ d < D are used to determine the specific elements in the convolutional kernel. It should also be noted that the set of convolutional kernels f refers to the same object as the symbol w L above. To simplify the derivation process, the symbols are slightly changed. It is also clear that even if the mini-batch strategy is used, the convolutional kernels remain unchanged.

[0084] As long as the convolutional kernel is larger than 1×1, the spatial range of the output is smaller than that of the input. Sometimes it is necessary for the input image and the output image to have the same height and width, and a simple padding technique can be used.

[0085] For each input channel, if padding (i.e., insertion) is added above the first line of input. Line, fill below the last line. Line, and fill Move the column to the left of the first column and fill. If the column is moved to the right of the last column, the size of the convolution output will be H. l ×W l ×D, which means having the same spatial range as the input. It is a floor function. Although the elements of the rows and columns being filled are usually set to 0, other values ​​are also possible.

[0086] Stride is another important concept in convolution. At each possible spatial location, the kernel is convolved with the input, which corresponds to a stride of s = 1. However, if s > 1, each slide of the kernel skips s-1 pixel locations (i.e., convolution is performed once every s pixels in both the horizontal and vertical directions).

[0087] In this section, we consider the simple case of a step size of 1 and no padding. Therefore, in There is y (or x) l+1 ), where H l+1 = l -+1,W l+1 =W l -+1, and D l+1 =. In precise mathematics, the convolution process can be represented as an equation:

[0088]

[0089] For all 0 ≤ d ≤ D = D l+1 and satisfying 0≤i l+1 <H l –H+1=H l+1 ,0≤j l+1 <W l –W+1=W l+1 Any spatial location ( l+1 (l+1) Repeat equation 2. In this equation, It refers to the triple (i l+1 +i,j l+1 +j,d l ) index x l Element. Usually for Add bias term b d For clarity, this item has been omitted here.

[0090] Pooling functions replace the network output at a given location with a generalized statistic of the outputs in the vicinity. For example, max pooling reports the maximum output within a rectangular neighborhood of the table. Other popular pooling functions include averaging over a rectangular neighborhood, using the L2 norm of a rectangular neighborhood, or a weighted average based on the distance to the center pixel. In all cases, pooling helps to make the representation approximately invariant to small translations of the input. Invariance to translation means that if the input is translated slightly, the value of the pooled output remains unchanged.

[0091] Because pooling generalizes the response over the entire neighborhood, fewer pooling units can be used than detector units by reporting generalized statistics of pooling regions spaced k pixels apart instead of one pixel. This improves the computational efficiency of the network, as the next layer processes approximately k times less input.

[0092] Assuming all parameters w of the CNN model have been learned 1 ,…,w L-1 Then, we are ready to use the model for prediction. Prediction only involves forward running of the CNN model, that is, running in the direction of the arrow in Equation 1. Let's take an image classification problem as an example. From the input x... 1 To begin, make it pass through the first layer (with parameter w). 1 The processing of the bounding box, and obtaining x 2 Furthermore, x 2 Passed to the second layer, and so on. Finally, received. Its estimate x 1 The posterior probability belonging to class C. The CNN prediction can be output as:

[0093]

[0094] Now, the question is: how do we learn the model parameters?

[0095] As with many other learning systems, the parameters of a CNN model are optimized to minimize the loss z, i.e., to ensure that the CNN model's predictions match the truth labels. Suppose we are given a training example x. 1 To train these parameters, the training process involves running the CNN network in two directions. First, the network is run in the forward propagation to obtain x. L This is done to achieve prediction using the current CNN parameters. Instead of outputting the prediction, the prediction needs to be compared with the prediction corresponding to x. 1 The objective t is compared, meaning the forward propagation continues until the last loss layer. Finally, the loss z is obtained. Loss z serves as a supervision signal, guiding how the model's parameters should be corrected (updated).

[0096] There are several algorithms for optimizing the loss function, and CNNs are not limited to a particular algorithm. An example algorithm is called stochastic gradient descent (SGD). This means that the parameters are updated by using the gradient estimated from a (typically) small subset of the training examples.

[0097]

[0098] In equation 4, the notation "←" implicitly means that the parameter w i is updated from time t to t+1. If the time index t is used explicitly, then this equation would be written as:

[0099]

[0100] In equation 4, the partial derivative measures the rate of change of z with respect to the change of w i in different dimensions. This vector of partial derivatives is called the gradient in mathematical optimization. Thus, in a small local region around the current value of w i , moving w i in the direction determined by the gradient will increase the target value z. To minimize the loss function, w i should be updated in the opposite direction of the gradient. This update rule is called gradient descent.

[0101] However, if one moves too far in the direction of the negative gradient, the loss function can increase. Therefore, at each update, only a small fraction of the negative gradient (controlled by η (the learning rate)) is changed in the parameters. Typically, η > 0 is set to a small number (e.g., η = 0.001). If the learning rate is not too high, one update based on x 1 will make the loss for this particular training example smaller. However, it is very likely that it will make the loss for some other training examples larger. Therefore, it is necessary to use all training examples to update the parameters. When all training examples have been used to update the parameters, it is said that one epoch has been processed. Typically, one epoch will reduce the average loss over the training set until the learning system fits the training data. Therefore, the gradient descent update epochs can be repeated and terminated at some point to obtain the CNN parameters (e.g., one can terminate when the average loss on a validation set increases).

[0102] The partial derivative of the last layer is easy to compute. Since x L is directly connected to z under the control of the parameter w L , it is easy to compute This step needs to be performed only when w L is not empty. Based on the same spirit, it is also easy to compute For example, if the squared loss L2 is used, one gets and

[0103] In fact, for each layer, two sets of gradients are computed: the partial derivative of z with respect to the layer parameters w i , and the input x i to the layer. As shown in Equation 4, the term can be used to update the parameters of the current (i-th) layer. The term can be used to update the parameters in the reverse direction, e.g., to the (i-1)-th layer. The intuitive explanation is that x i is the output of the (i-1)-th layer, while indicates how x i should be changed to reduce the loss function. Thus, one can think of as part of the "error" supervisory information that is propagated backwards from z layer by layer to the current layer. Thus, one can continue the backpropagation process and use to propagate the error backwards to the (i-1)-th layer. This layer-by-layer back update process makes learning a CNN easier.

[0104] Take the i-th layer as an example. When updating the i-th layer, the backpropagation process for the (i+1)-th layer must have already been completed. That is, the terms and and have both been computed and stored in memory and are ready for use. The task at this point is to compute and Using the chain rule, one obtains:

[0105]

[0106]

[0107] Since has already been computed and stored in memory, one only needs matrix reshaping operations (vec) and an additional transpose operation to obtain which is the first term in the right-hand side (RHS) of both equations. As long as one can compute and one can easily obtain the desired values (the left-hand side of both equations).

[0108] and are much easier to compute than and directly, because x i is directly related to x i through a function with parameters w i+1 .

[0109] In the context of neural networks, activation functions act as a transfer function between the input and output of a neuron. The activation function defines under which conditions an activation node, i.e. maps an input value to an output, which in turn is used as one of the inputs for subsequent neurons in a hidden layer. There are a large number of different activation functions with different properties.

[0110] Loss functions quantify the modeling effect of an algorithm on a given data. In order to learn from the data and change the weights of the network, the loss function has to be minimized. Generally, a distinction can be made between regression loss and classification loss. Classification predicts the output according to a finite set of classification values (classification labels), on the other hand, regression deals with the prediction of continuous values.

[0111] In the following mathematical formula, the following parameters are defined as:

[0112] • n is the number of training examples

[0113] • i is the i-th training example in the dataset

[0114] • y i is the true label of the i-th training example

[0115] is the prediction of the i-th training example

[0116] The most prevalent setting for classification problems is the cross-entropy loss. The cross-entropy loss increases with the deviation of the predicted probabilities from the actual labels. The logarithm of the actual predicted probabilities is multiplied with the true class. The important aspect is that the cross-entropy loss heavily penalizes confident but wrong predictions. The mathematical formula can be described as:

[0117]

[0118] A typical example of a regression loss is the mean squared error or L2 loss. As the name suggests, the mean squared error is the average of the squared differences between the predicted values and the actual observed values. The mean squared error only involves the magnitude of the average error, not their direction. However, due to the squaring, predictions far from the actual values suffer a heavy penalty compared to predictions with a small deviation. Additionally, the MSE has good mathematical properties that make it easier to compute the gradient. The formula for the MSE is as follows:

[0119]

[0120] The following document contains information on the functioning of convolutional neural networks:

[0121] The chapter on convolutional networks in Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning, 2016, see http: / / www.deeplearningbook.org.

[0122] Wu, J. (2017). Introduction to convolutional neural networks. Retrieved from https: / / pdfs.semanticscholar.org / 450c / a19932fcef1ca6d0442cbf52fec38fb9d1e5.pdf

[0123] Krizhevsky, A., Sutskever, I., & Hinton, G. (2009). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105).

[0124] Wu, J. (2017). Introduction to convolutional neural networks. Retrieved from https: / / pdfs.semanticscholar.org / 450c / a19932fcef1ca6d0442cbf52fec38fb9d1e5.pdf

[0125] Wu, J. (2017). Introduction to convolutional neural networks. Retrieved from https: / / pdfs.semanticscholar.org / 450c / a19932fcef1ca6d0442cbf52fec38fb9d1e5.pdf

[0126] Wu, J. (2017). Introduction to convolutional neural networks. Retrieved from https: / / pdfs.semanticscholar.org / 450c / a19932fcef1ca6d0442cbf52fec38fb9d1e5.pdf

[0127] Fully convolutional networks for semantic segmentation by Jonathan Long, Evan Shelhamer, and Trevor Darrell, https: / / www.cv-foundation.org / openaccess / content_cvpr_2015 / papers / Long_Fully_Convolutional_Networks_2015_CVPR_paper.pdf.

[0128] In a fourth aspect, the present application relates to a program which, when running on a computer or when loaded onto a computer, causes the computer to perform the method steps of the method according to the first to third aspects; and / or a (e.g. non-transitory) program storage medium having stored thereon a program or a program storage medium having stored thereon data defining model parameters and architecture of a learning algorithm trained by performing the method according to the first aspect; and / or a data carrier signal carrying the above-mentioned program and / or a data carrier signal carrying data defining model parameters and architecture of a learning algorithm trained by performing the method according to the first aspect; and / or a data stream carrying the above-mentioned program and / or a data stream carrying data defining model parameters and architecture of a learning algorithm trained by performing the method according to the first aspect; and / or at least one computer comprising at least one processor and a memory, wherein the above-mentioned program is running on the at least one processor or is loaded into the memory of said computer.

[0129] Alternatively or additionally, the present application can relate to (physical, e.g. technically generated, e.g. electrical) signal waves, e.g. digital signal waves, like electromagnetic carrier waves, carrying information representing a program, e.g. the above-mentioned program, e.g. comprising code means adapted to perform any or all steps of the method according to the first aspect. In an example, the signal waves are data carrier signals carrying the above-mentioned computer program. A computer program stored on a disk is a data file which, when read and transferred, becomes a data stream in the form of, e.g., (physical, e.g. technically generated, e.g. electrical) signal waves. The signal can be implemented as signal waves, e.g. electromagnetic carrier waves, as described herein. For example, the signal, e.g. signal waves, are structured to be transmitted via computer networks and mobile networks, e.g. the Internet, e.g. local area networks (LAN), wireless local area networks (WLAN) and wide area networks (WAN). For example, the signal, e.g. signal waves, are structured to be transmitted by optical or acoustic data transmission. Thus, alternatively or additionally, the present application according to the second aspect can relate to a data stream representing the above-mentioned program, i.e. comprising the program.

[0130] In a fifth aspect, the present application relates to a system for determining whether a plurality, for example two, or exactly two medical image acquisitions are of the same patient, the system comprising:

[0131] a) at least one computer according to the fourth aspect;

[0132] b) at least one electronic data storage device for storing first patient image data and second patient image data; and

[0133] c) a program storage medium according to the fourth aspect,

[0134] wherein the at least one computer is operatively coupled to the at least one electronic data storage device to: retrieve the first patient image data and the second patient image data from the at least one electronic data storage device; and store at least patient similarity data in the at least one electronic data storage device.

[0135] In a sixth aspect, the present application relates to a system for determining whether a plurality, for example two, or exactly two medical image acquisitions are of the same patient, the system comprising:

[0136] a) at least one computer according to the fourth aspect;

[0137] b) at least one electronic data storage device for storing registration data; and

[0138] c) a program storage medium according to the fourth aspect,

[0139] wherein the at least one computer is operatively coupled to:

[0140] - the at least one electronic data storage device to: retrieve the specific registration data from the at least one electronic data storage device, and store at least specific image similarity data in the at least one electronic data storage device; and

[0141] - the program storage medium to retrieve data defining the architecture and model parameters of the learning algorithm from the program storage medium.

[0142] For example, the present application does not relate to, in particular does not comprise or involve, an invasive step, which represents a substantial physical intervention into the body, which requires professional medical measures and which can still entail significant health risks to the body even when the required professional care and measures are taken while performing the professional medical measures.

[0143] Definitions

[0144] In this part, definitions of specific terms used in the present disclosure are provided, which also form part of the present disclosure.

[0145] The method according to the present application is for example a computer- implemented method. For example, all steps or only some steps (i.e. less than the total number of steps) of the method according to the present application can be performed by a computer (e.g. at least one computer). Embodiments of the computer-implemented method use a computer to perform a data processing method. Embodiments of the computer-implemented method are methods involving the operation of a computer such that the computer is operated to perform one, more or all steps of the method.

[0146] A computer for example comprises at least one processor and for example at least one memory in order to (technically) process data, for example electronically and / or optically. The processor is for example made of a substance or composition that is a semiconductor, for example at least partially n-doped and / or p-doped semiconductor, for example at least one of a II, III, IV, V, VI type semiconductor material, for example (doped) silicon and / or gallium arsenide. The described computing steps or determining steps are for example performed by a computer. The determining steps or computing steps are for example steps of determining data within the framework of a technical method, for example within the framework of a program. The computer is for example any type of data processing device, for example an electronic data processing device. The computer can be a device that is generally regarded as a computer, for example a desktop personal computer, a laptop computer, a netbook, etc., but also any programmable apparatus, for example a mobile phone or an embedded processor. The computer can for example comprise a "subcomputer" system (network), wherein each subcomputer represents a computer in itself. The term "computer" includes a cloud computer, for example a cloud server. The term "computer" includes a server resource. The term "cloud computer" includes a cloud computer system, for example a system comprising at least one cloud computer, and for example a system comprising a plurality of operatively interconnected cloud computers, like a server farm. Such a cloud computer is preferably connected to a wide area network, such as the World Wide Web (WWW), and is located in the so-called cloud of computers all connected to the World Wide Web. Such an infrastructure is used for "cloud computing", which describes those computing, software, data access and storage services that do not require the end user to know the physical location and / or configuration of the computer providing a specific service. For example, the term "cloud" is used here to metaphorically refer to the Internet (World Wide Web). For example, the cloud provides computing infrastructure as a service (IaaS). The cloud computer can be used as a virtual host for operating systems and / or data processing applications for performing the inventive method. The cloud computer is for example the Elastic Compute Cloud (EC2) provided by Amazon Web Services™. The computer for example comprises an interface in order to receive or output data and / or to perform an analog-digital conversion. The data is for example data representing a physical property and / or generated by a technical signal. The technical signal is for example generated by a (technical) detection apparatus, for example an apparatus for detecting a marker apparatus, and / or a (technical) analysis apparatus, for example an apparatus for performing a (medical) imaging method, wherein the technical signal is for example an electrical signal or an optical signal. The technical signal is for example representative of the data received or output by the computer. The computer is preferably operatively coupled to a display apparatus that allows the information output by the computer to be displayed to for example a user. An example of a display apparatus is a virtual reality apparatus or an augmented reality apparatus (also known as virtual reality glasses or augmented reality glasses), which can be used as "goggles" for navigation. A specific example of such augmented reality glasses is Google Glass (a trademark brand under Google, Inc.).An augmented reality device or a virtual reality device can be used both for inputting information into a computer by user interaction and for displaying information output by the computer. Another example of a display device is a standard computer monitor, e.g. comprising a liquid crystal display, which is operably coupled to a computer for receiving display control data from the computer generating a signal to display image information content on the display device. A specific embodiment of such a computer monitor is a digital light box. An example of such a digital light box is the Brainlab® Digital Light Box, which is a product of Brainlab AG. The display can also be a portable device, e.g. handheld, such as a smartphone or a personal digital assistant or a digital media player.

[0147] The present application also relates to a computer program comprising instructions, which, when executed on a computer, cause the computer to carry out one or more of the methods described herein, e.g. the steps of one or more of the methods; and / or the present application relates to a computer-readable storage medium (e.g. a non-transitory computer-readable storage medium) having stored thereon such a program; and / or the present application relates to a computer comprising such a storage medium; and / or the present application relates to a (e.g. physically (e.g. electrically), e.g. in technical means, generated) signal wave, e.g. a digital signal wave such as an electromagnetic carrier wave, carrying information which represents a program (e.g. such as described above), e.g. including code means which are adapted to perform any or all of the steps of any or all of the methods described herein. In an example, the signal wave is a data carrier signal carrying the computer program described above. The present application also relates to a computer comprising at least one processor and / or the above-mentioned computer-readable storage medium and, e.g. a memory, wherein the program is executed by the processor.

[0148] ​Within the framework of this invention, computer program elements can be embodied in hardware and / or software (including firmware, resident software, microcode, etc.). Within the framework of this invention, computer program elements can take the form of a computer program product, which can be embodied in a computer-usable (e.g., computer-readable) data storage medium comprising computer-usable (e.g., computer-readable) program instructions, wherein the "code" or "computer program" embodied in the data storage medium is used for or in conjunction with an instruction execution system. Such a system can be a computer; the computer can be a data processing apparatus including components for executing the computer program elements and / or programs according to the invention, such as a data processing apparatus including a digital processor (central processing unit or CPU) for executing the computer program elements, and optionally including a data processing apparatus storing volatile memory (e.g., random access memory or RAM) for the computer program elements and / or data generated by executing the computer program elements. Within the framework of this invention, a computer-usable (e.g., computer-readable) data storage medium can be any data storage medium that can include, store, communicate, propagate, or transmit programs for or in conjunction with an instruction execution system, device, or apparatus. Computer-usable (e.g., computer-readable) data storage media can be, for example, but not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or apparatuses, or propagation media such as the Internet. Computer-usable or computer-readable data storage media can even be, for example, paper or other suitable media on which a program is printed, since the program can be obtained electronically, for example by optically scanning the paper or other suitable medium, and then compiled, decoded, or otherwise processed as appropriate. The data storage medium is preferably a non-volatile data storage medium. The computer program products described herein and any software and / or hardware form various components for performing the functions of the invention in the exemplary embodiments. Computer and / or data processing apparatuses may, for example, include guidance information devices comprising components for outputting guidance information. For example, guidance information can be output to a user visually via visual indication components (e.g., displays and / or lights), and / or audibly via auditory indication components (e.g., speakers and / or digital voice output devices), and / or tactilely via tactile indication components (e.g., vibrating elements or vibrating elements incorporated into an instrument). For the purposes of this document, a computer is a professional computer, which includes, for example, specialized (e.g., tangible) components, such as mechanical and / or electronic components. Any device referred to herein is a professional device and, for example, a tangible device.

[0149] The expression "acquiring data" (within the framework of a computer-implemented method) for example comprises the scenario of determining data by means of a computer-implemented method or program. Determining data for example comprises measuring a physical quantity and converting the measured value into data (e.g. digital data), and / or calculating (and e.g. outputting) the data by means of a computer and e.g. within the framework of a method according to the application. The "determining" steps described herein for example comprise the step of issuing a command for performing the determination described herein. For example, the step comprises the step of issuing a command for causing a computer (e.g. a remote computer, e.g. a remote server, e.g. a computer in the cloud) to perform the determination. Alternatively or additionally, the "determining" steps described herein for example comprise the step of receiving data resulting from the determination described herein, e.g. receiving the result data from a remote computer (which remote computer was e.g. caused to perform the determination). The meaning of "acquiring data" for example also comprises the scenario of receiving or retrieving data (e.g. into) a computer-implemented method or program, e.g. from another program, a previous method step or a data storage medium, e.g. for further processing by the computer-implemented method or program. The generation of the data to be acquired can but need not be part of the method according to the application. Thus, the expression "acquiring data" for example can also mean waiting for receiving data and / or receiving data. The received data can be input, e.g. via an interface. The expression "acquiring data" can also mean that the computer-implemented method or program performs a step in order to (actively) receive or retrieve data from a data source, e.g. a data storage medium (such as a ROM, a RAM, a database, a hard drive, etc.) or via an interface (e.g. from another computer or a network). The data acquired by the method or device of the present disclosure, respectively, can be acquired from a database located in a data storage device which is operatively connected to a computer for data transfer between the database and the computer (e.g. from the database to the computer). The computer acquires the data for use as input for the step of "determining data". The determined data can be output again to the same or another database storage for subsequent use. The database or the database for implementing the method of the present disclosure can be located in a network data storage device or a network server (e.g. a cloud data storage device or a cloud server) or a local data storage device (e.g. a mass storage device which is operatively connected to at least one computer performing the method of the present disclosure). The data can be "ready" by performing an additional step prior to the acquiring step. According to the additional step, the data is generated for acquisition. For example, the data is detected or acquired (e.g. by an analysis device). Alternatively or additionally, according to the additional step, the data is input, e.g. via an interface. For example, the generated data can be input (e.g. into a computer). According to the additional step (which is performed prior to the acquiring step), the data can also be provided by performing an additional step of storing the data in a data storage medium (such as a ROM, a RAM, a CD and / or a hard drive), such that the data is ready within the framework of the method or program according to the application.Accordingly, the step of "acquiring data" can also involve commanding the device to obtain and / or provide the data to be acquired. In particular, the step of acquiring data does not involve an invasive step, which represents a substantial physical intervention into the body, which requires taking professional medical measures on the body, and which can expose the body to significant health risks even if the required professional care and measures are taken. In particular, the step of acquiring data (e.g. determining data) does not involve a surgical step, in particular a step which makes use of a surgical or therapeutic treatment of the human or animal body. In order to distinguish between the different data used by the method, the data are denoted (i.e. referred to) as "XY data" or the like and defined in terms of the information they describe, and are then preferably referred to as "XY information" or the like.

[0150] The function of the marker is to be detected by a marker detection device (e.g. a camera or an ultrasound receiver or an analysis device such as a CT or MRI device) so that its spatial position (i.e. its spatial positioning and / or alignment) can be determined. The detection device is for example part of a navigation system. The marker can be an active marker. The active marker can for example emit electromagnetic radiation and / or waves which can be in the infrared, visible and / or ultraviolet spectral range. However, the marker can also be passive, i.e. can for example reflect electromagnetic radiation in the infrared, visible and / or ultraviolet spectral range or can block x-ray radiation. To this end, the marker can be provided with a surface having corresponding reflection properties or can be made of metal to block x-ray radiation. The marker can also reflect and / or emit electromagnetic radiation and / or waves in the radio frequency range or in the ultrasound wavelength range. The marker is preferably spherical and / or ellipsoidal in shape and can therefore be referred to as a marker sphere. However, the marker can also exhibit an angular (e.g. cubic) shape.

[0151] The marker device can be for example a reference star or a pointer or a single marker or a plurality of (independent) markers, the plurality of markers preferably being in a predetermined spatial relationship. The marker device comprises one, two, three or more markers, wherein two or more of such markers are in a predetermined spatial relationship. This predetermined spatial relationship is for example known to the navigation system and for example stored in a computer of the navigation system.

[0152] In another embodiment, the marker device comprises an optical pattern on for example a two-dimensional surface. The optical pattern can comprise a plurality of geometric shapes like circles, rectangles and / or triangles. The optical pattern can be identified in an image captured by a camera and the position of the marker device relative to the camera can be determined from the size of the pattern in the image, the orientation of the pattern in the image and the distortion of the pattern in the image. This enables to determine a relative position in up to three rotational dimensions and up to three translational dimensions from a single two-dimensional image.

[0153] The position of the marker device can be determined, for example, by a medical navigation system. If the marker device is attached to an object, such as a bone or a medical instrument, the position of the object can be determined from the position of the marker device and the relative position between the marker device and the object. Determining this relative position is also referred to as registering the marker device and the object. The marker device or the object can be tracked, which means that the position of the marker device or the object is determined two or more times over time.

[0154] Preferably, atlas data is acquired, which atlas data describes (e.g. defines, more specifically represents and / or is) a general three-dimensional shape of an anatomical body part. Thus, the atlas data represents an atlas of the anatomical body part. The atlas typically consists of a plurality of generic object models, wherein these generic object models together constitute a composite structure. For example, the atlas constitutes a statistical model of a patient's body (e.g. a part of the body) which is generated from anatomical information collected from a plurality of human bodies (e.g. from medical image data comprising images of these human bodies). Thus, in principle, the atlas data represents a statistical analysis result of such medical image data of a plurality of human bodies. This result can be output as an image, thus the atlas data contains or is comparable to the medical image data. This comparison can be performed, for example, by applying an image fusion algorithm which performs image fusion between the atlas data and the medical image data. The result of the comparison can be a measure of similarity between the atlas data and the medical image data. The atlas data comprises image information (e.g. positional image information) which can be matched (e.g. by applying an elastic or rigid image fusion algorithm) to image information (e.g. positional image information) contained in the medical image data in order to, for example, compare the atlas data with the medical image data in order to determine the position of an anatomical structure in the medical image data which corresponds to an anatomical structure defined by the atlas data.

[0155] The human bodies (of which the anatomical structures are used as input for generating the atlas data) advantageously share common characteristics such as at least one of gender, age, race, body measurements (e.g. height and / or weight) and pathological state. The anatomical information describes, for example, the anatomical structure of the human body and is extracted, for example, from medical image information about the human body. For example, an atlas of a femur can comprise the femoral head, the femoral neck, the femoral shaft, the greater trochanter, the lesser trochanter and the lower extremity which are objects together constituting this complete structure. For example, an atlas of the brain can comprise the telencephalon, the cerebellum, the diencephalon, the pons, the mesencephalon and the medulla oblongata which are objects together constituting this complex structure. One application of such an atlas is in medical image segmentation, wherein the atlas is matched to medical image data and the image data is compared to the matched atlas in order to assign points (pixels or voxels) of the image data to objects of the matched atlas, thereby segmenting the image data into objects.

[0156] For example, the atlas data comprises information of an anatomical body part. The information is, for example, at least one of patient-specific, non-patient-specific, indication-specific, or non-indication-specific. Thus, the atlas data describes, for example, at least one of a patient-specific, non-patient-specific, indication-specific, or non-indication-specific atlas. For example, the atlas data comprises movement information indicative of a degree of freedom of movement of the anatomical body part relative to a given reference (e.g., another anatomical body part). For example, the atlas is a multi-modal atlas which defines atlas information for a plurality (i.e., at least two) of imaging modalities and contains a mapping between the atlas information under the different imaging modalities (e.g., a mapping between all modalities) such that the atlas can be used to transform medical image information from its image depiction under a first imaging modality to its image depiction under a second imaging modality different from the first imaging modality, or to compare (e.g., match or register) images of different imaging modalities with each other.

[0157] In the medical field, imaging methods (also referred to as imaging modalities and / or medical imaging modalities) are used to generate image data (e.g. two- or three-dimensional image data) of anatomical structures (e.g. soft tissue, bones, organs, etc.) of a human body. The term “medical imaging method” is to be understood to mean an (advantageously device-based) imaging method (e.g. a so-called medical imaging modality and / or a radiological imaging method) such as, for example, computed tomography (CT) and cone-beam computed tomography (CBCT, e.g. volumetric CBCT), x-ray tomography, magnetic resonance tomography (MRT or MRI), conventional x-ray, ultrasound scanning and / or sonography, and positron emission tomography. The medical imaging method is performed, for example, by an analysis device. Examples of medical imaging modalities to which the medical imaging method applies are: X-ray radiography, magnetic resonance imaging, medical ultrasonography or ultrasound, endoscopy, elastography, tactile imaging, thermal infrared imaging, medical photography, and nuclear medicine functional imaging techniques such as positron emission tomography (PET) and single-photon emission computed tomography (SPECT). The image data thus generated is also referred to as “medical imaging data”. The analysis device is used, for example, to generate image data in a device-based imaging method. These imaging methods are used, for example, for medical diagnostics, for analyzing anatomical body structures to generate images described by the image data. These imaging methods are also used, for example, to detect pathological changes in the human body. However, some changes in the anatomical structure (such as pathological changes in the structure (tissue)) can not be detected and, for example, can not be visible in the images generated by the imaging method. Tumors represent an example of changes in the anatomical structure. If a tumor grows, it can be considered to represent an expanded anatomical structure. This expanded anatomical structure can not be detected, for example, only a part of the expanded anatomical structure can be detected. When using a contrast agent to infiltrate the tumor, for example, primary / high-grade brain tumors are usually visible in an MRI scan. The MRI scan represents an example of an imaging method. In the case of an MRI scan of such a brain tumor, a signal enhancement in the MRI image (caused by the contrast agent infiltrating the tumor) is considered to represent a solid tumor mass. Thus, the tumor can be detected and, for example, discernible in the images generated by the imaging method. In addition to these tumors, which are referred to as “enhancing” tumors, approximately 10% of brain tumors are not discernible in the scan and, for example, not visible to a user observing the images generated by the imaging method. BRIEF DESCRIPTION OF DRAWINGS

[0158] In the following, the application is described with reference to the figures, which give a background description and represent specific embodiments of the application. The scope of the application is not limited to the specific features disclosed in the context of the figures, in which:

[0159] Figure 1 A neuron of a neural network is shown;

[0160] Figure 2 shows the basic flow of the method according to the first aspect;

[0161] Figure 3 shows the basic flow of the method according to the second aspect;

[0162] Figure 4 shows the basic flow of the method according to the third aspect;

[0163] Figures 5a to 5e shows the application of the methods according to the first to third aspects; and

[0164] Figure 6 is a schematic diagram of a system according to the fifth aspect. DETAILED DESCRIPTION

[0165] Figure 1 shows the structure of a neuron as part of a neural network (e.g. a convolutional neural network), wherein inputs are assigned certain weights in order to be processed by an activation function generating a neuron output.

[0166] Figure 2 shows the basic flow of the method according to the first aspect, which starts with acquiring first patient image data in step S21, proceeds to step S22, which comprises acquiring second patient image data, and then advances to determining registration data in step S23. On this basis, step S24 calculates patient image comparison data, followed by acquiring similarity threshold data in step S25. Finally, patient similarity data is determined in step S26.

[0167] Figure 3 illustrates the basic steps of the method according to the third aspect, wherein step S31 comprises acquiring training patient image data, and step S32 acquires patient identifier data. Subsequent step S33 determines training registration data, followed by determining patient association data in step S34. Image similarity data is then determined in step S35.

[0168] Figure 4 illustrates the basic steps of the method according to the second aspect, wherein step S41 comprises acquiring specific registration data, and step S42 determines specific image similarity data.

[0169] Figure 5a shows how the first patient image 1 corresponding to a first medical image and the second patient image 2 corresponding to a second medical image are directly registered by registration 3. Figure 5b illustrates the indirect registration of the first patient image 1 and the second patient image 2, both of which are registered to the atlas 3 by a first registration 4 and a second registration 5. According to Figure 5c , in addition to Figure 5bIn addition to the features shown in the middle, the atlas 3 applies the variability model 6 to the first registration 4 and the second registration 5 to determine whether any differences between the registrations 4, 5 are due to anatomical adaptation or adaptation. Figure 5d The first patient image 1 is shown registered to the atlas 3 by a first registration 4. The second patient image 2 is registered to the same atlas 3 by a second registration 5. The patient identity information 6 associated with the patient of the first patient image 1 and the patient identity information 7 associated with the patient of the second patient image 2 are compared 11 to determine whether the patient associated with the first patient image 1 and the patient associated with the second patient image 2 are the same. The registration 4 is input 9 into a machine learning algorithm 12. The registration 5 is input 10 into the machine learning algorithm 12. In addition, information describing whether the two patients and / or the first patient image 1 and the second patient image 2 are the same 11 is input 13 into the machine learning algorithm. Based on this input, the machine learning algorithm is trained as described with respect to the second aspect of the disclosure. Figure 5e The use of the trained machine learning algorithm 12 is shown: the first patient image 1 is registered to the atlas 3 by a first registration 4, and the second patient image 2 is registered to the same atlas 3 by a second registration 5. The first registration 4 is input 9 into the trained machine learning algorithm 12. The second registration 5 is also input 10 into the trained machine learning algorithm 12. The machine learning algorithm determines whether the first patient image 1 and the second patient image 2 are from the same patient 11 or from different patients 13 and outputs the corresponding information.

[0170] Figure 6 is a schematic diagram of a medical system 61 according to the fifth aspect. The system as a whole is identified by reference numeral 61 and comprises: a computer 62; an electronic data storage device (such as a hard disk) 64 for storing at least data stored by the system according to the fifth aspect; and a program storage medium 63 for storing the program according to the fourth aspect. The components of the medical system 61 have the functions and properties explained above with respect to the fifth and sixth aspects of the disclosure.

[0171] The following examples form part of the present application:

[0172] A. A computer-implemented method for determining whether a plurality of medical image acquisitions are of the same patient, the method comprising the steps of:

[0173] a) obtaining first patient image data (S21), the first patient image data describing a first medical image of a first anatomical body part of a first patient;

[0174] b) obtaining second patient image data (S22), the second patient image data describing a second medical image of a second anatomical body part of a second patient, the second anatomical body part corresponding to the first anatomical body part;

[0175] c) determining registration data (S23) based on the first patient image data and the second patient image data, wherein the registration data describe a registration between the first medical image and the second medical image;

[0176] d) determining patient image comparison data (S24) based on the registration data, wherein the patient image comparison data describe a measure of similarity between the first medical image and the second medical image;

[0177] e) obtaining similarity threshold data (S25) describing a threshold value for the measure of similarity; and

[0178] f) determining patient similarity data (S26) based on the patient image comparison data, wherein the patient similarity data describe that the first medical image and the second medical image were taken of the same patient if the measure of similarity reaches a predetermined relationship with respect to the threshold value.

[0179] B. The method according to embodiment A, comprising the step of:

[0180] obtaining statistical threshold data describing a threshold value for a statistical value describing a statistical quantity characterizing the registration,

[0181] wherein the measure of similarity is determined by determining whether the statistical value reaches a predetermined relationship with respect to the threshold value, wherein the first medical image and the second medical image are determined to be similar if the statistical value reaches a predetermined relationship with respect to the threshold value.

[0182] C. The method according to any one of embodiments A to B, comprising the step of:

[0183] determining registration analysis data based on the registration data,

[0184] wherein the registration analysis data describe a statistical value of a set of registration vectors characterizing at least one of an anatomically invariant or an anatomically variant portion of the first and second anatomical body parts,

[0185] wherein the patient image comparison data are determined based on the registration analysis data.

[0186] D. The method according to embodiment C, comprising the step of:

[0187] obtaining atlas data describing an image-based model of the first and second anatomical body parts,

[0188] wherein the determining registration data comprises:

[0189] determining first registration data based on the first patient image data and the atlas data, wherein the first registration data describes a first registration between the first medical image and the image-based model; and

[0190] determining second registration data based on the second patient image data and the atlas data, wherein the second registration data describes a second registration between the second medical image and the image-based model,

[0191] wherein the registration analysis data is determined for each of the first registration data and the second registration data, respectively,

[0192] wherein determining the patient image comparison data comprises:

[0193] determining registration comparison data based on the first registration data and the second registration data, wherein the registration comparison data describes a measure of similarity between the first registration and the second registration,

[0194] wherein the patient image comparison data is further determined based on the registration comparison data by comparing statistical values of the first registration with statistical values of the second registration.

[0195] E. The method according to embodiment C, comprising the step of:

[0196] obtaining atlas data describing an image-based model of the first and second anatomical body part,

[0197] wherein the determining registration data comprises:

[0198] determining first registration data based on the first patient image data and the atlas data, wherein the first registration data describes a first registration between the first medical image and the image-based model; and

[0199] determining second registration data based on the second patient image data and the atlas data, wherein the second registration data describes a second registration between the second medical image and the image-based model,

[0200] wherein the determining patient image comparison data comprises:

[0201] determining registration comparison data based on the first registration data and the second registration data, wherein the registration comparison data describes a measure of similarity between the first registration and the second registration,

[0202] wherein said determining registration comparison data comprises determining an inverse of said second registration and multiplying the inverse of said second registration with said first registration, and determining that said first registration and said second registration are similar if said multiplying results in at least substantially 1 within an anatomically invariant region.

[0203] F. The method according to embodiment D or E, comprising the step of:

[0204] obtaining biomechanical model data, said biomechanical model data describing a biomechanical model of the first and second anatomical body part;

[0205] wherein said determining registration comparison data comprises determining registration difference data describing a difference between said first registration and said second registration, wherein said difference is determined by determining an inverse of said second registration and multiplying the inverse of said second registration with said first registration, and wherein said multiplying results in not at least substantially 1,

[0206] wherein at least a part of said first and second anatomical body part is anatomically variable, and

[0207] wherein based on said biomechanical model data and said registration difference data determining variability data, wherein said variability data describes a variability, e.g. a deformation capability, of the biomechanical model according to a difference between said first registration and said second registration, and wherein said first registration and said second registration are determined to be similar if said variability data describes that the biomechanical model is variable according to a difference between said first registration and said second registration.

[0208] G. The method according to embodiment A, comprising the step of:

[0209] obtaining atlas data, said atlas data describing an image-based model of said first and second anatomical body part,

[0210] wherein said determining registration data comprises:

[0211] determining first registration data based on said first patient image data and said atlas data, wherein said first registration data describes a first registration between said first medical image and said image-based model; and

[0212] determining second registration data based on said second patient image data and said atlas data, wherein said second registration data describes a second registration between said second medical image and said image-based model,

[0213] wherein said determining patient image comparison data comprises:

[0214] determining the registration comparison data based on the first registration data and the second registration data, wherein the registration comparison data describes a measure of similarity between the first registration and the second registration,

[0215] wherein the patient image comparison data is further determined based on the registration comparison data.

[0216] H. A computer-implemented method for training a learning algorithm to determine whether a plurality of medical image shots are of the same patient, the method comprising the steps of:

[0217] a) obtaining training patient image data (S31) describing a plurality of medical patient training images showing an anatomical body part;

[0218] b) obtaining patient identifier data (S32) describing, for each medical patient training image, a patient identifier identifying a patient of which the respective medical patient image is a shot;

[0219] c) determining training registration data (S33) based on the training patient image data, wherein the training registration data describes registrations between pairs of images from the plurality of medical patient training images;

[0220] d) determining patient association data (S34) for images belonging to the plurality of medical training images, wherein the patient association data is determined based on the training patient image data and the patient identifier data and comprises patient association information describing whether the images belonging to each pair of medical images are shots of the same patient; and

[0221] e) determining image similarity data (S35) describing model parameters of a learning algorithm for establishing a relationship between the registrations and the patient association information, wherein the image similarity data is determined by inputting the training registration data and the patient association data into a function establishing the relationship.

[0222] I. The method according to embodiment H, wherein the registrations are established directly between the plurality of medical patient training images.

[0223] J. The method according to claim 8, comprising the steps of:

[0224] obtaining atlas data describing image-based models of first and second anatomical body parts,

[0225] determining first registration data based on a first image of each pair of medical patient training images, wherein the first registration data describes a first registration between the first image and the image-based model;

[0226] determining second registration data based on a second image of each pair of medical patient training images, wherein the second registration data describes a second registration between the second image and the image-based model,

[0227] wherein the learning algorithm is trained to determine a statement indicating whether the first image and the second image were taken of the same patient based on the first registration data and the second registration data, wherein the statement can be associated with a certain level of confidence.

[0228] K. The method according to embodiment J, wherein the anatomical body part is anatomically invariant or anatomically variant, the method comprising the steps of:

[0229] determining registration analysis data for each of the first registration and the second registration based on the registration data, wherein the registration analysis data describes statistical values of a set of registration vectors characterizing at least one anatomically invariant part of the first and second anatomical body parts,

[0230] wherein the learning algorithm is trained to determine a statement indicating whether the first image and the second image were taken of the same patient based on the first registration data and the second registration data, wherein the statement can be associated with a certain level of confidence.

[0231] L. The method according to embodiment H, comprising the steps of:

[0232] obtaining atlas data describing image-based models of the first and second anatomical body parts,

[0233] determining first registration data based on a first image of each pair of medical patient training images, wherein the first registration data describes a first registration between the first image and the image-based model;

[0234] determining second registration data based on a second image of each pair of medical patient training images, wherein the second registration data describes a second registration between the second image and the image-based model,

[0235] wherein the learning algorithm is trained to determine a statement indicating whether the first image and the second image were taken of the same patient based on the first registration data and the second registration data, wherein the statement can be associated with a certain level of confidence.

[0236] M. The method according to embodiment L, wherein the learning algorithm is trained such that it determines that the first and second image were taken of the same patient if the multiplication results in at least substantially one.

[0237] N. A computer-implemented method for determining whether a plurality of medical images were taken of the same patient, the method comprising the steps of:

[0238] a) obtaining specific registration data (S41) describing a specific registration between a plurality of medical images of an anatomical body part; and

[0239] b) determining specific image similarity data (S42) describing a relationship between patient association data and the specific registration data, the patient association data describing whether image takings belonging to the plurality of medical images were taken of the same patient, wherein the specific image similarity data is determined by inputting the specific registration data into a function establishing the relationship, the function being part of a learning algorithm trained by performing the method according to any one of the six preceding claims or any one of claims 15 to 20 (backtraceable to any one of the six preceding claims).

[0240] O. The method according to any one of embodiments H to N, wherein the learning algorithm comprises or consists of a machine learning algorithm.

[0241] P. The method according to any one of embodiments H to O, wherein the learning algorithm comprises or consists of a convolutional neural network.

[0242] Q. The method according to any one of embodiments H to P, wherein the model parameters define learnable parameters, such as weights, of the learning algorithm.

[0243] R. A program which, when running on a computer (62) or loaded onto a computer (62), causes the computer (62) to perform the method steps of the method according to any of the preceding claims, and / or a program storage medium (63) having a program stored thereon, or a program storage medium (63) having data stored thereon defining model parameters and architecture of a learning algorithm trained by performing the method according to any of embodiments H to Q or O to Q (with retroactive dependency to any of embodiments H to M), and / or a data carrier signal carrying the aforementioned program, and / or a data carrier signal carrying data defining model parameters and architecture of a learning algorithm trained by performing the method according to any of embodiments H to M or O to Q (with retroactive dependency to any of embodiments H to M), and / or a data stream carrying the aforementioned program, and / or a data stream carrying data defining model parameters and architecture of a learning algorithm trained by performing the method according to any of embodiments H to M or O to Q (with retroactive dependency to any of embodiments H to M), and / or at least one computer (62) comprising at least one processor and a memory, wherein the aforementioned program is running on the at least one processor or loaded into the memory of the computer (62).

[0244] S. A system (61) for determining whether a plurality of medical image recordings are of the same patient, the system (61) comprising:

[0245] a) at least one computer (62) according to embodiment R (with retroactive dependency to any of embodiments A to G);

[0246] b) at least one electronic data storage (64) storing the first patient image data and the second patient image; and

[0247] c) a program storage medium (63) according to embodiment R (with retroactive dependency to any of embodiments A to G), wherein the at least one computer (62) is operably coupled to the at least one electronic data storage (64) to: retrieve the first patient image data and the second patient image data from the at least one electronic data storage (64); and / or store at least the patient similarity data in the at least one electronic data storage (64).

[0248] T. A system (61) for determining whether a plurality of medical image recordings are of the same patient, the system (61) comprising:

[0249] a) at least one computer (62) according to embodiment R (backtraceable to dependent on claim 14);

[0250] b) at least one electronic data storage device (64) for storing registration data; and

[0251] c) a program storage medium (63) according to embodiment R (backtraceable to dependent on any one of claims 14),

[0252] wherein the at least one computer (62) is operatively coupled to:

[0253] - the at least one electronic data storage device (64) to retrieve the specific registration data from the at least one electronic data storage device (64) and / or to store at least the specific image similarity data in the at least one electronic data storage device (64); and

[0254] - the program storage medium (64) to retrieve data defining model parameters and architecture of the learning algorithm from the program storage medium (63).

Claims

1. A computer-implemented method for determining whether multiple medical images were taken from the same patient, the method comprising the following steps: a) Acquire first patient image data, the first patient image data being a first medical image describing a first anatomical body part of the first patient; b) Acquire second patient image data, the second patient image data being a second medical image describing a second anatomical body part of the second patient, wherein the second anatomical body part corresponds to the first anatomical body part; c) Determine registration data based on the first patient image data and the second patient image data, wherein the registration data describes the registration between the first medical image and the second medical image; d) Determine patient image comparison data based on the registration data, wherein the patient image comparison data describes a measure of similarity between the first medical image and the second medical image; e) Obtain similarity threshold data, which describes a threshold for the measure of similarity; f) Determine patient similarity data based on the patient image comparison data, wherein if the measure of similarity reaches a predetermined relationship relative to the threshold, the patient similarity data describes that the first medical image and the second medical image were taken of the same patient; g) Acquire atlas data, which describes image-based models of the first anatomical body part and the second anatomical body part. The registration data to be determined includes: h) Determine first registration data based on the first patient image data and the atlas data, wherein the first registration data describes a first registration between the first medical image and the image-based model; and i) Determine second registration data based on the second patient image data and the atlas data, wherein the second registration data describes a second registration between the second medical image and the image-based model; j) Determine registration comparison data based on the first registration data and the second registration data, wherein the registration comparison data describes a measure of similarity between the first registration and the second registration, and wherein the patient image comparison data is determined based on the registration comparison data.

2. The method according to claim 1, wherein, The determination of registration comparison data includes acquiring variability model data, which describes the variability models of the first anatomical body part and the second anatomical body part. The determination of registration comparison data also includes determining registration difference data, which describes the difference between the first registration and the second registration. This difference is determined by determining the inverse of the second registration and concatenating it with the first registration. Both the first and second registrations are defined as displacement vector fields. The result of the concatenation and the difference from the variability transformation result in a displacement vector field, which describes the configuration of the variability model.

3. The method according to claim 1, wherein, The determination of registration comparison data includes the determination of registration difference data, which describes the difference between the first registration and the second registration. The difference is determined by determining the inverse of the second registration and concatenating the inverse of the second registration with the first registration. Both the first registration and the second registration are defined as displacement vector fields.

4. The method according to claim 1, wherein, The determination of registration comparison data includes determining registration difference data, which describes the difference between the first registration and the second registration, wherein the difference is determined by: determining the inverse of the second registration and concatenating the inverse of the second registration with the first registration, wherein both the first registration and the second registration are defined as displacement vector fields; and subtracting a rigid transformation from the concatenation, wherein the rigid transformation describes rotation and / or translation.

5. The method according to any one of claims 2 to 4, wherein, The first anatomical body part is an anatomically invariant or anatomically variable anatomical body part.

6. A computer-implemented method for training a learning algorithm to determine whether multiple medical images were taken from the same patient, the method comprising the following steps: a) Acquire training patient image data, wherein the training patient image data describes multiple medical patient training images displaying anatomical body parts; b) Obtain patient identifier data, wherein the patient identifier data describes a patient identifier for each medical patient training image, and the patient identifier identifies the patient in the corresponding medical patient image; c) Determine training registration data based on the training patient image data, wherein the training registration data describes the registration between image pairs from the plurality of medical patient training images; d) Determine patient association data for images belonging to the plurality of medical patient training images, wherein the patient association data is determined based on the training patient image data and the patient identifier data, and includes patient association information describing whether images belonging to each pair of medical images were taken from the same patient; and e) Determine image similarity data, which describes model parameters of a learning algorithm used to establish a relationship between the registration and the patient association information, wherein the image similarity data is determined by inputting the training registration data and the patient association data into a function that establishes the relationship.

7. The method according to claim 6, comprising the following steps: Acquire atlas data, which describes image-based models of a first anatomical body part and a second anatomical body part. First registration data is determined based on a first image of each pair of medical patient training images, wherein the first registration data describes a first registration between the first image and the image-based model; A second registration data is determined based on a second image of each pair of medical patient training images, wherein the second registration data describes a second registration between the second image and the image-based model. The learning algorithm is trained to determine a statement indicating whether the first image and the second image were taken from the same patient, based on the first registration data and the second registration data, wherein the statement is associated with a certain confidence level.

8. The method according to claim 7, wherein, The anatomical body part is either anatomically invariant or anatomically variable, and the method includes the following steps: Based on the registration data, registration analysis data is determined for each of the first and second registrations, wherein the registration analysis data is described by the following statistical values: the statistical values ​​characterize the registration vector set of at least one anatomically invariant part of the first anatomical body part and the second anatomical body part. The learning algorithm is trained to determine a statement indicating whether the first image and the second image were taken from the same patient, based on the first registration data and the second registration data, wherein the statement is associated with a certain confidence level.

9. A computer-implemented method for determining whether multiple medical images were taken from the same patient, the method comprising the following steps: a) Acquire specific registration data, which describes a specific registration between multiple medical images of an anatomical body part; as well as (b) Determine specific image similarity data, which describes the relationship between patient association data and specific registration data, wherein the patient association data describes whether images belonging to the plurality of medical images were taken from the same patient, wherein the specific image similarity data is determined by inputting the specific registration data into a function that establishes the relationship, the function being part of a learning algorithm trained by performing the method according to any one of claims 6 to 8.

10. A program storage medium storing a program that, when run on or loaded onto a computer, causes the computer to perform the method steps of any one of claims 1 to 9.

11. A computer comprising at least one processor and a memory, wherein, The program stored in the program storage medium according to claim 10 is executed on the at least one processor of the computer or loaded into the memory of the computer.

12. A system for determining whether multiple medical images were taken from the same patient, the system comprising: a) At least one computer according to claim 11, wherein the program, when run on or loaded onto the computer, causes the computer to perform the method according to any one of claims 1 to 5; b) At least one electronic data storage device for storing first patient image data and second patient images; as well as c) The program storage medium according to claim 10, The at least one computer is operatively coupled to the at least one electronic data storage device to acquire first patient image data and second patient image data from the at least one electronic data storage device; And / or at least the patient similarity data is stored in the at least one electronic data storage device.

13. A system for determining that multiple medical images are taken of the same patient, the system comprising: a) At least one computer according to claim 11, wherein the program, when run on or loaded onto the computer, causes the computer to perform the method according to claim 9; b) At least one electronic data storage device for storing the registration data; The at least one computer is operatively coupled to: The at least one electronic data storage device is used to obtain the specific registration data from the at least one electronic data storage device; and / or at least store the specific image similarity data in the at least one electronic data storage device; and The program storage medium is used to retrieve data defining the model parameters and architecture of the learning algorithm.

Images