Method for recognizing branches in a vascular tree, related methods and devices

Abstract

Method, related method and apparatus for recognizing branches in a vascular tree. The present invention relates to the field of analyzing data included in a vascular tree. For this purpose, the present invention proposes to use branch smartening to generate appropriate synthetic data. Such synthetic data makes it possible to form a training data set for training an artificial intelligence algorithm adapted to recognize branches in a vascular tree. The present invention thus makes it possible to obtain a better recognition of branches. Such better recognition can be advantageously used in diagnostic, follow-up, and prognostic methods.

Description

Technical Field

[0001] The technical field of the present invention The present invention relates to a method for recognizing at least one branch of a vascular tree in a real image of a subject's vascular tree. The present invention relates to a method for predicting that a subject is at risk of developing an aneurysm. The present invention also relates to a method for diagnosing an aneurysm. The present invention also relates to a method for identifying a therapeutic target for preventing and / or treating an aneurysm. The present invention also relates to a method for identifying a biomarker that is a diagnostic biomarker for an aneurysm, a susceptibility biomarker for an aneurysm, a prognostic biomarker for an aneurysm, or a predictive biomarker for responding to treatment of an aneurysm. The present invention also relates to a method for screening for a compound useful as a drug having an effect on a known therapeutic target for preventing and / or treating an aneurysm. The present invention also relates to a related computer program product and computer-readable medium.

[0002] The background of the present invention The cardiovascular system (also called the circulatory system) is composed of all blood vessels that carry blood and lymph throughout the body. The purpose of this organ system is to transport nutrients, oxygen, and carbon dioxide between body tissues. In certain organs such as the heart, liver, kidneys, lungs, or brain, the vascular system becomes denser. When reaching these organs, arteries, capillaries, or veins branch into several branches, which form a vascular tree.

[0003] The circulatory system can be affected by various vascular diseases such as arteriosclerosis, thrombosis, inflammation, or several genetic diseases. Several factors such as smoking habits, high blood pressure, cardiovascular history, or certain treatments can lead to the vulnerability of the vascular system. Vascular diseases can occur in various arteries of the human body and thus can induce various effects (such as coronary artery disease, thoracic vascular disease, abdominal aortic aneurysm, etc.).

[0004] Weakening of the blood vessel wall can lead to the formation of aneurysms. In the brain, aneurysms often take the form of dissecting aneurysms (where blood leaks from the inner layer of the arterial wall), fusiform aneurysms (localized bulges of arteries characterized by dilation of the blood vessel, i.e., a local increase in diameter), or saccular (sometimes called berry-shaped) aneurysms (bulges that occur on one side of an artery). Ninety percent of brain aneurysms belong to this latter type.

[0005] In many cases, aneurysms remain benign and do not progress to a dangerous state. The main complication caused by an aneurysm is that when it ruptures, blood can flow out into the surrounding tissue, causing a subarachnoid hemorrhage that can lead to death or permanent disability. Rupture causes a decrease in blood flow downstream and the resulting ischemia. Due to the high risk of rupture, the ICA should be carefully monitored. The risk of rupture is higher along a subset of arteries called the circle of Willis, which is located in the center of the brain. Eighty-five percent of saccular ICAs occur along the circle of Willis. ICA aneurysms are very common, affecting 2-5% of the adult population worldwide. ICA ruptures occur at a rate of approximately 8-10 per 100,000 people per year in white people and approximately 20 per 100,000 people per year in Japanese or Finnish people.

[0006] Therefore, it is desirable to detect such pathologies, particularly cerebrovascular diseases, and more particularly the formation of intracranial aneurysms (ICA).

[0007] Several studies have been conducted on vascular segmentation or aneurysm detection. Few studies have focused on aneurysm segmentation, even fewer have focused on the detection of bifurcations, and no studies have been conducted on the recognition of brain bifurcations.

[0008] Due to the recent remarkable progress achieved by medical image analysis using deep learning techniques, it naturally emerges as the most obvious approach to tackle any of the previously mentioned tasks.

[0009] However, in the case of deep learning techniques, manual annotation of unlabeled data is almost inevitable. Generally, research on segmentation or detection related to vascular system segmentation / detection relies on hundreds (at most) of manually segmented images to train neural networks.

[0010] This results in a set of data that provides insufficient learning for the neural network. Therefore, this neural network exhibits low performance such as insufficient robustness, inaccurate predictions, etc.

Summary of the Invention

[0011] Summary of the Invention An object of the present invention is to provide a method for recognizing at least one branch of a vascular tree in a real image of a subject's vascular tree, which exhibits better accuracy and better robustness.

[0012] For this purpose, this specification is a method for recognizing at least one branch of a vascular tree in a real image of a subject, particularly the vascular tree of the brain, and is a method implemented by a computer, The method is - generating a synthetic image of at least one branch of the vascular tree, and the generating step here includes the following steps for each synthetic image, 〇 receiving a real image including at least one branch of the vascular tree, 〇 modeling the real image by an image model having a specific set of values for a set of parameters, where the image model includes at least a geometric model of the branch, The geometric model is a three-dimensional model of the branch and includes a graph of the vascular tree, where the graph is a set of nodes connected by weighted branches, and the geometric model is obtained by partitioning the real image, 〇 generating an image corresponding to the image model having a modified set of values, where the generated image is a synthetic image. - To obtain a trained recognition prediction factor, training a recognition prediction factor adapted to obtain branch recognition data in an input image, wherein the training step includes the following steps: 〇 Forming a training data set based on the synthetic image; 〇 Training the recognition prediction factor by using the training data set; - A step of speculation, which includes the following steps: 〇 Forming a training data set based on the synthetic image; 〇 Training the recognition prediction factor by using the training data set; Including.

[0013] According to a further aspect which is advantageous but not mandatory, the recognition method may incorporate one or more of the following features employed in any technically acceptable combination: 〇 The image model includes a noise model, the noise model models the noise of the image by Gaussian noise having a standard deviation, the standard deviation of the Gaussian noise is one of the parameters of the model, and the standard deviation is equal to a first value; During the generating step, a Gaussian filter having a standard deviation is applied to the real image to obtain an image having Gaussian noise having a standard deviation with a second value, the second value being different from the first value, and the standard deviation of the Gaussian filter depends on the first value and the second value. 〇 The parameters of the geometric model further include the diameter of the branch; The value of the diameter is obtained by applying a convolution kernel to the real image. 〇 During the generating step, geometric distortion is applied to the geometric model. 〇 During the generating step, geometric distortion is applied to the geometric model. 〇 The geometric model defines a reference point of the branch, and the geometric model includes an interpolation function connecting the reference points. Each interpolation function is a function defined by coefficients, and the coefficients are parameters of a set of parameters, and the coefficients are modified during the generating step. 〇 Each interpolation function is a B-spline function defined by B-spline coefficients, and the coefficients are B-spline coefficients. 〇 The value of the coefficient is modified by adding a random value weighted by a specific value. 〇 The image model includes a background model, and the background model includes a shape having two distinct values. 〇 Each image is taken by MRA-TOF technology. 〇 The branch recognition data is selected from the following elements. 〇 The class or type of the branch, 〇 The existence of the branch, 〇 The position of the branch, 〇 The branch angle of the branch, 〇 The geodesic distance between two branches, 〇 The cross-sectional area of the detected branch, and 〇 The twist parameter of the branch. 〇 The recognition predictor is a neural network. 〇 The neural network is a convolutional neural network.

[0014] This specification further relates to a method including performing steps of a method for recognizing at least one branch of a vascular tree in a real image of a subject's vascular tree. The method is the method according to any one of claims 1 to 8, and the method is - A method for predicting that a subject is at risk of developing an aneurysm, the predicting method including at least the following steps: 〇 Performing the steps of the recognizing method to obtain branch recognition data, 〇 Predicting that the subject is at risk of developing an aneurysm based on the obtained branch recognition data. - A method for diagnosing an aneurysm, wherein the diagnosing method includes at least the following steps: 〇 Executing the steps of the recognizing method to obtain branch recognition data; 〇 Diagnosing an aneurysm based on the obtained branch recognition data; - A method for identifying a treatment target for preventing and / or treating an aneurysm, wherein the method includes at least the following steps: 〇 Executing the steps of the identifying method on a first subject to obtain first obtained branch recognition data, wherein the first subject is a subject suffering from an aneurysm; 〇 Executing the steps of the identifying method on a second subject to obtain second obtained branch recognition data, wherein the second subject is a subject not suffering from an aneurysm; 〇 Selecting a treatment target based on a comparison between the first obtained branch recognition data and the second obtained branch recognition data; - A method for identifying a biomarker, wherein the biomarker is a diagnostic biomarker for an aneurysm, a susceptibility biomarker for an aneurysm, a prognostic biomarker for an aneurysm, or a predictive biomarker for the treatment of an aneurysm, and the method includes at least the following steps: 〇 Executing the steps of the identifying method on a first subject to obtain first obtained branch recognition data, wherein the first subject is a subject suffering from an aneurysm; 〇 Executing the steps of the identifying method on a second subject to obtain second obtained branch recognition data, wherein the second subject is a subject not suffering from an aneurysm; 〇 Selecting a biomarker based on a comparison between the first obtained branch recognition data and the second obtained branch recognition data; and - A method for screening for compounds useful as probiotics, prebiotics or pharmaceuticals, wherein said compounds have an effect on known therapeutic targets for preventing and / or treating aneurysms, the method comprising at least the following steps: 〇 Performing the step of a method of identifying a first subject to obtain first obtained branch recognition data, wherein said first subject is a subject suffering from an aneurysm and administered a compound; 〇 Performing the step of a method of identifying a second subject to obtain second obtained branch recognition data, wherein said second subject is a subject suffering from an aneurysm and not administered a compound; 〇 Selecting a compound based on a comparison of said first determined parameter and said second determined parameter; Selected from the list consisting of:

[0015] This specification further relates to a computer program product comprising instructions for performing the steps of the previously described method when the computer program product is executed on a suitable computer device.

[0016] This specification also relates to a computer-readable medium encoded with the previously described computer program.

Brief Description of the Drawings

[0017] The present invention will be better understood based on the following description, given by way of example and corresponding to the accompanying drawings without limiting the object of the present invention. In the accompanying drawings:

Figure 1

Figure 2

Figure 3

Figure 4

DETAILED DESCRIPTION OF THE INVENTION

[0018] DETAILED DESCRIPTION OF THE EMBODIMENT DESCRIPTION OF THE SYSTEM System 10 and computer program product 12 are shown in FIG. 1. The interaction between computer program product 12 and system 10 makes it possible to execute a method for recognizing at least one branch of the vascular tree in a real image of the vascular tree of a subject.

[0019] System 10 is a computer. In this case, system 10 is a laptop.

[0020] More generally, system 10 is a computer, a computer system, or a similar electronic computing device that manipulates and / or transforms data represented as a physical quantity such as an electron in the registers and memories of the computer system into other data represented as a physical quantity in the memories, registers, or other such information storage devices, transmission, or display devices of the computer system.

[0021] System 10 includes a processor 14, a keyboard 22, and a display unit 24.

[0022] Processor 14 includes a data processing unit 16, a memory 18, and a reader 20. Reader 20 is adapted to read a computer-readable medium.

[0023] Computer program product 12 includes a computer-readable medium.

[0024] A computer-readable medium is a medium that can be read by the reader of the processor. A computer-readable medium is a medium suitable for storing electronic instructions and can be coupled to a computer system bus.

[0025] Such a computer-readable storage medium may be, for example, a disk, a floppy disk, an optical disk, a CD-ROM, a magneto-optical disk, a read-only memory (ROM), a random access memory (RAM), an electrically programmable read-only memory (EPROM), an electrically erasable and programmable read-only memory (EEPROM), a magnetic card or an optical card, or any other type of medium suitable for storing electronic instructions and may be coupled to a computer system bus.

[0026] A computer program is stored in a computer-readable storage medium. The computer program includes one or more stored sequences of program instructions.

[0027] The computer program is loadable into a data processing unit and is adapted to cause the execution of a method to be recognized when the computer program is executed by the data processing unit.

[0028] Operation of the system Next, the operation of system 10 is described by showing an example of implementing a method for recognizing at least one branch of a blood vessel tree in a real image of a subject's blood vessel tree, as shown in the flowchart of FIG. 2.

[0029] Such a method of recognition aims to identify, i.e., localize and / or characterize, at least one branch of the blood vessel tree.

[0030] In this example, it is assumed that the method of recognition is specialized for a classification task.

[0031] More precisely, some types of branches are judged to be interesting because they apparently lead to aneurysms, and the method of recognition seeks to appropriately identify whether the imaged branch is one of these branches of interest.

[0032] For example, the number of types of branches is included between 10 and 20.

[0033] The vascular tree is a set of blood vessels within a region. The cerebral vascular tree is the vascular tree of the subject's brain.

[0034] Here, it should be noted that the present method can be applied to any vascular tree, and the cerebral vascular tree is given only as a specific example.

[0035] Also, unlike many other methods belonging to the prior art, it should be mentioned that the described recognition method can handle three-dimensional vascular trees.

[0036] The subject is an animal, particularly a mammal.

[0037] In particular, the subject is a mouse or a human.

[0038] A branch is where the parent artery divides into two or more daughter arteries.

[0039] In particular, it is not uncommon to find a branch where the parent artery divides into three daughter arteries.

[0040] The following meaning of a branch is either where the parent artery divides into two or more daughter arteries (broad meaning) or where the parent artery divides exactly into two daughter arteries (strict meaning) depending on the context. The branch is shown in Figure 3.

[0041] For clarity, it should be noted that only the use of the recognition method for one branch is presented, keeping in mind that the recognition method is preferably applied to each branch present in the cerebral vascular tree.

[0042] Also, the branches to which the recognition method is applied are assumed to be branches located along the circle of Willis arteries. In fact, this location is where 85% of saccular intracranial aneurysms occur, leading to a more accurate prediction of aneurysms.

[0043] In this embodiment, the recognition method includes three steps: a generation step, a training step, and an inference step.

[0044] During the generation step, the system 10 generates a synthetic image of at least one branch of the blood vessel tree.

[0045] To this end, the system 10 performs several steps for each synthetic image to be generated.

[0046] Here, the system 1 performs a receiving step, a modeling step, and a generating step.

[0047] During the receiving step, the system 10 receives at least one image including at least one branch of the cerebral blood vessel tree.

[0048] Such an image is named a "real image" in contrast to the image generated at the end of the generation step.

[0049] This image is an image of the subject's brain, and the image was acquired by an imaging technique.

[0050] This imaging technique is, for example, the MRA-TOF technique. MRA is the abbreviation of magnetic resonance angiography, and TOF is the abbreviation of time of flight.

[0051] However, other imaging techniques such as magnetic resonance imaging (MRI), digital subtraction angiography (DSA), or computed tomography angiography (CTA) may be considered.

[0052] In any case, the image forms a three-dimensional (3D) image of the cerebral blood vessel tree.

[0053] In the modeling step, the system 10 attempts to represent the real image by an imaging model having a specific set of values for a set of parameters.

[0054] In other words, the real image is represented by a model that depends on parameters, and system 10 searches for the values of the parameters.

[0055] The idea is to generate a synthetic image by changing one or several parameters.

[0056] This means that the model represents the image and its content relatively accurately, and the parameters can be easily changed to obtain a new image that is realistic even if it is synthetically generated.

[0057] For this purpose, the model is selected here, especially for this purpose.

[0058] As shown below, several models can be considered at this stage.

[0059] However, all of these models share at least some features in that the image model includes at least a geometric model of the branches. This geometric model is a three-dimensional model of the branches and includes a graph of a vascular tree that is a set of nodes connected by weighted branches.

[0060] The geometric model is obtained by extracting a three-dimensional model of the branches from the real image.

[0061] Next, an example of an operation that enables the step of extracting the three-dimensional model of the branches is described.

[0062] The step of extraction includes the operation of binarizing the image.

[0063] In such an operation, a threshold is used to obtain voxels with high values (indicating the presence of blood) and low values (no detected elements).

[0064] More generally, several other types of segmentation operations can be considered here. In particular, the segmentation operation enables classifying voxels into two or more categories.

[0065] The extraction step further includes a skeletonization operation.

[0066] This operation consists of using an octree structure (a set of 3*3*3 pixels).

[0067] When the octree structure is filled with voxels having high values, the voxels with high values connected to the octree structure are set as candidates for removal.

[0068] After testing that their removal does not affect the connectivity of the skeleton, the candidate voxels are removed. If their removal does not maintain the connectivity of the skeleton, the candidate voxels are not removed.

[0069] By repeatedly performing this operation and changing the position of the octree structure so as to sweep across the entire image, a skeleton is obtained.

[0070] The extraction step also includes an analysis operation applied to the skeleton by using a second technique.

[0071] The second technique makes it possible to obtain a graph of a combined and non - directional blood vessel tree.

[0072] By definition, a graph is a set of nodes connected by branches.

[0073] The second technique consists of detecting the branches of the skeleton and setting the ends of each branch to be nodes.

[0074] At the end of the extraction step, a three - dimensional model of the branches is thus obtained.

[0075] Some more sophisticated imaging models may be considered.

[0076] As a first example, the parameters of the geometric model further include the diameter of the branch.

[0077] Arteries consist of various cell layers. The innermost layer is called the intima. The innermost layer is in direct contact with the blood flow.

[0078] The applicant has chosen to define the artery diameter as the artery diameter within the inner arterial wall.

[0079] System 10 obtains the diameter value by applying a convolution kernel to the real image.

[0080] In this case, this means that System 10 performs convolution with a three-dimensional spherical kernel on each branch and all branches. The spherical radius of the applied spherical kernel varies to model the diameters of various branches. As a second example, the geometric model defines a reference point of the branch, and the geometric model includes an interpolation function that connects the reference points. Each interpolation function is a function defined by coefficients, and the coefficients are parameters of a set of parameters.

[0081] Such reference points are the three-dimensional coordinates of the skeleton of the artery, that is, the centerline of the three-dimensional tube obtained in the dividing step.

[0082] System 10 interpolates these reference points by any appropriate function. In particular, a completely linear branch is obtained by fitting with a linear function.

[0083] As a specific example, System 10 interpolates the branch by a B-spline function. This means that System 10 searches for the fitting of the centerline using a three-dimensional spline function.

[0084] The B-spline function can be represented by three different characteristics, the knot points that define the intervals of the chunks where the polynomials are defined, the B-spline (or polynomial) coefficients, and the degree of the spline (i.e., the degree to which the fit is performed).

[0085] Here, the variable parameter of the B-spline function is the coefficient of the B-spline.

[0086] As a third example, the image model includes a noise model, and the noise model models the noise of the image by Gaussian noise with a given mean value and standard deviation.

[0087] In such a case, the standard deviation of the Gaussian noise is one of the parameters of the model.

[0088] To generate plausible background noise, it is desirable to use a Gaussian mixture model, for example, to collect the statistical characteristics of the ambient noise on a plurality of images.

[0089] Such collection can be performed by collecting the characteristics within the image. This is relevant when it has been determined to generate the image very accurately.

[0090] Another means is to provide an expected value. Such an expected value can be determined, in particular, by the type of device that enabled the acquisition of the image. The first type of device corresponds to a first set of expected values, the second type of device corresponds to a second set of expected values, and the second set of expected values is different from the first set of expected values.

[0091] As a fourth example, the image model includes a background model, and the background model includes shapes with several different values.

[0092] More precisely, when having two values, the background model aims to model the various components of the brain, namely white matter / gray matter, cerebrospinal fluid (CSF), ventricles or corpus callosum.

[0093] These different elements within the brain exhibit various radiopacity and, thus, various gray-level intensities.

[0094] In the model, it is considered as the background noise of two materials. The darker material (corresponding to cerebrospinal fluid, ventricles and / or corpus callosum) and the brighter material (white matter / gray matter). The creation of this dual background noise was performed via two steps.

[0095] First, the geometry of the darker part was modeled. For this task, system 20 generates a strong distortion of a branching shape (using elastic deformation as described below), and this highly distorted pattern functions as a mask for the darker noise.

[0096] In the second step, all voxels that do not belong to either the darker noise part or the arteries were assigned to the brighter noise region. Proceeding in this way, the model is forced to immerse the arteries in the darker material, and in fact most of the arteries of the circle of Willis can be found there.

[0097] Synthetic aneurysms can also be easily modeled by constructing binary spheres incorporated into binary bifurcations. The ICA can withstand geometric deformations, but in the human brain the ICA generally exhibits a high level of distortion. Regarding the spatial position of the ICA, we aim to place the ICA approximately on the angle bisector. Thanks to the 3D graph extraction from the vascular skeleton, system 10 can start from the bifurcation center and collect the 3D direction vectors following each of the three arteries (branches of the bifurcation).

[0098] Starting from the bifurcation center (3D graph node), three vectors tangent to the centerlines of the three branches

Number

Number

[0099] The system 10 arranges the ICA on the axis of the parent branch as much as possible. Therefore, the system 10 calculates the direction

Number

[0100] The system 10 uses the following formula. The system 10 determines the distance D separating the branch center and the aneurysm center (

Number

Number

Number

Number

[0101] System 10 is considered such that the angle cannot be smaller than 5° (π / 36 rad). Thus, the smaller the angle between the two branches, the deeper the ICA is embedded. Conversely, in the case of the maximum angle configuration (180°), the ICA is approximately from the branch center to [Number] arranged in a voxel, i.e., the aneurysm is in contact with the branch wall.

[0102] According to another embodiment, the distance D separating the branch center and the aneurysm center is given by the following equation. [Number] where ● [Number] represents the angle formed by the two daughter arteries, and here it is assumed to be as follows [Number] ● R is the average radius of the branches forming the bifurcation. Thus, R is [Number] equal to.

[0103] In a more sophisticated model, the growth parameter γ is considered and the distance D is as follows. [Number]

[0104] The growth parameter γ is composed between 0 and 1.

[0105] The use of growth parameters makes it possible to simulate different behaviors of aneurysms.

[0106] Alternatively, if the distance D is defined as the distance from the center of the aneurysm to the junction point (blood vessel wall) of the two branches, the formula is as follows.

Number

[0107] Such calculations make it possible to automatically adjust the shift from the aneurysm center and the blood vessel wall where the daughter artery divides.

[0108] In a variant, for each calculation of the distance D, the direction is the angle bisector

Number

Number

[0109] It can be easily understood that the previous examples can be freely combined as needed to form a new model.

[0110] In particular, combinations of the previous four examples can be considered to obtain a more detailed model.

[0111] During the generating step, the system 10 generates an image corresponding to the image model using the set of modified values. The image generated by the system 10 is a composite image.

[0112] Here too, depending on the image model, the generating step is performed in different ways.

[0113] For each example given in the modeling step, some examples are given below.

[0114] In a first example, during the generation step, geometric distortion is applied to the geometric model. This means in particular that the geometric distortion is not applied to other models, in particular noise models.

[0115] For this purpose, elastic deformations such as can be found in the elastic deformation library (https: / / elasticdeform.readthedocs.io / ) can be applied.

[0116] Furthermore, it is also possible to apply various kernels (or more precisely, various elastic deformations to the same kernel) along the centerline of the blood vessel.

[0117] In a second example, when the geometric model includes an interpolation function, in particular a B-spline function, the values of the coefficients are modified. In fact, slightly changing the values of these parameters distorts the position of the centerline coordinates.

[0118] Specifically, here the system 10 changes the coefficients of the polynomial. When the centerline of the blood vessel is adjusted via the change of the spline function, the system 10 collects the diameters of all the arteries considered within the three-dimensional crop. Each centerline (morphological skeleton) first adjusted by the change of the spline can thus pass through a convolution with a spherical kernel sized to fit the corresponding observed diameter. The system 10 then thickens each artery according to the measured anatomical characteristics.

[0119] Such an approach not only allows to substantially control the shape of the arteries, but also to adjust their thickness and to maintain a good balance between the various branches of a given bifurcation.

[0120] An example of B-spline interpolation is shown in FIG. 4. The left panel shows the 3D representation of three different aspects for a given bifurcation. The grey solid line represents the actual coordinates of the branches of the bifurcation as collected within the MRA-TOF acquisition, the black dashed line represents the spline function that best represents the artery, and finally, the black dotted curve represents the modified spline function (the new centerline of the bifurcation).

[0121] The right panel represents the B-spline coefficients, the gray bars represent the coefficients returned by interpolation (i.e., the best fit for modeling the artery), and the black bars represent the coefficients modified according to a constant intensity parameter, i.e., the coefficients multiplied by some weights (here, the weights are set to 5 in the upper panel and 15 in the lower panel).

[0122] Specifically, the weights are applied as follows. To the first B-spline coefficient (the gray bar in FIG. 4), the system 10 adds a random value (in the range) multiplied by a weight (e.g., in the range [5, 20] where reasonable geometric distortion is seen).

Number

[0123] According to another embodiment, the B-spline coefficient is the result of multiplying the initial B-spline coefficient by a constant greater than 1.

[0124] For example, the constant is equal to the modified B-spline coefficient being the initial B-spline coefficient plus X% of the initial B-spline coefficient, where X is between 5% and 30%.

[0125] In a third example corresponding to the noise model, the standard deviation changes from a first value

Number

Number

[0126] Therefore, the system 10 adjusts the filter standard deviation applied to the real image

Number

Number

Number

Number

[0127] More precisely, in this example, the filter standard deviation

Number

Number

Number

Number

[0128] The fact that such a value enables the desired second value

Number

[0129] When passing through the Gaussian blur, the input image I(x, y) is filtered as follows.

Number

[0130] Bienaymé's identity is stated as follows.

Mathematics

[0131] Therefore, the variance of the linear combination is as follows.

Mathematics

[0132] However, when X i ,..., X n are mutually independent integrable random variables, that is

Mathematics

Mathematics

[0133] Therefore, the variance of the input image is as follows.

Mathematics

[0134] The goal here is to estimate the variance of the output (filtered) image.

Mathematics

[0135] Therefore,

Mathematics

[0136] [Number] When it is large, the square of the Gaussian distribution is smooth, and its sum can be approximated as follows. [Number]

[0137] Therefore, [Number]

[0138] In summary, the standard deviation [Number] An image composed of Gaussian noise with [Number] When filtered by a Gaussian filter with [Number] has

[0139] In the fourth example using the background model, the region can be changed by geometric transformation.

[0140] Here too, it can be easily understood that any combination of the previously described modifications can be considered.

[0141] At the end of the modification step, the system 10 obtains an image based on the same model that is a real image but has different parameters.

[0142] In some cases, this synthetic image can be improved by the modification step.

[0143] For example, it may happen that the tip of a branch located on a bifurcation can be slightly separated from the other two arteries. In other words, at the bifurcation node, any one of the three arteries no longer combines with the other arteries, and the system 10 addresses this problem simply by positioning the new tip coordinates and centering the entire set of coordinates.

[0144] At the end of the generating stage, a set of synthetic images is thus obtained. Such a set can include a very large number of synthetic images thanks to the ease of changing one or some values of the parameters of the image model.

[0145] The training stage is a stage of training a recognition predictor adapted to obtain bifurcation recognition data in the input image.

[0146] This makes it possible to obtain a trained recognition predictor.

[0147] Here, the trained recognition predictor is adapted to determine whether a bifurcation belongs to one type of bifurcation of interest and to indicate which bifurcation it belongs to.

[0148] The recognition predictor is, for example, a neural network.

[0149] In particular, the neural network can be a three-dimensional convolutional neural network such as a Unet network.

[0150] This makes it possible to obtain a trained recognition predictor.

[0151] The training stage includes a forming step and a training step.

[0152] During the forming step, the system 10 forms a training data set based on the synthetic images.

[0153] Such formation can be done by using only synthetic images or can be added to the dataset from real images.

[0154] For example, system 10 randomly selects synthetic images from all synthetic images to form a training dataset.

[0155] Annotations can be obtained by getting values from the model.

[0156] This makes it possible to generate a training dataset that is automatically annotated.

[0157] During the training step, system 10 trains recognition predictors by using the training dataset according to any unsupervised learning technique.

[0158] System 10 thus obtains trained recognition predictors.

[0159] During the inference step, system 10 receives the real image to be analyzed, and the real image to be analyzed is an image of the subject's vascular tree.

[0160] The image is taken, for example, by MRA-TOF technology.

[0161] System 10 then applies the recognition predictors trained on the image to be analyzed to obtain branch recognition data.

[0162] The method thus corresponds to a complete synthetic model of the three-dimensional cerebral arteries and branches. By constructing this model, the goal is to provide a substantial dataset of cerebral arteries that can be used by a three-dimensional neural network to distinguish or detect / recognize some components of the cerebral vascular system.

[0163] This method reduces manual labeling as much as possible and, in some cases, enables people to be freed from manual labeling. In other words, using thousands or tens of thousands of modeled branches to train recognition predictors can provide excellent performance using only about 100 or 200 actual TOF segments. This is especially because there are large variations in the anatomical structures of the vascular system.

[0164] Therefore, the data augmentation performed here is not simple because simple data augmentation can distort geometric or statistical properties in an undesirable way, i.e., it can move the augmented images far from their actual images. Thus, the proposed data augmentation is made smart by using a specific model.

[0165] It should also be noted that such an effect is obtained independently of the nature of the branch recognition data. This means that the predictors can be adapted to predict other data.

[0166] In particular, the data for branch recognition can be selected here from among the following elements. - The presence of a branch - The position of a branch - The branch angle of a branch: The branch angle can be defined as a combination of two angles named A1 and A2 in Figure 3. The first angle A1 is the angle between the main artery and the first artery, and the second angle A2 is the angle between the main artery and the second artery. - The geodesic distance between two branches: By definition, the geodesic distance between two consecutive branches is the geodesic distance between the centers of the two branches. The geodesic distance between two points of a branch is the length of the path traced in the graph to connect both, and is the number of voxels between these two points. - The cross-sectional area of the detected branch - The twist parameter of a branch.

[0167] Therefore, the method of recognition is an accurate method for obtaining branch recognition data.

[0168] This was confirmed by experiments conducted by the applicant.

[0169] Many applications of this recognition method are conceivable. Some of them are shown below.

[0170] An application connected to an aneurysm has been specifically developed. If appropriate, the application is an in-vitro application.

[0171] The first example of an application is a method of prediction.

[0172] For this application, a method of predicting that a subject has a risk of developing an aneurysm is proposed.

[0173] The method of prediction includes the step of performing the steps of the recognition method to obtain branch recognition data.

[0174] The method of prediction also includes the step of predicting, based on the obtained data, that the subject has a risk of developing an aneurysm.

[0175] The second example of an application is a method of diagnosis.

[0176] This application corresponds to a method of diagnosing an aneurysm in a subject.

[0177] The method of diagnosis includes the step of performing the steps of the recognition method to obtain branch recognition data.

[0178] The method of diagnosis also includes the step of performing the step of diagnosing an aneurysm based on the obtained data.

[0179] The third example of an application is a method of treatment.

[0180] This application corresponds to a method of treating an aneurysm.

[0181] The method of treatment includes the step of performing the steps of the recognition method to obtain branch recognition data.

[0182] The method of treatment also includes performing a step of administering an agent for treating an aneurysm determined based on the obtained data.

[0183] A fourth example of an application is a method of identifying a treatment target.

[0184] This application corresponds to a method of identifying a treatment target for preventing and / or treating an aneurysm.

[0185] The method of identification includes performing a step of a method of recognizing at least one branch of the vascular tree of a first subject in order to obtain first branch recognition data, where the first subject is a subject suffering from an aneurysm.

[0186] The method of identification also includes performing a step of a method of recognizing at least one branch of the vascular tree of a second subject in order to obtain second branch recognition data, where the second subject is a subject not suffering from an aneurysm.

[0187] The method of identification further includes a step of selecting a treatment target based on a comparison between the first branch recognition data and the second branch recognition data.

[0188] A fifth example of an application is a method of identifying a biomarker.

[0189] The biomarker can be one of a diagnostic biomarker for an aneurysm, a sensitivity biomarker for an aneurysm, a prognostic biomarker for an aneurysm, or a predictive biomarker for responding to treatment of an aneurysm.

[0190] The method of identification includes performing a step of a method of recognizing at least one branch of the vascular tree of a first subject in order to obtain first branch recognition data, where the first subject is a subject suffering from an aneurysm.

[0191] The method of identification includes performing the steps of a method for recognizing at least one branch of the vascular tree of a second subject in order to obtain second branch recognition data, where the second subject is a subject not suffering from an aneurysm.

[0192] The method of identification includes the step of selecting a biomarker based on a comparison between the first branch recognition data and the second branch recognition data.

[0193] The sixth example of application corresponds to a method for screening a compound.

[0194] The medicament has an effect on a known therapeutic target for preventing and / or treating an aneurysm.

[0195] The method of screening includes performing the steps of a method for recognizing at least one branch of the vascular tree of a first subject in order to obtain first branch recognition data, where the first subject is a subject suffering from an aneurysm and administered with a compound.

[0196] Here, the term "administer" includes any method of administering a medicament.

[0197] The method of screening also includes performing the steps of a method for recognizing at least one branch of the vascular tree of a second subject in order to obtain second branch recognition data, where the second subject is a subject suffering from an aneurysm and not administered with a compound.

[0198] The method of screening further includes selecting a compound based on a comparison between the first branch recognition data and the second branch recognition data.

Claims

1. A method for recognizing at least one branch of a vascular tree in a real image of a subject's vascular tree, particularly in the brain, and a method performed by a computer, - A step of generating a composite image of at least one branch of the vascular tree, wherein the step of generating said image includes the following steps for each composite image: 〇 A step of receiving a real image that includes at least one branch of the vascular tree, 〇 A step of modeling the real image by an image model having a specific set of values ​​for a set of parameters, wherein the image model includes at least the geometric model of the branching, the geometric model being a three-dimensional model of the branching, and including a graph of the vascular tree, the graph being a set of nodes connected by weighted branches, and the geometric model being obtained by segmenting the real image. 〇 A step of generating an image corresponding to the image model having a set of modified values, wherein the generated image is the composite image. - To obtain a trained recognition predictor, the process involves training a recognition predictor adapted to obtain branch recognition data in an input image, wherein the training step includes the following steps: ○ A step of forming a training dataset based on the aforementioned composite image, ○ A step of training the recognition predictor by using the aforementioned training dataset, - The estimation stage, and the estimation stage includes the following steps: 〇 A step of receiving the actual image to be analyzed, wherein the actual image to be analyzed is an image of the vascular tree of the subject. ○ To obtain branch recognition data, the trained recognition predictor is applied to the image to be analyzed. A method of recognition that includes this.

2. The image model includes a noise model, the noise model models the noise in the image using Gaussian noise having a standard deviation, the standard deviation of the Gaussian noise is one of the parameters of the model, and the standard deviation is equal to a first value. The recognition method according to claim 1, wherein during the generating step, a Gaussian filter having a standard deviation is applied to the real image to obtain an image having Gaussian noise having a standard deviation having a second value, the second value being different from the first value, and the standard deviation of the Gaussian filter depending on the first value and the second value.

3. The parameters of the geometric model further include the diameter of the branch, The aforementioned diameter value is obtained by applying a convolution kernel to the real image. The recognition method according to claim 1.

4. During the generation step, geometric distortion is applied to the geometric model. The recognition method according to claim 1.

5. The recognition method according to claim 1, wherein the geometric model defines the reference points of the branching, the geometric model includes an interpolation function that connects the reference points, each interpolation function is a function defined by coefficients, the coefficients are parameters of the set of parameters, and the coefficients are modified during the generating step.

6. The recognition method according to claim 5, wherein each interpolation function is a B-spline function defined by a B-spline coefficient, and the coefficient is the B-spline coefficient.

7. The recognition method according to claim 5, wherein the value of the coefficient is modified by adding a random value that is weighted to the specific value.

8. The recognition method according to claim 1, wherein the image model includes a background model, and the background model includes a shape having two distinct values.

9. The recognition method according to claim 1, wherein each image is captured by MRA-TOF technology.

10. The branch recognition data is selected from the following elements: ○ Branching class or type, ○ The existence of the aforementioned branch, ○ The location of the aforementioned branch, ○ The branching angle of the aforementioned branch, 〇 Geodetic distance between two branches, ○ The cross-sectional area of ​​the detected branch, and ○ Torsion parameter of the aforementioned branch, The recognition method according to claim 1.

11. The recognition method according to claim 1, wherein the recognition predictor is a neural network.

12. The recognition method according to claim 11, wherein the neural network is a convolutional neural network.

13. A method comprising performing the steps of a method for recognizing at least one branch of a vascular tree in a real image of a subject's vascular tree, wherein the method is the method according to any one of claims 1 to 12. - A method for predicting that a subject is at risk of developing an aneurysm, the method of prediction comprising at least the following steps: ○ To obtain branch recognition data, the step of performing the steps of the recognition method, ○ A step of predicting that the subject is at risk of developing the aneurysm based on the branch recognition data obtained above. - A method for diagnosing an aneurysm, the method for diagnosing the aneurysm comprising at least the following steps: ○ To obtain branch recognition data, the step of performing the steps of the recognition method, ○ A step of diagnosing the aneurysm based on the branch recognition data obtained above, - A method for identifying therapeutic targets for preventing and / or treating aneurysms, the method comprising at least the following steps: 〇 To obtain the first branch recognition data, the steps of the identification method are performed on a first subject, where the first subject is a subject suffering from an aneurysm. ○ To obtain second obtained branch recognition data, the steps of the identification method are performed on a second subject, wherein the second subject is a subject that does not have an aneurysm. ○ A step of selecting a treatment target based on a comparison of the first obtained branch recognition data and the second obtained branch recognition data, - A method for identifying a biomarker, wherein the biomarker is a diagnostic biomarker for an aneurysm, a susceptibility biomarker for an aneurysm, a prognostic biomarker for an aneurysm, or a predictive biomarker for the treatment of an aneurysm, and the method comprises at least the following steps: 〇 To obtain the first obtained branch recognition data, the step of performing the steps of the identification method on a first subject, where the first subject is a subject suffering from an aneurysm. ○ To obtain a second branch recognition data, the step of performing the steps of the identification method on a second subject, wherein the second subject is a subject that does not have an aneurysm. 〇 A step of selecting a biomarker based on a comparison between the first determined parameter and the second determined parameter, and - A method for screening compounds useful as probiotics, prebiotics, or pharmaceuticals, wherein the compounds are effective against known therapeutic targets for the prevention and / or treatment of aneurysms, and the method comprises at least the following steps: 〇 To obtain the first obtained branch recognition data, the step of performing the steps of the identification method on a first subject, wherein the first subject is a subject suffering from the aneurysm and administered the compound. 〇 To obtain a second obtained branch recognition data, the step of performing the steps of the identification method on a second subject, wherein the second subject is a subject suffering from the aneurysm and not administered the compound. ○ A step of selecting a compound based on a comparison of the first determined parameter and the second determined parameter, A method of selecting from a list composed of these elements.

14. The computer program product includes instructions for performing the steps of the method according to any one of claims 1 to 12 when the computer program product is executed on a suitable computer device.

15. A computer-readable medium on which the computer program described in claim 14 is encoded.

Images