Method for reconstructing an image and medical imaging system

Reconstructing CT images using one-dimensional homography resampling transformation solves the problem of difficulty in implementing geometric structure models on large-scale parallel computing architectures, improves the accuracy and efficiency of image reconstruction, is suitable for deep learning image reconstruction, and promotes the accuracy of medical diagnosis.

CN114067013BActive Publication Date: 2024-11-08GE PRECISION HEALTHCARE LLC
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202110788677.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-07-31
Filing Date
2021-07-13
Publication Date
2024-11-08
Estimated Expiration
2041-07-13

AI Technical Summary

Technical Problem

Existing CT imaging reconstruction models based on geometry are difficult to parallelize on large-scale parallel computing architectures, resulting in limited computational efficiency and accuracy, and making it difficult to meet the needs of deep learning image reconstruction.

Method used

One-dimensional homography resampling transformation is used to reconstruct images. The homography transformation reconstructs the distribution samples of the object of interest with the interaction weights of the X-ray detector interval, avoiding geometric representation and making it suitable for large-scale parallel computing architectures.

Benefits of technology

It improves the accuracy and computational efficiency of image reconstruction, is suitable for deep learning image reconstruction, generates higher fidelity reconstructed images, and promotes the accuracy of medical diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114067013B_ABST
    Figure CN114067013B_ABST
Patent Text Reader

Abstract

The present application relates to a method and a medical imaging system for reconstructing an image. The present invention provides a system and method for reprojecting and backprojecting an object of interest via a homography transformation and specifically a one-dimensional homography transformation. In one example, a method may include: acquiring imaging data corresponding to a plurality of divergent X-rays, assigning a single functional form to the plurality of divergent X-rays, determining weights of interactions between a plurality of distribution samples and a plurality of X-ray detector intervals based on the single functional form via a homography transformation, and reconstructing an image based on the weights of the interactions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the subject matter disclosed herein relate to medical imaging, and more particularly to reprojecting and backprojecting an object of interest via a one-dimensional homography resampling transform. Background Art

[0002] Computed tomography (CT) can be used as a non-invasive medical imaging technique. Specifically, CT imaging data acquisition can include passing an X-ray beam through an object, such as a patient, causing the X-ray beam to attenuate, and then collecting the attenuated X-ray beam at an X-ray detector array. The acquired CT imaging data can be a set of line integral measurements corresponding to an attenuation coefficient distribution of the object. The distribution can be reconstructed into a visible image from the set of line integral measurements via a backprojection or forward projection step in an analytical or iterative reconstruction algorithm. In the case of an iterative reconstruction algorithm, the reconstructed distribution can then be mapped onto a set of updated calculated line integrals via a reprojection or forward projection step, after which the backprojection step and the reprojection step can be iterated as needed.

[0003] The tomographic operator or projector may be selected to encode a geometric representation of the integral line corresponding to the CT imaging data. For the backprojection step, this can be expressed algebraically as:

[0004]

[0005] where p i is the ith line integral measurement value of the I line integral measurement value, and r ij is the element of the matrix representation R of the tomography operator, which backprojects the i-th measurement value to the sampling point f corresponding to the j-th attenuation coefficient (i.e., the attenuation coefficient at the j-th sampling point among the J sampling points) j The adjoint (transpose, since R is real) operation constitutes the reprojection step, expressed as:

[0006]

[0007] There are many such geometry-based models for image reconstruction, and the choice of which model determines the accuracy and therefore the final image quality. Computational speed is an additional factor, which can be achieved through massive parallelization, such as on a graphics processing unit (GPU). However, geometry-based models can be difficult to implement and scale on a GPU computing framework. Both accuracy and computational speed can be key factors in driving image reconstruction via deep learning techniques. Summary of the invention

[0008] In one embodiment, a method may include: acquiring imaging data corresponding to multiple divergent X-rays, assigning a single functional form to the multiple divergent X-rays, determining weights of interactions between multiple distribution samples and multiple X-ray detector intervals based on the single functional form via a homography transformation, and reconstructing an image based on the interaction weights. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The present disclosure will be better understood by reading the following description of non-limiting embodiments with reference to the accompanying drawings, in which:

[0010] Figure 1 shows a pictorial view of an exemplary medical imaging system according to one embodiment;

[0011] Figure 2 shows a schematic block diagram of an exemplary medical imaging system according to one embodiment;

[0012] Figure 3 A schematic diagram of an exemplary geometry-based distance-driven image reconstruction model is shown;

[0013] Figure 4A A schematic diagram showing an exemplary flat panel detector based cone beam tomography geometry according to one embodiment;

[0014] Figure 4B A schematic diagram showing an exemplary one-dimensional homography resampling transformation according to one embodiment is shown;

[0015] Figure 5 A schematic diagram showing a basis set approximation of a continuous function resampled using an exemplary one-dimensional homography resampling transform according to one embodiment; and

[0016] Figure 6 A flow chart of a method for iterative image reconstruction via an exemplary homography-based model is shown according to one embodiment. DETAILED DESCRIPTION

[0017] The following description relates to various embodiments of medical imaging systems and methods for reprojecting and backprojecting an object of interest via a symmetric projector / backprojector operator pair. One such medical imaging system configured to acquire computed tomography (CT) medical imaging data is shown in Figure 1 and Figure 2 and one such method for (iterative) reprojection and backprojection via a symmetric projector / backprojector operator pair is provided in Figure 6 In the present invention, the reconstructed image can be accurately and efficiently obtained on a massively parallel computing architecture (e.g., multiple graphics processing units (GPUs)).

[0018] An image of an object of interest can be reconstructed from multiple line integral measurements via analytical or iterative reconstruction methods. Each method utilizes a backprojection operator or backprojector to map multiple line integral measurements onto an attenuation coefficient distribution of the object of interest. Iterative reconstruction methods can also utilize an adjoint of a backprojector (referred to as a projection operator or projector) to reshape the distribution of attenuation coefficients into multiple calculated line integrals that can be compared to the line integral measurements.

[0019] The projector / backprojector pair may encode a geometric representation of a plurality of integration lines along which a plurality of line integral measurements are acquired. A geometry-based image reconstruction model may utilize the geometric representation to reconstruct the attenuation coefficient distribution of a given object of interest from which a visible image may be obtained. An example of such a geometry-based image reconstruction model (a so-called range-driven model) is schematically shown in Figure 3 middle.

[0020] Depending on the image reconstruction scheme (e.g., analytical or iterative), mathematical constraints on the symmetry of the projector / backprojector pair may limit the choice of a suitable model. Although geometry-based image reconstruction models satisfy such constraints, computational scalability may be hampered due to inherent difficulties in implementation on massively parallel computing architectures.

[0021] Thus, in embodiments described herein, systems and methods are provided for resampling continuous functions via homography transformations. Figure 4A Aspects of one such homography-based image reconstruction model are discussed in the context of an exemplary flat panel detector-based cone beam tomography geometry of , and an exemplary one-dimensional homography resampling transform is schematically illustrated in Figure 4B In. Figure 5 It is further schematically shown in that utilizing homography transformations in this manner will provide an arbitrary level of accuracy via basis set expansion of continuous functions without sacrificing the symmetry of the projector / back-projector pair. Therefore, no assumptions have been made about the geometric representation of the integral line. By reshaping or completely avoiding the geometric representation, the problem of large-scale parallel implementation peculiar to geometric structure-based image reconstruction models can be avoided. In fact, the image reconstruction model based on homography can be advantageously applied to the parallelization of geometric structure-based image reconstruction models. Therefore, the image reconstruction model based on homography can be easily adapted to the deep learning paradigm of image reconstruction. In this way, a more generalized image reconstruction model is provided, which is used to symmetrically reproject and back-project an object of interest via a homography transformation.

[0022] An exemplary embodiment of the imaging system may include a massively parallel GPU architecture configured to determine the coefficients of the matrix representation of the projector / backprojector operator pair substantially in parallel. Thus, the homography-based image reconstruction model described herein may enable more efficient processing of imaging data compared to geometry-based image reconstruction models (which may utilize loops that are difficult to parallelize, may be parallelized to a more limited extent, or may not be parallelized at all).

[0023] In addition, arbitrary levels of precision with which reprojection and backprojection can be performed by varying the size and structure of the basis set extension can result in improved precision of homography-based image reconstruction models relative to geometry-based image reconstruction models, even while maintaining improved processing efficiency via easily and massively parallelized coefficient determination (as described above). In this way, the homography-based image reconstruction model described herein can provide technical improvements in computer functionality by simultaneously improving processing efficiency and precision of image reconstruction algorithms. Furthermore, specific implementations on massively parallel GPU architectures can be well suited for interacting with deep learning image reconstruction techniques, which can further improve image fidelity. Thus, in some embodiments, a higher fidelity reconstructed image can be generated, which can simultaneously promote higher accuracy in the diagnosis of medical problems depicted by the reconstructed image.

[0024] Reference now Figure 1 , depicts an exemplary imaging system 100 according to one embodiment. In the illustrated embodiment, the imaging system 100 is an X-ray imaging system configured to perform CT imaging. Although the illustrated embodiment actively acquires medical images, it should be understood that other embodiments do not actively acquire medical images. Instead, embodiments may retrieve images or imaging data previously acquired by the imaging system and process the imaging data as described herein.

[0025] The imaging system 100 may be configured to image a subject 112 (such as a patient, an inanimate object, one or more manufactured parts) and / or a foreign object (such as a dental implant, a stent, and / or a contrast agent present in the body). In one embodiment, the imaging system 100 may include a gantry 102, which in turn may further include at least one X-ray source 104 configured to project an X-ray radiation beam 106 (see Figure 2 ) for imaging a subject 112 placed on an examination table 114. Specifically, the X-ray source 104 may be configured to project an X-ray radiation beam 106 toward a detector array 108 positioned on an opposite side of the gantry 102. Figure 1 A curved detector array 108 is shown, but in some embodiments, a flat panel detector may be used. Figure 1A single X-ray source 104 is shown, but in certain embodiments, multiple X-ray sources and / or detectors may be used to project multiple X-ray radiation beams 106 for collecting projection data corresponding to the subject 112 at different energy levels or angular orientations. In some CT imaging embodiments, the X-ray source 104 can achieve dual energy imaging by fast peak kilovolt (kVp) switching. In some embodiments, the X-ray detectors used are photon counting detectors capable of distinguishing X-ray photons of different energies. In other embodiments, two sets of X-ray sources and detectors are used to generate dual energy projections, one set of which is collected at a low kVp setting and the other set of which is collected at a high kVp setting. It should therefore be understood that the methods described herein can be implemented using single energy acquisition techniques as well as dual energy acquisition techniques.

[0026] In some embodiments, the imaging system 100 also includes an image processor unit 110 that is configured to reconstruct an image of a target volume of a subject 112 using an iterative or analytical image reconstruction method or a combination of both. For example, in some CT imaging applications, the image processor unit 110 may use an analytical image reconstruction method such as filtered back projection (FBP) to reconstruct an image of a patient's target volume. As another example, the image processor unit 110 may use an iterative image reconstruction method such as advanced statistical iterative reconstruction (ASIR) or model-based iterative reconstruction (MBIR) to reconstruct an image of a target volume of a subject 112. In some examples, in addition to the iterative image reconstruction method, the image processor unit 110 may also use an analytical image reconstruction method such as FBP. In one embodiment, and as discussed in detail below, the image processor unit 110 may use an iterative image reconstruction method that utilizes a one-dimensional homography resampling transform.

[0027] In some CT imaging system configurations, an X-ray source projects a cone-shaped X-ray radiation beam that is collimated to lie in an XYZ plane of a Cartesian coordinate system (commonly referred to as an "imaging plane"). The X-ray radiation beam passes through an object being imaged, such as a patient or subject. The X-ray radiation beam, after being attenuated by the object, impinges upon a radiation detector array. The intensity of the attenuated X-ray radiation beam received at the detector array depends on the attenuation of the X-ray radiation beam by the object. Each detector element of the array produces a separate electrical signal that is a measure of the attenuation of the X-ray beam at the detector location (e.g., a line integral measurement). The attenuation measurements from all detector elements are collected individually to produce a transmission profile.

[0028] In some CT imaging systems, a gantry is used to rotate an X-ray source and a detector array around an imaging plane around an object to be imaged, so that the angle at which the radiation beam intersects the object is constantly changing. A set of X-ray radiation attenuation measurements (e.g., projection data) from a detector array at one angular position of the gantry is referred to as a "view". A "scan" of an object includes a set of views obtained at different angular positions or viewing angles during one rotation of the X-ray source and detector around the object. It is conceivable that the benefits of the methods described herein are derived from many medical imaging modalities, so as used herein, the term "view" is not limited to the use described above with respect to projection data from one gantry angle. The term "view" is used to mean a data acquisition whenever there are multiple data acquisitions from different angles (whether from CT, X-ray radiography, positron emission tomography (PET) or single photon emission CT (SPECT) acquisitions), and / or any other modality (including modalities yet to be developed) and their combination in fusion embodiments.

[0029] The projection data is processed to reconstruct an image corresponding to one or more two-dimensional slices acquired through the object, or in some examples where the projection data includes an extended axial coverage (e.g., Z-axis illumination), an image corresponding to a three-dimensional image volume of the object is reconstructed. A method for reconstructing an image based on a projection data set is referred to in the art as a filtered back-projection technique. Transmission and emission tomography reconstruction techniques also include statistical iterative methods such as maximum likelihood expectation maximization (MLEM) and ordered subset expectation maximization reconstruction techniques and iterative reconstruction techniques. The method converts the attenuation measurements from the scan into an integer (referred to as a "CT number" or "Hounsfield unit" in terms of a CT imaging system), which is used to control the brightness of the corresponding pixel on the display device.

[0030] To reduce the total scan time, a "helical" scan can be performed. To perform a "helical" scan, the patient is moved while acquiring data for a specified axial coverage. Such systems produce a single spiral from a cone beam helical scan. The spiral mapped out by the cone beam produces projection data from which an image in each specified slice can be reconstructed.

[0031] As used herein, the phrase "reconstructing an image" is not intended to exclude embodiments of the present disclosure in which data representing an image is generated rather than a visible image. Thus, as used herein, the term "image" broadly refers to both a visible image and data representing a visible image. However, many embodiments generate (or are configured to generate) at least one visible image.

[0032] Reference now Figure 2 , depicting something like Figure 11. An exemplary imaging system 200 of the imaging system 100 of FIG. 10 is shown. As shown, the imaging system 200 may include a plurality of components. The components may be coupled to each other to form a single structure, may be separate but located in a common room, or may be remote relative to each other. For example, one or more of the modules described herein may operate in a data server that has a different and remote location relative to the other components of the imaging system 200.

[0033] According to aspects of the present disclosure, the imaging system 200 can be configured to image a subject 204 (eg, Figure 1 In one embodiment, the imaging system 200 may include a detector array 108 (see Figure 1 ). The detector array 108 may also include a plurality of detector elements 202 that together sense the X-ray radiation beam 106 passing through a subject 204 (such as a patient) to acquire corresponding projection data. Thus, in one embodiment, the detector array 108 may be fabricated including a multi-row unit or multi-slice configuration of detector elements 202. In such a configuration, one or more additional rows of detector elements 202 may be arranged in a parallel configuration to acquire projection data.

[0034] The gantry 102 movably supports the X-ray source 104 and the detector array 108 mounted on opposite ends relative to each other. The subject 204 can thus be supported by the examination table 114 and disposed between the X-ray source 104 and the detector array 108 .

[0035] It should be understood that in some embodiments, the examination table 114 may also be movable to achieve the desired image acquisition. During such image data acquisition, the gantry 102 may be movable to change the position and / or orientation of the x-ray source 104 and / or the detector array 108 relative to the subject 204.

[0036] Thus, in some embodiments, the gantry 102 may remain stationary during a given imaging session in order to image a single 2D projection of the subject 204. In such embodiments, the position of the gantry 102 and / or the table 114 may be adjusted between imaging sessions in order to image another view of the subject 204.

[0037] In other embodiments, such as in CT imaging applications, the imaging system 200 can be configured to traverse different angular positions around the subject 204 to collect the desired projection data. Thus, the gantry 102 and components mounted thereon can be configured to rotate about the rotation center 206 to collect, for example, projection data at different energy levels. Alternatively, in embodiments where the projection angle relative to the subject 204 changes over time, the mounted components can be configured to move along a generally curved line rather than along a segment of a circle.

[0038] In such embodiments, the detector array 108 may collect data of the attenuated X-ray beam as the X-ray source 104 and the detector array 108 rotate. The data collected by the detector array 108 may undergo pre-processing and calibration to condition and process the data to represent the line integral of the attenuation coefficient of the scanned subject 204. The processed data is often referred to as a projection.

[0039] In some examples, individual detectors or detector elements 202 in detector array 108 may include photon counting detectors that register interactions of individual photons into one or more energy bins. It should be appreciated that the methods described herein may also be implemented using energy integrating detectors.

[0040] The acquired projection data sets may be used for base material decomposition (BMD). During BMD, the measured projections may be converted into a set of material density projections. The material density projections may be reconstructed to form a pair or set of material density maps or images (such as bone, soft tissue, and / or contrast agent maps) for each corresponding base material. The material density maps or images may then be correlated to form a volume rendering of the base material (e.g., bone, soft tissue, and / or contrast agent) in the imaging volume.

[0041] Once reconstructed, the base material image produced by the imaging system 200 can reveal internal features of the subject 204 represented by the density of the two base materials. The density image or a combination of multiple density images can be displayed to show these features. In traditional methods of diagnosing medical conditions (such as disease states), and more generally diagnosing medical events, a radiologist or physician will consider a hard copy or display of a density image or a combination thereof to discern characteristic features of interest. Such features can include lesions, size and shape of specific anatomical structures or organs, as well as other features that should be discernible in the image based on the skills and knowledge of the individual practitioner.

[0042] In one embodiment, the imaging system 200 may include a control mechanism 208 to control movement of components, such as rotation of the gantry 102 and operation of the x-ray source 104. In certain embodiments, the control mechanism 208 may also include an x-ray controller 210 configured to provide power and timing signals to the x-ray source 104. Additionally, the control mechanism 208 may include a gantry motor controller 212 configured to control the rotational speed and / or position of the gantry 102 or various components thereof (e.g., the x-ray source 104, the detector array 108, etc.) based on imaging requirements.

[0043] In certain embodiments, the control mechanism 208 may also include a data acquisition system (DAS) 214 that is configured to sample analog data received from the detector elements 202 and convert the analog data to digital signals for subsequent processing. For a photon counting imaging system, the DAS 214 may download measured photon counts in one or more energy bins from the detector array 108. The DAS 214 may also be configured to selectively aggregate analog data from a subset of the detector elements 202 into a so-called macro-detector, as further described herein.

[0044] The data sampled and digitized by the DAS 214 may be transmitted to a computer or computing device 216. In the illustrated embodiment, the computing device 216 may be configured to interact with various components of the imaging system 200. As such, the computing device 216 may be configured to control the operation of the imaging system 200. In various embodiments, the computing device 216 may take the form of a mainframe computer, a server computer, a desktop computer, a laptop computer, a tablet device, a network computing device, a mobile computing device, a mobile communication device, etc. In one embodiment, the computing device 216 may take the form of an edge device for use in Figure 2 In some embodiments, one or more components of the imaging system 200 configured to collect X-ray radiation may be considered an X-ray imaging subsystem (e.g., X-ray source 104, detector array 108, etc.) of the entire imaging system 200, which may be a computing system further configured to interact with a user and perform various computing processes (e.g., imaging or non-imaging). Thus, other components of the imaging system 200 (e.g., computing device 216, etc.) can be communicatively coupled to the X-ray imaging subsystem.

[0045] In some embodiments, the computing device 216 may store data in a storage device or mass storage device (e.g., storage device 218) that is included in the computing device 216 (in such examples, the computing device 216 may be referred to as a controller) or is a separate device that can be communicatively coupled to the computing device 216 (in such examples, the computing device 216 may be referred to as a processor). The storage device 218 may include removable media and / or built-in devices. Specifically, the storage device 218 may include one or more physical, non-transitory devices that are configured to store data and / or instructions that can be executed by the computing device 216 to implement the methods described herein. Therefore, when implementing such methods, the state of the storage device 218 may be converted (e.g., to save different or changed data). The storage device 218 may include, for example, a magnetoresistive random access memory (MRAM), a hard disk drive, a floppy disk drive, a tape drive, a compact disk-read / write (CD-R / W) drive, a digital versatile disk (DVD) drive, a high-definition DVD (HD-DVD) drive, a Blu-ray drive, a flash drive, and / or a solid-state storage drive. It should be appreciated that the storage device 218 may be a non-transitory storage medium.

[0046] In addition, the computing device 216 can provide commands and parameters to one or more of the DAS 214, the X-ray controller 210, and the gantry motor controller 212 to control system operations, such as data acquisition and / or processing. In certain embodiments, the computing device 216 controls system operations based on operator input, such as via a user interface 234. The computing device 216 receives operator input, such as commands and / or scan parameters, via an operator console 220 operably coupled to the computing device 216. The operator console 220 can include a physical keyboard, a mouse, a touch pad, and / or a touch screen to allow an operator to specify commands and / or scan parameters.

[0047] In some embodiments, the computing device 216 may include or be coupled to one or more multi-core CPUs or multiple general-purpose GPUs (GPGPUs) 236, where the multiple GPGPUs 236 may be configured to execute instructions stored in a non-volatile memory (e.g., storage device 218) of the computing device 216 via highly parallelized data flows and computation flows.

[0048] Although Figure 2Only one operator console 220 is shown, but more than one operator console 220 may be coupled to the imaging system 200, e.g., for inputting or outputting system parameters, requesting exams, plotting data, and / or viewing images. In addition, in certain embodiments, the imaging system 200 may be coupled to multiple displays, printers, workstations, and / or the like located locally or remotely, e.g., within an institution or hospital or at an entirely different location, via one or more configurable wired and / or wireless networks (such as the Internet and / or a virtual private network, a wireless telephone network, a wireless local area network, a wired local area network, a wireless wide area network, a wired wide area network, etc.).

[0049] In one embodiment, for example, the imaging system 200 may include or may be coupled to a picture archiving and communication system (PACS) 224. In an exemplary embodiment, the PACS 224 may be further coupled to remote systems, such as a radiology information system (e.g., RIS), an electronic health or medical record and / or a hospital information system (e.g., EHR / HIS), and / or to an internal or external network (not shown) to allow operators located in different locations to provide commands and parameters and / or have access to image data.

[0050] The computing device 216 can use operator-supplied and / or system-defined commands and parameters to operate a table motor controller 226, which in turn can control the table 114, which can be a motorized table. Specifically, the table motor controller 226 can move the table 114 to properly position the subject 204 in the gantry 102 to acquire projection data corresponding to a target volume of the subject 204.

[0051] As previously described, the DAS 214 samples and digitizes the projection data acquired by the detector elements 202. Subsequently, the image reconstructor 230 uses the sampled and digitized X-ray data to perform high-speed reconstruction. Figure 2 The image reconstructor 230 is shown as a separate entity, but in certain embodiments, the image reconstructor 230 may form part of the computing device 216. Alternatively, the image reconstructor 230 may not be present in the imaging system 200, and instead the computing device 216 may perform one or more functions of the image reconstructor 230. Furthermore, the image reconstructor 230 may be located locally or remotely, and may be operably connected to the imaging system 200 using a wired or wireless network. For example, one embodiment may use computing resources in a "cloud" network cluster for the image reconstructor 230.

[0052] In one embodiment, the image reconstructor 230 may be configured as follows: Figure 2The computing device 216 shown may store the reconstructed image in a storage device 218, or via a direct connection (not shown). Alternatively, the image reconstructor 230 may transmit the reconstructed image to the computing device 216 to generate usable patient information for diagnosis and evaluation. In certain embodiments, the computing device 216 may transmit the reconstructed image and / or the patient information to a display or display device 232 that is communicatively coupled to the computing device 216 and / or the image reconstructor 230. In some embodiments, the reconstructed image may be transmitted from the computing device 216 or the image reconstructor 230 to the storage device 218 for short-term or long-term storage.

[0053] Various methods or processes further described herein (such as those described below in reference Figure 6 The methods described herein may be stored as executable instructions in a non-transitory memory on a computing device (or controller) in the imaging system 200 or communicate with the computing device (or processor). In one embodiment, the image reconstructor 230 may include such executable instructions in a non-transitory memory and may apply the methods described herein to reconstruct an image from the scan data. In another embodiment, the computing device 216 may include instructions in a non-transitory memory and may apply the methods described herein, at least in part, to the reconstructed image after receiving the reconstructed image from the image reconstructor 230. In yet another embodiment, the methods and processes described herein may be distributed across the image reconstructor 230 and the computing device 216.

[0054] In operation, the computing device 216 may acquire imaging data and other medical data, which may be converted for display to a user (e.g., a medical professional) on a display device 232 via, for example, a user interface 234. As an example, the medical data may be converted and displayed at the display device 232 as a user-oriented graphical and / or textual format, which may be standardized in all embodiments of the imaging system 200, or may be specific to a given institution, department, professional, or individual user. As another example, the imaging data (e.g., a three-dimensional (3D) volume data set, a two-dimensional (2D) imaging slice, etc.) may be used to generate one or more images at the computing device 216, which may then be displayed to an operator or user at the display device 232. In this way, the display device 232 may allow an operator to evaluate the imaged anatomical structure. The display device 232 may also allow an operator to select a volume of interest (VOI) and / or request patient information, for example, via a graphical user interface (GUI), for subsequent scanning or processing.

[0055] The image reconstruction algorithm for the CT imaging system can be structured around the discretization of the tomographic operator or projector. The CT imaging data can be collected in the form of multiple line integral measurements corresponding to the attenuation coefficient distribution of the object of interest. Specifically, an X-ray imaging system such as a CT imaging system can collect measurements collected at a set of lines belonging to a geometric object (such as a cone). For example, an X-ray beam can diverge from the focus of an X-ray source (e.g., an X-ray tube) and can be shaped into a cone through a collimation window when leaving the X-ray source. After passing through the object of interest, the divergent X-ray beam cone can then be completely captured by the X-ray detector. Therefore, the overall shape and configuration of the X-ray detector can define a surface that intersects with the divergent X-ray beam cone. The image formed at the intersection surface can be referred to as a projection (e.g., multiple line integral measurements). The intersection surface can present a variety of geometric structures (e.g., curved, flat, etc.), but in principle can be arbitrary. In some examples, during the acquisition of CT imaging data, multiple projections can be collected at multiple positions of the X-ray imaging system.

[0056] The tomography operator may be selected to encode a geometric representation of a plurality of integration lines (also referred to as rays in some examples) along which a plurality of line-integrated measurements are acquired. However, it should be appreciated that additional or alternative factors may be considered in obtaining the plurality of line-integrated measurements, such as depth-dependent resolution distortion, hardening of the X-ray spectrum along the rays (e.g., via selective attenuation of low-energy photons), and / or considerations specific to emission tomography modes, where the integrated measurements may represent accumulated radioactivity along a given ray.

[0057] In the implementation, a finite number of projections p of I can be collected i , where the finite number I may be the product of a finite number of locations on the X-ray detector and a finite number of measurement points (also referred to herein as intervals). The distribution of interest (e.g., the distribution of the attenuation coefficient of the object of interest) may also be sampled (e.g., calculated) at a finite number of points to obtain a finite number J of distribution samples f j However, with the projection p i In contrast, a finite number of J distribution samples f j and the corresponding locations at which the distribution of interest is calculated can be arbitrary (since the true distribution of interest is continuous). As used herein, "continuous" when referring to a function, parameter, or other measure may mean defined everywhere (e.g., at an infinite set of locations) and is in contrast to "discrete," which when referring to a function, parameter, or other measure herein may mean defined at a finite set of locations.

[0058] Without loss of generality, we consider the case of a single position in an X-ray imaging system. Discretization of the projectors can be done by decomposing the continuous quantity into a sum of basis objects scaled by the corresponding sample values ​​(e.g., basis set expansion) to represent a finite number of projections p i Coupled to a finite number J of distribution samples f j .

[0059] The discretization can be viewed from two perspectives: the X-ray detector and the distribution of interest. With respect to the X-ray detector, and as described above, detection can occur at a finite number of intervals, each corresponding to a subset of the total volume of interest. Since each of these intervals can have a relatively small size, the projection p measured at the ith interval is i It can be derived from the position u set as the center of the i-th interval i The numerical integration algorithm can decompose the i-th line into S segments, where each segment can have a length of l s and distribution value f s , so that:

[0060]

[0061] where f(l) is the (continuous) distribution of interest along the i-th line.

[0062] A finite number J of distribution samples or grid samples f of the distribution of interest j (e.g., a discrete set) may be defined at J regularly spaced locations x j 3D Cartesian grid (however, in other examples, position x j can be variably spaced apart). Thus, in the case of a diverging X-ray beam cone, there may be almost no j Specifically, on a Cartesian grid, position x j can be aligned along at least three directions (e.g., degrees of freedom in each dimension and possible additional lines, such as 45° angles in the plane, etc.); however, a set of scattered rays may cover all directions between the X-ray source and the detector, so that it is not possible to achieve a uniform alignment with position x j To align the j The grid samples at f j Coupling to the distribution value f s , which can be obtained by transforming each grid sample x j Considered as uniform intensity f j (For example, corresponding to the distribution sample f j) to expand the modeling of the distribution of interest. The voxel can be shaped as a geometric object, such as a cube (one dimension of which can be the same or different from the two remaining equal dimensions of the plane defining the square pixel). Numerical integration can be performed by combining multiple fragments l ij It is limited to the intersection point between the i-th ray and the j-th voxel, resulting in:

[0063]

[0064] In algebraic formulas, all projections p i The vector p can be expressed as the operator R and all distribution samples f j The product of the vectors f such that p = Rf. Therefore, R can be represented as a matrix with I rows and J columns, which stores the coefficients r 1≤i≤I,1≤j≤J , and can be called a projector or a forward projector. The matrix-vector product of the forward projection operation can be written as:

[0065]

[0066] Since the coefficient r ij is a real number, and the adjoint operator of R is the transposed matrix and is called the backprojector. The matrix-vector product of the backprojection operation can be written as:

[0067]

[0068] One class of discretization models that can be used includes the so-called ray-driven models. Ray-driven models can use a geometric representation where multiple lines are drawn from multiple positions u. i I rays are formed at and intersecting the focus of the X-ray source.

[0069] The ray-driven model can use one of the numerical integration schemes known as ray-driven multilinear interpolation. For example, a cubic sampling grid can be used with a sampling rate of 1 along each grid axis. The integration length can also be set to 1 so that fixed-length segments can be obtained. Therefore, formula (4) can be simplified to:

[0070]

[0071] The distribution value f s can be regularly spaced on the i-th ray. Therefore, the distribution value f s A linear interpolation may be performed from the two nearest samples in the direction of each grid axis. It should be noted that no shape is assumed for voxels in the ray-driven model. In addition, since multilinear interpolation may be separable, d linear interpolations may be formulated for a sampling grid with d dimensions.

[0072] Joseph's method is a simplification of ray-driven multilinear interpolation. In Joseph's method, the numerical integration still relies on fixed-length segments; however, larger-sized segments can be used. To define larger-sized segments, the distribution of interest can be projected as a set of parallel hyperplanes orthogonal to one axis of the sampling grid, where the set of parallel hyperplanes can be regularly spaced along the grid axis with a spacing δ (e.g., δ=1). According to the angle θ between the i-th line and the grid axis, i , the length between two consecutive hyperplanes along the i-th line (ray) The grid axes can be chosen to produce an angle θ i , so that

[0073] Therefore, the distribution value f s The position of can be compared with the distribution samples f along the grid axis orthogonal to the hyperplane j The positions of . For each other grid axis, a separable linear interpolation may be employed. Thus, for a d-dimensional grid, Joseph's method may utilize d-1 linear interpolations (e.g., one less than the ray-driven multilinear interpolation described above). Joseph's model may be well suited for CT imaging, e.g., when the redundancy introduced by ray-driven multilinear interpolation may be desirable for graphics and display.

[0074] One desirable property of a ray-driven model is to utilize 100% (e.g., for each i-th interval) of the intervals computed by the sampling grid (e.g., during reprojection) or partitioned on the sampling grid (e.g., during backprojection), which can be mathematically represented as follows:

[0075]

[0076] However, it may not be guaranteed that voxels are similarly 100% considered. Mathematically, this can be expressed as:

[0077]

[0078] Depending on the value of number I, voxels can be undersampled or oversampled (resulting in holes or redundancy, respectively).

[0079] In order to consider voxels, a voxel-driven model can be used. Compared with the ray-driven method for discretization (as discussed in detail above), the voxel-driven model utilizes each distribution sample f j 100% contribution to the X-ray detector. Therefore, the voxel-driven model can be determined by how each distribution sample f j It is characterized by its contribution to the measured value (interval).

[0080] In the voxel-driven model, linear interpolation is considered across each j-th distribution sample f j The line integral of, for example, the line connecting the focus of the X-ray source to the jth grid sample x j The intersection of the line with the X-ray detector may be defined by a continuous function u(x), which may not necessarily correspond to the position u at the center of the i-th interval. i , on the contrary, can be more generally related to the position u(x j ) corresponds to (therefore, this line may be referred to as line u(x j ), because the line can be relative to the X-ray source by the position u(x j ) is limited). Since the line u(x j ) can be the number of grid samples x through the jth grid for a given position of the X-ray source and the X-ray detector. j and the only line of focus of the X-ray source, so the distribution sample f j It can be a line u(x j ) contributes 100% (e.g., distribution sample f j can be the time between the X-ray source and the position u(x j ) on the line defined by the connection of j ) contributes 100%). Then the j ) and the two nearest neighboring positions u i The derived linear interpolation coefficients partition the distribution sample f j To determine the adjacent position u i Therefore, due to the given line u(x j ) can be modeled in a certain way as a function of multiple locations u i Interact so that given position u i It is also modeled as multiple lines u(x j ) interactions, so asymmetric linear dependencies can exist in voxel-driven models (as encoded by the operator R).

[0081] The geometric model may be selected to describe a plurality of voxels. For example, and as described above, each voxel in the plurality of voxels may be shaped to have a uniform intensity f j In a voxel-driven model, the projection or footprint of a voxel on the X-ray detector can be determined by the shape of each voxel, which for a cube is a trapezoid with a width that varies depending on the projection angle. A geometric model can also be selected for the X-ray detector. For example, the X-ray detector can be divided into uniform intervals. Therefore, the (discretization) coefficient r ij can be considered as being at position x jquantification of the interaction between the footprint of a given voxel at and the support set from the ith bin of the X-ray detector. This interaction forms the so-called range-driven model.

[0082] See now Figure 3 , a schematic diagram 300 is shown illustrating interactions in an exemplary distance-driven model. An exemplary distance-driven model as a voxel-driven model based on an integral line geometric representation may consider two sets of lines: (i) a plurality of source lines 302 extending from a focus 324 of an X-ray source to respective edges 306 of the plurality of voxels 304 (e.g., at dashed lines 308); and (ii) a plurality of target lines 312 extending from the focus 324 of the X-ray source to respective edges 316 of a plurality of bins 314 of an X-ray detector (for clarity, the plurality of target lines 312 are not shown extending completely to the focus 324 of the X-ray source). The projection axis 322 may be used as a reference for the relative positions of the plurality of voxels 304 and the plurality of bins 314 by defining a set of line segments (e.g., a set of line segments of lines 302, 312, also referred to herein as a support set), the length of which may weight the values ​​of the respective projections (e.g., footprints) of the plurality of voxels 304 and the plurality of bins 314. Additionally, since the plurality of voxels 304 and the plurality of bins 314 may be projected as relatively simple geometric objects, the projection along the projection axis 322 may be completely described from a geometric perspective. However, since the two sets of lines 302, 312 are based on different sampling schemes (the plurality of voxels 304 and the plurality of bins 314, respectively), the plurality of source lines 302 may have a different information density than the plurality of target lines 312. Therefore, an interpolation scheme may be necessary to merge the two projections, thereby removing the asymmetric linear dependence between the two by taking into account the different information densities.

[0083] Compared to the ray-driven model, in the voxel-driven model, 100% of the voxels can be detected (e.g., for every j-th voxel) and shared over these intervals, which can be mathematically expressed as:

[0084]

[0085] However, also in contrast to the ray-driven model, in the voxel-driven model it may not be guaranteed that the intervals are similarly 100% considered. Mathematically, this can be expressed as:

[0086]

[0087] Depending on the value of the number J, the intervals may be undersampled or oversampled (resulting in holes or redundancy, respectively).

[0088] Due to poor interval sampling, linear interpolation in the voxel-driven model may be an inaccurate projector model and is not suitable for reprojection and iterative image reconstruction (compare Eq. (2)). However, 100% voxel sampling is suitable for voxel-driven models used for backprojection and analytical image reconstruction (the latter uses a single backprojection and no reprojection). In addition, with the distribution sample f j Sample x in the grid j In contrast to the partitioning of contributions to neighboring intervals, linear interpolation in the voxel-driven model can be viewed as the sum of the contributions along the line u(x j ) to the distribution sample f j The sum of the detector measurements (assuming that all detector measurements form a continuous function u(x), from which the linear interpolation along the line u(x) can be determined j ) of the detector measurement).

[0089] The range-driven model can achieve both conditions of 100% voxel sampling (e.g., Equation (8)) and 100% interval sampling (because the interaction is chosen to verify Equation (6)). Therefore, the range-driven model is equally suitable for reprojection and backprojection, and can therefore be successfully used in iterative image reconstruction algorithms. However, accurate imaging and massively parallel implementation can be excluded by omitting the continuous function u(x), so that the range-driven model can instead rely on an ordered determination of the intersection positions via sequential scanning of the (sampling) grid axes.

[0090] According to the embodiments described herein, a one-dimensional resampling transform can provide a more general framework for separable interpolation. Since a one-dimensional resampling transform may not present a multi-dimensional geometry, a one-dimensional subset of samples may be considered instead, each sample having a sampling position and having a one-dimensional basis function instead of a geometry.

[0091] In the context of image transformations, resampling transformations can be used as translations, rotations, and magnifications. Such resampling transformations are special cases of homography transformations. Therefore, homography transformations may be ideally suited for resampling projections obtained in cone-beam tomography geometries based on flat-panel detectors, since such projections may be associated with magnification and rotation. In addition, the invertibility of all homography transformations may be desirable for resampling.

[0092] The homography resampling transformation can be viewed as a change in a variable. For example, the homography resampling transformation h can replace the variable x with the variable u, such as u=h(x). Conversely, the variable u can be replaced by x=h -1 (u) is replaced by the variable x.

[0093] In this paper, the notation used in the context of resampling transforms can be simplified by taking u(x) and x(u) such that x(u(x)) = x and u(x(u)) = u, respectively. Therefore, the projection operation is linear and can be decomposed into summing projections p(u) computed as resamplings of the distribution f to one axis (e.g., x of f(x)) via a homographic resampling transform u(x) (such that For example, p(u(x))=f(x)).

[0094] The (real-valued) function u(x) of x∈[–∞,+∞] is a one-dimensional homography, defined by:

[0095]

[0096] where h 11 、h 12 、h 21 and h 22 are the elements of the transformation matrix H, for example,

[0097]

[0098] Use the intermediate coordinate s(x)=h 21 x+h 22 , we can obtain a real-valued matrix relationship where the homogeneous coordinates su(x) = s(x)u(x) and s(x) are as follows:

[0099]

[0100] Consider the matrix H λ =λH, where λ is a real, non-zero scalar value, the same resampling is obtained:

[0101]

[0102] However, H λ For example, different intermediate coordinates s can be used. λ (x) = λs(x). Since λ is chosen to be non-zero, the transformation matrix H can be assumed to be invertible and is therefore suitable for resampling.

[0103] It is worth noting that, without loss of generality, the transformation matrix H can also be chosen to be a unimodular matrix, for example, such that h 11 h 22 –h 12 h 21 =1 or |H|=1.

[0104] Since the transformation matrix H is invertible, the inverse resampling (e.g., function x(u)) can be written as -1Associated, the matrix can be given as:

[0105]

[0106] Since the transformation matrix H is a 2×2 matrix, similar to the intermediate coordinate s(x) defined for u(x), the intermediate coordinate t(u) = –h 21 u+h 11 , such that (and similar to the discussion above):

[0107]

[0108] (Thus, it follows that tx(u)=t(u)x(u)).

[0109] The derivatives u'(x) and x'(u) provide a local approximation of the homography and can be expressed as:

[0110]

[0111] as well as

[0112]

[0113] Therefore, the magnification factor in x can be s 2 (x), and the reciprocal magnification factor in u can be [t 2 (u)] -1 In addition, in h 21 = 0, |H| = h 11 h 22 = 1, and the magnification factors in x and u are constant, for example,

[0114]

[0115] In cone-beam tomography geometry based on flat-panel detectors, the relationship between a point in space (x,y,z) and its projection on the detection plane (u(x,y,z),v(x,y,z)) can therefore be described by the homogeneous coordinates s(x,y,z), su(x,y,z)=s(x,y,z)u(x,y,z) and sv(x,y,z)=s(x,y,z)v(x,y,z) via a 3×4 projection matrix P:

[0116]

[0117] This form can be adapted to the interpolation scheme described above. For example, a ray-driven model can loop through the lines (rays) to obtain the formula for the i-th line in space via the following formula:

[0118]

[0119] Similarly, the voxel-driven model can loop through the voxels to convert the voxel (x j ,y j ,z j ) of the projection (u(x j ,y j ,z j ),v(x j ,y j ,z j )) Direct calculation, for example:

[0120]

[0121] In contrast, by employing a one-dimensional resampling transform, the projection operation can be decomposed into a linear combination of transforms, each of which relates a free variable in space (e.g., x) to a free variable on the surface of the X-ray detector (e.g., u). For example, the resampling (e.g., via the u(x) form, as described above) can be expressed as:

[0122]

[0123] Thus, for any pair (i, j), there may be six possible resampling transforms: u(x), v(x), u(y), v(y), u(z), and v(z). The projective decomposition may involve looping over (i, j) and selecting one of the six resampling transforms for each loop. The resulting subset of resampling transforms may be selected so as to have a reversible homography. For example, a looping scheme in a voxel-driven model may include first looping through j and then first fixing z=z j , then y = y j , then for all i, split the interpolation between u(x) and v(x) (which are independent and separable from each other). The axes x, y and z can be swapped in the aforementioned loop scheme as needed for looping over voxels along the corresponding axes x, y and z. If i and j are swapped, a loop for a ray-driven model can be obtained.

[0124] Therefore, without loss of generality, the following discussion uses u(x) and x(u) as examples. Thus, among a plurality of voxels (indexed by j and having a size equal to a finite number J), a subset lying on axis x may be indexed by l and having a size equal to a finite number L. Likewise, among a plurality of bins (indexed by i and having a size equal to a finite number I), a subset lying on axis u may be indexed by k and having a size equal to a finite number K. Thus, index k may subsample index i, and index l may subsample index j, such that according to u(x), r lk =r ij .

[0125] In image reconstruction, forward projection (reprojection) can be described as mapping a volume onto a surface, and back projection can be described as inversely mapping a surface onto a volume. Such mappings can be decomposed into a sum of simpler, separable mappings, so that a given image reconstruction model can be described as the projection of a single line of the volume onto a line of projection (surface), and vice versa.

[0126] For example, and as referenced below Figure 4A and Figure 4B As described, the homography-based model of the present disclosure can be completely described by a single (one-dimensional) homography resampling transformation on a single axis. Therefore, a set of one-dimensional homography resampling transformations can form a symmetrical projector / backprojector pair (e.g., as a matrix of voxel-interval interactions at the surface), which can be used in image reconstruction algorithms.

[0127] See now Figure 4A , shows a schematic diagram 400 illustrating an exemplary flat panel detector based cone beam tomography geometry. Two sets of lines can geometrically represent the integral lines being sampled: (i) a plurality of dashed lines 402 extending from a focal point 424 of an X-ray source to respective centers 410 of a subset of voxels 404 (e.g., along dashed lines 408 parallel to axis x); and (ii) a plurality of lines 412 extending from a focal point 424 of an X-ray source to respective centers 420 of a subset of bins 414 (e.g., along lines 418 on an X-ray detector, surface 418 parallel to axis u). However, unlike reference Figure 3 Compared to the geometry-based perspective described above, for each sampled integral line in the homography-based perspective (rather than the two-dimensional projection planes of voxels and bins), only a single one-dimensional point (e.g., the one-dimensional point of the centers 410, 420) needs to be considered. Specifically, each one-dimensional point corresponding to the integral line sampled at the center 410 of the subset of voxels 404 and the center 420 of the subset of bins 414, respectively, can be mapped to a single continuous function via a one-dimensional homography resampling transformation (as described below with reference to Figure 4B As described above). Therefore, in a homography-based perspective, the geometric representation of the integrated lines (eg, lines 402, 412) may not be considered at all.

[0128] Reference now Figure 4B, a schematic diagram 450 is shown that illustrates an exemplary one-dimensional homography resampling transform 476. A plurality of points 460 (e.g., centers of a subset of voxels) may represent corresponding locations of a plurality of samples 462 of a distribution 474 along an axis 458. Similarly, a plurality of points 470 (e.g., centers of a subset of bins of an x-ray detector) may represent corresponding locations of a plurality of samples 452 of a distribution 478 along an axis 468. As shown, the distribution 474 may be mapped (e.g., transformed) into the distribution 478 via the one-dimensional homography resampling transform 476. In this way, the one-dimensional homography resampling transform may completely replace the geometric representation of the integral line in the image reconstruction algorithm.

[0129] In order to get from a fixed known position x to an arbitrary position x m M samples g at m (called “source sample”) resamples a single continuous function g, which can be used for each x m Introduce the basis function β(x, m) defined for x∈[-∞,+∞] to m to describe the function g. Therefore, a continuous approximation of g can be given by a finite sum of M basis functions:

[0130] g(x)=∑ m g m β(x,m) (23)

[0131] In some examples, each basis function β(x, m) can be effectively expressed as β(x, m)=β(x m ).

[0132] In the linear interpolation scheme, the basis function β(x, m) may be the basis function β1(x), which is defined as:

[0133]

[0134] In the nearest neighbor interpolation scheme, the basis function β(x, m) may be the basis function β0(x), which is defined as:

[0135]

[0136] The number indices "0" and "1" index the basis functions according to a family of functions known as B-splines of order n. Specifically, the interpolation basis functions of order n>0 can be recursively defined based on B-splines of order n as:

[0137] β n =β n-1 *β0 (26)

[0138] Where "*" indicates convolution. For example:

[0139] β1=β0*β0 (27)

[0140] In addition, each basis function in the B-spline of order n can be verified:

[0141] β n (x) = β n (-x) (28)

[0142] In the continuous domain, homographies u(x) and x(u) may describe the same relationship between a given line (subset) of voxels of distribution f and a given line (subset) of intervals of projection p, such that p(u) = f(x). Although from a geometric perspective, the lines of voxels and intervals are located at different locations in space, the functions u(x) and x(u) map f(x) and p(u) to the same axis by replacing one variable with the other (e.g., replacing x with u or vice versa). Specifically, function u(x) maps both functions f(x) and p(u) along axis u, and function x(u) maps both functions f(x) and p(u) along axis x. However, since sampling along axis x may be different from sampling along axis u, equivalent discretization may not be guaranteed. Along axis x, L points (e.g., the centers of a subset of voxels along axis x, as described above with reference to FIG. 4 ) may be spaced δ apart. x Along axis u, K points (eg, centers of a subset of intervals along axis u, as described above with reference to FIG. 4 ) may be spaced equidistantly apart. u Spaced equally.

[0143] Therefore, the basis function β can be introduced n,δ (x) to specify the spacing δ, which can be expressed differently as:

[0144]

[0145] Using the basis function β n,δ The basis set expansion of (x) can obtain one of the two continuous models:

[0146]

[0147] as well as

[0148]

[0149] For projection p and distribution f (where β n,δ The sign of (x) is related to the spacing δ x Simplified to And for the spacing δ u Simplified to

[0150] The spacing δ can be understood as the sampling rate δ -1 Therefore, a change in the spacing δ can result in a sampling rate δ-1 , and can be quantified as a magnification, e.g., δ x / δ u In some examples, the sampling rate may be higher (e.g., 1 / δ x or 1 / δ u ) to select a continuous model (e.g., formula (29) or (30)).

[0151] See now Figure 5 , shows a schematic diagram 500, which shows an exemplary one-dimensional homography resampling transformation (such as the above reference Figure 4B The exemplary one-dimensional homography resampling transform 476) resamples the continuous function of the basis set approximation. From the perspective of resampling, there exists a function h(x) such that Thus, the function f(x) (e.g., the distribution f) and the function p(u) (e.g., the projection p) can be the same function but at different sampling locations, e.g., the function f(x) of x l and function p(u) of u k , and along different axes. However, the functions f(x) and p(u) can be placed along a single axis because the position x l can be transformed into position h(x) by function h(x) l ), the function h(x) can be placed at position u k The function p(u) can then be modeled as the set {u 1<k≤K ,h(x 1<l≤L )} defined. The interpolation weights for modeling can then be obtained by calculating the values ​​of the basis functions at the selected positions. Since the basis functions can be globally continuous, their calculations can be performed independently at any point, at any time, and therefore in parallel. In this way, algorithms that utilize one-dimensional homography resampling transforms can be easily adapted to large-scale parallel computing architectures.

[0152] As shown in schematic diagram 500, exemplary position 510 may represent a position h(x l ), and exemplary position 520 may represent position u k Thus, positions 510, 520 are plotted along the same axis 532, and 512, 522 are mapped to the same function 530, respectively, where function 530 can be is expressed as the sum of basis functions. Figure 5 The single axis representation of the one-dimensional homography resampling transformation described with reference Figure 4B 4, which shows that function 474 sampled along axis 458 is transformed 476 into function 478 sampled along axis 468. Note that Figure 4BIn FIG. 4 , functions 474 and 478 are depicted as having similar shapes, showing their equivalent forms. In this way, Figure 3 The geometric representation shown can be replaced by a single function along a single axis (in Figure 5 Axis 532 is shown as detector coordinates u, but using axes of sampling coordinates x provides an equivalent framework).

[0153] It should be noted that the above basis set expansion can be intermediate in nature, since it can be obtained by the (discrete) coefficient set r kl Construct the resulting model. For example, in a voxel-driven model, each voxel position x l A basis function β can be set n,δ (x), can be obtained by the basis function at position x(u k ) to obtain the following coefficients:

[0154]

[0155] For example, in the ray-driven model, each interval position u k- A basis function β can be set n,δ (x), can be obtained by using the basis function at position u(x l ) to obtain the following coefficients:

[0156]

[0157] Of course, adapting the homography-driven model to a ray-driven model and a voxel-driven model may not remedy any of its limitations. However, from an alternative perspective, the homography representation can advantageously provide valuable insights into such limitations. For example, in the homography representation, the ray-driven model and the voxel-driven model can select one sampling axis (e.g., axis x or axis u) that is sufficient for the corresponding unilateral view of image reconstruction. However, and when the homography representation is clarified, a similar model can be induced on another axis. For example, the function p(u) can be expressed as:

[0158]

[0159] as well as

[0160]

[0161] (Compare formulas (30) and (31) respectively).

[0162] Since the homography can act locally as a magnification on the axis x (as described above), the basis function It can be expressed as:

[0163]

[0164] The fixed spacing δ x However, on axis u, the induced basis function Can have a continuously variable spacing δ u(x) ,For example,

[0165]

[0166] Due to the continuously variable spacing δ u(x) It may not be related to the measured value p l (which may have a fixed spacing δ imposed by the detector design u ) coincide with each other, so the basis function It is possible not to use evenly spaced measurements p l Provides a fully resampled extension of the function p(u); however, the sample f l Through the basis function of the function p(u) Provides sufficient resampling expansion. Conversely, the basis function It may not be possible to use sample f k Provides sufficient resampling expansion for function (x); however, the sample p can be used k Through the basis functions of the function f(x) Provides adequate resampling extensions.

[0167] Thus, when resampling a given function, the homography-driven model can be a sufficient resampling extension when used for destination-driven resampling. Specifically, the resampling extension can be selected for an axis describing uniformly spaced known samples (source samples, as described above), and the resampling position (called the "destination") can be given by the resampling transform (as described above with reference to Figure 5 Detailed description). Source-driven resampling can be defined from the opposite perspective. For example, a ray-driven model can be destination-driven in the case of reprojection (thereby verifying the condition of formula (6)), and can be source-driven in the case of backprojection. For another example, a voxel-driven model can be destination-driven in the case of backprojection (thereby verifying the condition of formula (8)), and can be source-driven in the case of reprojection.

[0168] In some examples, the above-described one-dimensional homography resampling transform may be known as a homography in a flat-panel detector-based cone beam tomography geometry (e.g., where the X-ray detector is flat and all rays converge toward the focus of the X-ray source, as described above with reference to FIG. 4 ), and may be implemented thereby. Such geometry is inherent to interventional X-ray C-arm systems and all systems for performing cone beam CT (CBCT) utilizing a flat-panel X-ray detector. Specific applications of homography generalization may include transforms describing projections in parallel geometries (nuclear medicine) or digital breast tomosynthesis (DBT).

[0169] In other examples, such as in a CT scanner with a curved X-ray detector, the function h(x) may not be a homography. One possible adaptation of the homography-driven model to such a system includes performing a pre-sampling step of the CT imaging data, wherein a virtual X-ray detector may be constructed for which the function h(x) is a homography. Another possible adaptation of the homography-driven model to such a system includes defining a self-organizing function that models a geometric structure for which the function h(x) is (at least partially) a homography. Advantageously, defining the self-organizing function in this manner may be generalized to include imaging systems other than flat-panel CT imaging systems.

[0170] Thus, in the embodiments described herein, a more generalized and universal extension for homography sampling is provided that can alleviate the limitations of geometry-based models by providing a discretization scheme based on a single set of basis functions that provide sufficient extensions for both reprojection and backprojection (e.g., sufficient resampling transforms can be implemented in both source-driven and destination-driven uses of the extensions). In this way, the homography-based models described herein can enable the reshaping of geometry-based models to implement similar methods while avoiding geometric representations entirely. Such reshaping can increase the ease of implementation on massively parallelized computing frameworks (e.g., GPU-based architectures).

[0171] In particular, while some geometry-based models such as range-driven models may be desirable for favorable image quality, highly ordered memory access patterns, and low arithmetic cost, easy massively parallel implementation may be difficult to achieve. For example, a range-driven algorithm may include internal branches to adjust the calculation based on the relative position between a voxel and a detector cell (bin) edge (see Figure 3 ). However, when implemented on a massively parallel computing device such as a GPU, the irregularity of the resulting branching behavior can lead to inefficiencies.

[0172] Homography-based models can provide a more computationally efficient alternative by avoiding internal branching (either by reshaping the distance-driven model or by replacing the distance-driven model entirely). In addition, homography-based models can outperform GPU implementations of geometry-based models in terms of accuracy (e.g., via high-order interpolation) and speed (e.g., via easy parallelization). Such precise and GPU-friendly computational schemes may be desirable for deep learning-based image reconstruction algorithms (e.g., via convolutional neural networks, which are typically trained on GPU-based architectures).

[0173] It should be noted that geometric considerations are not completely absent in homography-driven interpolation. For example, the basis functions β0(x) of the nearest neighbor interpolation scheme may be uniform over a finite support set. Since detection occurs at a finite detector surface, the basis functions β0(x) may be well suited to describe the tomographic geometry from a detector-level perspective. From a voxel-level perspective, the basis functions β0(x) may similarly present a flat profile along a selected voxel line. However, the geometric interpretation in this case is a 0th-order discretization of the distribution - smaller than the discretization that can be assumed using a linear interpolation scheme (e.g., with an interpolation order of 1).

[0174] In some embodiments, homography-based models may also be coupled to footprint-based models (e.g., distance-driven models). Such coupling may be achieved by bypassing sampling restrictions defined based on one sampling axis (e.g., X-ray detectors in ray-driven models, sampling grids in voxel-driven models), while the other sampling axis is continuous and computed at a set of locations.

[0175] As mentioned above, if a set of L functions (voxels) is defined along the sampling axis of the sampling grid, then the given model remains unchanged regardless of the number K of bins of the X-ray detector. In the footprint-based model, the interaction between voxels and bins can be explicitly quantified. In the homography perspective, the sampling positions can be projected instead of the geometry. For example, in the function u(x), one can s 2 The sampling rate of (x) varies continuously along axis u relative to x (e.g., according to the magnification factor). Thus, when at position x l When modeling the grid samples, we can assume a fixed spacing δx and ignore the sampling rate that is converted by u(x) to the local sampling rate of the spacing Δu, such as:

[0176]

[0177] In order to include the local sampling rate of the spacing Δu, a basis function ξn(x) can be defined, where the basis function ξn(x) can be the basis function of the fixed sampling δx Basis functions of the (transformed) local sampling rate with spacing Δu Without loss of generality, the fixed spacing δx can be set to 1, and the local sampling rate of spacing Δl can be defined as:

[0178] Δ l =u′(x l )δ x (39)

[0179] Then, the L basis functions ξ n,l Each of (x) (at position x1) can be defined as:

[0180]

[0181] L basis functions ξ n,l (x) can expand p(u) and f(x) to:

[0182] p(u)=f(x)=∑ l f l ξ n,l (x(u k )-x l ) (41)

[0183] And the K×L coefficient r kl Each of can thus be obtained as:

[0184] r kl =ξ n,l (x(u k )-x l ) (42)

[0185] Alternatively, the K basis functions χ n,k (u)(at position u k Each of these can be defined as:

[0186]

[0187] The spacing Δ k The local sampling rate can be defined as:

[0188] Δ k = x′(u k )δ u (44)

[0189] and the (fixed) spacing δ u is set to 1. K basis functions χ n,k (u) can expand f(x) and p(u) to:

[0190] f(x)=p(u)=∑ k p k χ n,k (uu k ) (45)

[0191] And the K×L coefficient r kl Each of can thus be obtained as:

[0192] n kl = x n,k (u(x l )-u k ) (46)

[0193] Therefore, and as described above, two similar models are obtained (one in x and one in u), each modeling a fixed sample and its homography transformation.

[0194] The special case of n=0 can be developed by considering the properties and definitions of equations (24)-(28). Assume any sampling rate Δ>0 and define:

[0195]

[0196] as well as

[0197]

[0198] Then the convolution β 0,Δ *β0(t) can be given as:

[0199]

[0200] In addition, the convolution β 0,Δ *β0(t) can be defined as:

[0201]

[0202] So that the coefficient r kl It can be calculated as:

[0203]

[0204] Taking into account equation (25), the integral of equation (51) is equivalent to the value corresponding to the position x l The support set of the function β0 centered at position x(u k ) is centered and has a width Δ l =u'(x l ) The intersection of the coverage area (e.g., the projection of the interval) of the support set of (compare formula (39)). Therefore, the position u k is centered and has a width δ u The projection of the X-ray detector interval is modeled as a k ) is centered and has a width Δ l =u'(x l ) of the projection interval.

[0205] The above description of the distance-driven model assumes a unique geometric representation with cubic voxels and uniform voxels with basis functions of order 0 on each side. The interaction of the distance-driven model with the homography-driven model increases the interpolation order to 1 (e.g., that of a linear interpolation scheme). To continue to increase accuracy, coverage (rather than support sets) can be considered.

[0206] Specifically, the index m can be introduced to convert the basis function ξ n,l (x) is generalized to the basis function ξ n,m,l (x), given as:

[0207]

[0208] Thus, each axis can independently have an arbitrarily high level of accuracy.In practice, for each method of distribution sampling, an interpolation order of at least 1 may be desirable, so that any resulting model may have an interpolation order of at least 2.

[0209] In the distance-driven model, uniform intervals and uniform voxels can be further assumed. The distance-driven model can further consider the edges of the intervals and the voxels along each (separable) direction. Figure 3 As mentioned above, the edge position can also be projected onto any common projection axis to form a line segment set (support set) of intervals and voxels. kl can be set equal to the intersection on the projection axis between the support set defined by the projection of the kth interval on the projection axis and the support set defined by the projection of the lth voxel on the projection axis.

[0210] Coupling considerations of a footprint-based model to a homography-based model as described above can provide a model that is equally applicable to reprojection and backprojection, since voxels, detectors, and interactions between them can be represented simultaneously. Such a model can be correspondingly more computationally demanding.

[0211] To facilitate parallelization, the above interpolation scheme considering both the geometric representation and the homography representation can be rewritten, assuming that the projection axis is the axis u, where the position u k The center of the interval for the X-ray detector, and the position u(x l ) is used to project the center of the voxel onto the axis u. In this case, the function u(x) may not be assumed to be a homography and therefore may not be approximated by magnification. Instead, the function u(x) may be used to determine a vector with width δ x The projection support of the voxel can form a set of L continuous line segments S l , which can be given as

[0212]

[0213] Similarly, the projected support of the interval can form a set S of K continuous line segments k , which can be given as:

[0214]

[0215] Therefore, the (distance-driven) coefficient r kl Can be given as:

[0216] r kl =S k ∩S l (55)

[0217] The intersection point can be determined based on the segment set S l ,S k However, apart from separability considerations, determining such bounds may not be amenable to vectorization and parallelization.

[0218] The basis function β0 can be used to develop alternative methods where the fragment S k It can be a function β 0,1 (u–u k ) and the fragment S l can be of width δ given by l Function The support set of:

[0219]

[0220] The position u l Can be defined as:

[0221]

[0222] Therefore, r kl Can be given as:

[0223]

[0224] By adopting the formula (58) Analytical definitions of , sorting operations can be avoided, and algorithms with easily parallelizable function calls can be developed.

[0225] In the case of cone beam tomography geometries based on flat panel detectors, and as described above, the resampling transform is a homography that can be locally approximated by a magnification factor given by u'(x) or x'(u). However, in one example, if the magnification factor is instead defined as the segment S k and S lThe above homography-driven interpolation can be an accurate reconstruction of the range-driven model of the cone-beam tomography geometry based on a flat-panel detector.

[0226] Although the above discussion of extensions representing both a set of uniformly spaced samples and a corresponding set of non-uniformly spaced samples is in the context of B-splines, in other examples, other families of basis functions (e.g., O-MOMS) may be used. In further examples, more complex extensions in more than one dimension may be designed (albeit with less separability).

[0227] See now Figure 6 , a flow chart is shown, which illustrates a method 600 for iterative reconstruction via an exemplary homography-based model. As described below, the sampled imaging data can be resampled using a one-dimensional homography resampling transform and a symmetric projector / backprojector operator pair can be formed. Specifically, the symmetric projector / backprojector operator pair can be decomposed into a linear combination of transformations, each of which maps a subset of the sampled imaging data to a single continuous function, from which unknown sampling points can be determined by interpolation of known sampling points. In contrast to a geometry-based model (e.g., a model based on mapping the sampled imaging data to a set of geometric objects), an exemplary homography-based model reshapes the resampled imaging data into a one-dimensional basis function that can extend and approximate a single continuous function to an arbitrary level of accuracy. In some examples, the coefficients (e.g., weights) associated with each basis function can be determined substantially independently. In this way, a set of coefficients encoding the resampled imaging data can be determined in parallel, thereby overcoming the computational limitations of the geometry-based model.

[0228] See below Figure 1 and Figure 2 The systems and components shown in FIG. 6 illustrate the method 600. For example, in some embodiments, the method 600 may be performed in Figure 2 However, it should be understood that the method 600 may be implemented with other systems and components without departing from the scope of the present disclosure. In some embodiments, the method 600 may be implemented as described above with reference to Figure 1 and Figure 2 In one embodiment, method 600 may be implemented in a non-volatile memory of a computing device, such as a Figure 2 The computing device 216 of the imaging system 200. It should also be understood that the various steps discussed with reference to the method 600 may be added, removed, replaced, or interchanged within the scope of the present disclosure.

[0229] At 605, method 600 may include determining one or more imaging system operating parameters. In one example, the one or more imaging system operating parameters may include an imaging protocol that determines an initial configuration of an imaging source (e.g., an emission device, an injected dye, etc.) and a detector for receiving imaging data. The imaging protocol may also determine movement of the imaging source and the detector relative to the object of interest so that multiple images can be constructed sequentially.

[0230] In additional or alternative examples, one or more imaging system operating parameters may include imaging geometry. For example, the imaging geometry may include the relative positioning of the imaging source and detector relative to the object of interest (e.g., as determined by an imaging protocol), as well as the specific imaging modality utilized. In some examples, the computing device may be coupled to a single imaging modality. In other examples, the computing device may be coupled to multiple imaging modalities (e.g., multiple imaging subsystems located at the same location or at different locations).

[0231] In additional or alternative examples, one or more imaging system operating parameters may include a predefined homography-based model. Specifically, prior to acquiring imaging data via the imaging system, a one-dimensional homography resampling transformation may be selected (eg, based on an imaging protocol and / or imaging geometry).

[0232] At 610, method 610 may include determining whether the imaging geometry is a flat panel detector based cone beam tomography geometry. For example, the computing device may determine the imaging geometry based on one or more imaging system operating parameters. If the imaging geometry is not a flat panel detector based cone beam tomography geometry, method 600 may proceed to 615, where method 600 may include performing a pre-sampling routine or defining a self-organizing model based on a homography. In all imaging systems, the one-dimensional homography resampling transform described above may not be guaranteed to be a homography. For example, the one-dimensional homography resampling transform may not be a homography for a CT imaging system with a curved detector. Therefore, pre-sampling of imaging data on a virtual detector may be performed for which a homography may be selected and implemented, or a self-organizing transform may be defined that at least partially models the desired imaging geometry.

[0233] If the imaging geometry is a cone beam tomography geometry based on a flat panel detector, or once a pre-sampling routine is performed at 615 or a homography-based self-organizing model is defined, the method 600 may proceed to 620 where the method 600 may include acquiring sampled imaging data using a sampling grid and a detector. Specifically, the sampling grid may be divided into a plurality of voxels, and the detector may be divided into a plurality of bins. A plurality of line integral measurements may be determined for each of the plurality of voxels and the plurality of bins for which an image may be reconstructed via an image reconstruction algorithm.

[0234] At 625, method 600 may include selecting a set of basis functions for one-dimensional homography resampling transformation based on the desired level of accuracy. Specifically, the one-dimensional homography resampling transformation can resample the sampled imaging data by mapping the sampled imaging data to a single continuous function to form a symmetrical projector / reverse projector operator pair. The group of basis functions may include a set of one-dimensional basis functions, which can extend and approximate a single continuous function to an arbitrary level of accuracy. In one example, the size of the group of basis functions can be increased to correspondingly improve the level of accuracy. An interpolation scheme may also be selected to determine unknown points along a single continuous function based on known points from the sampled imaging data. In this way, a symmetrical projector / reverse projector operator pair may provide a generalized framework, wherein the group of basis functions and the interpolation scheme may be selected to reshape or match previously implemented algorithms or provide a novel process for sampled imaging data.

[0235] At 630, method 600 may include determining a set of coefficients (e.g., weights) of a selected set of basis functions. Specifically, each basis function in the set of basis functions may be associated with a corresponding coefficient in the set of coefficients. Each coefficient in the set of coefficients may encode an allocation sample (e.g., sampled at a point in a corresponding voxel) in a plurality of allocation samples and an interaction between an interval in a plurality of intervals in a matrix representation of a projector or a back-projector of a given projector / back-projector operator pair. In some examples, each coefficient in the set of coefficients may be determined substantially independently, so that each coefficient in the set of coefficients may be substantially calculated simultaneously using a massively parallelized algorithm (e.g., implemented on a GPU-based architecture). Therefore, the computational limitations of the geometry-based model may be improved via the homography-based model described in the embodiments herein.

[0236] At 635, method 600 may include reconstructing an image via iterative reprojection (e.g., forward projection) and backprojection (e.g., reverse projection) based on a one-dimensional homography resampling transform. The symmetry of the symmetric projector / backprojector operator pair may be well suited for such iterative image reconstruction while also allowing arbitrary precision and computational efficiency by selecting the set of basis functions, interpolation schemes, and independent determination of the set of coefficients. Thus, the homography-based model described by the embodiments herein may be suitable for deep learning-based image reconstruction techniques that may utilize a high-precision, massively parallel algorithm to generate accurately reconstructed images.

[0237] At 640, method 600 may include displaying the reconstructed image. For example, the reconstructed image may be displayed to a user at a display device of the medical imaging system, such as the one described above with reference to FIG. Figure 2 The display device 232 of the imaging system 200 is described in detail.

[0238] At 645, method 600 may include diagnosing a medical problem based on the (displayed) reconstructed image. For example, method 600 may include identifying a lesion depicted by the reconstructed image. In some examples, since the accuracy level of the imaging data can be arbitrarily improved by selecting a basis set and an interpolation scheme, the lesion can be more accurately identified relative to an image reconstruction algorithm with a fixed accuracy level. After identification, the lesion itself can be diagnosed as benign or malignant, after which treatment can be recommended. It should be understood that the diagnosis of the lesion is provided as an illustrative and non-limiting example, and any of a plurality of tissues, organs, or foreign bodies in the subject can be identified based on the reconstructed image. It will also be understood that other embodiments disclosed herein may relate to other applications other than medical diagnosis, such as the identification of one or more components of an inanimate object.

[0239] In this way, an iterative image reconstruction method is provided, which utilizes a one-dimensional homography resampling transform. Specifically, a symmetric projector / backprojector pair made from a given set of one-dimensional homography resampling transforms can be used to satisfy the mathematical constraints placed on the iterative reprojection and backprojection paradigms, while providing an arbitrary level of accuracy via a basis set expansion of the resampling function. The technical effect of performing iterative reconstruction via such a homography-based model is that the geometric representation of the integral line can be reshaped (e.g., generalized) or avoided entirely, resulting in an easier and highly parallelized algorithm. In some examples, an implementation of the iterative reconstruction model method can therefore be implemented as a massively parallel graphics processing unit architecture, which can also be suitable for deep learning techniques for image reconstruction.

[0240] In one embodiment, a method is provided, the method comprising acquiring imaging data corresponding to a plurality of divergent X-rays, assigning a single functional form to the plurality of divergent X-rays, determining weights of interactions between a plurality of distribution samples and a plurality of X-ray detector intervals based on the single functional form via a homography transformation, and reconstructing an image based on the weights of the interactions. The first example of the method also includes sampling the single functional form to obtain a desired level of accuracy. The second example of the method optionally includes the first example of the method, and further includes wherein sampling the single functional form includes extending the single functional form using a set of one-dimensional basis functions. The third example of the method optionally includes one or more of the first and second examples of the method, and further includes wherein each weight of the interaction weights is associated with a corresponding one-dimensional basis function in a set of one-dimensional basis functions. The fourth example of the method optionally includes one or more of the first to third examples of the method, and further includes wherein sampling the single functional form includes determining unknown points corresponding to the single functional form by interpolating known points from the acquired imaging data. The fifth example of the method optionally includes one or more of the first to fourth examples of the method, and further includes wherein the imaging data is sampled in a cone beam tomography geometry based on a flat panel detector. The sixth example of the method optionally includes one or more of the first to fifth examples of the method, and also includes wherein a homography transformation is defined before acquiring imaging data. The seventh example of the method optionally includes one or more of the first to sixth examples of the method, and also includes wherein reconstructing the image includes iteratively reconstructing the image with a symmetric projector / back-projector operator pair determined based on interaction weights. The eighth example of the method optionally includes one or more of the first to seventh examples of the method, and also includes wherein the interaction weights are determined in parallel. The ninth example of the method optionally includes one or more of the first to eighth examples of the method, and also includes diagnosing a medical problem based on the reconstructed image.

[0241] In another embodiment, a medical imaging system is provided, the medical imaging system comprising an X-ray source, an X-ray detector array configured to detect X-ray radiation emitted by the X-ray source and attenuated by an object of interest, and a controller communicatively coupled to the X-ray detector array and storing instructions in a non-volatile memory, the instructions executable to acquire sampled imaging data sampled on a sampling grid and interval of the X-ray detector array, select a set of basis functions for a set of homography resampling transforms based on a desired level of accuracy, determine in parallel each coefficient in a set of coefficients for a corresponding one of the set of basis functions, reconstruct an image depicting the object of interest by iteratively reprojecting and backprojecting the sampled imaging data based on the set of coefficients, and display the reconstructed image. The first example of the system also includes wherein each basis function in the set of basis functions is globally continuous and independently calculated. The second example of the system optionally includes the first example of the system, and also includes wherein the instructions are also executable to select an interpolation scheme for the set of basis functions based on the desired level of accuracy. The third example of the system optionally includes one or more of the first and second examples of the system, and further includes wherein the controller is further communicatively coupled to a plurality of GPGPUs, the plurality of GPGPUs being configured to execute at least some of the instructions in parallel. The fourth example of the system optionally includes one or more of the first to third examples of the system, and further includes wherein the image is reconstructed using a deep learning-based image reconstruction algorithm.

[0242] In yet another embodiment, a method for reconstructing an image from sampled imaging data is provided, the method comprising resampling the sampled imaging data via a set of one-dimensional homography resampling transformations to obtain resampled imaging data as a plurality of known sampling points, determining a plurality of unknown sampling points via interpolation of the plurality of known sampling points, forming a symmetric projector / backprojector operator pair based on the plurality of known sampling points and the plurality of unknown sampling points, and reconstructing the image by iteratively projecting and backprojecting the resampled imaging data via the symmetric projector / backprojector operator pair. The first example of the method also includes where determining the plurality of unknown sampling points via interpolation of the known sampling points includes associating the plurality of known sampling points with the plurality of unknown sampling points to form a set of one-dimensional basis functions. The second example of the method optionally includes the first example of the method, and further includes where forming the symmetric projector / backprojector operator pair includes determining a set of coefficients respectively associated with a set of one-dimensional basis functions. The third example of the method optionally includes one or more of the first and second examples of the method, and further includes where the determination of the set of coefficients is performed analytically. A fourth example of the method optionally includes one or more of the first to third examples of the method, further comprising wherein each one-dimensional basis function in a set of one-dimensional basis functions is formed as a convolution of a B-spline function.

[0243] As used herein, the elements or steps listed in the singular and beginning with the word "one" or "a kind of" should be understood as not excluding a plurality of the elements or steps, unless such exclusion is explicitly stated. In addition, the reference to "an embodiment" of the present invention is not intended to be interpreted as excluding the existence of additional embodiments that also include the cited features. In addition, unless explicitly stated to the contrary, the embodiment of "including", "comprising" or "having" an element or multiple elements with a specific characteristic may include additional such elements without the characteristic. The terms "including" and "in..." are used as the concise language equivalents of the corresponding terms "including" and "wherein". In addition, the terms "first", "second" and "third" etc. are only used as marks, and are not intended to impose numerical requirements or specific positional order on their objects.

[0244] This written description uses examples to disclose the invention, including the best mode, and also to enable a person of ordinary skill in the relevant art to practice the invention, including making and using any devices or systems and performing any included methods. The patentable scope of the invention is defined by the claims, and may include other examples that occur to a person of ordinary skill in the art. Such other examples are intended to fall within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insignificant differences from the literal language of the claims.

Claims

1. A method for reconstructing an image, the method comprising: collecting imaging data corresponding to a plurality of divergent X-rays; assigning a single functional form to the plurality of divergent X-rays; determining weights of interactions between a plurality of distribution samples and a plurality of X-ray detector bins based on the single functional form via a homography transformation; reconstructing an image based on the weights of the interactions; as well as The single functional form is sampled to obtain a desired level of accuracy, wherein a set of basis functions is selected for a set of homography transforms based on the desired level of accuracy, and wherein a size of the set of basis functions is increased to correspondingly increase the level of accuracy. 2 . The method of claim 1 , wherein sampling the single functional form comprises expanding the single functional form using a set of one-dimensional basis functions.

3. The method of claim 2, wherein each of the weights of the interaction is associated with a corresponding one-dimensional basis function in the set of one-dimensional basis functions. 4 . The method of claim 1 , wherein sampling the single functional form comprises determining unknown points corresponding to the single functional form by interpolating known points from the acquired imaging data.

5. The method of claim 1, wherein the imaging data is sampled in a cone beam tomography geometry based on a flat panel detector. The method of claim 1 , wherein the homography transformation is defined prior to acquiring the imaging data.

7. The method of claim 1, wherein reconstructing the image comprises iteratively reconstructing the image using a symmetric projector / backprojector operator pair determined based on weights of the interactions. The method of claim 1 , wherein the weights of the interactions are determined in parallel.

9. The method of claim 1, further comprising diagnosing a medical problem based on the reconstructed image.

10. A medical imaging system comprising: X-ray source; an X-ray detector array configured to detect X-ray radiation emitted by the X-ray source and attenuated by an object of interest; and a controller communicatively coupled to the X-ray detector array and storing instructions in a non-transitory memory, the instructions executable to: Collecting sampled imaging data sampled on a sampling grid and interval of the X-ray detector array; selecting a set of basis functions for a set of homography resampling transformations based on a desired level of accuracy; determining in parallel each coefficient in a set of coefficients for a corresponding one of the set of basis functions; reconstructing an image depicting the object of interest by iteratively reprojecting and backprojecting the sampled imaging data based on the set of coefficients; and Display the reconstructed image, The controller is further communicatively coupled to one or more CPUs or GPUs configured to execute at least some of the instructions in parallel.

11. The system of claim 10, wherein each basis function in the set of basis functions is globally continuous and independently computed.

12. The system of claim 10, wherein the instructions are further executable to select an interpolation scheme for the set of basis functions based on the desired level of accuracy.

13. The system of claim 10, wherein the image is reconstructed using a deep learning based image reconstruction algorithm.

14. A method for reconstructing an image from sampled imaging data, the method comprising: resampling the sampled imaging data via a set of one-dimensional homography resampling transformations to obtain resampled imaging data as a plurality of known sampling points; determining a plurality of unknown sampling points via interpolation of the plurality of known sampling points; forming a symmetric projector / backprojector operator pair based on the plurality of known sampling points and the plurality of unknown sampling points; as well as reconstructing the image by iteratively projecting and back-projecting the resampled imaging data via the symmetric projector / back-projector operator pair, The image is reconstructed using a deep learning-based image reconstruction algorithm.

15. The method of claim 14, wherein determining the plurality of unknown sampling points via interpolation of the known sampling points comprises correlating the plurality of known sampling points with the plurality of unknown sampling points to form a set of one-dimensional basis functions.

16. The method of claim 15, wherein forming the symmetric projector / backprojector operator pair comprises determining a set of coefficients respectively associated with the set of one-dimensional basis functions.

17. The method of claim 16, wherein the determination of the set of coefficients is performed analytically.

18. The method of claim 15, wherein each one-dimensional basis function in the set of one-dimensional basis functions is formed as a convolution of a B-spline function.

Citation Information

Patent Citations

  • GPU acceleration method of CT image reconstruction

    CN101283913A

  • Method and apparatus for reconstructing three-dimensional computerized tomography (CT) image of object from incomplete cone heam projection data

    CN1062609A

  • Local adaptive mismatching point elimination method based on radial basis function fitting

    CN110047032A

  • Dynamic tomographic image reconstruction and rendering on-demand

    US20090274354A1