Local spectral covariance or local spectral covariance depletion used to highlight relevant material transitions in spectral CT and MR.
By calculating the local covariance matrix and performing Gaussian convolution operations, combined with a pre-trained classifier, the problem of information compression in multi-energy spectral imaging is solved, enabling efficient material transformation recognition and display under conventional image reading habits.
Patent Information
- Application Number
- CN202180080678.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-12-01
- Filing Date
- 2021-11-28
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2041-11-28
AI Technical Summary
Existing multi-energy spectral imaging techniques struggle to efficiently compress rich multi-energy spectral information into a single image, and conventional image reading habits may favor single-energy images, making material transformations difficult to discern in conventional images.
By calculating the values of the local covariance matrix and the local covariance matrix, using Gaussian convolution operations and a pre-trained classifier, combined with edge detection, the material transition locations are identified and superimposed to generate superimposed images to highlight the material transitions. Color or grayscale values are used for encoding to ensure compatibility with conventional image reading habits.
It enables efficient identification and display of multi-energy spectrum information without changing conventional image reading habits, highlighting material transformation, reducing computational workload, and improving the identifiability of material transformation.
Smart Images

Figure CN116710970B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates generally to multi-energy spectral imaging, and more particularly to two alternative apparatuses and methods, medical imaging systems, and computer program units for processing image data of an object of interest comprising at least two single-energy images acquired from different spectral channels. Background Technology
[0002] Due to the rapid development of computer technology and a growing body of literature demonstrating significant advantages over conventional single-energy imaging, multi-spectral imaging is becoming increasingly common in clinical practice. For example, spectral CT generates multiple spectral channels ranging from low to high energy levels, controlled to varying degrees by the Compton effect and photoelectric effect. These spectral channels can also be converted into various representations with pure Compton images, pure photoelectric images, and pure (pseudo / virtual) single-energy images. Typical dual-slice or dual-energy CT scanners generate two independent spectral channels, while photon-counting CT scanners, MR scanners, or ultrasound (US) probes can generate N independent spectral channels.
[0003] However, the rich multi-spectral information provided by each voxel in an energy spectrum image may be difficult to compress into a single view for efficient reading. Furthermore, due to existing reading habits and accumulated expertise, conventional images are generally preferred for viewing. Summary of the Invention
[0004] It may be necessary to compress the multi-spectral information into a single image.
[0005] The object of the invention is achieved through the subject matter of the independent claims, wherein further embodiments are incorporated in the dependent claims. It should be noted that the aspects described below also apply to the two alternative devices and methods, medical imaging systems, and computer program units.
[0006] According to a first aspect of the present invention, an apparatus is provided for processing image data of an object of interest, the image data comprising at least two single-energy images acquired from different energy spectral channels. The apparatus includes an input module, a calculation module, a rendering module, and an output module. The input module is configured to receive the image data of the object of interest. The calculation module is configured to determine local covariance matrices at multiple image locations in the at least two single-energy images. Each local covariance matrix is a matrix of the local variance and local covariance between image intensities at one of the multiple image locations in the at least two single-energy images. The rendering module is configured to superimpose the values of the local variance and / or local covariance at the multiple image locations with one or more of the at least two single-energy images. The output module is configured to provide the superposition result.
[0007] In the example, image data can be acquired by a dual-energy CT scanner or a multi-energy CT scanner. For example, in dual-energy CT, additional attenuation measurements are obtained using a second X-ray energy spectrum (i.e., a second "energy"), allowing for the differentiation of multiple materials. Alternatively, this allows for the quantification of the mass density of two or three materials in a mixture with a known elemental composition.
[0008] In another example, image data can be acquired via multiparameter magnetic resonance imaging (MRI), which includes the following sequences: T1-weighted images, T2-weighted images, diffusion-weighted images (DWI), and dynamic contrast-enhanced imaging (DCEI).
[0009] In another example, image data can be acquired using a set of ultrasound (US) modes, such as A / B / C / M modes or a set of harmonic modes.
[0010] Determine the local covariance matrix at multiple image locations (such as pixel locations or voxel locations) in the first and second single-energy images. Each local covariance matrix is a matrix of the local variance and local covariance between the image intensity at one of the multiple image locations in the first and second single-energy images. For each image location, the local variance and covariance can be calculated using image patches of adjacent pixels or voxels. An "image patch" refers to a block or group of pixels or voxels having a specific size, shape, and location corresponding to an image. For example, an image patch may have a predetermined pixel width / height (e.g., 7×7, 9×9, 11×11, etc.) and the location of each image patch may be defined based on one or more center pixels or voxels.
[0011] Local variance and covariance are positive definite and reach their maximum values at the location of material transitions. The local maxima of the variance and / or covariance of the intensity in multichannel images can be used to locate material transitions. These material transitions can be induced by contrast agents and indicate blood perfusion.
[0012] The covariance matrix values characterize certain material transformation types, such as adipose tissue, muscle tissue, air, and contrast-labeled materials. For example, Figure 2A-2C The results show that the local variance of the energy spectrum in the 'photoelectric' and 'scattering' images is sensitive to the colonic folds immersed in the contrast-labeled fecal residue, while the local covariance of the energy spectrum specifically represents the material transition from label to air.
[0013] The values of local variance and / or local covariance at multiple image locations can be superimposed with a first single-energy image and / or a second single-energy image to generate a superimposed representation of the first single-energy image and / or the second single-energy image. The superimposed representation of the first single-energy image and / or the second single-energy image can be visualized on a display.
[0014] Local variance and / or covariance can be selected discretely only for a specific type of transformation (e.g., the transformation between air and tissue), and the trace of the matrix can be used as weights for display. Alternatively, local variance and / or covariance can be weighted continuously for application-specific correlations.
[0015] In the example, the values of local variance and / or local covariance can be overlaid on a standard slice image using color coding.
[0016] In another example, grayscale values (such as gradual black-to-white transitions) can be used to convey cues of material transitions, ensuring a familiar viewing experience similar to regular images and the possibility of deployment on a widely distributed network of black-and-white monitors. The presence / absence and weight of the overlay can be interactively controlled or toggled by the operator (referred to as the "user" in this document).
[0017] Another visualization option is to plot the local variance magnitudes or regression values in three dimensions (3D) onto a virtual display, which can be interactively rotatable.
[0018] In other words, relevant material transitions can be superimposed as colors or grayscale values to convey additional energy-spectral information. In this way, users can evaluate only one image type, i.e., one energy-spectral band, instead of reading two or even an entire series of different image representations. Some material transitions that may be indistinguishable in conventional images are identifiable in energy-spectral image channels. Furthermore, because only transition lines (rather than entire regions) are marked, standard images are not cluttered by region-by-region occlusion superposition, and standard reading expertise does not need to be altered.
[0019] According to an embodiment of the present invention, the computation module is configured to compute each local variance matrix by means of a Gaussian convolution operation.
[0020] In other words, local variance and covariance can be computed using Gaussian convolution operations (i.e., smoothing operations). Gaussian filtering can be computed for various kernel widths corresponding to the scale space, and therefore the entire vector of the covariance matrix can be computed for various values of the scale space tensor of the covariance.
[0021] According to an embodiment of the invention, the calculation module is further configured to determine the number of eigenvalues of at least one local covariance matrix in the local covariance matrix, for the purpose of determining the amount of material that undergoes a material transition therebetween.
[0022] In other words, the calculation module is also configured to determine the number of eigenvalues of at least one local covariance matrix in the local covariance matrix, and classify at least one local covariance matrix in the local covariance matrix into material transitions between N materials based on the number of eigenvalues, where N is greater than or equal to 2.
[0023] In the example, a single eigenvalue of the local variance matrix can represent a bimaterial transition (e.g., between air and soft tissue).
[0024] In the example, the two eigenvalues of the local variance matrix can represent the three material transitions (e.g., between air, soft tissue, and contrast agent-labeled materials).
[0025] According to an embodiment of the present invention, the apparatus further includes a classifier module configured to apply a pre-trained classifier to classify each local covariance matrix in the local covariance matrix into a material transition type.
[0026] Examples of classifiers may include, but are not limited to, linear, maximum likelihood, random forest, support vector machine, or artificial neural network, which can be used to classify the covariance at any location into the material transition type.
[0027] The classification result can be one or more type-specific scalar values output from each local covariance matrix or covariance scale space tensor.
[0028] According to embodiments of the present invention, the pre-trained classifier includes at least one of the following: artificial neural network, support vector machine, maximum likelihood or random forest.
[0029] According to an embodiment of the present invention, the plurality of image locations include locations of material transitions.
[0030] Edge detection can be used to identify locations of material transitions. Edge detection is the name of a set of mathematical methods designed to identify points in digital images where image brightness changes abruptly or, more formally, where there are discontinuities. Points with abrupt changes in image brightness are typically organized into a set of curve segments called edges. Various techniques can be used for edge detection, such as Prewitt edge detection, Laplacian edge detection, LOG edge detection, Canny edge detection, etc.
[0031] In this way, the local covariance matrix is calculated only at the location of the material transition to identify the relevant multi-spectral material transition. Therefore, the computational workload can be reduced.
[0032] According to an embodiment of the present invention, the apparatus further includes a display device configured to display the superposition result.
[0033] In the example, the image data includes at least one of image data acquired by energy-dispersive computed tomography (CT) or image data acquired by energy-dispersive magnetic resonance imaging (MRI).
[0034] According to a second aspect of the invention, an apparatus is provided for processing image data of an object of interest, the image data comprising at least two single-energy images acquired from different energy spectral channels. The apparatus includes an input module, a computation module, and an output module. The input module is configured to receive the image data of the object of interest. The computation module is configured to determine local covariance matrices at multiple image locations in the at least two single-energy images, and to determine a local covariance deficit based on a comparison between the product of the local variances across all energy spectral channel combinations and the product of the local covariances. Each local covariance matrix is a matrix of local variances and local covariances between image intensities at one of the multiple image locations in the at least two single-energy images. The output module is configured to provide the local covariance deficit, which is preferably used for superimposing with one or more of the single-energy images by means of graphical cues, and / or as a machine learning feature for performing material classification.
[0035] In other words, an alternative method for processing image data is proposed. Unlike the method for determining local variance and / or local covariance specified in the first aspect of the invention, the alternative method in the second aspect of the invention performs voxel-by-voxel or pixel-by-pixel calculation of the inter-channel covariance loss. The multi-energy spectral covariance loss can be mapped to a color overlay or so-called conventional slice image on individual energy spectral channels. Alternatively or additionally, the inter-channel covariance loss can be used as a machine learning feature for performing material classification.
[0036] The proposed method is independent of spatial drift across image volumes. As will be discussed below, it also allows for fast computation of standard filters, which can be adapted for massively parallel processing, for example, on GPUs. In this alternative method, eigenvalue or singular value decomposition is not required.
[0037] According to an embodiment of the present invention, the apparatus further includes a rendering module configured to superimpose the values of the local covariance deficit at the plurality of image locations with one or more of the at least two single-energy images.
[0038] Multi-energy spectral covariance deficit can be conveyed to the reader using graphical cues (e.g., slice-by-slice color overlay, maximum intensity projection (MIP), direct volume rendering (DVR), 3D texture mapping, etc.).
[0039] By overlaying the values of local covariance deficits onto one or more single-energy images, salient locations in the energy spectrum within a multi-energy spectral image volume, such as that generated by photon-counting CT, can be identified and navigated. Therefore, users can evaluate only one image type, i.e., one energy band, saving the time of reading all channel image volumes separately. Furthermore, this allows for an unobstructed view over smooth areas of vanishing variance, while simultaneously marking different material transitions, when various covariance deficits are overlaid with different color hues and opacities according to their amplitudes. Thus, although differences between different spectral channels may be subtle and may be further obscured by the different overall dynamic scaling and offset of the multi-energy spectral channels (the proposed overlay image), users can still identify different material transitions in the overlay image. In particular, users can be able to identify subtle variations in the injected contrast agent that indicate local differences in blood perfusion.
[0040] According to an embodiment of the present invention, the apparatus further includes a display device configured to display the superposition result.
[0041] According to an embodiment of the present invention, the apparatus further includes a classifier module configured to apply a pre-trained classifier to perform material classification at one or more of the plurality of image locations based on the local covariance deficit.
[0042] Examples of pre-trained classifiers may include, but are not limited to, classifiers based on linear discriminant analysis (LDA), support vector machines (SVM), random forest classifiers, and convolutional neural networks (CNN).
[0043] Classification accuracy can be increased by adding a local covariance loss. For two spectral channels, voxel-by-voxel or pixel-by-pixel material classification can rely on two single energy values per voxel / pixel and the value of the local covariance loss, resulting in three features per voxel / pixel. For multi-band spectral images, the local covariance loss can be a computationally efficient summary of gradient magnitude and correlation.
[0044] According to an embodiment of the present invention, the calculation module is configured to determine local covariance deficit in the image space or projection space.
[0045] According to a third aspect of the present invention, a medical imaging system is provided. The medical imaging system includes a scanner and means according to the first aspect and any related examples, or according to the second aspect and any related examples. The scanner is configured to scan an object of interest to acquire image data of the object of interest. The means is configured to process the image data of the object of interest.
[0046] According to a fourth aspect of the present invention, a computer-implemented method is provided for processing image data of an object of interest, said image data comprising at least two single-energy images acquired from different energy spectral channels. The computer-implemented method includes:
[0047] - Receive the image data of the object of interest;
[0048] - Determine the local covariance matrix at multiple image locations in the at least two single-energy images, wherein each local covariance matrix is a matrix of the local variance and local covariance between the image intensities at one of the multiple image locations in the at least two single-energy images;
[0049] - The values of local variance and / or local covariance at the plurality of image locations are superimposed with one or more of the at least two single-energy images; and
[0050] - Provides overlay results.
[0051] In the example, the step of determining the local covariance matrix also includes the step of computing each local variance matrix through a Gaussian convolution operation.
[0052] In the example, the step of determining the local covariance matrix further includes determining the number of eigenvalues of at least one local covariance matrix in the local covariance matrix, in order to determine the amount of material that undergoes a material transition therebetween.
[0053] In the example, the method further includes the step of applying a pre-trained classifier to classify each local covariance matrix in the local covariance matrix into a material transition type.
[0054] According to a fifth aspect of the present invention, a computer-implemented method is provided for processing image data of an object of interest, the image data comprising at least two single-energy images acquired from different spectral channels at different energies, the computer-implemented method comprising:
[0055] - Receive the image data of the object of interest;
[0056] - Determine the local covariance matrix at multiple image locations in the at least two single-energy images, wherein each local covariance matrix is a matrix of the local variance and local covariance between the image intensities at one of the multiple image locations in the at least two single-energy images;
[0057] - The local multispectral covariance deficit is determined by comparing the product of the local variances of all spectral channel combinations with the product of the local covariances; and
[0058] - Provide the local covariance loss, which is preferably used to communicate with the aid of graphical cues and / or as a machine learning feature.
[0059] In the example, the computer-implemented method further includes superimposing the values of the local covariance deficit at the plurality of image locations with one or more of the at least two single-energy images.
[0060] In the example, the computer-implemented method also includes displaying the overlay results.
[0061] In the example, the computer-implemented method further includes applying a pre-trained classifier to perform material classification at one or more of the plurality of image locations based on the local covariance loss.
[0062] In the example, the computer-implemented method further includes determining a local covariance deficit in the image space or projection space.
[0063] According to another aspect of the invention, a computer program unit is provided, which is configured to execute method steps of the third aspect and any associated examples during operation.
[0064] In the example, a computer-readable medium including the computer program unit is provided.
[0065] Advantageously, the benefits provided by any of the foregoing aspects also apply to all other aspects, and vice versa.
[0066] As used in this article, the term "learning" in the context of machine learning refers to identifying and training suitable algorithms to accomplish a task of interest. The term "learning" includes, but is not limited to, association learning, classification learning, clustering, and number prediction.
[0067] As used in this article, the term “machine learning” refers to the field of computer science that studies the design of computer programs that can derive patterns, rules, or criteria from past experience to develop appropriate responses to future data or to describe data in a meaningful way.
[0068] As used in this article, the term "data-driven model" in the context of machine learning refers to a suitable algorithm that learns based on appropriate training data.
[0069] As used herein, the term “model” may refer to, or may include, an application-specific integrated circuit (ASIC), electronic circuitry, a processor (shared, dedicated, or grouped) and / or memory (shared, dedicated, or grouped) that executes one or more software or firmware programs, combinational logic circuitry, and / or other suitable components that provide the described functionality.
[0070] It should be understood that all combinations of the foregoing and additional concepts discussed in more detail below (assuming these concepts are not inconsistent with each other) are contemplated as part of the inventive subject matter disclosed herein. Specifically, all combinations of the claimed subject matter appearing at the end of this disclosure are contemplated as part of the inventive subject matter disclosed herein.
[0071] These and other aspects of the invention will become apparent with reference to one or more embodiments described below, and will be illustrated with reference to one or more embodiments described below. Attached Figure Description
[0072] In the accompanying drawings, similar reference numerals throughout different views generally refer to the same parts. Furthermore, the drawings are not necessarily drawn to scale, but typically focus on illustrating the principles of the invention.
[0073] Figure 1 A flowchart of a computer-implemented method is shown.
[0074] Figure 2A An exemplary 'photoelectric' image with local variance amplitude superimposed is shown.
[0075] Figure 2B An exemplary 'scattering' image is shown with local variance amplitude superimposed.
[0076] Figure 2C An exemplary 'scattering' image is shown with local covariance amplitudes superimposed.
[0077] Figure 3A An exemplary composite image of a 'photoelectric' image and a 'scattering' image with local variance amplitude is shown.
[0078] Figure 3B It shows how to get from Figure 3A Another exemplary synthesized image is obtained by subtracting the local covariance of the energy spectrum from the exemplary synthesized image.
[0079] Figure 4 A flowchart of an alternative computer-implemented method is shown.
[0080] Figure 5A An exemplary covariance-deficient image volume is shown, wherein the multi-energy spectral covariance deficit is superimposed using a 3D text mapping.
[0081] Figure 5B An exemplary covariance-deficient image volume is shown, where multi-spectral covariance deficit is represented using rotated maximum intensity projection (MIP).
[0082] Figure 6 An example of the device is shown schematically.
[0083] Figure 7 An alternative example of the device is illustrated schematically.
[0084] Figure 8 An example of a medical imaging system is illustrated schematically. Detailed Implementation
[0085] In the following description, the method is combined with abdominal spectral CT scans of contrast-labeled fecal residue in the colon. Although the following detailed description uses the application of CT virtual colonoscopy for illustrative purposes, those skilled in the art will understand that the methods, apparatus, and medical imaging systems described above and below can be applied to other objects of interest (e.g., blood vessel walls) and other imaging modalities (e.g., MRI). Therefore, examples described below are presented without loss of generality and without imposing limitations on the claimed invention.
[0086] Figure 1 A flowchart of a computer-implemented method 200 according to some embodiments of the present disclosure is shown. A computer-implemented method 200 for processing image data of an object of interest is proposed, the image data including at least two single-energy images, such as a first single-energy image acquired at a first energy and a second single-energy image acquired at a second energy different from the first energy.
[0087] The computer-implemented method 200 can be implemented as a device, module, or associated component stored in a non-transient machine or computer-readable storage medium (such as random access memory (RAM), read-only memory (ROM), programmable ROM (PROM), firmware, flash memory, etc.), stored in configurable logic (e.g., programmable logic array (PLA), field-programmable gate array (FPGA), complex programmable logic device (CPLD)), or stored in a set of logic instructions in fixed-function hardware logic using circuit technology (e.g., application-specific integrated circuit (ASIC), complementary metal-oxide-semiconductor (CMOS), or transistor-transistor logic (TTL) technology, or any combination thereof). For example, the computer program code performing the operations shown in method 200 can be written in any combination of one or more programming languages (including object-oriented programming languages such as JAVA, SMALLTALK, C++, Python, etc., and conventional programming languages such as the "C" programming language or similar programming languages).
[0088] In step 210 (i.e., step a), image data of the object of interest is received.
[0089] Medical images can be two-dimensional images that include image pixel data or three-dimensional images that include image voxel data.
[0090] In the example, the object of interest is the colon. In the example, the object of interest is the lumen of the vascular system.
[0091] In the example, image data can be acquired via spectral CT. For instance, the basic idea behind dual-energy imaging in CT is to acquire two datasets at low and high energy levels and use the paired datasets to infer additional information about the patient. The physics of dual-energy imaging involves two main mechanisms of X-ray-matter interaction within the clinically relevant diagnostic energy range of 30 keV to 140 keV, namely photoelectric absorption and Compton scattering, each having its own functional dependence on X-ray energy. Photoelectric absorption is a rapidly decreasing function of the energy used to generate a 'photoelectric' image, while Compton scattering is a moderate function of the energy used to generate a 'scattered' image.
[0092] In another example, image data can be acquired via multi-parameter MRI.
[0093] In step 220 (i.e., step b), local covariance matrices are determined at multiple image locations in the first and second single-energy images. Each local covariance matrix is a matrix of the local variance and local covariance between the image intensity at one of the multiple image locations in the first and second single-energy images. For the local variance at each pixel or voxel location, we use statistics from small patches of neighboring pixels or voxels (also referred to as image patches). Each image patch represents one or a group of pixels in a two-dimensional medical image, or one or a group of voxels in a three-dimensional medical image. For example, an image patch may have a predetermined pixel width / height (e.g., 7×7, 8×8, 9×9, etc.), and the location of each image patch may be defined based on one or more center pixels or voxels.
[0094] In the example, for two energy spectrum channels, the local covariance matrix has a 2×2 covariance matrix, where the three are independent.
[0095] In the example, for N (N>2) energy spectrum channels, the local covariance matrix has N×N covariance matrices.
[0096] Alternatively, the local variance and covariance can be calculated using a Gaussian convolution operation:
[0097] var ii (x)=G σ *(I i (x)·I i (x))-(G σ *(I i (x)·G*(I i (x))
[0098] cov ik (x)=G σ *(I i (x)·I k (x))-(G σ *(I i (x)·G*(I k (x))
[0099] Among them G σ It is a Gaussian convolution, and σ is the kernel width of the Gaussian filter, I i (x) is the image intensity at a certain image position x in the energy spectrum band i, and I k (x) is the image intensity at a certain image position x at the energy spectrum band k.
[0100] Gaussian filter G σ It is possible to compute for various kernel widths σ corresponding to the scale space, and therefore the entire vector of the covariance matrix can be computed for various values of σ to obtain the covariance scale space tensor.
[0101] The local variance and covariance are positive definite, and reach their maximum values at the material transition points. The softness of the filter response is affected by the Gaussian filter width σ.
[0102] Compared to gradient vectors, local variance is undirected and less susceptible to noise.
[0103] Optionally, the method further includes the step of: determining the number of eigenvalues of at least one local covariance matrix in the local covariance matrix, for the purpose of determining the amount of material that undergoes a material transition therebetween.
[0104] For example, in dual-energy CT, the two energy bands ('photoelectric' image and 'scattering' image) produce a 2×2 local covariance matrix. A single eigenvalue of the local covariance matrix represents a two-material transition, while two eigenvalues represent a three-material transition.
[0105] In step 230 (i.e., step c), the values of local variance and / or local covariance at multiple image locations are superimposed on the first single-energy image and / or the second single-energy image.
[0106] In the example, the values of local variance and / or local covariance can be overlaid on a standard slice image using color coding.
[0107] Another visualization option is to plot the local variance magnitude or regression values in 3D onto an interactively rotatable virtual display.
[0108] As an alternative, grayscale values (gradual black-to-white transitions, rather than color overlays) can be used to convey transition cues, ensuring a familiar viewing experience similar to regular images and feasible deployment on a wide range of black-and-white monitors. The presence / absence of the overlay and its weight can be interactively controlled / toggled by the user.
[0109] For example, Figures 2A to 2C An example is shown in an abdominal spectral CT scan of contrast-labeled fecal residue in the colon with two spectral bands ('photoelectric' image and 'scattering' image).
[0110] In particular, Figure 2A An exemplary 'photoelectric' image, superimposed with local variance amplitude, is shown as a brighter outline. This image demonstrates that the local variance of the energy spectrum in the 'photoelectric' image responds sensitively to colonic folds immersed in contrast-labeled fecal residue.
[0111] Figure 2B An exemplary 'scattering' image, superimposed with local variance amplitude, is shown as a brighter outline. This image demonstrates that...
[0112] The local variance of the energy spectrum in the 'scattering' image also responds sensitively to the colonic folds immersed in contrast-labeled fecal residue.
[0113] Figure 2C An exemplary 'scattering' image, superimposed with local covariance amplitudes, is shown as a brighter outline. This image demonstrates that the local covariance of the energy spectrum specifically represents the material transition from label to air.
[0114] Because the local variance of the energy spectrum in both the 'photoelectric' and 'scattering' images responds sensitively to the colonic folds submerged in contrast-labeled fecal residue, these two images can be synthesized to highlight the colonic folds submerged in contrast-labeled fecal residue. Figure 3A The figure shows an example of a composite image.
[0115] Additionally, by means of Figure 3A Subtracting the local covariance of the energy spectrum, which specifically represents the material transition from marker to air, from the synthesized image can further highlight the contours of the colonic folds immersed in contrast-labeled fecal residue. Figure 3B The figure below illustrates an example of subtraction.
[0116] Return to Figure 1 In step 240 (i.e. step d), the overlay result is provided to, for example, a display (e.g., a built-in screen, a connected monitor, or a projector) or a file storage device (e.g., a hard disk drive or a solid-state drive).
[0117] It should also be understood that, unless expressly stated to the contrary, in any method claimed herein that includes more than one step or action, the order of the steps or actions of the method is not necessarily limited to the order in which the steps or actions of the method are described.
[0118] Using the proposed method, users can evaluate only one image type, namely a single energy spectrum band, without having to read two or even an entire series of different image representations. Furthermore, associated material transitions can be superimposed as color or grayscale values, thereby conveying additional energy spectrum information. Some material transitions that may be indistinguishable in conventional images are discernible in energy spectrum image channels. Moreover, because only transition lines (rather than entire regions) are labeled, standard images are not cluttered by region-by-region occlusion superposition, and standard reading expertise need not be altered.
[0119] Optionally, a pre-trained classifier can be applied to classify each local covariance matrix into a material transition type.
[0120] In this example, the classifier is a pre-trained machine learning model. The machine learning model has been pre-trained on historical patient data retrievable from image repositories from the same or other hospitals. Preferably, a supervised learning scheme is used, where the historical images are pre-labeled by experienced clinicians. The labeling can provide data including local covariance matrices at multiple image locations across multiple images and the corresponding classification of material transition types (e.g., transitions between air-tissue, tissue contents, reagent-labeled materials, etc.).
[0121] Training a classifier can include the following steps: receiving training data in one or more iterations, applying the classifier to the training data. As a result of this application, a pre-trained classifier is obtained, which can then be used in deployment. In deployment, new image data can be applied to the pre-trained classifier to obtain classification results for that new data.
[0122] For example, neural network models (also known as artificial neural networks (ANNs)) can be used as examples of classifiers. However, other machine learning techniques such as support vector machines, maximum likelihood, and random forests can be used instead of neural networks.
[0123] The trained classifier attempts to approximate the correlation between the local covariance matrix and the material transition type. This approximation can be achieved during the learning or training process, where the parameters that form the high-dimensional space are adjusted based on the training data in the optimization scheme.
[0124] More specifically, the classifier can be implemented as an artificial neural network (“ANN”). An ANN can operate in two modes: a “training mode / phase” and a “deployment mode / phase”. In training mode, an initial model of the ANN is trained based on a set of training data to produce a trained ANN model. In deployment mode, the pre-trained ANN model is fed new, untrained data to operate during normal use. Training mode can be a one-time operation, or it can continue in repeated training phases to improve performance. Everything described so far regarding the two modes applies to any kind of machine learning algorithm and is not limited to ANNs.
[0125] An ANN consists of a set of interconnected nodes organized in layers. An ANN includes an output layer and an input layer. The input layer can be a matrix whose size (rows and columns) matches the size of the local covariance matrix of the training input. The output layer can be a vector or matrix with a size that matches the size chosen for the material transformation type.
[0126] ANNs preferably have a deep learning architecture, that is, there is at least one, preferably two or more, hidden layers between the output layer and the input layer.
[0127] Nodes are associated with numbers called “weights”, which represent how a node responds to inputs from earlier nodes in the previous layer.
[0128] The set of all weights defines the configuration of the ANN. During the learning phase, the initial configuration is adjusted based on the training data using a learning algorithm such as forward-backward (“FB”)-propagation or other optimization schemes or other gradient descent methods. Gradients are obtained with respect to the parameters of the objective function.
[0129] The training mode is preferably supervised, i.e., based on annotated training data. Annotated training data consists of pairs of training data items. For each pair, one item is the training input data, and the other item is target training data that is known prior to be correctly associated with its training input data item. This association defines the annotation and is preferably provided by a human expert. The training pair includes a local covariance matrix as training input and is associated with the material transformation type as output.
[0130] In training mode, preferably, multiple such pairs are applied to the input layer to propagate through the ANN until the output appears at the output layer. Initially, the output is typically different from the target. During optimization, the initial configuration is readjusted to achieve a good match between the input training data and their respective targets for all pairs. The matching is measured by a similarity metric, which can be formulated according to an objective function or a cost function. The aim is to tune the parameters to produce a low-cost, i.e., good match.
[0131] The training dataset is applied to the initially configured ANN and then processed according to a learning algorithm (such as the FB-propagation algorithm described above). At the end of the training phase, the pre-trained ANN can then be used in the deployment phase to compute the material transformation type for new data (i.e., newly acquired images not present in the training data).
[0132] The classification result can be one or more type-specific scalar values output from each local covariance matrix or covariance scale space tensor.
[0133] In the example, the local variance can be selected discretely only for a specific transformation type, and the trace of the matrix can be used as weights for display.
[0134] As an alternative to "hard" classification, for applications with specific relevance, local variances and / or covariances can be continuously weighted, for example, through regression rather than classification.
[0135] Figure 4 The illustration shows a flowchart of an alternative embodiment of a computer-implemented method 400 for processing image data of an object of interest, the image data comprising at least two single-energy images acquired from different spectral channels at different energies.
[0136] In step 410, image data of the object of interest is received.
[0137] In step 420, local covariance matrices are determined at multiple image locations in the first and second single-energy images. Each local covariance matrix is a matrix of the local variance and local covariance between the image intensity at one of the multiple image locations in the first and second single-energy images. This can be based on the above... Figure 1 The method described in step 220 is used to determine the local covariance matrix.
[0138] In step 430, the local multi-spectral covariance deficit is determined by comparing the product of the local variances of all energy spectrum channel combinations with the product of the local covariances.
[0139] As mentioned above, local variance and covariance can be calculated using spatially symmetric Gaussian or binomial convolution filters (or multiplication in Fourier space).
[0140] For two spectral channels i and k, the local variance and covariance can be calculated using the following Gaussian convolution operation:
[0141] var ii (x)=G σ *(I i (x)·I i (x))-(Gσ *(I i (x)·G*(I i (x))
[0142] cov ik (x)=G σ *(I i (x)·I k (x))-(G σ *(I i (x)·G*(I k (x))
[0143] Among them G σ It is a Gaussian convolution, and σ is the kernel width of the Gaussian filter, I i (x) is the image intensity at a certain image position x in the energy spectrum band i, and I k (x) is the image intensity at a certain image position x at the energy spectrum band k.
[0144] The correlation loss between two spectral channels can be defined as:
[0145] Correlation loss = 1 - (cov) ik (x)·cov ik (x)) / (var ii (x)·var kk (x))
[0146] However, when the denominator approaches zero, the correlation deficit is numerically unstable for locally vanishing variance.
[0147] Therefore, this disclosure proposes to consider the local covariance loss, which is positive semidefinite. The local covariance loss at a certain image location x can be defined as...
[0148] Covariance loss = cov ik (x)·cov ik (x)-var ii (x)·var kk (x)≥0
[0149] Alternatively, the signs of the mutual covariances can be multiplied to produce the following signed covariance deficit:
[0150] Signed covariance loss = sign(cov ik (x)·Covariance Loss
[0151] The covariance deficit is positive semidefinite and vanishes smoothly at the vanishing local variances without singularities.
[0152] For N (N>2) energy spectrum channels, we consider the following quadratic matrix of covariance:
[0153] C = {[I i (x)-m i ]·[I k (x)-m k ] T}
[0154] Among them, I i (x) is the image intensity at a certain image position x at the energy spectral band i in N energy spectral channels, I k (x) is the image intensity at a certain image position x in the energy spectrum band k of N energy spectrum channels, where m is the image intensity at a certain image position x. i It is the local average value at band i in the energy spectrum, and m k It is the local average value at the k-th position of the energy spectrum.
[0155] This produces N(N-1) / 2 independent covariance loss terms that add up to a multi-spectral covariance loss, which respond independently of which spectral channel indicates the added information.
[0156] Covariance loss can be defined as
[0157] Covariance loss = 1 t (diag(C)·diag(C) t -C⊙C)1
[0158] Where ⊙ is the Hadamard matrix product, and the diag operator extracts the main diagonal of the matrix, and 1 represents a vector consisting entirely of ones.
[0159] As an alternative formula, we consider the Kronecker product. That is, the product of the outer matrices of the quadratic covariance matrix C, and comparing the traces tr:
[0160]
[0161] The second term corresponds to the squared Frobenius matrix norm of C.
[0162] In some implementations, summation can be efficiently performed directly interleaved with Gaussian convolution operations, for example, using an overlapped addition scheme to leverage inherent parallelism and a small memory footprint. Therefore, the computational workload is O(N). 2 M) runtime and O(M) memory, where M is the number of voxels in the image.
[0163] In the example above, the multi-energy spectral covariance loss is calculated in the image space. Alternatively, the multi-energy spectral covariance loss can be calculated in the projection space before reconstructing the image volume. In this case, the local variance and covariance can be calculated using an adapted anisotropic smoothing filter shape. Finally, the projection covariance loss is reconstructed into the image space.
[0164] In step 440, the local covariance deficit is preferably provided for overlay with one or more of the at least two single-energy images and / or as a machine learning feature for performing material classification.
[0165] In one option, the values of local covariance loss at multiple image locations can be superimposed with one or more of the single-energy images from at least two single-energy images. For example, multi-energy spectral covariance loss can be superimposed using graphical cues such as slice-by-slice color overlay, maximum intensity projection (MIP), direct volume rendering (DVR), 3D texture mapping, etc.
[0166] Of particular interest is mapping the multi-energy spectral covariance loss to a color overlay, or so-called conventional slice image, on each energy channel. By construction, the covariance loss is highest at the location of material transitions. Therefore, when various covariance losses are overlaid with different color hues and opacities according to their magnitudes, this allows for an unobstructed view over smooth areas of vanishing variance, while simultaneously marking different material transitions.
[0167] Figure 5A and 5B Two exemplary covariance-deficient image volumes using a 9×9×9 binomial smoothing kernel are shown.
[0168] exist Figure 5A In the example shown, the multi-energy spectral covariance deficit is superimposed using a 3D text mapping. Figure 5A The dark gray area in the image corresponds to the blue area, which represents the vanishing covariance deficit. Figure 5A The light gray area in the image corresponds to yellow, which represents a significant covariance loss corresponding to the added multi-energy spectral information.
[0169] exist Figure 5B In the example shown, the rotated maximum intensity projection (MIP) is used to represent the covariance deficit.
[0170] For some applications, it may be beneficial to assess or present covariance deficit only in selected regions of interest (e.g., myocardium). For example, organ shape models can be used to perform enhancement.
[0171] In another option, a pre-trained classifier (e.g., SVM, CNN, etc.) can be applied to perform material classification at one or more image locations across multiple image locations based on a local covariance loss. The local covariance loss can incorporate energy spectral information to increase classification accuracy. For example, classification can rely on a single energy value for each pixel / voxel (e.g., two single energy values for a dual-channel energy spectral CT) and the value of the local covariance loss, resulting in multiple features per pixel. Of course, one or more other features can be added to each pixel / voxel, such as local mean values, local covariance matrices, etc.
[0172] Figure 6 An example of an apparatus 100 for processing image data of an object of interest is illustrated schematically, the image data including a first single-energy image acquired at a first energy and a second single-energy image acquired at a second energy different from the first energy.
[0173] The device 100 includes an input module 110, a calculation module 120, a drawing module 130, and an output module 140. Each module may be part of an application-specific integrated circuit (ASIC), electronic circuitry, a processor (shared, dedicated, or grouped) and / or memory (shared, dedicated, or grouped) executing one or more software or firmware programs, combinational logic circuitry, and / or other suitable components that provide the described functions, or may include an application-specific integrated circuit (ASIC), electronic circuitry, a processor (shared, dedicated, or grouped) and / or memory (shared, dedicated, or grouped) executing one or more software or firmware programs, combinational logic circuitry, and / or other suitable components that provide the described functions.
[0174] Device 100 can be any computing device suitable for processing image data, such as mobile devices, laptops and desktop computers, wearable computing devices and other computing devices.
[0175] Input module 110 is configured to receive image data of an object of interest. Medical images can be two-dimensional images including image pixel data or three-dimensional images including image voxel data. Examples of imaging modalities may include, but are not limited to, spectral CT or multiparametric MRI.
[0176] The calculation module 120 is configured to determine the local covariance matrix at multiple image locations in the first and second single-energy images. These multiple image locations include locations of material transitions. Each local covariance matrix is a matrix of the local variance and local covariance between image intensities at one of the multiple image locations in the first and second single-energy images. For the local variance at each pixel or voxel location, statistics from small patches of adjacent pixels or voxels (i.e., image patches) can be used for calculation.
[0177] Optionally, the computation module 120 can be configured to compute each local variance matrix via a Gaussian convolution operation.
[0178] Optionally, the calculation module 120 may also be configured to determine the number of eigenvalues of at least one local covariance matrix in the local covariance matrix, for the purpose of determining the amount of material that undergoes a material transition therebetween.
[0179] In the example, a single eigenvalue of the local variance matrix can represent, for example, a bimaterial transition between air and soft tissue.
[0180] In the example, the two eigenvalues of the local variance matrix can represent, for example, the three-material transition between air, soft tissue, and contrast agent-labeled materials.
[0181] The rendering module 130 is configured to overlay the values of local variance and / or local covariates (e.g., in color or grayscale values) at multiple image locations with a first single-energy image and / or a second single-energy image. Figure 2A-2C and Figure 3A An exemplary overlay result is shown in -3C.
[0182] The output module 140 is configured to provide the overlay result, for example, to a display (e.g., a built-in screen, a connected monitor, or a projector) or a file storage device (e.g., a hard disk drive or a solid-state drive).
[0183] Optionally, the device 100 may also include a classifier module configured to apply a pre-trained classifier (such as an artificial neural network, support vector machine, maximum likelihood, or random forest) to classify the local covariance matrix into a material conversion type.
[0184] Optionally, the apparatus may include a display device (not shown) configured to display the overlay result.
[0185] Figure 7 An alternative embodiment of the device 500 is schematically shown, which includes an input module 510, a calculation module 520, and an output module 530.
[0186] Input module 510 can be essentially equivalent in function to Figure 6 The input module 110 shown.
[0187] The calculation module 520 is configured to determine the local covariance matrix at multiple image locations in the first and second single-energy images. Each local covariance matrix is a matrix of the local variance and local covariance between image intensities at one of the multiple image locations in the at least two single-energy images. The local covariance matrix can be determined based on... Figure 1The method described in step 220 is used to determine this.
[0188] The calculation module 520 is also configured to determine the local covariance deficit based on a comparison between the product of the local variances of all energy spectrum channel combinations and the product of the local covariances. The local covariance deficit can be determined based on... Figure 4 The method described in step 430 is used to determine this. The calculation module can be configured to determine the local covariance loss in the image space or projection space.
[0189] The output module 530 is configured to provide a local covariance loss, which is preferably used to overlay with one or more of the at least two single-energy images and / or as a machine learning feature for performing material classification.
[0190] In some examples, the apparatus 500 may also include a rendering module (not shown) configured to superimpose the values of the local covariance deficit at the plurality of image locations with one or more of the at least two single-energy images. Figures 5A-5B An exemplary overlay result is shown. The device 500 may also include a display (not shown) for displaying the overlay result.
[0191] In some examples, the apparatus 500 may also include a classifier module (not shown) configured to apply a pre-trained classifier to perform material classification at one or more image locations among a plurality of image locations based on local covariance loss.
[0192] Figure 8 A medical imaging system 300 according to some embodiments of the present disclosure is illustrated schematically.
[0193] The medical imaging system 300 includes a scanner 310 configured to scan an object of interest to acquire image data of that object. The scanner 310 may be a spectral CT scanner or an MRI scanner.
[0194] The medical imaging system 300 also includes, according to the relevant provisions Figure 6 Examples of descriptions or about Figure 7 The described example is an apparatus 100 for processing image data of an object of interest.
[0195] All definitions defined and used herein should be understood to be governed by the dictionary definition, the definition in references incorporated herein by reference, and / or the general meaning of the defined terms.
[0196] The words “a” and “an” as used herein in the specification and claims should be understood to mean “at least one” unless explicitly indicated to the contrary.
[0197] The phrase “and / or” as used in this specification and claims should be understood to mean “any one or both” of the elements so joined together, i.e., elements that exist together in some cases and separately in others. Multiple elements listed using “and / or” should be interpreted in the same way, i.e., “one or more” of the elements so joined together. Other elements may optionally be present, whether or not they are related to those specifically identified, except for those elements specifically identified by the “and / or” clause.
[0198] As used in this specification and claims, “or” should be understood to have the same meaning as “and / or” as defined above. For example, when items are separated in a list, “or” or “and / or” should be interpreted as inclusive, that is, including at least one element from multiple elements or a list of elements, as well as optional other unlisted items, but also including more than one element. Only terms clearly indicating the opposite, such as “only one” or “exactly one”, or, when used in the claims, “consisting of…”, will refer to including exactly one element from multiple elements or a list of elements. Generally, the term “or” as used herein should only be interpreted as indicating an exclusive alternative (i.e., “one or another but not both”) when preceded by an exclusive term (e.g., “any one,” “one of,” “only one,” or “exactly one”).
[0199] As used herein in the specification and claims, the phrase "at least one" relating to a list of one or more elements should be understood to mean at least one element selected from any one or more elements in the list, but not necessarily every single element specifically listed in the list, or at least one of each element, and does not exclude any combination of elements in the list. This definition also allows for the optional presence of elements other than those specifically identified in the list of elements referred to by the phrase "at least one," whether related to or unrelated to those specifically identified elements.
[0200] In the claims and the description above, all transitional phrases (e.g., "comprising," "including," "carrying," "having," "containing," "involving," "holding," "with," etc.) should be understood as open-ended, that is, meaning including but not limited to. Only the transitional phrases "consisting of" and "consisting substantially of" should be closed or semi-closed transitional phrases, respectively.
[0201] In another exemplary embodiment of the invention, a computer program or computer program unit is provided, characterized in that it is adapted to run on a suitable system the method steps of one of the methods described in the foregoing embodiments.
[0202] The computer program unit can therefore be stored on the computer unit, which may also be part of an embodiment of the present invention. The computing unit can be adapted to perform the steps of the described method or to induce the execution of the steps of the described method. Furthermore, it can be adapted to operate components of the apparatus described above. The computing unit is adapted to operate automatically and / or to execute user commands. The computer program can be loaded into the working memory of the data processor. Therefore, the data processor can be equipped to perform the method of the present invention.
[0203] This exemplary embodiment of the invention covers both computer programs that initially use the invention and computer programs that, through updates, transform existing programs into programs that use the invention.
[0204] Furthermore, the computer program unit may be able to provide all the necessary steps to perform the process of an exemplary embodiment of the method as described above.
[0205] According to yet another exemplary embodiment of the present invention, a computer-readable medium, such as a CD-ROM, is provided, wherein the computer-readable medium has computer program units stored thereon, the computer program units being described in the preceding sections.
[0206] Computer programs can be stored and / or distributed on suitable media, such as optical storage media or solid-state media provided with or as part of other hardware, but the computer programs can also be distributed in other forms, such as via the Internet or other wired or wireless long-distance communication systems.
[0207] However, computer programs can also be provided on networks such as the World Wide Web and can be downloaded from such networks to the working memory of a data processor. According to yet another exemplary embodiment of the invention, a medium is provided for making computer program units available for download, said computer program units being arranged to perform a method according to one of the embodiments described above.
[0208] Although numerous embodiments of the invention have been described and illustrated herein, those skilled in the art will readily conceive of a variety of other means and / or structures for performing functions and / or obtaining results and / or one or more of the advantages described herein, and each such variation and / or modification should be considered within the scope of the embodiments of the invention described herein. More generally, those skilled in the art will readily recognize that all parameters, dimensions, materials, and constructions described herein are meant to be exemplary, and actual parameters, dimensions, materials, and / or constructions will depend on the specific application or application in which the teachings of this invention are used. Those skilled in the art will recognize, or can determine, many equivalent forms of the specific embodiments of the invention described herein using only conventional experimentation. Therefore, it should be understood that the foregoing embodiments are presented by way of example only, and other embodiments of the invention besides those specifically described and claimed can be implemented within the scope of the appended claims and their equivalents. The embodiments of the invention disclosed herein relate to each individual feature, system, article, material, kit, and / or method described herein. Furthermore, any combination of two or more such features, systems, articles, materials, kits, and / or methods (provided that these features, systems, articles, materials, kits, and / or methods are not inconsistent with each other) is included within the scope of the invention disclosed herein.
Claims
1. An apparatus (100) for processing image data of an object of interest, said image data comprising at least two single-energy images acquired from different energy spectral channels, said apparatus comprising: Input module (110), configured to receive the image data of the object of interest; A calculation module (120) is configured to determine local covariance matrices at multiple image locations in the at least two single-energy images, wherein each local covariance matrix is a matrix of local variance and local covariance between image intensities at one of the multiple image locations in the at least two single-energy images; A rendering module (130) is configured to superimpose the values of local variance and / or local covariance at the plurality of image locations with one or more of the at least two single-energy images; and Output module (140) is configured to provide overlay results.
2. The apparatus according to claim 1, in, The computation module is configured to compute each local variance matrix through Gaussian convolution operations.
3. The apparatus according to claim 1 or 2, in, The calculation module is also configured to determine the number of eigenvalues of at least one local covariance matrix in the local covariance matrix, in order to determine the amount of material that undergoes a material transition therebetween.
4. The apparatus according to claim 1 or 2, further comprising: A classifier module is configured to apply a pre-trained classifier to classify each local covariance matrix in the local covariance matrix into a material transition type.
5. The apparatus according to claim 1 or 2, in, The multiple image locations include the locations of material transitions.
6. The apparatus according to claim 1 or 2, further comprising: A display device is configured to display the overlay result.
7. An apparatus (500) for processing image data of an object of interest, said image data comprising at least two single-energy images acquired from different energy spectral channels, said apparatus comprising: Input module (510), configured to receive the image data of the object of interest; A calculation module (520) is configured to determine the local covariance matrix at multiple image locations in the at least two single-energy images, and to determine the local covariance deficit based on a comparison between the product of the local variances across all energy channel combinations and the product of the local covariances. Each local covariance matrix is a matrix of the local variance and local covariance between the image intensities at one of the plurality of image locations in the at least two single-energy images; as well as An output module (530) is configured to provide the local covariance loss, which is used to superimpose with one or more of the at least two single-energy images and / or as a machine learning feature for performing material classification.
8. The apparatus according to claim 7, further comprising: A rendering module is configured to superimpose the values of the local covariance deficit at the plurality of image locations onto one or more of the at least two single-energy images.
9. The apparatus according to claim 8, further comprising: Display device, which is configured to display the overlay result.
10. The apparatus according to any one of claims 7 to 9, further comprising: A classifier module is configured to apply a pre-trained classifier to perform material classification at one or more of the plurality of image locations based on the local covariance loss.
11. The apparatus according to any one of claims 7 to 9, in, The computation module is configured to determine the local covariance deficit in the image space or projection space.
12. A medical imaging system (300), comprising: A scanner configured to scan an object of interest to acquire image data of the object of interest; as well as The apparatus for processing image data of the object of interest according to any one of claims 1 to 2 or any one of claims 7 to 9.
13. A computer-implemented method (200) for processing image data of an object of interest, said image data comprising at least two single-energy images acquired from different energy spectral channels, said computer-implemented method comprising: Receive the image data of the object of interest (210); Determine (220) the local covariance matrix at multiple image locations in the at least two single-energy images, wherein each local covariance matrix is a matrix of the local variance and local covariance between the image intensities at one of the multiple image locations in the at least two single-energy images; The values of local variance and / or local covariance at the plurality of image locations are superimposed with one or more of the at least two single-energy images (230); and Provide (240) superposition results.
14. A computer-implemented method (400) for processing image data of an object of interest, said image data comprising at least two single-energy images acquired from different energy spectral channels, said computer-implemented method comprising: Receive the image data of the object of interest (410); Determine (420) the local covariance matrix at multiple image locations in the at least two single-energy images, wherein each local covariance matrix is a matrix of the local variance and local covariance between the image intensity at one of the multiple image locations in the at least two single-energy images; (430) Local multispectral covariance deficit is determined by comparing the product of the local variances of all spectral channel combinations with the product of the local covariances; and Provide (440) the local covariance loss, which is used to superimpose with one or more of the at least two single-energy images and / or as a machine learning feature for performing material classification.
15. A computer program product comprising instructions that, when executed by a processing unit, cause the processing unit to perform the steps of the method according to claim 13 or 14.
Citation Information
Patent Citations
Image processor, program and x-ray ct apparatus
JP2009153829A
