Method and system for calculating hemodynamic parameters

Abstract

Methods and systems for hemodynamic parameter estimation are described herein. In certain embodiments, a set of perfusion data for a region of interest is acquired using an imaging system. An arterial signal is obtained from the set of perfusion data. Tissue signals are obtained from the set of perfusion data. The arterial signal and the tissue signal are provided as inputs to one or more neural networks to determine one or more hemodynamic parameters of the region of interest. The one or more neural networks are trained using the one or more synthesized data.

Description

Background Art

[0001] The subject matter disclosed herein relates to obtaining hemodynamic parameters using a deep neural network to identify a non-parametric model from computed tomography (CT) perfusion.

[0002] Non-invasive imaging techniques (e.g., computed tomography (CT), magnetic resonance imaging (MRI), ultrasound (US), positron emission tomography (PET), single photon emission computed tomography (SPECT)) enable the acquisition of images of the internal structures or characteristics of a patient or object without performing invasive procedures on the patient or object. Specifically, such non-invasive imaging techniques rely on various physical principles (such as differential transmission of X-rays through a target volume, reflection of sound waves within a volume, paramagnetism of different tissues and materials within a volume, decomposition of a target radionuclide within the body, etc.) to collect data and construct images or otherwise represent the observed internal characteristics of the patient / object.

[0003] By way of example, computed tomography (CT) perfusion is an imaging modality used to evaluate microcirculation in tissues. Absolute regional measurements of hemodynamic parameters (e.g., blood flow (BF), blood volume (BV), mean transit time (MTT), time to peak (TMAX)) can be determined by computed tomography perfusion imaging. Visual encoded (e.g., color-coded, grayscale, annotated, etc.) maps (e.g., hemodynamic parameter maps) of these hemodynamic parameters can be generated for comparison with normal values. Thresholds or baseline values of these hemodynamic parameters can be established to monitor changes in microcirculation, which can be used to characterize various pathologies, such as local ischemia of organs (e.g., brain, myocardium, lung), tumor angiogenesis status and changes, characteristics of specific organs (e.g., liver, kidney, lung), etc.

[0004] Generating hemodynamic parameter maps from four-dimensional (4D) computed tomography perfusion acquisitions involves using a deconvolution algorithm to retrieve hemodynamic features from voxel-wise one-dimensional (1D) time signals. However, current implementations (e.g., parametric models of perfusion using least squares (LSQ) regression) perform poorly and exhibit low signal-to-noise ratios. In addition, poor registration and / or superposition of residues from previous boluses may introduce errors. Thus, a better and more robust technique is needed to generate and evaluate such hemodynamic maps and data. Summary of the Invention

[0005] An overview of certain embodiments disclosed herein is presented below. It should be understood that these aspects are provided merely to give the reader a brief overview of these particular embodiments and are not intended to limit the scope of the present disclosure. Indeed, the present disclosure may cover various aspects that may not be shown below.

[0006] As discussed herein, techniques are described that relate to estimating hemodynamic parameters from four-dimensional (4D) computed tomography perfusion data using deep learning (DL) methods. In one embodiment, during computed tomography acquisition, a series of images of a region of interest (e.g., tissue) are acquired, which include images taken before, during, and after injection of a contrast agent (e.g., tracer bolus or otherwise (e.g., ASL) labeling of blood) into the region of interest. A deep learning algorithm is trained using synthetic (e.g., simulated) data generated based on the 4D computed tomography perfusion data to obtain the residue impulse function Q(t) of the region of interest. The deep learning algorithm is also trained to reduce / mitigate image non-ideal factors in the 4D computed tomography perfusion data. The residue impulse function Q(t) of the region of interest is estimated using a neural network trained in this manner, and the corresponding hemodynamic parameters of the region of interest are determined using the residue impulse function.

[0007] In one embodiment, a method for computing hemodynamic parameters is provided. According to this embodiment, a set of perfusion data of a region of interest is acquired using an imaging system. An arterial signal is obtained from the set of perfusion data. A tissue signal is obtained from the set of perfusion data. The arterial signal and the tissue signal are provided as inputs to one or more neural networks to determine one or more hemodynamic parameters of the region of interest. The one or more neural networks are trained using synthetic data.

[0008] According to a further aspect, in such a method, the set of perfusion data may include computed tomography (CT) perfusion data. Alternatively, in other embodiments, the set of perfusion data may include magnetic resonance imaging (MRI) perfusion data, positron emission tomography (PET) perfusion data, single photon emission computed tomography (SPECT) data, or ultrasound imaging data. In the same or other embodiments, one or more synthetic data are generated based on a defined ground truth model. In the same or other embodiments, the tissue signal is a convolution of the arterial signal and the residue impulse function of the region of interest. In such an embodiment, one or more hemodynamic parameters can be determined from the residue impulse function. In the same or other embodiments, the one or more neural networks are trained to correct image non-ideal factors in the set of perfusion data. In the same or other embodiments, the one or more hemodynamic parameters include at least one of blood flow (BF), blood volume (BV), mean transit time (MTT), or time to peak (TMAX).

[0009] In yet another embodiment, a system is provided. According to this embodiment, the system includes one or more processors and a memory accessible by the one or more processors, and the memory stores instructions. When the instructions are executed by the one or more processors, the one or more processors are caused to perform operations that include: receiving a set of perfusion data acquired using an imaging system to image a region of interest; obtaining an arterial signal from the set of perfusion data; obtaining a tissue signal from the set of perfusion data; and providing the arterial signal and the tissue signal as inputs to one or more neural networks to determine one or more hemodynamic parameters of the region of interest, wherein the one or more neural networks are trained using one or more synthetic data.

[0010] According to a further aspect, in a specific implementation of such a system, the set of perfusion data may include computed tomography (CT) perfusion data. Alternatively, in other embodiments, the set of perfusion data may include magnetic resonance imaging (MRI) perfusion data, positron emission tomography (PET) perfusion data, single photon emission computed tomography (SPECT) data, or ultrasound imaging data. In the same or other embodiments, one or more synthetic data are generated based on a defined ground truth model. In the same or other embodiments, the tissue signal is a convolution of the arterial signal and a residual impulse function of the region of interest. In such an embodiment, one or more hemodynamic parameters may be determined from the residual impulse function. In the same or other embodiments, the one or more neural networks are trained to correct image non-idealities in the set of perfusion data. In the same or other embodiments, the one or more hemodynamic parameters include at least one of blood flow (BF), blood volume (BV), mean transit time (MTT), or time to peak (TMAX).

[0011] In an additional embodiment, a method for training one or more neural networks is provided. According to this embodiment, a set of synthetic residual impulse functions of a region of interest is generated based on a defined ground truth model. An arterial signal is obtained from a set of perfusion data. A synthetic tissue signal is generated based on the set of synthetic residual impulse functions and the arterial signal. The one or more neural networks are trained using the signals generated using the synthetic tissue signal and the arterial signal.

[0012] According to another aspect, in such a method, the synthetic tissue signal may include perturbations associated with perturbations in the perfusion data. The perturbations may be associated with registration errors (e.g., patient movement during acquisition) that may occur during perfusion data acquisition, or may be associated with noise (e.g., Gaussian noise, artifacts, speckle noise, bolus superposition) originating from the image acquisition techniques used. In the same or other embodiments, the set of perfusion data may include computed tomography (CT) perfusion data. In other embodiments, the set of perfusion data may include magnetic resonance imaging (MRI) perfusion data, positron emission tomography (PET) perfusion data, single photon emission computed tomography (SPECT) data, or ultrasound imaging data. In the same or other embodiments, the loss is used as a bias for training one or more neural networks. In such a specific implementation, the loss is determined based on a comparison of a first set of parameters derived from an estimated residual impulse function output from one or more neural networks with a second set of parameters derived from a synthetic residual impulse function. The first set of parameters may include a first set of hemodynamic parameters, and the second set of parameters may include a second set of hemodynamic parameters. In the same or other embodiments, when the loss is less than a threshold or after a plurality of time periods, it is determined that the training of one or more neural networks has ended. In the same or other embodiments, regularization is used as a bias for training one or more neural networks. In such a specific implementation, the regularization is associated with the characteristics of the estimated residual impulse function output from the one or more neural networks. In the same or other embodiments, the regularization may be weighted by a deconvolution error. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] These and other features, aspects, and advantages of the present invention will be better understood when the following detailed description is read with reference to the accompanying drawings, in which like reference symbols represent like parts throughout the drawings, wherein:

[0014] Figure 1 An example of an artificial neural network for training a deep learning model according to aspects of the present disclosure is depicted;

[0015] Figure 2 is a block diagram depicting components of a computed tomography (CT) imaging system according to aspects of the present disclosure;

[0016] Figure 3 depicts a computing system for analyzing images obtained from a Figure 2 computed tomography imaging system according to aspects of the present disclosure;

[0017] Figure 4 depicts a rendering of a simplified indicator dilution physical model according to aspects of the present disclosure;

[0018] Figure 5A flowchart depicting a method for obtaining a residual pulse function and corresponding hemodynamic parameters of a region of interest in accordance with aspects of the present disclosure;

[0019] Figure 6 A flowchart depicting a method for training a deep learning (DL) model in accordance with aspects of the present disclosure;

[0020] Figure 7 A flowchart depicting a method for predicting an estimated residual pulse function using a deep learning model in accordance with aspects of the present disclosure;

[0021] Figure 8 A flowchart depicting a method for training Figure 7 a deep learning model; and

[0022] Figure 9 A ground truth model in accordance with aspects of the present disclosure that can be used for Figure 6 synthetic data generation. Detailed Description

[0023] One or more specific embodiments will be described below. To provide a concise description of these embodiments, not all features of an actual implementation are described in the specification. It should be understood that, as in any engineering or design project, numerous implementation-specific decisions must be made in developing any such actual implementation to achieve the developers' specific goals, such as compliance with system-related and business-related constraints that may vary from one implementation to another. In addition, it should be understood that such development efforts may be complex and time-consuming, but would still be a routine task for those of ordinary skill in the art who would benefit from the present disclosure in terms of design, fabrication, and manufacture.

[0024] Although aspects of the following discussion are provided in the context of medical imaging, it should be understood that aspects of the disclosed techniques are applicable to other contexts and are thus not limited to such medical examples. Indeed, the examples and explanations are provided in such medical contexts merely to facilitate explanation by providing instances of real-world implementations and applications, and should not be construed as limiting the applicability of the inventive methods to other suitable uses, such as for other non-destructive and / or non-invasive imaging contexts.

[0025] As discussed herein, perfusion imaging is an imaging modality used to evaluate microcirculation in tissues. Microcirculation is associated with supplying oxygen and nutrients (including drugs and toxins), removing CO2 and other metabolic waste products (e.g., catabolites, toxins), releasing and / or capturing modulators (e.g., hormones, neurotransmitters), generating immune responses and inflammation, regulating tissue fluid, temperature, and core body temperature, controlling blood pressure, etc. Perfusion generally involves an intravenous bolus injection of a contrast agent (e.g., a substance or composition that enhances the visibility of tissues such as blood or other media that may be difficult to observe in images generated using a given imaging modality) into the tissue, and acquisition of multiple temporal phases of the tissue using an imaging modality (such as a computed tomography (CT) scanner) after the bolus. Absolute regional measurements of hemodynamic parameters (e.g., blood flow (BF), blood volume (BV), mean transit time (MTT), time to peak (TMAX)) can be determined by computed tomography (CT) perfusion imaging. Color or other visually encoded maps of these hemodynamic parameters (e.g., hemodynamic parameter maps) can be generated for comparison with normal or baseline values of an individual (e.g., longitudinal study) or a related group or subgroup. Thresholds for these hemodynamic parameters can be established to monitor changes in microcirculation, which can be used to characterize various pathologies, such as local ischemia of organs (e.g., brain, myocardium, lung), tumor neovascularization status and changes, characteristics of specific organs (e.g., liver, kidney, lung), etc. Computed tomography perfusion acquisition data can include three-dimensional spatial data and one-dimensional temporal data, which together constitute four-dimensional (4D) computed tomography perfusion data. Generating hemodynamic parameter maps from four-dimensional (4D) computed tomography perfusion acquisitions involves using a deconvolution algorithm to retrieve hemodynamic features from voxel-wise one-dimensional (1D) time signals. This discussion relates to using deep learning (DL) methods to resolve the deconvolution algorithm and generate hemodynamic parameter maps or other comparable data outputs.

[0026] Although CT examples are mainly provided herein, it should be understood that the techniques disclosed in the present invention can be used in other imaging modalities. For example, the currently described methods can also be used for data acquired by other types of tomographic scanners (including but not limited to ultrasound (US) scanners, positron emission tomography (PET) scanners, single photon emission computed tomography (SPECT) scanners, magnetic resonance imaging (MRI) scanners, and / or other X-ray-based imaging techniques such as C-arm-based techniques). The techniques disclosed in the present invention can also be used for the processing of combined computed tomography angiography (CTA) and perfusion acquisitions.

[0027] As background, several imaging modalities, such as X-ray computed tomography (e.g., multislice CT, helical CT, cone-beam CT) and X-ray C-arm systems (e.g., cone-beam imaging), measure projections of a scanned object or patient, where depending on the technology, the projections correspond to Radon transform data, fan-beam transform data, cone-beam transform data, or non-uniform Fourier transform. In other cases, the scan data can be magnetic resonance data (e.g., magnetic resonance imaging (MRI) data) generated in response to an applied magnetic field and RF pulses, etc.

[0028] In other cases, single-photon emission computed tomography (SPECT) and positron emission tomography (PET) can utilize radiopharmaceuticals that are administered to a patient and whose decay results in the emission of positrons of gamma rays at locations within the patient. The radiopharmaceutical is typically chosen so as to preferentially or differentially distribute in the body based on physiological or biochemical processes in the body. For example, a radiopharmaceutical that is preferentially processed or absorbed by tumor tissue can be selected. In such an example, the radiopharmaceutical will typically be disposed at a greater concentration around the tumor tissue within the patient's body.

[0029] In other cases, an ultrasound imaging system can acquire ultrasound data of a patient. In some embodiments, the ultrasound system can be a digital acquisition and beamformer system, but in other embodiments, the ultrasound system can be any suitable type of ultrasound system. Such an ultrasound system can include an ultrasound probe and a workstation (e.g., monitor, console, user interface) that can control the operation of the ultrasound probe and can process the image data acquired by the ultrasound probe. The ultrasound probe can be coupled to the workstation by any suitable technique for transmitting image data and control signals between the ultrasound probe and the workstation, such as wireless, optical, coaxial, or other suitable connections.

[0030] Reconstruction routines, along with associated correction and calibration routines, are used in conjunction with these imaging modalities to generate useful clinical images and / or data, which can then be used, for example, to derive or measure hemodynamic parameters of interest by using deep learning (DL) techniques, as discussed herein.

[0031] The deep learning (DL) methods discussed in this document can be based on artificial neural networks and, therefore, can include one or more of deep neural networks such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformers, generative adversarial networks (GANs), etc. As discussed herein, deep learning techniques (which may also be referred to as deep machine learning, hierarchical learning, or deep structured learning) are a branch of machine learning techniques that employ mathematical representations of data and artificial neural networks for learning and processing such representations. A neural network can include multiple layers, such as an input layer, hidden layers, and an output layer. The basic unit of computation in a neural network is a neuron / node. Each neuron / node receives inputs from some other nodes or from an external source and computes an output. The input layer can include neurons / nodes that receive external inputs such as input data. Each hidden layer consists of a set of neurons / nodes with learnable weights and biases, and each neuron / node in a hidden layer can receive inputs from upstream-connected nodes or layers and perform operations on the inputs to compute an output provided to downstream-connected nodes or layers. The output layer can include neurons / nodes that receive inputs from the hidden layer and output results.

[0032] For example, deep learning (DL) methods can be characterized in that they use one or more algorithms to extract or model highly abstract concepts of a class of data of interest. This can be done using one or more processing layers, where each layer typically corresponds to a different level of abstract concept and, therefore, may adopt or utilize different aspects of the initial data or the output of the previous layer (i.e., the hierarchical or cascading structure of the layers) as the target for the process or algorithm of a given layer. In the context of image processing or reconstruction, this can be characterized as different layers corresponding to different feature levels or resolutions in the data. Generally speaking, the processing from one representation space to the next-level representation space can be regarded as a "phase" of the process. Each phase of the process can be performed by a separate neural network or by different parts of a larger neural network.

[0033] Accordingly, the techniques discussed in this document utilize deep learning (DL) methods to estimate hemodynamic parameters from computed tomography perfusion data. In certain specific implementations discussed herein, synthetic (e.g., simulated) data generated based on clinical data is used as training data to train deep learning algorithms, rather than clinical, real-world data or geometric constructs. As discussed herein, training one or more deep learning algorithms using synthetic data is very different from directly using clinical data for such training purposes, which may involve estimating the true state or acquiring additional data representing the true state and registering the additional data with the clinical data for the purpose of assembling training data.

[0034] As discussed herein, as part of the initial training of a deep learning process to solve a particular problem, a training data set with known initial values and known (i.e., true) values of the final output of the deep learning process can be employed. In this way, real training data can be used to train the network to provide known correct outputs in response to known inputs. As discussed in more detail below, according to the method of the present invention, synthetic data is used as training data, where the synthetic data is simulated or synthesized or derived from clinical data and / or simple geometric constructs, but is different from the clinical data. Additionally, due to its synthetic nature, the synthetic training data discussed herein is associated with known true attributes without having to estimate or measure such true values or perform additional invasive operations to derive such true attributes.

[0035] For example, a single stage of training can have known input values corresponding to one representation space and known output values corresponding to the next level of representation space. In this way, the deep learning algorithm can process the known or training data set (either in a supervised or guided manner or in an unsupervised or unguided manner) until a mathematical relationship between the initial data and one or more desired outputs is seen and / or a mathematical relationship between the inputs and outputs of each layer is seen and characterized. Similarly, a separate validation data set can be employed, where the initial and desired target values are known, but only the initial values are provided to the trained deep learning algorithm and the output of the deep learning algorithm is compared with the desired target values to verify the previous training and / or prevent overtraining.

[0036] In view of the foregoing, Figure 1 An example of an artificial neural network 50 is schematically depicted, which can be trained as a deep learning model as discussed herein. In this example, the network 50 is multi-layered, having training inputs 52 (e.g., synthetic data) and a plurality of layers present in the network 50, which includes an input layer 54, hidden layers 58A, 58B, etc., an output layer 60, and a training target 64. In certain specific implementations, the input layer 54 can also be characterized as or understood as a hidden layer. In this example, each layer is composed of a plurality of "neurons" or nodes 56. The number of neurons 56 can be constant between layers or, as shown, can vary from layer to layer. The neurons 56 of each layer generate corresponding outputs, which are used as inputs to the neurons 56 of the next hierarchical layer. In practice, a weighted sum of the inputs with added biases is computed to "excite" or "activate" each corresponding neuron of the layers according to an activation function, such as a rectified linear unit (ReLU), sigmoid function, hyperbolic tangent function, or otherwise specified or programmed function. The output of the last layer constitutes the network output 60, which together with the target image or parameter set 64 is used by a loss function or error function 62 to generate an error signal, which will be backpropagated to guide the network training.

[0037] A loss function or error function 62 measures the difference between the network output and the training target. In some specific implementations, the loss function can be the mean squared error (MSE) of voxel-level values or partial line integral values and / or can account for differences involving other image features such as image gradients or other image statistics. Alternatively, the loss function 62 can be defined by other metrics associated with the specific task under consideration, such as the softmax function or the DICE value (where DICE refers to the ratio and A ∩ B represents the intersection of regions A and B, and |·| represents the area of the region).

[0038] To facilitate the explanation of the inventive methods using deep learning techniques, this disclosure mainly discusses these methods in the context of CT or C-arm systems. However, it should be understood that the following discussion is also applicable to other image modalities and systems, including but not limited to multi-energy CT and MRI, and any context in which tomographic reconstruction is employed to reconstruct images from which hemodynamic parameters can be discerned and / or measured.

[0039] With this in mind, Figure 2 An example of an imaging system 110 (i.e., a scanner) is depicted. In the depicted example, the imaging system 110 is a computed tomography imaging system that is designed to acquire scan data (e.g., X-ray attenuation data) in various radial views around a patient (or other subject or object of interest) and is suitable for performing image reconstruction using tomographic reconstruction techniques. In Figure 2 the illustrated embodiment, the imaging system 110 includes an X-ray radiation source 112 positioned adjacent to a collimator 114. The X-ray source 112 can be an X-ray tube, a distributed X-ray source such as a solid-state or thermionic X-ray source, or any other X-ray radiation source suitable for acquiring medical or other images. Conversely, in an MRI embodiment, the measurements are samples in Fourier space and can be applied directly as input to a neural network or can first be transformed to line integrals in the sinogram space.

[0040] In the depicted example, the collimator 114 shapes or confines a beam of X-rays 116 that enters the region where the patient / object 118 is positioned. In the depicted example, the X-rays 116 are collimated into a cone beam (i.e., a conical beam) or a fan beam (i.e., a fan-shaped beam) that passes through the imaging volume. A portion of the X-ray radiation 120 passes through the patient / object 118 (or other subject of interest) or through its vicinity and impinges on a detector array, generally denoted by reference numeral 122. The detector elements of the array generate electrical signals representing the intensity of the incident X-rays 120. These signals are acquired and processed to reconstruct an image of the features within the patient / object 118.

[0041] Source 112 is controlled by system controller 124, which provides power and control signals for a computed tomography examination sequence. In the depicted embodiment, system controller 124 controls source 112 via X-ray controller 126, which may be a component of system controller 124. In such embodiments, X-ray controller 126 may be configured to provide power and timing signals to X-ray source 112.

[0042] In addition, detector 122 is coupled to system controller 124, which controls the acquisition of signals generated in detector 122. In the depicted embodiment, system controller 124 uses data acquisition system 128 to acquire signals generated by the detector. Data acquisition system 128 receives data collected by the readout electronics of detector 122. Data acquisition system 128 may receive sampled analog signals from detector 122 and convert the data to digital signals for subsequent processing by processing component 130 discussed below. Alternatively, in other embodiments, digital-to-analog conversion (DAC) may be performed by circuitry disposed on detector 122 itself. System controller 124 may also perform various signal processing and filtering functions on the acquired signals, such as initial adjustment for dynamic range, interleaving of digital data, and the like.

[0043] In Figure 2 the depicted embodiment, system controller 124 is coupled to rotation subsystem 132 and linear positioning subsystem 134. Rotation subsystem 132 enables X-ray source 112, collimator 114, and detector 122 to rotate one or more times around patient / object 118, such as rotating around the patient primarily in the x,y plane. It should be noted that rotation subsystem 132 may include a gantry or C-arm on which the corresponding X-ray emission and detection components are disposed. Thus, in such embodiments, system controller 124 may be used to operate the gantry or C-arm.

[0044] Linear positioning subsystem 134 may enable patient / object 118 or more specifically the table supporting the patient to be displaced within the bore of CT system 110, such as displaced in the z direction relative to the rotation of the gantry. Thus, the table may be linearly moved (either continuously or stepwise) within the gantry to generate images of a particular region of interest of patient 118. In the depicted embodiment, system controller 124 controls the movement of rotation subsystem 132 and / or linear positioning subsystem 134 via motor controller 136.

[0045] Generally, system controller 124 commands the operation of imaging system 110 (such as via the operation of source 112, detector 122, and the above-described positioning system) to perform an examination protocol such as a computed tomography perfusion protocol and process the acquired data. For example, system controller 124 can cause the gantry supporting source 112 and detector 122 to rotate around the subject of interest via the above-described systems and controllers, such that X-ray attenuation data can be obtained at one or more angular positions relative to the subject. In this background, system controller 124 can also include signal processing circuitry, an associated memory circuitry for storing programs and routines executed by a computer (such as routines for performing the vascular property estimation techniques described herein), and configuration parameters, image data, etc.

[0046] In the depicted embodiment, the signals acquired and processed by system controller 124 are provided to processing component 130, which can perform image reconstruction. Processing component 130 can be one or more general-purpose or special-purpose microprocessors. The data collected by data acquisition system 128 can be transmitted directly to processing component 130 or after being stored in memory 138. Any type of memory suitable for storing data can be utilized by such an exemplary system 110. For example, memory 138 can include one or more optical, magnetic, and / or solid-state memory storage structures. Additionally, memory 138 can be located at the acquisition system site and / or can include remote storage devices for storing data, processing parameters, and / or routines for tomographic image reconstruction, as described below.

[0047] Processing component 130 can be configured to receive commands and scan parameters from an operator via operator workstation 140, which is typically equipped with a keyboard and / or other input devices. The operator can control system 110 via operator workstation 140. Thus, the operator can use operator workstation 140 to view the reconstructed images and / or otherwise operate system 110. For example, the display 142 coupled to operator workstation 140 can be used to view the reconstructed images and control imaging. Additionally, the images can also be printed by printer 144, which can be coupled to operator workstation 140.

[0048] In addition, the processing component 130 and the operator workstation 140 can be coupled to other output devices, which can include standard or dedicated computer monitors and associated processing circuitry. One or more operator workstations 140 can be further linked in the system for outputting system parameters, requesting examinations, viewing images, and the like. Generally speaking, the displays, printers, workstations, and similar devices provided within the system can be local to the data acquisition components, or can be remote from these components, such as in other locations within an institution or hospital, or located in completely different locations, and linked to the image acquisition system via one or more configurable networks (such as the Internet, virtual private networks, etc.).

[0049] It should also be noted that the operator workstation 140 can also be coupled to a Picture Archiving and Communication System (PACS) 146. The PACS 146 can in turn be coupled to a remote client 148, a Radiology Information System (RIS), a Hospital Information System (HIS), or coupled to an internal or external network, such that others at different locations can access the raw or processed image data. By way of example, in the context of the present invention, previously or recently acquired computed tomography perfusion images or image sets can subsequently be accessed from such an archival system for processing according to the techniques discussed herein for hemodynamic property estimation or longitudinal tracking.

[0050] Although the foregoing discussion has addressed the various exemplary components of the imaging system 110 separately, these various components can be provided within a common platform or in an interconnected platform. For example, the processing component 130, the memory 138, and the operator workstation 140 can be jointly provided as a general-purpose or dedicated computer or workstation that is configured to operate in accordance with aspects of the present disclosure. In such embodiments, the general-purpose or dedicated computer can be provided as a separate component relative to the data acquisition components of the system 110, or can be provided within a common platform with such components. Similarly, the system controller 124 can be provided as part of such a computer or workstation, or as part of a separate system dedicated to image acquisition.

[0051] Figure 2 The system can be used to acquire X-ray projection data (or other scan data of other modalities) of various views of a vasculated region of interest of a patient to reconstruct an image of the imaged region (e.g., a perfusion image or map) using the scan data. The projection (or other) data acquired by a system such as the imaging system 110 can be reconstructed as discussed herein to perform tomographic reconstruction. Although Figure 2The system shows a rotating subsystem 132 for rotating an X-ray source 112 and a detector 122 around an object or a subject, but such a rotating subsystem may include non-planar rotation aspects (e.g., complex rotation trajectories or other motions, including motions in other dimensions so as not to rotate strictly within a single plane), such as may be applicable for use with certain C-arm type imaging systems.

[0052] Figure 3 is a block diagram showing a computing system 150 that may be used in a remote client 148. Although the following description details some example components that make up the computing system 150, it should be understood that the computing system 150 may include additional or fewer components. The computing system 150 may include a communication component 152, a processor 154, a memory 156, a storage device 158, an input / output (I / O) port 160, a display 162, etc. The communication component 152 may be a wireless or wired communication component, which may facilitate communication between the computing system 150 and various types of devices or resources (e.g., databases, servers) directly or via a network. Additionally, the communication component 152 may facilitate data transmission to the computing system 150 such that the computing system 150 may receive data from Figure 2 the components depicted therein (e.g., PACS 146), etc. The communication component 152 may use a variety of communication protocols, such as Open Database Connectivity (ODBC), TCP / IP protocol, Distributed Relational Database Architecture (DRDA) protocol, Database Change Protocol (DCP), HTTP protocol, other suitable current or future protocols, or combinations thereof.

[0053] The processor 154 may include a single-threaded processor, a multi-threaded processor, or both. The processor 154 may process instructions stored in the memory 156. The processor 154 may also include hardware-based processors, each including one or more cores. The processor 154 may include a general-purpose processor, a dedicated processor, or both. The processor 154 may be communicatively coupled to other internal components (such as the communication component 152, the storage device 158, the I / O port 160, and the display 162).

[0054] Memory 156 and storage device 158 can be any suitable article of manufacture that can serve as a medium for storing processor-executable code, data, and the like. These articles of manufacture can represent computer-readable media (e.g., any suitable form of memory or storage device) that can store the processor-executable code used by processor 154 to execute the presently disclosed techniques. As used herein, an application can include any suitable computer software or program that can be installed on computing system 150 and executed by processor 154. Memory 156 and storage device 158 can represent non-transitory computer-readable media (e.g., any suitable form of memory or storage device) that can store the processor-executable code used by processor 154 to perform the various techniques described herein. It should be noted that non-transitory only indicates that the medium is tangible and not a signal.

[0055] I / O port 160 can be an interface that can be coupled to other peripheral components such as input devices (e.g., keyboard, mouse), sensors, input / output (I / O) modules, etc. Display 162 can serve as a human-machine interface (HMI) for presenting visualizations associated with software or executable code being processed by processor 154. Display 162 can be operated to present a representation of a three-dimensional (3D) augmented reality (AR) or virtual reality (VR) visualization associated with software or executable code being processed by processor 154. In one embodiment, display 162 can be a touch display capable of receiving input from an operator of computing system 150. Display 162 can be any suitable type of display, such as a liquid crystal display (LCD), a plasma display, or an organic light-emitting diode (OLED) display. Additionally, in one embodiment, display 162 can be provided in combination with a touch-sensitive mechanism (e.g., a touch screen) that can be used as part of the control interface of computing system 150.

[0056] The computer system 150 may also include a prediction engine 164, which may include a training component 166 and a prediction component 168. The training component 166 may receive training data (e.g., synthetic data) stored in the database 170 and use the training data to train a machine learning model. For example, a deep learning (DL) model may be trained in a supervised or guided manner (e.g., trained with training data including input data and expected prediction outputs (e.g., labeled data sets)). The deep learning model may also be trained in an unsupervised or unguided manner (e.g., trained with training data including input data but no expected prediction outputs (e.g., unlabeled data sets)). The prediction component 168 may use a set of machine learning models (e.g., functions, algorithms) trained by the training data to predict an output (e.g., hemodynamic parameters) of an initial value (e.g., clinical data) provided to the prediction component 168. In some embodiments, the predicted output may be supervised (e.g., by a user) to monitor or confirm the accuracy of the output, and the training data may be updated, which may be used by the training component 166 to retrain the machine learning model. The prediction engine 164 and / or the database 170 may be located in the local environment of the remote client 148 or in a cloud computing environment (e.g., a data center).

[0057] In view of the foregoing background and context discussion, the present disclosure relates to estimating hemodynamic parameters from 4D computed tomography perfusion data using deep learning methods. As previously described, computed tomography (CT) perfusion typically includes injecting an intravenous bolus of a contrast agent into tissue and acquiring multiple phases of the tissue using a computed tomography (CT) scanner after the bolus. To measure the response of the tissue after the bolus, indicator dilution techniques have been used during physiological measurements. Figure 4 is a block diagram of a simplified indicator dilution physical model 200 for illustrating a computed tomography perfusion process. In the simplified indicator dilution physical model 200, a liquid with a constant flow rate F flows from the inflow 204 (e.g., artery) to the outflow 206 (e.g., vein) through the internal compartment B of the region of interest 202. The dilution of the indicator in the inflow 204 (e.g., artery) is indicated by the arterial signal C a (t), and the response of the region of interest 202 to a single pulse of the tracer in the inflow 204 is indicated by the residue pulse function Q(t). The response of the region of interest 202 to the arterial signal C a (t) is indicated by the tissue signal C r (t) in the outflow 206 (e.g., vein). The tissue signal C r (t) is the convolution of the arterial signal C a (t) and the residue pulse function Q(t), as shown in Equation (1).

[0058]

[0059] Therefore, based on formula (1), the tissue signal C r (t) can be deconvolved to obtain the residue impulse function Q(t). The residue impulse function Q(t) can be used to obtain hemodynamic parameters of the region of interest 202, such as blood flow (BF), blood volume (BV), mean transit time (MTT), time to peak (TMAX), etc. Tissue blood flow (BF) corresponds to the blood flow entering / leaving a certain volume of tissue (e.g., expressed in ml / min / 100 ml). Blood volume (BV) corresponds to the volume of capillary blood contained in a certain volume of tissue (e.g., expressed in ml / 100 ml or as a percentage). MTT is the average time taken for blood to pass through the capillary network (the time between arterial inflow and venous outflow) (expressed in seconds). The arterial signal C a (t) and the tissue signal C r (t) can be acquired or measured from computed tomography scans using other imaging modalities (e.g., MRI). For example, during computed tomography acquisition, a series of images of the region of interest 202 can be acquired, which may include images taken before injecting a contrast agent (e.g., tracer bolus or otherwise (e.g., ASL) labeling the blood), during the injection of the contrast agent, and after the injection of the contrast agent. Therefore, computed tomography perfusion acquisition data can include three-dimensional spatial data and one-dimensional time data, which together constitute four-dimensional (4D) computed tomography perfusion data. This series of images can be used to study the microcirculation during the injection of the contrast agent. For example, the images acquired before injecting the contrast agent can be used as reference or baseline images, and the images acquired during and after the injection of the contrast agent can be used to study the injection effect relative to the reference or baseline. Therefore, the changes in the residue impulse function Q(t) and the tissue signal C r (t) caused by the injection of the contrast agent can be obtained from a series of images acquired during computed tomography acquisition.

[0060] Figure 5It is a flowchart of a method 220 for obtaining a residual impulse function Q(t) and corresponding hemodynamic parameters of a region of interest. At block 222, a volume sequence can be obtained by using a series of images acquired using various modalities (e.g., CT, MR, etc.) applicable to the region of interest. As previously mentioned, the series of images can include images acquired before, during, and after intravenous injection of a contrast agent (e.g., a tracer bolus or otherwise labeling blood, such as ASL). For example, sequential acquisitions can be performed at the level of slices or volumes before, during, and after injection of the contrast agent, and the images acquired before the start of injection of the contrast agent can be used as reference images. These images can be segmented to produce a geometric representation of the true underlying lumen geometry. Segmentation of geometric features (such as plaque components, adjacent structures, etc.) is envisioned. These geometric representations can be voxelized (converted to or represented by a volume representation where each voxel corresponds to a specific tissue type or combination of tissue types based on its position relative to the geometric representation) or characterized as polygonal surfaces, NURBS (non-uniform rational B-splines), or any number of other representations. Due to noise, resolution limitations, and other image non-idealities, these representations may not exactly match the initial shape of the true lumen, but they are close enough such that when taken together, a large series of these representations extracted from a large set of corresponding images can represent the types of geometric features common in clinical practice. At block 224, the tissue signal for all voxels of the volume sequence can be obtained using the volume representation produced at block 222. At block 226, a sample tissue time signal for voxels or small spatial regions in the volume sequence can be obtained. At block 228, an arterial signal can be obtained from the series of images acquired at block 222. Then at block 230, the sample tissue time signal can be deconvolved with respect to the arterial signal based on formula (1) to obtain the residual impulse function Q(t) for the voxels or small spatial regions. At block 232, the results of the deconvolution can be used to obtain the values of the parameter maps in the corresponding voxels / small spatial regions.

[0061] Figure 6 It is a flowchart of a method 260 for training a deep learning model for deconvolution in Figure 5 block 230. At block 262, an arterial signal C a (t) of a tissue region can be obtained from 4D computed tomography perfusion data. At block 264, a synthetic residual impulse function Q S(t). The tissue region for calculating tissue signals may include many capillaries, and capillary parameters may be described by a probability distribution. A model can be developed to consider the distribution of capillaries in the tissue and capillary parameters. For a defined model with a known parametric probability distribution, a set of residual impulse functions Q(t) can be generated and averaged to obtain a synthetic residual impulse function Q S (t). Then, the synthetic residual impulse function Q S (t) and the arterial signal C a (t) can be used to generate the tissue signal C r (t) based on Equation (1). In some specific implementations, the synthetic arterial signal C a (t) with known real data can be used to generate the tissue signal C r (t). In some specific implementations, the arterial signal C a (t) can be partially based on clinical image data (e.g., 4D computed tomography perfusion data in block 262) or derived from clinical image data. In some embodiments, due to noise, resolution limitations, and other image non-ideal factors in the 4D computed tomography perfusion data, as well as registration errors (e.g., patient movement during acquisition) that may occur during perfusion data acquisition, perturbations (e.g., additive noise) may be added to the tissue signal C r (t). The generated synthetic residual impulse function Q S (t) and the tissue signal C r (t) can be stored as training data (e.g., stored in database 170). At block 266, the arterial signal C a (t) generated at block 262 and the tissue signal C r (t) generated at block 264 can be input into a deep learning model to calculate the estimated residual impulse function Q e (t) (e.g., using prediction component 168), as Figure 7 shown in detail. At block 268, the loss function can be calculated using the estimated residual impulse function Q e (t) and the synthetic residual impulse function Q S (t), and the loss function can be backpropagated to block 266 to guide the deep learning model training, as Figure 8 shown in detail. For example, the loss function can be used to calculate the learnable weights and biases of the processing layers (e.g., hidden layers) in the deep learning model, and blocks 266 and 268 can be repeated until the value of the loss function is less than a threshold.

[0062] Figure 7 FIG. is a diagram showing a method for predicting an estimated residual impulse function Q eFlowchart of method 320 for (t). At block 324, the arterial signal C obtained from the CT perfusion 4D data can be a (t) and the tissue signal C r (t) are input into the network (e.g., the input layer 54 of the network) of the deep learning model 322 at block 328. At block 330, the deep learning model 322 can output parameters that can be transformed into the estimated residual impulse function Q e (t). At block 332, hemodynamic parameters can be determined by using the estimated residual impulse function Q e (t) obtained at block 330. Although one deep learning model is shown in Figure 7 , multiple deep learning models can be used alone or together to predict the estimated residual impulse function Q e (t).

[0063] Figure 8 is a flowchart of method 340 for training the deep learning model 322 using a loss function. The arterial signal C obtained from the CT perfusion 4D data at block 262 can be a (t) and the synthetic tissue signal C obtained at block 264 r (t) are input into the network (e.g., the input layer 54 of the network) of the deep learning model 322 at block 328. At block 330, the deep learning model 322 can output parameters that can be transformed into the estimated residual impulse function Q e (t). At block 342, various parameters (e.g., hemodynamic parameters) can be calculated using the estimated residual impulse function Q e (t) obtained by the deep learning model 322 at block 330, and various features can be obtained. These features can be compared with the corresponding parameters and features calculated or derived using the synthetic residual impulse function Q S (t), and then the difference obtained from the comparison can be used to determine the loss function A, which can be used as the deviation for all or part of the training of the deep learning model 322. The loss function A can also include the mean square error (MSE) of the voxel-level values or partial line integral values and / or can account for differences involving other image features, such as image gradients or other image statistical values. At block 344, the estimated residual impulse function Q e (t) obtained by the deep learning model 322 at block 330 can be used to determine the regularization deviation B, which can be related to the characteristics of the estimated residual impulse function Q e (t) (e.g., Q e(Second derivative of (t)). At block 346, a training weight α (e.g., any real number) for the loss function A can be determined, and a training weight β (e.g., any real number) for the regularization bias B can be determined, and the weighted loss function A and the weighted regularization bias B can be backpropagated into the network of the deep learning model 322 (e.g., the hidden layers 58A, 58B of the network) to guide network training. Blocks 328, 330, 342, 344, and 346 can be repeated until the value of the loss function is less than a threshold, and then the deep learning training of the deep learning model 322 can be ended. Then, the clinical arterial signal Ca(t) and the tissue signal Cr(t) obtained from the computed tomography perfusion 4D data can be input into the trained deep learning model 322 to determine the estimated residue impulse function Q e (t). In these embodiments where a perturbation (e.g., additive noise) is added to the tissue signal C r (t), the noise in the output of the trained deep learning model 322 due to image non-ideal factors in the 4D computed tomography perfusion data or registration errors that may occur during perfusion data acquisition (e.g., patient movement during acquisition) can be reduced / mitigated because the deep learning model 322 has been trained for noise, resolution limitations, and other image non-ideal factors as well as registration errors in the 4D computed tomography perfusion data by adding perturbations to the training data. The clinical arterial signal Ca(t) and the tissue signal Cr(t) can be sampled by multi-phase image acquisition, and their sampling does not require equality or regularity. Additionally, the support for sampling may not be the same in different applications. A fixed support can be used as the minimum acquisition duration for input. Additionally, the signals (e.g., Ca(t), Cr(t)) can be interpolated using a fixed step size (e.g., ≤0.5s).

[0064] Figure 9 Shows an embodiment of a ground truth model 380 that can be used for Figure 6 synthetic data generation. In the ground truth model 380, the value of the residue impulse function Q(t) is zero before time T0. The residue impulse function Q(t) is equal to the relative flow F at time T0 and decreases from the extravascular relative flow FE after time T0 + W, where W is the mean transit time (MTT). Additionally, the residue impulse function Q(t) is non-negative during acquisition. The above characteristics of the ground truth model 380 can be used to determine the regularization bias B.

[0065] The technical effects of the present invention include estimating hemodynamic parameters from four-dimensional (4D) computed tomography perfusion data using deep learning (DL) methods. During computed tomography acquisition, a series of images of an area of interest (e.g., tissue) are acquired, which include images taken before, during, and after injection of a contrast agent (e.g., tracer bolus or otherwise (e.g., ASL) labeling the blood) into the area of interest. A deep learning algorithm is trained using synthetic (e.g., simulated) data generated based on the 4D computed tomography perfusion data to obtain the residue impulse function Q(t) of the area of interest. Additionally, the deep learning algorithm can also be trained to reduce / mitigate image non-ideal factors in the 4D computed tomography perfusion data. The residue impulse function Q(t) of the area of interest is estimated using a neural network trained in this way, and the corresponding hemodynamic parameters of the area of interest, such as blood flow (BF), blood volume (BV), mean transit time (MTT), etc., are determined using the residue impulse function. In certain specific embodiments, one or more neural networks are trained using synthetic data known for real data for hemodynamic parameter evaluation. In certain specific embodiments, the synthetic data can be partially based on or derived from clinical image data where the real data is unknown or unavailable.

[0066] This written description uses examples to disclose the present invention, including the best mode, and also enables any person skilled in the art to practice the present invention, including making and using any device or system and performing any included method. However, it should be understood that this disclosure is not intended to be limited to the specific forms disclosed. Instead, this disclosure is intended to cover all modifications, equivalents, and alternatives falling within the spirit and scope of this disclosure as defined by the appended claims. The patent scope of the present invention is defined by the claims and may include other examples that occur to those skilled in the art. If such other examples have structural elements that are not different from the literal language of the claims, or if they include equivalent structural elements with minor differences from the literal language of the claims, then such other examples are intended to fall within the scope of the claims.

[0067] Referring to the technology presented herein and protected by the claims and applying it to physical objects and specific examples having practical natures, the practical natures clearly improve the current technical field and are thus not abstract, intangible, or purely theoretical. Additionally, if any claim appended to the end of this specification contains one or more elements designated as "means for [performing]... function" or "steps for [performing]... function", then such elements are intended to be interpreted in accordance with 35 U.S.C. 112(f). However, for any claim containing elements designated in any other way, such elements are not intended to be interpreted in accordance with 35 U.S.C. 112(f).

Claims

1. A method, the method comprising: Acquiring a set of perfusion data of a region of interest using an imaging system; Obtaining an arterial signal from the set of perfusion data; Obtaining a tissue signal from the set of perfusion data; And Providing the arterial signal and the tissue signal as inputs to one or more neural networks to determine one or more hemodynamic parameters of the region of interest, wherein the one or more neural networks are trained using one or more clinical perfusion data and one or more synthetic data.

2. The method according to claim 1, wherein the set of perfusion data includes at least one of computed tomography (CT) perfusion data, magnetic resonance imaging (MRI) perfusion data, positron emission tomography (PET) perfusion data, single photon emission computed tomography (SPECT) data, or ultrasound imaging data.

3. The method according to claim 1, wherein the one or more synthetic data are generated based on a defined ground truth model.

4. The method according to claim 1, wherein the tissue signal is a convolution of the arterial signal and a residual impulse function of the region of interest, and wherein the one or more hemodynamic parameters are determined from the residual impulse function.

5. The method according to claim 1, the method comprising correcting non-ideal factors in the set of perfusion data based on an output from the one or more neural networks.

6. The method according to claim 1, wherein the one or more hemodynamic parameters include at least one of blood flow (BF), blood volume (BV), mean transit time (MTT), or time to peak (TMAX).

7. A system, the system comprising: One or more processors; And A memory accessible by the one or more processors, the memory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the method according to any one of claims 1 to 6.

8. A method for training one or more neural networks, the method comprising: Generating a set of synthetic residual impulse functions of a region of interest based on a defined ground truth model; Obtaining an arterial signal from a set of perfusion data; Generating a synthetic tissue signal based on the set of synthetic residual impulse functions and the arterial signal; And Training the one or more neural networks using signals generated using the synthetic tissue signal and the arterial signal.

9. The method according to claim 8, wherein the synthetic tissue signal includes perturbations associated with perturbations of the perfusion data.

10. The method according to claim 9, wherein the perturbations are associated with registration errors or acquisition errors.

11. The method according to claim 8, wherein the loss is used as a bias for the training of the one or more neural networks, and wherein the loss is determined based on a comparison of an estimated residual impulse function output from the one or more neural networks and a first set of parameters derived from the estimated residual impulse function with a combined set of synthetic residual impulse functions and a second set of parameters derived from the combined set of synthetic residual impulse functions.

12. The method according to claim 11, wherein the first set of parameters includes a first set of hemodynamic parameters, and the second set of parameters includes a second set of hemodynamic parameters.

13. The method according to claim 8, wherein regularization is used in the bias for the training of the one or more neural networks, and wherein the regularization is associated with a characteristic of the estimated residual impulse function output from the one or more neural networks.

14. The method according to claim 13, wherein the regularization includes taking a second derivative of the estimated residual impulse function.

Images