System and method for limited angle reconstruction with network determination

Abstract

A system and method include training an artificial neural network to generate an output three-dimensional image volume based on an input two-dimensional projection image, the training based on a plurality of subsets of two-dimensional projection images of each of a plurality of sets of two-dimensional projection images and an associated three-dimensional image volume of three-dimensional image volumes reconstructed from each of the plurality of sets of two-dimensional projection images.

Description

Background Technology

[0001] Conventional medical images can be generated via transmission tomography (TCT) or emission tomography (EPT). In TCT, the imaging source (e.g., an X-ray source) is outside the subject, and the source radiation (e.g., X-rays) is transmitted through the subject to the detector. In EPT, the imaging source (e.g., a radiopharmaceutical emitting gamma rays) is inside the subject (e.g., due to injection or uptake), and the source radiation (e.g., gamma rays) is emitted from within the subject to the detector. In either case, absorption or scattering within the subject's tissues attenuates the source radiation before it is received by the detector.

[0002] In some applications, emission imaging systems are unable to acquire the complete set of tomographic information (e.g., complete rotation) within the time required to adequately image the biological / physiological process of interest. Current practices in these applications involve performing planar imaging over complete rotation around the subject at a rate faster than sufficient data can be acquired at each location. Such images lack spatial and locational information related to tracer uptake and may therefore lead to inaccurate quantitative measurements and / or artifacts.

[0003] Compared to transmission tomography, emission tomography typically exhibits lower resolution, larger artifacts, and more pronounced partial volumetric effects. Current techniques for improving tomographic reconstruction of emission data utilize supplementary data obtained using other imaging modalities, such as computed tomography (CT) and magnetic resonance imaging (MR). Supplementary data can be obtained by segmenting CT / MR image data to, for example, identify tissue location and characteristics. This information can facilitate correction for resolution, partial volumetric effects, and attenuation. Attached Figure Description

[0004] Figure 1 This is a block diagram of a system for deploying a trained artificial neural network to generate image volumes from an incomplete set of projected images, according to some embodiments.

[0005] Figure 2 This is a block diagram of a system for deploying a trained artificial neural network to generate image volumes from an incomplete set of projected images and CT volumes, according to some embodiments.

[0006] Figure 3 This is a block diagram of a system for training an artificial neural network to generate image volumes from an incomplete set of projected images, according to some embodiments.

[0007] Figure 4 This is a flowchart illustrating the process of training an artificial neural network to generate image volumes from an incomplete set of projected images, according to some embodiments.

[0008] Figure 5 This is a block diagram of an artificial neural network training architecture according to some embodiments;

[0009] Figure 6 This is a block diagram of an artificial neural network training architecture according to some embodiments;

[0010] Figure 7 This is a block diagram of a system for deploying a trained artificial neural network to generate acquisition parameters from a set of projected images, according to some embodiments.

[0011] Figure 8 This is a block diagram of a system for training an artificial neural network to generate acquisition parameters, according to some embodiments;

[0012] Figure 9 This is a flowchart of a process for training an artificial neural network to generate acquisition parameters, according to some embodiments;

[0013] Figure 10 This is a block diagram of a system for deploying a trained artificial neural network to generate acquisition parameters from radiomics features, according to some embodiments.

[0014] Figure 11 This is a block diagram of a system for training an artificial neural network to generate acquisition parameters from radiomics features, according to some embodiments;

[0015] Figure 12 This is a flowchart of a process for training an artificial neural network to generate acquisition parameters from radiomics features, according to some embodiments;

[0016] Figure 13 This is a block diagram of a computational system used to train an artificial neural network to generate image volumes from an incomplete set of projected images; and

[0017] Figure 14 The illustration depicts a dual transmission and emission imaging SPECT / CT system with a trained neural network deployed according to some embodiments. Detailed Implementation

[0018] The following description is provided to enable any person skilled in the art to make and use the described embodiments, and illustrates the best mode contemplated for carrying out the described embodiments. However, various modifications will remain apparent to those skilled in the art.

[0019] Some embodiments provide for generating image volumes from an incomplete set of projected images. For example, embodiments may generate high-resolution reconstructed volumes based on a set of PET images obtained from a finite set of projection angles. The generation of the volume can be notified by CT scans acquired simultaneously with the projected images.

[0020] Figure 1This is a block diagram of a deployed system according to some embodiments. System 100 includes a trained network 110. The training of network 110 according to some embodiments will be described below. Although depicted as a neural network, network 110 may include any type of processing system to implement the functions generated by the training described below. For example, network 110 may include a software application programmed to implement the functions generated via prior neural network training.

[0021] During operation, the projected image 1-j is acquired via a first imaging modality. For example, the projected image 1-j may be acquired by a PET or SPECT scanner after a radioactive tracer has been injected into a subject volume (e.g., a patient or phantom). As is known in the art, the projected image 1-j may be acquired at several different projection angles.

[0022] According to some embodiments, the projected image 1-j is "incomplete" because the projection angle represented by the projected image is insufficient to generate a satisfactory reconstructed image. For example, the projection angle of the projected image 1-j can be defined as an arc less than 180 degrees.

[0023] The trained network 110 outputs a quantitatively reconstructed volume based on the input image. According to some embodiments, the trained artificial neural network 110 implements a function. This function can be characterized as a set of trained parameter values ​​associated with the various layers of the network nodes. As is known in the art, this function can be deployed on any computing device.

[0024] According to some embodiments, network 110 receives a 3D volume reconstructed from an incomplete projection image 1-j and generates a reconstructed volume from it. In such embodiments, and as described below, network 110 is trained using the 3D volume reconstructed from the incomplete projection image 1-j.

[0025] Figure 2 The diagram illustrates a deployment system 200 according to some embodiments. As described above with respect to system 100, an incomplete set of projected images 1-j is input into a training network 210. The input to network 210 also represents the CT volume of the same subject as imaged by projected images 1-j. The CT volume can be acquired simultaneously with the projected images 1-j to reduce registration errors between them.

[0026] The trained network 210 generates a reconstructed volume based on an incomplete set of projected images 1-j and the CT volume. The CT volume can improve the quality of the output reconstructed volume by providing attenuation information, which is... Figure 1 The deployment shown in the diagram does not exist. Other structural imaging modalities, such as, but not limited to, MR and PET, can replace CT to provide attenuation information.

[0027] Figure 3 The illustration depicts an architecture 300, according to some embodiments, for training network 110 to generate volumes based on an incomplete set of projected images. Network 110 may be temporally and / or geographically remote. Figure 1 and Figure 2 The deployment described in the diagram is used for training. For example, architecture 300 can be implemented in a data processing facility, while system 100 or 200 can be executed in an imaging room (theater) where the patient has just been imaged.

[0028] The training system 310 uses Q projected images. 1-k Collection, and in some embodiments, using CT volume. 1-Q To train the artificial neural network 110. The basic truth determination unit 320 also uses Q projected images. 1-k Collection and CT volume 1-Q This is used to generate baseline ground truth data for evaluating the performance of network 110 during training of training system 310. For example, for Q projected images... 1-k For set X in the set, training system 310 generates subsets 1-j and inputs these subsets (and, in some embodiments, the CT volumes corresponding to set X) into network 110 to generate volumes based on these subsets. Then, using quantitative and iterative reconstruction methods, this volume is compared with the volume generated by unit 320 based on the projected image. 1-k The complete set of X (and based on CT volume) X The generated underlying truth volumes are compared. This process is repeated for other subsets 1-j of set X, and also for Q projected images. 1-k This process is repeated for each other set in the set. Network 110 is modified based on this comparison, and the entire process is repeated until satisfactory network performance is achieved.

[0029] The artificial neural network 110 may include any type of network that can be trained to approximate a function. In some embodiments, the network 110 includes an implementation of a "u-net" convolutional network architecture as known in the art.

[0030] Generally, an artificial neural network 110 may include a network of neurons that receive input, change their internal state according to the input, and produce an output depending on the input and the internal state. The outputs of some neurons are connected to the inputs of other neurons to form a directed and weighted graph. The weights of the internal states and the function for computation can be modified through a training process based on underlying ground truth data. The artificial neural network 110 may include any one or more types of known or becoming known artificial neural networks, including but not limited to convolutional neural networks, recurrent neural networks, long short-term memory networks, deep reservoir computation and deep echo state networks, deep belief networks, and deep stacked networks.

[0031] According to some embodiments, the trained artificial neural network 110 implements a function of its input. This function can be characterized as a set of parameter values ​​associated with each network node. As is known in the art, this function can be deployed to external systems, such as… Figure 1 System 100. In one example, the kernel trained as a fully convolutional network generates parameter values. Another fully convolutional network including such parameterized kernels can be efficiently incorporated into a system such as System 100 to generate high-resolution volumes as described in this paper.

[0032] The training system 310 may include any (one or more) systems known or becoming known for training artificial neural networks. For example, the training system 310 may employ supervised learning, unsupervised learning, and / or reinforcement learning.

[0033] Q projected images 1-k Set and corresponding CT volume 1-Q It can represent many different patients, phantoms, or other subjects. Furthermore, Q projection images can be acquired at different locations using different contrast settings. 1-k Set and corresponding CT volume 1-Q Each of them. In general, training network 110 can be used to generate volumes based on input data of any modality, as long as those modalities are well represented in the training dataset.

[0034] Figure 4 This is a flowchart of a network training process according to some embodiments. Process 400 and other processes described herein can be performed using any suitable combination of hardware and software. The software program code embodying these processes can be stored by any non-transitory tangible medium, including but not limited to fixed disks, volatile or non-volatile random access memory, DVDs, flash drives, or magnetic tapes. Embodiments are not limited to the examples described below.

[0035] Initially, at S410, multiple two-dimensional projection datasets are acquired. The projection images can be acquired via nuclear imaging scans and / or any other imaging modality known or becoming known. Optionally, at S420, a three-dimensional CT volume associated with each two-dimensional projection dataset is acquired. According to some embodiments, as known in the art, each CT volume and its associated projection dataset are acquired substantially simultaneously. S410 and S420 can simply include access to a large repository of previously acquired imaging data.

[0036] Based on the data acquired at S410 and S420, an artificial neural network is trained at S430. The artificial neural network is trained to generate reconstructed 3D volumes based on multiple 2D projection datasets and optionally on corresponding 3D CT volumes. In some embodiments, network training involves determining a loss based on the network's output and iteratively modifying the network based on that loss until the loss reaches an acceptable level or training terminates in other ways (e.g., due to time constraints or because the loss asymptotically approaches a lower bound). Network training at S430 can occur long after the acquisition of training data and separately from the acquisition of training data. For example, training data can be acquired and accumulated in an image repository several months or years before performing S430.

[0037] Figure 5 The illustration shows training performed by training architecture 500 at S430 according to some embodiments. During training, reconstruction component 520 reconstructs Q projected images acquired at S410. 1-k Each element in the set generates a fundamental truth volume. Reconstruction component 520 can process the projected image. 1-k Reconstruction is performed using conjugate gradient, attenuation and scattering (CGAS), filtered back projection (FBP) reconstruction, or any other suitable technique. A subset of each of the Q sets is input to network 510, and in response, network 510 outputs a volume corresponding to each subset. According to some embodiments, instead of the input to each subset, or in addition to the input to each subset, the 3D volume can be reconstructed from each subset and input to network 510.

[0038] Loss layer component 530 determines the loss by comparing each output volume with its corresponding ground truth volume. More specifically, it compares the output volume based on a specific subset of the projected image set with the volume reconstructed by component 520 based on the same projected image set. During training at S430, any number of subsets of the specific projected image set can be used.

[0039] The total loss is backpropagated from the loss layer component 530 to the network 510. This loss may include L1 loss and L2 loss, or any other suitable measure of the total loss. The L1 loss is the sum of the absolute differences between each output volume and its corresponding ground truth volume, and the L2 loss is the sum of the squared differences between each output volume and its corresponding ground truth volume.

[0040] Network 510 modifies its internal weights or kernel parameter values ​​based on backpropagation loss as known in the art. Network 510 and loss layer 530 reprocess the training data as described above, and this process is repeated until it is determined that the loss has reached an acceptable level or training terminates in other ways. At termination, network 510 can be considered trained. In some embodiments, the function implemented by the now trained network 510 (e.g., embodied in the parameter values ​​of the trained convolutional kernel) can then be... Figure 1 Deployed as shown in the diagram.

[0041] Figure 6 The illustration shows training at S430 according to some embodiments. Training architecture 600 can be used to train network 610 for deployment in, for example, Figure 2 In architectures like architecture 200, similar to training architecture 500, the reconstruction component 620 is based on Q projected images acquired at S410. 1-k Each in the set generates the fundamental truth volume. However, for Q projected images... 1-k For each element in the set, component 620 also uses the corresponding CT volume q to generate an associated underlying ground truth volume. Structural image volumes other than CT can be used, including but not limited to MR and PET. According to some embodiments, component 620 segments and registers the corresponding CT volume q as known in the art, and based on this and the corresponding projected image... 1-k The set performs multimodal reconstruction to generate each underlying truth volume q.

[0042] During training, a subset of each of the Q sets is input to network 610, and network 610 outputs a volume corresponding to each subset. Any number of subsets of a particular set of projected images can be used during training, including any number of projected images. Again, the 3D volume can be reconstructed from each subset and input to network 610, either as an alternative to or in addition to the input of each subset.

[0043] The loss layer component 630 can determine the loss by comparing each output volume with the corresponding ground truth volume as described above, and modify the network 610 until it is determined that the loss has reached an acceptable level or training terminates in another manner. For example, as Figure 2 The functions shown, implemented by the now-trained network 160, can then be deployed.

[0044] The reconstruction component 420, the segmentation / reconstruction component 620, and each functional component described herein can be implemented in computer hardware, program code, and / or in one or more computing systems that execute program code as known in the art. Such a computing system may include one or more processing units that execute processor-executable program code stored in a memory system. Furthermore, networks 510 and 610 may include hardware and software specifically designed for executing algorithms based on a specified network architecture and trained kernel parameters.

[0045] Figure 7 This is a block diagram of a system 700 that deploys a trained artificial neural network 710 to generate image acquisition parameters 720 from a set of projected images, according to some embodiments. The training of the network 710 according to some embodiments will be described below. Although depicted as a categorized neural network, the network 710 may include any type of processing system to implement a learning function and output one or more image acquisition parameters (e.g., probabilities associated with each of the one or more acquisition parameters). For example, the network 710 may include a software application programmed to implement the function generated via previous neural network training.

[0046] During operation, projected images are acquired via a suitable imaging modality. 1-k Collection. For example, after a radioactive tracer is injected into the subject's volume, projection images can be acquired using a PET or SPECT scanner. 1-k Projected image 1-k It may include CT images as known in the art.

[0047] Reconstruction component 730 uses projected images 1-k The collection is used to reconstruct the volume, as is known in the art. The reconstruction technique applied to the reconstruction component 730 may depend on the method used to acquire the projected image. 1-k The type of modality of the set. Implementations of the reconstruction component 730 can employ any suitable reconstruction algorithm.

[0048] The trained network 710 receives the reconstructed volume and outputs an indication of acquisition parameters 720. Acquisition parameters 720 may include resolving the projected image. 1-k The defects in the reconstruction volume are thus identified, resulting in higher quality acquisition parameters for the reconstructed volume. Therefore, in some embodiments, the trained network 710 models the correlation between the undesirable characteristics of the reconstructed volume and the parameters used to acquire the projected image, which can be used to reconstruct a reduced volume exhibiting the undesirable characteristics.

[0049] Therefore, in some examples, the first projected image is obtained. 1-kThe first volume is reconstructed from the set of images obtained by the network 710 and the network 710 outputs acquisition parameters 720. A second set of projected images is obtained based on the output acquisition parameters 720, and a second volume is reconstructed from the second set of projected images. The second volume exhibits improved characteristics compared to the first volume.

[0050] The improvements depend on the data used to train the network 710. Furthermore, the "improvement" of a feature is relative to the desired use of the resulting reconstructed volume. For example, a high level of a particular image feature might be desirable for a type of diagnostic review, while a low level of a particular image feature might be desirable for a treatment plan. In the former case, "improving" a particular image feature includes increasing the level of the image feature, and in the latter case, improving the particular image feature by decreasing its level.

[0051] Figure 8 This is a block diagram of a system 800 for training an artificial neural network 710 to generate acquisition parameters, according to some embodiments. The training system 810 uses the reconstructed volume. 1-Q and remedial acquisition parameters 1-Q To train the artificial neural network 710. Each reconstructed volume 1-Q With a corresponding remedial acquisition parameter 1-Q Related.

[0052] In some embodiments, the remedial acquisition parameters associated with a given training volume are parameters for acquiring projected images that can address deficiencies in the given training volume. These parameters can be defined by humans when reviewing the given training volume. The remedial acquisition parameters can include any parameters related to the imaging modality used to acquire known or becoming known projected images. For example, in the case of SPECT imaging, remedial acquisition parameters 820 can include the duration of each projection, the number of frames per projection, the number of projections per scan, and the size of the acquisition matrix. In the case of CT imaging, remedial acquisition parameters 820 can include X-ray beam energy, X-ray tube current, integration time, the number of frames per projection, and acquisition time.

[0053] During the training of System 710, the reconstructed volume 1-Q The data is input into the training system 810, and the training system 810 outputs the reconstructed volume. 1-Q The network 710 is a set of acquisition parameters for each of the several possible acquisition parameters. As mentioned above, the output set of acquisition parameters may include a set of probabilities for each of the several possible acquisition parameters. The training system 810 compares each output set of the acquisition parameters with the corresponding set of remedial acquisition parameters stored in parameter 820. The total loss is calculated, and the network 710 is modified based on this loss. The training process is repeated until satisfactory network performance is achieved.

[0054] Figure 9 This is a flowchart of a process 900 for training an artificial neural network to generate acquisition parameters according to some embodiments. At S910, multiple reconstructed volumes of imaging data are acquired. As is known in the art, volumes can be reconstructed from a projected image. S910 may include access to a storage library of the reconstructed 3D image data.

[0055] At S920, remedial acquisition parameters associated with each reconstructed volume are determined. The remedial acquisition parameters associated with a given volume are projection image acquisition parameters, which, if used to acquire projection images for subsequent reconstruction, can resolve defects in the given volume. Radiologists can review each reconstructed volume to determine the remedial acquisition parameters associated with it.

[0056] Based on the acquired volume and determined remedial acquisition parameters, the artificial neural network is trained at S930. In some embodiments, network training involves determining a loss based on the network's output and iteratively modifying the network based on that loss until the loss reaches an acceptable level or training terminates in other ways (e.g., due to time constraints or because the loss asymptotically approaches a lower bound). Network training at S930 can occur long after and separately from the acquisition of training data at S910 and S920.

[0057] Figure 10 This is a block diagram of a system 1000 that deploys a trained artificial neural network 1010 to generate acquisition parameters 1020 from radiomics features 1030 according to some embodiments. The network 1010 may include any type of processing system to implement a learning function and output image acquisition parameters 1020 (e.g., probabilities associated with each of one or more acquisition parameters).

[0058] Radiomics refers to the extraction of features from radiographic medical images. This extraction is based on programmed and / or learned algorithms, and the features can provide insights into diagnosis, prognosis, and treatment response that may be incomprehensible to the naked eye.

[0059] The radiomics features 1030 of system 1000 can be obtained in any way that is known or becomes known. According to some embodiments, the radiomics features 1030 may include features based on size and shape, descriptors of image intensity histograms, textures derived from descriptors of relationships between image voxels (e.g., gray-level co-occurrence matrix (GLCM), run-length matrix (RLM), size-zone matrix (SZM), and neighborhood gray-level tonal difference matrix (NGTDM)), textures extracted from filtered images, and fractal features.

[0060] In operation, radiomics features 1030 are acquired based on one or more images of the subject. In some embodiments, a set of projected images is acquired and a volume is reconstructed from it, and radiomics features 1030 are extracted from that volume. A trained network 1010 receives the radiomics features and outputs acquisition parameters 1020. The output acquisition parameters 1020 may include acquisition parameters that address defects in the images(s) from which the radiomics features 1030 are extracted.

[0061] Therefore, in some embodiments, a first set of projected images is acquired, and a volume is reconstructed from them. This volume is input into network 1010, and network 1010 outputs acquisition parameters 1020. A second set of projected images is then acquired based on the output acquisition parameters 1020, and a second volume is reconstructed from the second set of projected images. Because the second set of projected images is acquired using the output acquisition parameters 1020, the second volume exhibits improved properties relative to the first volume. As mentioned above, the improved properties and the manner in which they are improved depend on the data used to train network 1010.

[0062] Figure 11 This is a block diagram of a system 1100 for training an artificial neural network 1010 to generate acquisition parameters, according to some embodiments. The training system 1110 uses radiomics features. 1-Q 1130 and remedial acquisition parameters 1-Q A set of 1140 is used to train the artificial neural network 1010. As is known or becomes known, the radiomics feature extraction component 1120 extracts features from a corresponding reconstructed volume. 1-Q Extracting radiomics features 1-Q Each set in set 1130. Radiomics characteristics. 1-Q Each set of 1130 is associated with a corresponding remedial acquisition parameter. 1-Q 1140 is associated with a corresponding remedial acquisition parameter. 1-Q 1140 corresponds to the reconstructed volume from which the set of radiomics features is extracted. In other words, the remedial acquisition parameters associated with a given set of radiomics features are parameters used for projected image acquisition that can address deficiencies in the image volume from which the given set of radiomics features is extracted. Remedial acquisition parameters 1-Q 1140 may include any parameter described herein or otherwise known.

[0063] Figure 12This is a flowchart of a process 1200 for training an artificial neural network to generate acquisition parameters according to some embodiments. Process 1200 will be described below with respect to system 1100, but embodiments are not limited thereto. At S1210, multiple reconstructed volumes of imaging data are acquired. As is known in the art, volumes can be reconstructed from projected images. In some embodiments, S1210 includes accessing a storage library of reconstructed three-dimensional image data.

[0064] Next, at S1220, multidimensional radiomics features are determined for each of the multiple reconstructed volumes of the image data. For example, radiomics feature extraction component 1120 can extract features from the reconstructed volumes as known or become known at S1220. 1-Q The corresponding reconstructed volume 1-Q Extracting radiomics features 1-Q 1130.

[0065] At S1230, remedial acquisition parameters associated with each reconstructed volume are determined. These remedial acquisition parameters can be determined via human review. For example, a radiologist can review each reconstructed volume to determine acquisition parameters that can remedy defects in the volume if used to acquire projection images for subsequent reconstruction.

[0066] Based on multidimensional radiomics features and determined salvage acquisition parameters, an artificial neural network is trained at S1240. In some embodiments, training the network involves inputting multidimensional radiomics features into a training system 1110, which outputs radiomics features. 1-Q Each set contains a set of acquisition parameters. The output set of acquisition parameters may include a set of probabilities for each of the several possible acquisition parameters. The training system 1110 compares each output set of the acquisition parameters with the corresponding set of remedial acquisition parameters stored in parameter 1140. The total loss is calculated, and the network 1010 is modified based on this loss. The training process is repeated until satisfactory network performance is achieved.

[0067] Figure 13 This is a block diagram of a computational system, according to some embodiments, for training an artificial neural network to generate image volumes from an incomplete set of projected images. System 1300 may include a computational system to facilitate the design and training of artificial neural networks as known in the art. Computational system 1300 may include a standalone system, or one or more elements of computational system 1300 may be located in the cloud.

[0068] System 1300 includes a communication interface 1310 for communicating with external devices via, for example, a network connection. One or more processing units 1320 may include one or more processors, processor cores, or other processing units to perform processor-executable process steps. In this regard, a storage system 1330 may include one or more memory devices (e.g., hard disk drives, solid-state drives) to store processor-executable process steps of a training program 1331, which can be executed by one or more processing units 1330 to train a network as described herein.

[0069] Training program 1331 may utilize node operator library 1332, which includes code that performs various operations associated with node operations. According to some embodiments, computing system 1300 provides interfaces and development software (not shown) to enable the development of training program 1331 and the generation of network definition 1335, which specifies the architecture of the neural network to be trained. Storage device 1330 may also include program code 1333 for reconstruction component 520 and / or segmentation / reconstruction component 620.

[0070] Data used to train the network can also be stored in storage device 1330, including but not limited to data such as those related to... Figure 5 The projection data 1334 is described. Once trained, the parameters of the neural network can be stored as trained network parameters 1336. As mentioned above, these trained parameters can be deployed in other systems known in the art to provide trained functions.

[0071] Figure 14 The illustration shows a SPECT-CT system 1400, which can be deployed with a trained network to generate high-resolution volumes based on CT data and lower-resolution nuclear imaging data as described in this paper.

[0072] System 1400 includes a stand 1402, to which two or more gamma cameras 1404a, 1404b are attached, although any number of gamma cameras may be used. Detectors within each gamma camera detect gamma photons (i.e., emission data) 1403 emitted by a radioactive isotope in the body of a patient 1406 lying on a bed 1408.

[0073] The bed 1408 is slidable along the axis of motion A. At the corresponding bed position (i.e., imaging position), a portion of the patient's body 1406 is positioned between gamma cameras 1404a and 1404b to capture emission data 1403 from that body portion. The gamma cameras 1404a and 1404b may include multifocal cone-beam collimators or parallel-aperture collimators as known in the art.

[0074] System 1400 also includes a CT housing 1410, which includes an X-ray imaging system (not shown) as known in the art. Generally, and according to some embodiments, the X-ray imaging system acquires two-dimensional X-ray images of the patient 1406 before, during, and / or after acquiring emission data using gamma cameras 1404a and 1404b.

[0075] The control system 1420 may include any general-purpose or special-purpose computing system. Therefore, the control system 1420 includes one or more processing units 1422 and a storage device 1430 for storing program code, said one or more processing units 1422 being configured to execute processor-executable program code to cause the system 1420 to operate as described herein. The storage device 1430 may include one or more fixed disks, solid-state random access memory, and / or removable media (e.g., thumb drives) mounted in a corresponding interface (e.g., a USB port).

[0076] Storage device 1430 stores the program code of system control program 1431. One or more processing units 1422 can execute system control program 1431 to control motors, servo systems, and encoders in conjunction with SPECT system interface 1440, thereby causing gamma cameras 1404a, 1404b to rotate along frame 1402 and acquire two-dimensional emission data 1403 at defined imaging positions during rotation. Acquired data 1432 may include projected images as described herein and may be stored in memory 1430. Reconstructed volume 1434 as described herein may be stored in memory 1430.

[0077] One or more processing units 1422 may also execute system control program 1431 to, in conjunction with CT system interface 1445, cause radiation sources within CT housing 1410 to emit radiation toward body 1406 from different projection angles, thereby controlling corresponding detectors to acquire two-dimensional CT images and reconstruct three-dimensional CT images based on the acquired images. As described above, CT images can be acquired substantially simultaneously with the emission data, and the volume reconstructed from the reconstructed images can be stored as CT data 1433.

[0078] The trained network parameters 1435 may include parameters of a neural network trained as described herein. For example, the projected image of the emission data 1432 and, optionally, the corresponding CT volume, may be input into the network implementing the trained network parameters 1435 to generate remedial acquisition parameters as described above.

[0079] Terminal 1450 may include a display device and an input device coupled to system 1420. Terminal 1450 may display any projected image, reconstructed volume, and remedial acquisition parameters, and may receive user input for controlling the display of data, operation of imaging system 1400, and / or the processing described herein. In some embodiments, terminal 1450 is a separate computing device, such as, but not limited to, a desktop computer, laptop computer, tablet computer, and smartphone.

[0080] Each component of system 1400 may include other elements necessary for its operation, as well as additional elements for providing functions beyond those described herein.

[0081] Those skilled in the art will appreciate that various adaptations and modifications to the above embodiments can be configured without departing from the claims. Therefore, it should be understood that the claims can be practiced in ways different from those specifically described herein.

Claims

1. A system for generating image volume, comprising: Storage devices; A processor for executing processor-executable process steps stored on a storage device to cause the system to: determine a set of multiple two-dimensional projected images; Reconstruct the volume of a 3D image based on each of multiple sets of 2D projected images; For each of a plurality of sets of two-dimensional projected images, a plurality of subsets of two-dimensional projected images are determined, wherein a plurality of subsets of two-dimensional projected images determined in one of the plurality of sets of two-dimensional projected images represents fewer projection angles than a two-dimensional projected image in one of the plurality of sets. as well as An artificial neural network is trained to generate an output three-dimensional image volume based on an input two-dimensional projected image. The training is based on multiple subsets of two-dimensional projected images from each of a plurality of sets of two-dimensional projected images and associated three-dimensional image volumes from the three-dimensional image volumes reconstructed from each of the plurality of sets of two-dimensional projected images.

2. The system of claim 1, wherein the artificial neural network is a convolutional network, and wherein the processor is to perform processor-executable process steps to cause the system to perform the following operations: The trained kernels of the trained network are output to the imaging system.

3. The system of claim 2, further comprising an imaging system, the imaging system being used for: Obtain a two-dimensional emission data set; The two-dimensional emission data set is input into a second convolutional network comprising trained kernels; and Store the first simulated reconstructed 3D volume generated by the second convolutional network based on the first and second 3D volumes of the input.

4. The system of claim 1, wherein the processor is configured to execute processor-executable process steps stored on a storage device to cause the system to perform the following operations: Determine the structure image volume associated with each of the multiple sets of two-dimensional projected images. The reconstruction of the 3D image volume based on one of the multiple sets of 2D projected images is based on the structural image volume associated with one of the multiple sets of 2D projected images and is based on one of the multiple sets of 2D projected images.

5. The system of claim 4, wherein the reconstruction of the three-dimensional image volume includes segmentation of the structural image volume.

6. A method for generating image volume, comprising: Determine a set of multiple two-dimensional projected images; Reconstruct the volume of a 3D image based on each of multiple sets of 2D projected images; For each of a plurality of sets of two-dimensional projected images, a plurality of subsets of two-dimensional projected images are determined, wherein a plurality of subsets of two-dimensional projected images determined in one of the plurality of sets of two-dimensional projected images represents fewer projection angles than a two-dimensional projected image in one of the plurality of sets. as well as An artificial neural network is trained to generate an output three-dimensional image volume based on an input two-dimensional projected image, wherein the training is based on multiple subsets of two-dimensional projected images from each of the plurality of sets of two-dimensional projected images and associated three-dimensional image volumes from the three-dimensional image volumes reconstructed from each of the plurality of sets of two-dimensional projected images.

7. The method of claim 6, wherein the artificial neural network is a convolutional network, and further comprises: The trained kernels of the trained network are output to the imaging system.

8. The method of claim 7, further comprising operating the imaging system to: Obtain the two-dimensional launch dataset; The two-dimensional emission dataset is input into a second convolutional network that includes trained kernels; and Store the first simulated reconstructed 3D volume generated by the second convolutional network based on the first and second 3D volumes of the input.

9. The method of claim 6, further comprising: Determine the structure image volume associated with each of the multiple sets of two-dimensional projected images. The reconstructed 3D image volume is based on one of a plurality of 2D projected image sets, and is based on a structural image volume associated with one of the plurality of 2D projected image sets.

10. The method of claim 9, wherein reconstructing the three-dimensional image volume includes segmenting the structured image volume.

11. A system for generating image volume, comprising: Storage devices, storage: A collection of multiple two-dimensional projected images; The volume of a three-dimensional image associated with each of a set of multiple two-dimensional projected images; and Nodes in an artificial neural network; as well as A processor, used to execute processor-executable process steps stored on a storage device, to cause the system to perform the following operations: For each of a plurality of sets of two-dimensional projected images, a plurality of subsets of two-dimensional projected images are determined, wherein a plurality of subsets of two-dimensional projected images determined in one of the plurality of sets of two-dimensional projected images represents fewer projection angles than a two-dimensional projected image in one of the plurality of sets. and An artificial neural network is trained to generate an output three-dimensional image volume based on an input two-dimensional projected image, wherein the training is based on multiple subsets of two-dimensional projected images from each of the plurality of sets of two-dimensional projected images and associated three-dimensional image volumes of three-dimensional image volumes reconstructed from each of the plurality of sets of two-dimensional projected images.

12. The system of claim 11, wherein the artificial neural network is a convolutional network, and wherein the processor is to perform processor-executable process steps to cause the system to perform the following operations: The trained kernels of the trained network are output to the imaging system.

13. The system of claim 12, further comprising an imaging system, the imaging system being used for: Obtain the two-dimensional launch dataset; The two-dimensional emission dataset is input into a second convolutional network that includes trained kernels; and Store the first simulated reconstructed 3D volume generated by the second convolutional network based on the first and second 3D volumes of the input.

14. The system of claim 11, wherein the storage device stores a structural image volume associated with each of a plurality of two-dimensional projected image sets, and The reconstructed 3D image volume is based on one of a plurality of 2D projected image sets, and is based on a structural image volume associated with one of the plurality of 2D projected image sets.

Images