Machine learning-based analysis of PET images

Abstract

An apparatus and method for image reconstruction for medical images is provided, the method including obtaining a PET image, dividing the PET image into local subset images, each subset image being analyzed by a trained machine learning system, obtaining an output for each of the subset images processed by the machine learning system, and determining a representation output based on the outputs.

Description

[Technical Field]

[0001] This invention relates to the analysis and reconstruction of medical images using neural networks and machine learning, and more particularly to apparatus and methods for the analysis and reconstruction of positron emission tomography (PET) images. [Background technology]

[0002] The use of image reconstruction, enhancement, and analysis is widely employed in images associated with medical scanners. For example, in PET scans, a radioactive tracer is injected into the patient, configured to allow cancer cells to accumulate the tracer material. A PET scanner is then used to record the radioactive decay, providing a reconstructed PET image that represents the in-situ distribution of tracer intensity. High tracer intensity can be an indication of a cancerous lesion.

[0003] Clinical oncology makes extensive use of positron emission tomography (PET) to detect, observe, and image tumors and metastases. Medical imaging techniques are increasingly utilized in diagnosis. Furthermore, PET scans are used as an important tool for the clinical diagnosis of brain diseases, and generally for mapping human cardiac function to the brain.

[0004] However, PET images are contaminated with blur and noise related to how the PET data was recorded. To remove blur and noise, and to enable the analysis of tracer intensity features in situ, a Bayesian approach to image reconstruction can be presented as a probabilistic inverse problem.

[0005] See the reference "Tarantola, A., & Valette, B. (1982). Inverse problems = quest for information. Journal of geophysics, 50(1), 159-170".

[0006] In the Bayesian approach, prior information, quantified using the prior probability distribution ρ(m) of the expected tracer intensity, is combined with a description of how well the model's forward response, quantified using the likelihood function L(m), fits the observed data. The probabilistic formulation of the inverse problem / the general solution to Bayes' formula is a probability distribution, i.e., the posterior probability distribution σ(m). σ(m) = ρ(m) * L(m) (1)

[0007] With the exception of the linear Gaussian inverse problem, an analytical description of σ(m) is not feasible. Instead, there are sampling methods based on Markov chain Monte Carlo methods, such as the Metropolis-Hastings algorithm, that enable sampling of σ(m). While these methods are guaranteed in principle to sample the correct posterior probability distribution σ(m), they are computationally intensive and the results are difficult to interpret, making them practically impractical.

[0008] While continuous efforts are being made to improve the image quality of medical images, particularly PET image processing techniques, further improvements in PET image quality are still desired.

[0009] Therefore, improved image reconstruction is beneficial, and in particular, more efficient and / or more reliable methods of image reconstruction would be beneficial. [Overview of the Initiative]

[0010] The present invention aims to provide improvements in the quality and speed of Bayesian image analysis and reconstruction of PET images.

[0011] In particular, it may be considered an objective of the present invention to solve some of the aforementioned problems of the prior art and to provide a method for improving tumor detectability by performing medical image analysis and reconstruction faster with image quality equivalent to known Bayesian methods.

[0012] Therefore, in the first aspect of the present invention, the above objective and several other objectives are intended to be achieved by providing a faster method for image reconstruction of medical images of a subject. The subject may be a human or an animal.

[0013] The present invention is particularly useful for improving the detection and identification of small tumors, but is not limited thereto.

[0014] The object of the present invention is to analyze and, in some cases, reconstruct images scanned by a PET scanner. PET images and reconstructed PET images (hereinafter referred to as PET image dobs) can be two-dimensional or three-dimensional images encompassing a portion of the human body. In this invention, the PET image is divided into multiple small localized images (hereinafter referred to as subset image dsso). A subset image dsso is a localized image that encompasses a small portion of the complete PET image dobs.

[0015] A machine learning system, preferably a neural network, is used in this invention to analyze PET images. The analysis of PET images is performed by using subset images (dsso) as input to the machine learning system and analyzing multiple subset images (dsso) one at a time.

[0016] To train a machine learning system, a training set (represented by ρ(m)) is created based on the collection of expert data. This data is then used to create a training set that will enable the machine learning system to improve the quality of PET images of humans or animals, and the characteristics of the posterior probability distribution σ(m) are evaluated to analyze potential cancers in PET images.

[0017] Here, we present a method for utilizing machine learning. A training set for a machine learning algorithm is generated based on the use of arbitrary composite prior information (hereafter referred to as prior information ρ(m)). This method is based on the selection of a model m from the prior information, which is used to create a data image d, and the data image d is used as input for the machine learning system to represent a subset image dsso.

[0018] Model m and data image d represent a subset similar to the subset image dsso from PET images, and are therefore called subset model mss and subset image dss.

[0019] forward problem This method relies on an understanding of the physical processes that generate the observed PET image dobs, including blur and noise. For example, model m represents the in-situ distribution of tracer intensity, and the acquired image d can be calculated by the following process. d = g(m) + n(m) (2) Here, "g" represents a function applied to blur or smoothing depending on the scanner and reconstruction method used, and "n" represents a function that generates noise according to a specific noise model. d represents an example of an image that might be observed for a particular model m.

[0020] For local subset image DSS, dss = g(mss) + n(mss) Here, mss is the in-situ distribution of tracer intensity in this local subset image dss.

[0021] "g" represents blur or smoothing resulting from a specific scanner type and reconstruction method, and is related to the point spread function. "g" is typically represented by the linear operator GPSF, and the forward problem can be described as follows: d = GPSF(m) + n(m) (3) Alternatively, in local subset images, dss = GPSF(mss) + n(mss)

[0022] In practice, the noise model can be obtained by scanning known targets and calculating the residuals between the PET image obtained from the PET scanner and the PET image calculated using GPSF(m). These residuals represent one manifestation of the noise, from which a statistical model of the noise can be inferred. Similarly, the averaging function GPSF(m) can be obtained by analyzing the residuals between the obtained PET image and the known target. Typically, the averaging function and the noise model are inferred simultaneously.

[0023] Therefore, if the subset model mss is known, the subset image dss can be calculated. In this invention, the subset model mss for PET images is not known, but the subset image dsso can be obtained from the PET images. Therefore, a method is needed to obtain the subset model mss, or the features of the subset model mss, from the subset image dsso. A machine learning system is created and trained to find the features of the subset model mss.

[0024] Prior information ρ(m) To obtain a machine learning system that can solve the inverse problem by starting with PET image data and ultimately acquiring output data, the first step is to generate data for training the machine learning system. For this purpose, the collected expert data is used to represent prior information ρ(m). Prior information represents prior knowledge about the model parameters. Prior knowledge can come from expert insights, previous research, and similar sources. The output data generated by the machine learning system, called the representation output, can be an image or statistical data that evaluates the features of the PET image.

[0025] Experts, such as physicians who diagnose cancer from PET images, can typically collaborate with data specialists to create expert data. For example, an expert might create or select a set of images for expert data. The images selected by the expert could be human images showing tumors or metastases, or other types of relevant cell structures. Based on the expert data, the data specialist can create statistical models that represent prior information describing the differences in those images. Alternatively, the data specialist can create prior information from previous studies, assumed statistical models, or other data sources.

[0026] The method of quantifying prior information using spatial correlation and sampling from that prior information is described in the references. Deutsch, CV, & Journel, AG (1992). Geostatistical software library and user's guide. New York, 119(147). Mariethoz, G., & Caers, J. (2014). Multiple-point geostatistics: stochastic modeling with training images. John Wiley & Sons.

[0027] Prior information ρ(m) does not need to exist as a mathematical model (although it may exist), but is instead represented by the selection of algorithms and statistical models. The explicit selection of prior information is quantified using the embodiment generated by the selected algorithms and statistical models.

[0028] Sample M*, model image m*, and data image d* The example images are generated as embodiments of prior information, are called model images, denoted by mss*, and form a sample M*. Upon completion, sample M* contains a large number of model images mss* distributed according to ρ(m). Sample M* may contain 1,000, 10,000, 100,000, or any other appropriate number of model images mss*.

[0029] mss* represents a model image derived from sample M*, embodying the prior information. In principle, the model image mss* is free from noise and blur.

[0030] For each model image mss*, the corresponding noise-free data image dss* is calculated by evaluating the forward model (Equation 3). dss*=Gpsf(mss*)

[0031] By adding noise to the data image dss*, a sim-data-image (dss,sim*) can be obtained using the noise model n. dss,sim*=Gpsf(mss*)+n(mss*)

[0032] dss* represents a data image based on the model image mss*. Multiple data images dss* form a data sample D*. The sim data image dss,sim* is obtained by adding noise to the data image dss*. The sim data image dss,sim* forms a pseudo-data sample Dsim*.

[0033] The data image dss* is obtained by quantifying and smoothing the original model image mss* using the smoothing operator g, i.e., GPSF. The data image dss* is one representation of the model image mss*, and blur has been added. The sim data image dss,sim* is another representation of the model image mss*, and both blur and noise have been added.

[0034] A model image mss* can represent NM pixels, which can be the same number of pixels as the data image dss*, sim data image dss,sim*, and subset image dsso. However, a model image mss can also have more pixels than the data image dss*, sim data image dss,sim*, and subset image dsso. In other words, a model image mss* can be represented using a resolution that is finer or coarser than the resolution of the data image dss* and sim data image dss,sim*, and a model image mss* can have more pixels than the data image dss*.

[0035] Model images (mss*), data images (dss*), sim data images (dss,sim*), and subset images (dsso) are representations of a single image containing 1, 10, 100, 1000, or any other appropriate number of pixels. Pixels can be arranged in a three-dimensional configuration, for example, a 9x9x9 image, or in a two-dimensional configuration, for example, a 9x9 image.

[0036] The number of pixels represented in the model image mss*, data image dss*, sim data image dss,sim*, and subset image dsso is denoted by np. The model image mss* = [mss1, mss2, ..., mssn] is a vector representing np model parameters that represent the actual tracer intensity within np pixels. In other words, the model image mss* represents an image consisting of a total of 81 pixels, for example, a 9x9 image, and therefore the model image mss* is a vector consisting of np = 81 in this example. The same applies to the subset image dsso, sim data image dss,sim*, and data image dsso.

[0037] The model generated from prior information is called a model image (mss*), and the data image generated from the model image (mss*) using a forward model is called a data image (dss*) if there is no noise, and is called a sim data image (dss,sim*) if it contains noise.

[0038] The generated set of multiple sim data images dss,sim* forms a pseudo-data sample Dsim*. Upon completion, the pseudo-data sample Dsim* contains as many sim data images dss,sim* as the model image represented by mss* in M*.

[0039] Sim data images dss,sim* are used to train a machine learning system, and the trained machine learning system is used to analyze PET images by evaluating the properties of σ(mss) from local subset images dsso of the PET images.

[0040] Machine learning systems When creating a training set that includes simulation data images dss,sim*, a mapping is performed from the simulation data images dss,sim* to the model images mss*, or from the simulation data images dss,sim* to the features of the model images mss*. The mapped features of the model images mss* are the expected output mf* for the machine learning system. Thus, the training set contains both the simulation data images dss,sim* and the expected output mf*, respectively.

[0041] The expected output sample Mf* is the set of all expected output mf*. The training data [Dsim*;Mf*] is the set of all training sets [dss,sim;mf*]. The training data [Dsim*;Mf*] is used to train the machine learning system to learn the mapping from the pseudo-data sample Dsim* to the expected output sample Mf*.

[0042] While training a machine learning system, the characteristic expected output mf* may be identical to the model image mss*, but it may also be a subset of the model image mss*, for example, the intensity of the central pixel in the model image mss*, or the probability that the central pixel belongs to a specific category, such as "cancer". The characteristic mf* is the expected output from the machine learning system when the input is sim data images dss,sim*. The expected output mf* may also be a statistical function, such as a normal distribution, in which case the expected output mf* is a vector with two values, for example, the mean and covariance of a normal distribution describing the probability of pixel intensity.

[0043] When a machine learning system is being trained, the output mo generated by the machine learning system is compared to the expected output mf*. During training, the output mo is calculated for all sim data images dss,sim* in the training set and compared to the expected output mf*. The machine learning system is trained to minimize the difference between the output mo and the expected output mf* according to the cost function C(mf*, mo). Training is complete when the cost function is minimized.

[0044] The cost function is typically chosen to represent the log-likelihood of the feature being evaluated. If mf* represents the mean and covariance (N(m0f*, Cf*)) of a pair of pixels in the model image, then minimize the following cost function: C(mf*, mo)=-0.5((m0f*-mo)Cf*-1(m0f*, mo)') This leads to an assessment of the posterior mean and covariance of σ(mf*). If mf* represents the probability of a particular outcome, σ(mf*) can be fully explained by using categorical cross-entropy.

[0045] The introduction of machine learning and the use of cost functions are explained, for example, below: Bishop, Christopher M. Pattern, Recognition and machine learning. Springer, 2006.

[0046] PET images are analyzed by a machine learning system and reconstructed by analyzing each of the subset images (dsso) extracted from the PET images. When the trained machine learning system analyzes the subset images (dsso) from the PET images, it seeks to map each subset image (dsso) to a subset model (mss), or to map the subset images (dsso) to features of the subset model (mss).

[0047] A trained machine learning system receives a subset image dsso as input. The trained machine learning system evaluates one subset image dsso at a time, and generates an output mo for each subset image dsso. The output mo can represent the complete subset model mss or features of the subset model mss. The subset model mss may not be completely determined, but the machine learning system can determine features of the subset model mss, such as the intensity of the central pixel or group of central pixels of the subset model mss, based on the subset image dsso.

[0048] The output mo can be a numerical value representing the intensity of a pixel or a group of pixels in the subset model mss. The output mo can also be a numerical value representing the number of pixels indicating cancer, or the number of pixels with a pixel intensity higher than the threshold pixel probability, that are concatenated to the central pixel of the subset model mss.

[0049] Alternatively, the output mo may represent a category, where if a certain condition is met, it is category A, and if a certain condition is not met, it is category B, or it may be the probability of being category A and the probability of being category B. The output mo may also be a probability value for a large number of categories. There may be two or more categories to which each category is assigned a probability. Category A may be the probability of "cancer", and category B may be the probability of "not having cancer".

[0050] Furthermore, the output mo may represent a statistical function, such as a normal distribution, in which case the output mo is a vector containing two values: the mean and variance of a normal distribution describing, for example, the probability of pixel intensity. Alternatively, mo may represent the mean and covariance of a multivariate normal distribution.

[0051] Therefore, in a first aspect of the present invention, the described objectives of the present invention and several other objectives are intended to be achieved by providing a computer-implemented method for image analysis and reconstruction of medical images of a subject, and this method is - Acquire a PET image and divide the PET image into multiple subset images (dsso), each of which is represented by one or more pixels. - Preparing a trained machine learning system, - Using subset image DSSOs as input to a trained machine learning system, - Obtain the output MO for each subset image DSSO from the machine learning system, - Determine and output the representation based on the output mo, Includes.

[0052] Each subset image dsso from the PET images is processed here by a machine learning system. For each subset image dsso, the output mo is obtained based on multiple output mos obtained for multiple analyzed subset images, and the representation output is determined and output. The representation output may be an image or a file of statistical data.

[0053] Therefore, preparing a trained machine learning system is - Obtaining a model image mss* from a sample M*, where the model image mss* is a manifestation of prior information ρ(m), where prior information ρ(m) is a statistical model based on expert prior data, and each model image mss* is an image representation consisting of one or more pixels. - The method involves obtaining a training set, each containing a sim data image dss,sim* and an expected output mf*, by determining the sim data image dss,sim* based on the model image mss*, and selecting the expected output mf* based on the model image mss* for each sim data image dss,sim*, wherein each sim data image dss,sim* is represented by one or more pixels. - The training set is used as input to train the machine learning system, and the output mo for each sim data image dss,sim* is obtained from the machine learning system. - Training the machine learning system until it converges based on a comparison of the output mo and the expected output mf*, Includes.

[0054] The machine learning system is trained using a training set. The training set is generated by obtaining sim data images dss,sim* and expected output mf*. For each model image mss* in which prior information ρ(m) is embodied, sim data images dss,sim* are generated by the equation dss,sim* = g(mss*) + n(mss*), where g(mss*) can be the function Gpsf(mss*).

[0055] The sim data images dss,sim* are used as input for training the machine learning system. Furthermore, the expected output mf* is an input to the machine learning system. The expected output mf* is generated from the model image mss* and can be a number representing a feature of the model image mss*, which is the intensity of the central pixel, or the number of pixels connected to the central pixel that indicate the probability of cancer, or the number of pixels with a pixel intensity higher than the threshold pixel intensity.

[0056] Alternatively, the expected output mf* may be the posterior probability of a certain category obtained from the model image mss*. There could be category A, which is the probability that the model image mss* indicates "cancer," and category B, which is the probability that the model image mss* indicates "not cancer." There may be more than two categories.

[0057] A machine learning system converges when the cost function of the training set for the machine learning system is minimized. The cost function is based on a comparison between the output mo and the expected output mf*.

[0058] Typically, a portion of the training data is extracted and called the test dataset. To ensure that the machine learning system performs well on new data, the system is trained on the training data [Dsim*;Mf*] and its performance is compared to that on the test dataset.

[0059] The training set is obtained by generating the model image mss* from prior information ρ(m), and then using a forward model to generate sim data images dss,sim* from the model image mss*. The training set is further obtained by generating the expected output mf* based on the model image mss*. Therefore, each training set contains the sim data images dss,sim* and the corresponding expected output mf*.

[0060] The training set may be stored in a table [Dsim*;Mf*], which serves as the training data. The table can also be stored in a data storage device, such as a database, to preserve it for future use so that the training data can be reused to train another machine learning system.

[0061] For example, a table may be generated with many different expected outputs for each sim data image dss,sim*. Therefore, the expected output used for training depends on the type of analysis required.

[0062] Therefore, preparing a trained machine learning system is - Selecting a machine learning system from multiple trained machine learning systems based on the type of output (mo) that should be obtained. Includes.

[0063] There may be several trained machine learning systems stored on or available on the computer, and the computer selects which of the trained machine learning systems to use based on the desired output to be obtained.

[0064] Therefore, the method further includes the fact that the sim data image dss,sim* is obtained from the model image mss* by a function of kind dss,sim* = g(mss*) + n(mss*), where g is the smoothing function and n is the noise function.

[0065] The training set simulation data images dss and sim* are obtained from the model image mss* and expected output mf*, depending on the type of output required.

[0066] Therefore, the method further includes the selected expected output mf* being the pixel intensity of the central pixel or group of central pixels in the model image mss*, or the probability of a particular feature associated with the model image mss*.

[0067] One type of expected output mf* can be the pixel intensity of the central pixel of the model image mss*. Alternatively, the expected output mf* can originate from the group of central pixels of the model image mss*. For example, if the model image mss* contains 9x9 pixels, the expected output mf* could be the nine central pixels forming a 3x3 matrix around the central pixel. In this case, the expected output mf* can be a vector consisting of the nine pixel intensities.

[0068] Therefore, the method further includes the selected expected output mf* being the number of pixels concatenated to the central pixel that have an intensity higher than a threshold intensity value or a probability of disease (e.g., cancer) higher than a threshold probability value.

[0069] The expected output mf* can be, for example, the number of pixels connected to the central pixel that have a pixel intensity greater than 12, given that the intensity is a value between 0 and 20.

[0070] Therefore, the method further includes the fact that the expected output mf* is a category of the central pixel or group of pixels in the model image mss*, or that the expected output mf* is a probability for one or more categories.

[0071] The expected output mf* can be a vector consisting of one or more categories, where category A can be, for example, the probability that the model image mss* represents cancer, and category B can be the probability that the model image mss* does not represent cancer.

[0072] Therefore, the method further includes the fact that the expected output mf*(21) is a vector with two values, namely the mean and covariance of a normal distribution.

[0073] Expression output When multiple subset image DSSOs are processed by a machine learning system, multiple output MOs, each consisting of one output MO for each subset image DSSO from the machine learning system, are used to determine the representation output of the data.

[0074] Therefore, determining the representation output involves arranging the output mo in order according to the location of the subset image dsso in the original PET image.

[0075] Therefore, the method further includes the fact that the output mo is a numerical value such as pixel intensity, the mean and covariance of the number of pixels, or the number of pixels, or that the expected output mf* is a category, or the probability for one or more categories.

[0076] The output mo of a machine learning system may be the pixel intensity for a single pixel, or the pixel intensity for several pixels, for example, nine central pixels. Alternatively, output mo may be the mean and covariance for a large number of pixels. Output mo may also be the number of pixels that have an intensity higher than a threshold intensity. Output mo may also be a category, for example, the category "cancer" or the category "not cancer". mo may be a single value, or a vector consisting of two or more values. For example, output mo may be a vector containing probabilities for two or more categories.

[0077] If the output mo is a numerical value, such as pixel intensity or average pixel intensity, the output can be used to generate a higher quality image of the original PET scan image, i.e., a purified version of the PET scan image.

[0078] Therefore, the method further includes arranging the output mo in order to form a representation output.

[0079] Of course, the representation output can be determined in many different ways using many different statistical calculation methods.

[0080] If the output mo is an intensity value, a purified version of the PET image can be created. The purified image is model mrestore, which is a solution to the inverse problem of creating a purified image from the original PET image dobs.

[0081] If the output mo represents the volume / region of a cancerous lesion, and is the number of pixels with a high probability of being cancerous that are concatenated, then a representation output will be created in which a large number of concatenated pixels with a high probability of being cancerous are darkened compared to when only a few pixels are concatenated, or when only one pixel has a high probability of being cancerous. The same is true if the output mo is the number of pixels with high intensity that are concatenated.

[0082] If the output mo has multiple categories, the output mo may be a vector with two or more numbers, where the numbers may be probabilities for each category. In this case, the representation output may be constructed to represent the probability for one of the multiple categories. If the category is the probability of cancer, the representation output image will be constructed with darker pixels for high probabilities, less dark pixels for lower probabilities, and white pixels for probabilities of zero. Such a representation output emphasizes the risk of cancer.

[0083] When a subset image dsso is selected from a PET image, a group of pixels is selected, not just a single pixel. This is because, using GPSF, the pixel intensity for a single pixel is determined by the intensity and noise of neighboring pixels; therefore, several rows of neighboring pixels are needed to obtain the correct result for a single pixel. Thus, the method for creating a subset image dsso is to move the frame around the subset image dsso by one pixel at a time, creating a subset image dsso for each pixel of the PET image. A subset image dsso can be, for example, 9 pixels × 9 pixels. The central pixel is surrounded by neighboring pixels within the subset image dsso, but when the subset image dsso is analyzed, for example, only the intensity of the central pixel of the corresponding subset model mss is used, and the other 80 pixels around the central pixel are ignored. The output mf* can be the intensity of the central pixel of the subset model mss, and the represented output image is created by arranging and combining the central pixel values ​​of all subset model mss in the same order as the subset image dsso of the original PET image, thereby creating a purified PET image.

[0084] In a 9x9 subset image, this calculation yields the correct central pixel value in a given case. Using a larger subset image allows for the use of more central pixels, potentially increasing efficiency. This efficiency improvement depends on the increased computational demand required to set up and train the machine learning system.

[0085] Therefore, the machine learning system includes a regression-type mapping if the expected output mf* represents a numerical value, or a classification-type mapping if the expected output mf* represents a category.

[0086] When creating a machine learning system using a neural network, the final layer of the neural network differs depending on whether the expected output mf* refers to a continuous value such as pixel intensity, an individual parameter, or a probability value for a specific category. If a continuous value is required, the final layer is a regression layer. If probabilities for different categories are required, the final layer is a classification layer.

[0087] Therefore, machine learning systems include neural networks.

[0088] A preferred machine learning system is a neural network, but it may also be a decision tree, a support vector machine, or any other machine learning algorithm capable of learning the mapping from Dsim* to Mf*.

[0089] Specifically, the machine learning system may use a feedforward neural network consisting of multiple layers, in which case the input layer consists of a number of neurons corresponding to the number of parameters in the subset image dsso, more specifically, the number of pixels in the subset image dsso, with each pixel represented by its pixel intensity. In principle, any network architecture can be used as long as the input layer reflects the subset image dsso and the output layer reflects the expected output mf*.

[0090] The output layer consists of one or more neurons, one for each element of output mo. The output layer may have one neuron with an output mo value, or it may have two or more neurons, in which case each neuron represents a category and is a probability value for a given category. The number of layers in the neural network is determined by which outputs are needed. The number of layers and nodes is selected to be sufficiently large so that the correspondence between subset images dsso and expected outputs mf* can be resolved, and sufficiently small so that the network does not overfit the training data during training.

[0091] Backpropagation is used to calculate weights and biases for a neural network, and the ADAM algorithm can be used, for example, to modify weights and biases. The ADAM algorithm is a standard optimization algorithm available, for example, via TensorFlow. Kingma, Diederik P., and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014).

[0092] The present invention relates to a computer program product comprising at least one computer connected to a data storage means, wherein the computer program includes instructions that, when executed by the computer, cause the computer to perform each step of an image reconstruction method.

[0093] According to a second aspect of the present invention, the present invention relates to a method for training a machine learning system for image reconstruction, wherein the method is - Obtaining model images mss*(16) from sample M*(14), where model images mss*(16) are a manifestation of prior information ρ(m)(12), prior information ρ(m)(12) is a statistical model based on expert prior data, and each model image mss*(16) is an image representation consisting of one or more pixels. - The method involves obtaining a training set in which each training set includes a sim data image dss,sim*(18) and an expected output mf*(21) by determining the sim data image dss,sim*(18) based on the model image mss*(16), and selecting the expected output mf*(21) for each sim data image dss,sim*(18) based on the model image mss*(16), wherein each sim data image dss,sim*(18) is represented by one or more pixels. - The training set is used as input to train the machine learning system (36), and the output mo (34) for each sim data image dss,sim* (18) is obtained from the machine learning system. - Training the machine learning system until it converges based on a comparison between the output mo(34) and the expected output mf*(21), Includes.

[0094] According to a third aspect of the present invention, the present invention relates to a medical imaging system comprising a scanner (112) and a control system (111) for scanning and recording PET images, wherein the processing device is - Acquire a PET image, and divide the PET image into multiple subset images dsso, each of which is represented by one or more pixels. - Prepare a trained machine learning system and use a subset of DSSO images as input to the trained machine learning system. - Obtain the output MO for each subset image DSSO from the machine learning system, - Determine and output the representation based on the output mo. It is structured in such a way.

[0095] Furthermore, according to a third aspect of the present invention, the present invention relates to a medical imaging system, and the processing apparatus further comprises: - Obtaining a model image mss* from a sample M*, where the model image mss* is a manifestation of prior information ρ(m), where prior information ρ(m) is a statistical model based on expert prior data, and each model image mss* is an image representation consisting of one or more pixels. - Each training set includes a sim data image dss,sim* and an expected output mf*, obtained by determining the sim data image dss,sim* based on the model image mss*, and selecting the expected output mf* based on the model image mss* for each sim data image dss,sim*, wherein each sim data image dss,sim* is represented by one or more pixels. - The training set is used as input to train the machine learning system, and the output mo for each sim data image dss,sim* is obtained from the machine learning system. - Training the machine learning system until it converges based on a comparison of the output mo and the expected output mf*, This is configured to provide a trained machine learning system.

[0096] In one aspect of the present invention, the present invention relates to a computer program product configured to control a medical imaging system, comprising at least one computer to which data storage means are connected, and which includes instructions that cause the computer to perform the method of the present invention when executed by the computer.

[0097] This aspect of the present invention is particularly useful in that it can be implemented by a computer program product that enables a computer system to perform the operations of the image reconstruction method of the present invention when downloaded or uploaded to a computer system. Such a computer program product may be provided on any type of computer-readable medium or via a network.

[0098] Each individual aspect of the present invention can be combined with any of the other aspects. These and other aspects of the present invention will become apparent from the following description relating to the embodiments described.

[0099] The method of the present invention will now be described in more detail with reference to the accompanying figures. The figures illustrate one method of carrying out the present invention and should not be construed as limiting other possible embodiments that fall within the scope of the accompanying set of claims. [Brief explanation of the drawing]

[0100] [Figure 1] A schematic diagram of a medical imaging system is shown. [Figure 2A] This shows a PET image observed in two dimensions. [Figure 2B] This shows a subset of PET images. [Figure 3] This demonstrates that when a machine learning system is running, a subset image DSSO is taken as input, and the output is MO. [Figure 4] The following shows the forward model, the smoothed model, and the noisy smoothed model representing the PET image, using the reference model mref. [Figure 5] This document demonstrates how to generate model images (mss*) from expert data. [Figure 6] This shows how to create a training set. [Figure 7] This demonstrates the training of a machine learning system. [Figure 8]Here are six example training sets for training a machine learning system on three different features. [Figure 9] Here is another example of six training sets for training a machine learning system. [Figure 10] This demonstrates a classification neural network using a multilayer perceptron (MLP). [Figure 11] This section demonstrates an implementation of the present invention using a convolutional neural network (CNN) and another possible neural network. [Figure 12A] Figure 2A shows possible representation output images obtained using MLP and CNN based on the PET image. [Figure 12B] Figure 2A shows possible representation output images obtained using MLP and CNN based on the PET image. [Figure 13A] This demonstrates the determination of the pixel intensity of a single pixel in an ideal reference model, mref. [Figure 13B] This demonstrates the determination of the pixel intensity of a single pixel in an ideal reference model, mref. [Figure 14] This shows the output image obtained using a CNN neural network and the output image obtained using an MLP neural network. [Figure 15] The output image shows the representation of the output mo when it contains two values, namely the mean and covariance, and the output mo represents a normal distribution. [Modes for carrying out the invention]

[0101] The figures illustrate one way of carrying out the present invention and should not be construed as limiting other possible embodiments that fall within the scope of the appended set of claims.

[0102] Figure 1 shows a schematic diagram of a medical imaging system 100 in one aspect of the present invention. The system 100 includes a scanning system 110 that can obtain data representing the contrast agent concentration as a function of time of the contrast agent injected into a human being 200, and a schematic cross-sectional view of the human head can be obtained. One or more parts of the human being, for example, the brain, can be imaged. The system includes a PET scanner 112 and a corresponding measurement and control system 111. The scanner acquires primary data DAT, which is transmitted to the measurement and control system 111 for further processing to become secondary data DAT', which is then transmitted to a processing unit 120.

[0103] In the processing unit 120, a method for evaluating the perfusion index of a human 200 is performed using acquired data DAT and DAT', which represent the contrast agent concentration as a function of time of the injected contrast agent.

[0104] The processing unit is operablely connected to a display device 130, such as a computer screen or monitor, and can display the obtained PET image in a schematic manner. Alternatively, the PET image can be transmitted to any suitable type of storage device 140 for subsequent analysis or diagnosis.

[0105] Figure 2A shows a two-dimensional PET image output from a PET scanner. This two-dimensional PET image is an example of a PET image 30. Figure 2B shows one subset image dsso32 within a marked square, where each tracer intensity in subset image dsso32 is shown as a pixel representing the tracer intensity. This representation by subset image dsso32 has 11 pixels × 11 pixels. According to the method of the present invention, the PET image 30 is divided into multiple subset images dsso32. In a preferred embodiment, a subset image can be created for each pixel of the PET image, and each pixel becomes the central pixel within the subset image dsso32 surrounded by adjacent pixels, creating a subset image dsso32 that can be an 11 pixels × 11 pixels image as shown in Figure 2B.

[0106] Figure 3 shows that when the machine learning system 36 is operating, a subset image dsso32 is taken as input and an output mo34 is output. The machine learning system 36 analyzes one subset image dsso32 at a time and outputs an output mo34. The subset image dsso32 and the output mo34 can be arranged in order in a table or database according to the location of the subset image dsso in the original PET image to construct the output mo, which forms a representational output image.

[0107] Figure 4 shows the forward model. Image 38 in Figure 4 is an example of the reference model mref38 for the PET image dobs30. The reference model is an image that appears to be completely free of noise and blur. Image 37 shows the forward response dref = Gpsf(mref). Image 30 is the PET image dobs, which is an embodiment of the noise model dobs = Gpsf(mref) + n(mref).

[0108] Figure 5 shows how to generate model images mss*16 from expert data 10. When creating a training set, prior information ρ(m)12 is created based on expert data 10, and sample M*14 is created from prior information ρ(m)12, which includes model images mss*16 that embody prior information ρ(m)12.

[0109] Figure 6 shows how to create a training set. From the model image mss*16 in sample M*14, the data image dss*17 and the simulation data image dss,sim*18 obtained from the simulation are obtained using the functions dss*=g(mss*) and dss,sim*=g(mss*)+n(mss*). The expected output mf*21 is also obtained from the model image mss*16. The simulation data image dss,sim*18 and the expected output mf*21 form the training set. The training set can be stored in a table with two columns, the pseudo-data sample Dsim*15 and the expected output sample Mf*19, as shown in Figure 4, or the training set can be stored in a database or another type of data storage device. It is also possible to generate a different expected output, which can be stored, for example, in another column in the table.

[0110] Figure 7 shows the training of the machine learning system 36. The sim data images dss,sim*18 and the expected output mf*21 are the training data used to train the machine learning system. The output is output mo. The training accuracy is determined by comparing output mo with the expected output mf*. If the accuracy is sufficient and convergence has occurred, training is complete. Otherwise, training continues with modified weights and biases for the nodes of the neural network, and training is performed again.

[0111] Figure 8 shows examples of six training sets for training a machine learning system, obtained by selecting model images mss*16 from samples M* sampled from prior information ρ(m). Data images dss*17 are obtained from model images mss*16 by the formula dss*=Gpsf(mss*), and sim data images dss,sim* are obtained by the formula dss,sim*=Gpsf(mss*)+n(mss*). Sim data images dss,sim* are used as input for training the machine learning system. Expected outputs mf* are also obtained for training. Figure 8 shows examples of three expected outputs mf*(21a, 21b, 21c). In the first column, the expected output mf*(21a) is a category, which can be "high" or "low" depending on the intensity of the central pixel. In the second column, the expected output mf*(21b) is the intensity value for the intensity of the central pixel of the model image mss*. In the third column, the expected output mf*(21c) is the 25 central pixels of the model image mss*, corresponding to one pixel of the data image dss* and the sim data image dss,sim*. Here, the expected output mf* should be understood as a vector consisting of 25 intensity values ​​representing the 25 central pixels of the model image mss*. The number of pixels in the data image dss* and the sim data image dss,sim* is reduced to minimize computation time. However, the 25 central pixels of the model image mss* can still be used as the expected output mf*.

[0112] Figure 9 shows six further training sets for training the machine learning system. In this case, the pixels in the data image dss* and the sim data images dss,sim* each represent four pixels in the model image mss*. Furthermore, the expected output mf* vector 21c is a vector consisting of nine central pixels in the model image mss*.

[0113] Figure 10 shows a classification neural network using a multilayer perceptron (MLP), which is one possible network configuration in the presented invention. The network consists of four layers, the first of which is the input layer, and in this case, it is a layer with 121 neurons. The input is the pixel intensity for each pixel in a subset image dsso having 11 pixels × 11 pixels, and 121 pixel intensities are given. Each neuron in the input regenerator represents a pixel in the subset image dsso and therefore receives an intensity value as input. The neural network then includes two hidden layers, namely layer 2 and layer 3, which are dense layers consisting of 20 neurons each, and the final layer is an output layer containing two neurons, one neuron giving the probability of class 1 and the other giving the probability of class 2. Class 1 may be the probability of cancer, and class 2 may be the probability of not having cancer. The fourth column gives the number of parameters used, and in this implementation a total of 2902 parameters are used.

[0114] Figure 11 shows an implementation of the present invention using a convolutional neural network (CNN), which is a classification neural network with seven layers. The first layer is the input layer, which in this case also contains 121 neurons for the pixel intensity of a subset image dsso having 11 pixels × 11 pixels. The second layer is a folding layer (conv2d) with 32 neurons, and the third layer is a pooling layer (max_pooling2d) with 32 neurons. The fourth layer is another folding layer (conv2d) with 64 neurons. The fifth layer is a pooling layer (max_pooling2d) with 64 neurons. The sixth layer is a flattening layer with 64 neurons, and the final layer is an output layer with 2 neurons, each neuron representing a class probability as shown in Figure 10. The fourth column gives the number of parameters used, and in this implementation a total of 18946 parameters are used.

[0115] The implementations in Figures 10 and 11 use standard routines for constructing neural networks, namely conv2d, Max_pooling2d, Flatten, and Dense. The important point in this invention is not how a particular neural network is constructed, as many different types of neural networks are available. What is important in this application are the input and output layers, as well as the cost function used.

[0116] Figures 12A and 12B show two possible representation output images based on the PET image in Figure 2A. Figure 12A is a one-to-one representation, where each pixel in Figure 12A represents a pixel in Figure 2A. In Figure 12B, each pixel in Figure 2A is represented by 16 pixels in a 4x4 square in Figure 12B. This is the case when the resolution of the model image mss* is 16 times the resolution of the PET image.

[0117] Figures 13A and 13B show the determination of pixel intensity 135 for the smallest region with high intensity in the ideal reference model mref. Figure 13B is an enlarged view of a portion of Figure 13A. Figures 13A and 13B show the shape along the X-axis at y=193 in Figures 12A and 12B. The ideal pixel value is shown by line 131, and line 132 shows the pixel value within the PET image dobs. A pixel has a pixel intensity of 16 kBq / ml in the ideal reference model with no blur and noise, but only about 10 kBq / ml in the original PET image with blur and noise. The method of the present invention shows that the pixel intensity is determined according to line 133 in a machine learning system operating at a finer scale than that used in the PET image, and according to line 134 in a machine learning system operating at the same scale as that used in the PET image, and the analysis results of the machine learning system are close to the pixel values ​​of the ideal reference model.

[0118] Figure 14 shows the representation output 142 obtained using a CNN neural network and the representation output 143 obtained using an MLP neural network. Figure 141 is the ideal reference model mref, which is a PET image dobs free of blur and noise, demonstrating that similar results can be obtained using either a CNN neural network or an MLP neural network.

[0119] Figure 15 shows the represented output images when the output mo for each subset image dsso contains two values, namely the mean and the variance, so that the output mo represents a normal distribution. Image 151 shows the mean of multiple output mo for multiple subset image dsso, where each pixel in the image represents the mean for one output mo. Image 152 shows the corresponding variances for multiple output mo.

[0120] The present invention can be implemented using hardware, software, firmware, or any combination thereof. The present invention, or some of its features, can also be implemented as software running on one or more data processing devices and / or digital signal processors.

[0121] Individual elements of embodiments of the present invention can be implemented in any suitable physical, functional, and logical manner, either in a single unit, in multiple units, or as part of multiple other functional units. The present invention can be implemented in a single unit or physically and functionally distributed among multiple different units and processing units.

[0122] While the present invention has been described in relation to specific embodiments, it should not be construed as being limited to the examples presented. The scope of the present invention should be interpreted in light of the accompanying set of claims. In relation to the claims, the terms “comprising” or “comprises” do not exclude other possible elements or steps. Nor should references to elements such as “a” or “an” be construed as excluding plurals. Nor should the use of reference numerals in the claims for elements shown in the figures be construed as limiting the scope of the present invention. Furthermore, individual features described in different claims may, in some cases, be advantageously combined, and references to these features in different claims do not preclude the possibility that such combinations of features are possible and beneficial.

[0123] symbol m Model or model image Data image related to dm, using the formula d=g(m)+n(m). dobs PET images mrestore: A model free of noise and blur obtained from PET images. mref Ideal reference model The ideal data image obtained from dref mref MSS: A subset model, which is a subset of a large model image. DSS is a subset image, which is a subset of a larger data image. A DSSO subset image is one of several subset images of a PET image. A subset image is a data image that is a subset of the PET image. mss* A model image that embodies the prior information ρ(m). The data image obtained from mss* using the formula dss* = g(mss*) Sim data image obtained from MSS* using the formula dss,sim* = g(mss*) + n(mss*). mf* Expected output from a machine learning system during training Output from the machine learning system M* mss* specimen D* dss* data sample Dsim* dss,sim* pseudo-data sample Sample of expected output for Mf*

[0124] References Tarantola, A., & Valette, B. Inverse problems= quest for information. Journal of geophysics, 50(1), 159-170 (1982). Deutsch, CV, & Journel, AG Geostatistical software library and user's guide. New York, 119(147) (1992). Mariethoz, G., & Caers, J. Multiple-point geostatistics: stochastic modeling with training images. John Wiley & Sons (2014). Bishop, Christopher M. Pattern, Recognition and machine learning. Springer (2006). Kingma, Diederik P., and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014).

[0125] The references listed above are incorporated herein by reference in their entirety. (Note 1) A computer-implemented method for image analysis and reconstruction of medical images of a subject, - Obtain PET images (30) and select each subset image d sso (32) Multiple subset images d, each represented by one or more pixels. sso The PET image (30) is divided into (32) and - Prepare a trained machine learning system (36), - The subset image d sso (32) is used as input to the trained machine learning system (36), - The subset image d sso Output m for each of (32) o (34) obtained from the trained machine learning system, - The output m o Based on (34), the representation output (38) is determined and output, Methods that include... (Note 2) Preparing a trained machine learning system (36) is - Specimen M * (14) Model image m ss * (16) is to obtain the model image m ss * (16) is a manifestation of the prior information ρ(m)(12), and the prior information ρ(m)(12) is a statistical model based on expert prior data, and the model image m ss * Each of (16) is an image representation consisting of one or more pixels, and to obtain, - Each training set is a sim data image d ss,sim * (18) and expected output m f * (21) The training set including the model image m ss * Based on (16), the shim data image d ss,sim * (18) is obtained, and the shim data image d ss,sim * (18) The model image m for each of the above ss * Based on (16), the expected output m f * (21) is obtained by selecting the shim data image d ss,sim * (18) Each of these is a representation of one or more pixels, and to obtain, - The training set is used as input for training the machine learning system (36), and the simulation data image d ss,sim * Output m for each of (18) o (34) obtained from the machine learning system, - The output m o (34) and the expected output m f * (21) Training the machine learning system until it converges based on a comparison with the above, A computer-implemented method for the analysis and reconstruction of images as described in (Appendix 1), including the following. (Note 3) Preparing a trained machine learning system (36) is - The output m to be obtained from multiple trained machine learning systems o Select the trained machine learning system (36) based on the type of (34), A computer-implemented method for the analysis and reconstruction of images as described in (Appendix 1) or (Appendix 2), including the following: (Note 4) The aforementioned shim data image d ss,sim * (18) The function d of the above type ss,sim * =g(m ss * ) + n(m ss * ) by the aforementioned model image m ss * A computer-implemented method for the analysis and reconstruction of images described in (Appendix 2) or (Appendix 3), obtained from (16), where g is the smoothing function and n is the noise function. (Note 5) The selected expected output m f * (21) is the aforementioned model image m ss * A computer-implemented method for analyzing and reconstructing an image as described in any one of the items (Appendix 2) to (Appendix 4), wherein the pixel intensity of the central pixel or group of central pixels in (16). (Note 6) The selected expected output m f * (21) is the aforementioned model image m ss * A computer-implemented method for analyzing and reconstructing an image according to any one of (Appendix 2) to (Appendix 4), wherein the number of pixels connected to the central pixel in (16) has an intensity higher than a threshold intensity value, or a probability of disease (e.g., cancer) higher than a threshold probability value. (Note 7) The expected output m f * (21) is the aforementioned model image m ss * (16) The category of the central pixel or group of central pixels, or the expected output m f * A computer-implemented method for the analysis and reconstruction of images as described in any one of (Appendix 2) to (Appendix 4), wherein is the probability for one or more categories. (Note 8) The expected output m f * (21) is a vector having two values, namely the mean and covariance of a normal distribution, a computer-implemented method for the analysis and reconstruction of an image as described in any one of (Appendix 2) to (Appendix 4). (Note 9) The aforementioned output m o (34) is a numerical value such as the average and covariance of the pixel intensity and number of pixels, or the number of pixels, or the expected output m f * A computer-implemented method for the analysis and reconstruction of images as described in any one of the items (Appendix 1) to (Appendix 8), wherein is a category, or a probability for one or more categories. (Note 10) The aforementioned output m o A computer-implemented method for analyzing and reconstructing an image as described in any one of the items (Appendix 1) to (Appendix 9), wherein (34) is arranged in order to form the representation output (38). (Note 11) The machine learning system (36) has the expected output m f * (21) If the value is a numerical value, it includes a regression-type correspondence, or the expected output m f * A computer-implemented method for the analysis and reconstruction of images as described in any one of the items (Appendix 2) to (Appendix 10), including a classification type correspondence if represents a category. (Note 12) The machine learning system (36) is a computer-implemented method for image analysis and reconstruction as described in any one of the items (Appendix 1) to (Appendix 11), including a neural network. (Note 13) A method for training a machine learning system for image analysis and reconstruction, - Specimen M * (14) Model image m ss * (16) is to obtain the model image m ss * (16) is a manifestation of the prior information ρ(m)(12), and the prior information ρ(m)(12) is a statistical model based on expert prior data, and the model image m ss * Each of (16) is an image representation consisting of one or more pixels, and to obtain, - Each of the aforementioned training sets is a sim data image d ss,sim * (18) and expected output m f * (21) The training set including the model image m ss * Based on (16), the shim data image d ss,sim * (18) is obtained, and the shim data image d ss,sim * (18) m for each of the following ss * Based on (16), the expected output m f * (21) is obtained by selecting the shim data image d ss,sim * (18) Each of these is a representation of one or more pixels, and to obtain, - The training set is used as input for training the machine learning system (36), and the simulation data image d ss,sim * Output m for each of (18) o (34) obtained from the machine learning system, - The output m o (34) and the expected output m f * (21) Training the machine learning system until it converges based on a comparison with the above, Methods that include... (Note 14) A medical imaging system comprising a scanner (112) and a control system (111) and a processing unit (120) for scanning and recording PET images, wherein the processing unit is - The PET image (30) is acquired, and each subset image d sso (32) Multiple subset images d, each represented by one or more pixels. sso The PET image (30) is divided into (32) and - Prepare a trained machine learning system (36) and process the subset image d sso (32) is used as input to the trained machine learning system (36), - The subset image d sso Output m for each of (32) o (34) is obtained from the machine learning system, - The output m o Based on this, the representation output (38) is determined and output. A medical imaging system configured in such a way. (Note 15) Computer program software, such as a computer program product, which, when executed by a computer, includes instructions to cause the computer to perform the method described in any one of the items (Appendix 1) to (Appendix 12), and is configured to control the medical imaging device described in (Appendix 14), comprising a computer system having at least one computer to which data storage means are connected.

Claims

1. A method performed by a computer for image analysis of medical images of a subject, The aforementioned method, - Prepare a trained machine learning system (36), - Acquire PET images (30) and each subset image d ｓｓｏ (32) Multiple subset images d, where each is represented by one or more pixels. ｓｓｏ Dividing the PET image (30) into (32), - The subset image d ｓｓｏ (32) is used as input to the trained machine learning system (36), - The subset image d ｓｓｏ (32) The respective features of the subset model m ss for each of the above are the output m ｏ (34) obtained from the trained machine learning system, - The aforementioned output m ｏ Based on (34), the representation output (38) is determined and output, Includes, Preparing the aforementioned trained machine learning system (36) is - To generate a plurality of model images m ss* (16) based on prior information ρ(m)(12), and to form a sample M* (14) containing the plurality of model images m ss* (16), wherein the prior information ρ(m)(12) represents prior knowledge about parameters relating to the model images m ss* (16), the prior knowledge originates from expert knowledge or previous investigations relating to the PET images, the prior information ρ(m)(12) represents a statistical model, and the model images m ss* (16) are realized by sampling based on the prior information ρ(m)(12), and the model images m ｓｓ ＊ Each of (16) is an image representation consisting of one or more pixels, forming and - obtaining a plurality of training sets, wherein one said training set corresponds to one said model image m ss * (16), one said training set comprises a simulated data image d ss,sim * (18) obtained based on said model image m ss * (16) corresponding to said training set, and an expected output m f * (21) obtained based on said model image m ss * (16) corresponding to said training set, said simulated data image d ss,sim * (18) is an image obtained by adding blur and noise to said model image m ss * (16), said simulated data image d ｓｓ，ｓｉｍ ＊ each of (18) is a representation by one or more pixels, the expected output m f * (21) represents features of the model image m ss * (16), and the obtaining step; - The sim data image d ss, sim * (18) included in the training set is used as input for training the machine learning system (36), and the training output m for the sim data image d ss, sim * (18) is used. ｏ (34) obtained from the machine learning system, - The training output m for the sim data image d ss, sim * (18) included in the training set. ｏ (34) and the expected output m included in the training set. ｆ ＊ Based on a comparison with (21), the machine learning system is trained to minimize the difference between the training output m o (34) and the expected output m f * (21), The methods a computer uses for image analysis, including [specific methods / techniques].

2. Preparing a trained machine learning system (36) is - The output m to be obtained from multiple trained machine learning systems ｏ Select the trained machine learning system (36) based on the type of (34), A method performed by a computer for the analysis of an image according to claim 1, including the following:

3. The aforementioned shim data image d ｓｓ，ｓｉｍ ＊ (18) is function d ｓｓ，ｓｉｍ ＊ = g(m) ｓｓ ＊ ) + n(m ｓｓ ＊ ) and the aforementioned model image m ｓｓ ＊ A method performed by a computer for the analysis of an image according to claim 1 or 2, wherein g is a smoothing function and n is a noise function, obtained from (16).

4. The expected output m ｆ ＊ (21) is the aforementioned model image m ｓｓ ＊ A method performed by a computer for analyzing an image according to any one of claims 1 to 3, wherein the pixel intensity of the central pixel or group of central pixels in (16).

5. The expected output m ｆ ＊ (21) is the aforementioned model image m ｓｓ ＊ A method for a computer to perform image analysis according to any one of claims 1 to 3, wherein the number of pixels in the group of pixels formed by concatenating pixels having an intensity higher than the threshold intensity value or pixels having a disease probability higher than the threshold probability value in (16) is the number of pixels in the group of pixels that includes the central pixel in the model image m ss * (16).

6. The expected output m ｆ ＊ (21) is The aforementioned model image m ｓｓ ＊ (16) Category of the central pixel or group of central pixels, A method performed by a computer for analyzing an image according to any one of claims 1 to 3, which is the probability of one or more of the central pixels or groups of central pixels in the model image m ss* (16) for the category.

7. The expected output m ｆ ＊ A method performed by a computer for the analysis of an image according to any one of claims 1 to 3, wherein (21) is a vector having a mean and covariance of a normal distribution that describes the probability of the pixel intensity of the central pixel or group of central pixels in the model image m ss * (16).

8. The aforementioned output m ｏ (34) is, The pixel intensity of the central pixel or group of central pixels in the subset model m ss corresponding to the subset image d sso (32), In the subset model m ss, the number of pixels is the number of pixels in the pixel group that includes the central pixel in the subset model m ss, which is formed by concatenating pixels that have an intensity higher than the threshold intensity value or pixels that have a disease probability higher than the threshold probability value. A vector having the mean and covariance of a normal distribution that describes the probability of the pixel intensity of the central pixel or group of central pixels of the subset model m ss, The category of the central pixel or group of central pixels in the subset model m ss, or A method performed by a computer for analyzing an image according to any one of claims 1 to 7, wherein the probability of a central pixel or group of central pixels in the subset model m ss for one or more of the categories.

9. The aforementioned output m ｏ A method for computer analysis of an image according to any one of claims 1 to 8, wherein (34) is arranged in order to form a representation output (38).

10. The machine learning system (36) has the expected output m ｆ ＊ (21) If the value of (21) is a numerical value, it includes a regression-type correspondence, and the expected output m ｆ ＊ A method performed by a computer for the analysis of an image according to any one of claims 1 to 9, including a classification type mapping if represents a category.

11. The machine learning system (36) includes a neural network, and the computer performs the image analysis according to any one of claims 1 to 10.

12. A medical imaging system comprising a scanner (112) and a control system (111) and a processing unit (120) for scanning and recording PET images, The aforementioned processing apparatus is - Prepare a trained machine learning system (36), - The PET image (30) is acquired, and each subset image d ｓｓｏ (32) Multiple subset images d, where each is represented by one or more pixels. ｓｓｏ The PET image (30) is divided into (32) and - The subset image d ｓｓｏ (32) is used as input to the trained machine learning system (36), - The subset image d ｓｓｏ (32) The respective features of the subset model m ss for each of the above are the output m ｏ (34) is obtained from the machine learning system, - The aforementioned output m ｏ Based on this, the representation output (38) is determined and output. It is structured in such a way, When the processing device prepares the trained machine learning system (36), - To generate a plurality of model images m ss* (16) based on prior information ρ(m)(12), and to form a sample M* (14) containing the plurality of model images m ss* (16), wherein the prior information ρ(m)(12) represents prior knowledge about parameters relating to the model images m ss* (16), the prior knowledge originates from expert knowledge or previous investigations relating to the PET images, the prior information ρ(m)(12) represents a statistical model, and the model images m ss* (16) are realized by sampling based on the prior information ρ(m)(12), and the model images m ｓｓ ＊ Each of (16) is an image representation consisting of one or more pixels, forming and - Obtaining multiple training sets, wherein one training set corresponds to one model image m ss* (16), and one training set includes a sim data image d ss, sim* (18) obtained based on the model image m ss* (16) corresponding to the training set, and an expected output m f* (21) obtained based on the model image m ss* (16) corresponding to the training set, and the sim data image d ss, sim* (18) is an image in which blur and noise have been added to the model image m ss* (16), and the sim data image d ｓｓ，ｓｉｍ ＊ Each of (18) is represented by one or more pixels, and the expected output m f * (21) represents the characteristics of the model image m ss * (16), and the acquisition of - The sim data image d ss, sim * (18) included in the training set is used as input for training the machine learning system (36), and the training output m for the sim data image d ss, sim * (18) is used. ｏ (34) obtained from the machine learning system, - The training output m for the sim data image d ss, sim * (18) included in the training set. ｏ (34) and the expected output m included in the training set. ｆ ＊ Based on a comparison with (21), the machine learning system is trained to minimize the difference between the training output m o (34) and the expected output m f * (21), A medical imaging system configured to perform the following actions.

13. A computer program, The computer program is a computer program that causes at least one computer connected to a data storage means to perform the method described in any one of claims 1 to 12.

Images