High resolution PET-CT imaging using PCCT and advanced generative model

Abstract

The invention relates to high resolution PET-CT imaging using PCCT and an advanced generative model. Systems and methods for generating a high resolution photon counting CT (PCCT) and PET image using a pair of high resolution PET images from the same patient. A machine learning network is trained on pairs of images from a patient. When the trained model is applied, new patient's PET and PCCT images may be used to generate high resolution PET images for medical diagnosis or further processing.

Description

[0001] Cross-reference to related applications

[0002] This application claims the benefit of U.S. Provisional Patent Application Serial No. 63 / 679674, filed August 6, 2024, with a filing date under 35 U.S. SC § 119(e), which is incorporated herein by reference. Technical Field

[0003] This disclosure relates to medical imaging. Background Technology

[0004] Positron emission tomography (PET) scans are an imaging test that uses a radioactive substance called a tracer to look for disease within the body. While PET imaging excels at functional imaging of metabolic processes, it inherently suffers from lower spatial resolution compared to computational tomography (CT) imaging. This difference can limit the diagnostic utility of PET / CT images, particularly in detecting small lesions or providing detailed anatomical context for metabolic activity. Higher resolution would certainly translate into several clinical and diagnostic advantages.

[0005] To date, this problem has been addressed through various image reconstruction techniques and hardware improvements. For example, xSPECT technology combines SPECT and CT data, leveraging the higher resolution of CT images to enhance the spatial resolution of SPECT images using conventional reconstruction methods. However, these methods have limitations, including the possibility of introducing artifacts or requirements for specific hardware configurations. Summary of the Invention

[0006] By way of introduction, the preferred embodiments described below include methods, systems, instructions, and / or computer-readable media that use advanced generative models to leverage the high resolution of CT images to enhance the resolution of corresponding PET images from the same patient. The generative models can include, for example, conditional diffusion models, which excel due to their ability to generate high-quality, detailed outputs under the guidance of high-resolution CT / PCCT data. CycleGAN, Pix2PixHD, and similar conditional GANs provide robust paired image-to-image translation, effectively leveraging the complementary nature of PET and CT / PCCT data. Attention-UNet and TransUNet architectures effectively incorporate multi-modal inputs, preserving spatial details of CT / PCCT while enhancing PET features. Meanwhile, multi-modal variational autoencoders (MM-VAE) and fusion models enable probabilistic and structural fusion of PET and CT / PCCT, generating high-resolution PET or infused PET-CT / PCCT outputs. Neural Radiance Fields (NeRF), while less common, can be adapted for fine-detail reconstruction using CT / PCCT priors, highlighting the vast resolution gap between PET and CT / PCCT and the need for such advanced methodologies.

[0007] In a first aspect, a system for generating a high-resolution PET image from a pair of CT and PCCT images, the system comprising: a medical imaging device configured to acquire a PET image of a patient region and a PCCT image of the patient region; a processing unit configured to input the PET image and the PCCT image into a machine learning network trained to generate a high-resolution PET-CT image comprising a higher resolution than the input PET image; and a display configured to display the high-resolution PET-CT image generated from the PET image and the PCCT image.

[0008] In a second aspect, a method for generating a high-resolution PET image, the method comprising: acquiring, by a medical imaging device, a PET image of a patient region; acquiring, by the medical imaging device, a PCCT image of the patient region; inputting the PET image and the PCCT image into a machine learning network trained to output a high-resolution PET image comprising a higher resolution than the input PET image; and outputting the high-resolution PET image of the patient region.

[0009] In a third aspect, a method for generating a fused high-resolution PET image from a pair of CT and PCCT images, the method comprising: acquiring, by a medical imaging device, a PET image of a patient region; acquiring, by the medical imaging device, a PCCT image of the patient region; inputting the PET image and the PCCT image into a machine learning network trained to output a high-resolution PET-CT image comprising a higher resolution than the input PET image; and outputting a high-resolution fused PET-CT image of the patient region.

[0010] Any one or more of the aspects described above can be used alone or in combination. These and other aspects, features, and advantages will become apparent to those of ordinary skill in the art from the following detailed description, taken in conjunction with the accompanying drawings. The application is defined by the claims and no limitation of the scope of those claims shall be inferred from anything herein, including this section. Other aspects and advantages of the present application are discussed in connection with the preferred embodiments, as described below, and drawings attached hereto. Claims may be sought independent or in combination. BRIEF DESCRIPTION OF DRAWINGS

[0011] The components and various diagrams are not necessarily drawn to scale, but emphasis is placed on illustrating the principles of the embodiments. In addition, in the various drawings, like reference numerals designate corresponding parts throughout the several views. The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0012] Figure 1 An example system is depicted that is used to generate a high-resolution PET image using a pair of high-resolution photon-counting CT (PCCT) and PET images from the same patient, in accordance with an embodiment.

[0013] Figure 2 An example PET / PCCT imaging device is depicted that is used to generate a high-resolution PET image using a pair of high-resolution photon-counting CT (PCCT) and PET images from the same patient, in accordance with an embodiment.

[0014] Figure 3 An example machine learning network is depicted.

[0015] Figure 4 An example convolutional neural network (CNN) is depicted.

[0016] Figure 5 An example generative adversarial network (GAN) is depicted, in accordance with an embodiment.

[0017] Figure 6An example CycleGAN according to an embodiment is depicted for generating high resolution PET images using paired high resolution photon counting CT (PCCT) and PET images from the same patient.

[0018] Figure 7A and Figure 7B An example CNN according to an embodiment is depicted for generating high resolution PET images using paired high resolution photon counting CT (PCCT) and PET images from the same patient.

[0019] Figure 8 An example autoencoder network according to an embodiment is depicted for generating high resolution PET images using paired high resolution photon counting CT (PCCT) and PET images from the same patient.

[0020] Figure 9 An example network with a transducer according to an embodiment is depicted for generating high resolution PET images using paired high resolution photon counting CT (PCCT) and PET images from the same patient.

[0021] Figure 10 An example method according to an embodiment is depicted for generating high resolution fused PET / CT images using paired high resolution photon counting CT (PCCT) and PET images from the same patient.

[0022] Figure 11 An example according to an embodiment is depicted for generating high resolution PET images using paired high resolution photon counting CT (PCCT) and PET images from the same patient. DETAILED DESCRIPTION

[0023] Embodiments described herein provide systems and methods for generating high resolution PET images using paired high resolution photon counting CT (PCCT) and PET images from the same patient. Machine learning networks / models are trained on paired images from a patient. When the trained model is applied, PET and PCCT images of a new patient can be used to generate high resolution PET images for medical diagnosis or further processing.

[0024] Positron emission tomography (PET) is a type of nuclear medicine procedure that measures the metabolic activity of cells in body tissues. PET differs from other nuclear medicine examinations in that it detects metabolism within body tissues, while other types of nuclear medicine examinations detect the amount of radioactive material collected in a specific location of body tissue to examine tissue function. PET can also be used in conjunction with other diagnostic tests, such as computational computed tomography (CT) or magnetic resonance imaging (MRI), to provide more definitive information, such as information about tumors and other lesions. Newer technologies combine PET and CT into a single scanner, called a PET / CT scanner, which can perform both scans on a patient during an imaging session. Clinically used PET scanners typically have a spatial resolution of approximately 4–5 mm. This low resolution is caused by detector size, the free path of positrons, and non-collinear uncertainties in annihilation photon pairs. This limitation in resolution significantly restricts the sensitivity of PET in imaging small lesions, such as early-stage cancer or small metastases. Furthermore, clinical PET scans typically take 15–20 minutes. The potential to provide high spatial resolution imaging using various techniques may require extended scan times, which can be inefficient and impractical.

[0025] Computational photon counting computed tomography (PCCT) is an advanced CT imaging technique. Conventional CT equipment uses an energy integral detector (EID) equipped with a scintillator element and a reflective layer. The scintillator element layer converts incident X-ray photons into low-energy secondary photons in the visible spectrum. These photons are then absorbed by an array of photodiodes made of semiconductor material, which generates an electrical signal proportional to the total deposited energy and adds it to electronic thermal noise. Finally, the electrical signal is amplified and then converted into a digital signal, allowing it to be processed for tomographic image reconstruction. In contrast to typical CT imaging, in PCCT, the photon counting detector decomposes the number of photons and the incident X-ray energy spectrum into multiple energy bins. Compared to conventional CT techniques, PCCT offers improved spatial and contrast resolution, reduced image noise and artifacts, reduced radiation exposure, and the advantages of multi-energy / multi-parameter imaging based on the atomic properties of tissues, allowing the use of different contrast agents and improved quantitative imaging.

[0026] The embodiments described herein combine PCCT imaging (which offers excellent resolution (up to 0.11 mm) and sensitivity) with the application of PET imaging. In alternative embodiments, CT images with high resolution can be used. The resulting images improve the resolution of PET images, thereby enhancing disease detection, staging, and treatment monitoring.

[0027] Figure 1A system is described for generating high-resolution PET images using paired high-resolution photon-counting CT (PCCT) 120 and PET images 115 from the same patient. Images are acquired from the same patient to ensure accuracy and alignment. While the same model can be used to generate synthetic high-resolution PET or PET-CT images, these synthetic outputs do not represent real patient data. The system includes a medical imaging device 105, an image processing unit 130, and an operator interface 140. Fewer or more devices may be included or excluded. For example, if real patient images are acquired from other sources, such as medical imaging databases, the medical imaging device 105 may not be used. The image processing unit 130 is configured to generate high-resolution PET images 150 and / or fused images 160 using a machine learning network 110 based on paired PCCT and PET images 115 from patient 225. The medical imaging device 105 can be used to acquire real patient images to train the model. Figure 1 Two examples of generating high-resolution PET images are depicted. The first example depicts the generation of a high-resolution PET image 150 from PET images 115 and PCCT images 120 of patient 225 acquired during a single session. The second example depicts the generation of a fused PET / PCCT image 160 from PET images 115 and PCCT images 120 of patient 225 acquired during a single session. The output image may be part of or included in the xSPECT system. By using CT as a reference frame for image reconstruction, the xSPECT system can extract zonation maps with different tissue segments to indicate the boundaries of nuclear uptake during SPECT reconstruction.

[0028] Operator interface 140 includes input and output devices. Inputs may be interfaces, such as those connecting to a computer network, memory, database, medical image storage device, or other input data sources. Inputs may be user input devices, such as a mouse, trackpad, keyboard, ball, touchpad, touchscreen, or another device for receiving user input. Outputs are display devices, but may also be interfaces. They display raw images, fused images, and / or higher-resolution images from scans. For example, a high-resolution image of a region of patient 225 may be displayed. The display may be a CRT, LCD, plasma, projector, printer, or other display device. The display is configured by loading images onto a display plane or buffer. The display is configured to display images of a region of patient 225. The operator interface may include a graphical user interface (GUI) that enables users to interact with medical imaging equipment 105 and allows users to make modifications or selections substantially in real time.

[0029] The system includes a medical imaging device 105 (PET / PCCT imaging device) configured with both PET and PCCT positron emission tomography / CT (PET / CT) units, allowing the medical imaging device 105 to perform both examinations simultaneously, for example, during the same imaging session. The PCCT system 220 and the PET system 210 are housed in the same enclosure. Alternatively, in embodiments, the PET and PCCT scanners may be separate devices, but still acquire images from the same patient at different times. Figure 2 An example medical imaging device 105 is depicted, configured to acquire both PET and PCCT data. Medical imaging device 105 is merely exemplary, and a wide variety of CT scanning systems can be used to collect PCCT and PET data with different configurations. Figure 2 In the medical imaging apparatus 105, a subject 225 (e.g., a patient 225) is positioned on a table 205, which is configured to move to multiple positions via a power system through a circular opening 230 in the PET / PCCT scanner. An X-ray source (or other radiation source) and one or more detector elements 250 are part of the PET / PCCT scanner and are configured to rotate on the gantry around the subject 225 while the subject is within the opening / hole 230. Rotation may be combined with movement of the table 205 to scan along a longitudinal extent of the patient 225. Alternatively, the gantry moves the source and detector 250 around the patient 225 in a helical path. In a PET / PCCT scanner, a single rotation may take approximately one second or less. During the rotation of the X-ray source and / or detector, the X-ray source generates a narrow fan-shaped (or cone-shaped) X-ray beam that passes through the target portion of the body being imaged from the subject 225. One or more detector elements 250 are positioned opposite an X-ray source and register the X-rays passing through the body of the subject 225 being imaged, recording snapshots in the process for image reconstruction. Numerous different snapshots are collected from many angles through the subject by one or more rotations of the X-ray source and / or one or more detector elements 250. Data generated from the collected snapshots is transmitted to an image processing unit 130, which stores or processes the acquired data based on the snapshots into one or more cross-sectional images or volumes of the subject's interior (e.g., internal organs or tissues) scanned by a PET / PCCT scanner. Any PET / PCCT scanner now known or developed later can be used. Other X-ray scanners, such as CT-type C-arm scanners, can also be used.

[0030] Conventional medical CT systems are equipped with solid-state scintillation detector elements (such as energy integrating detectors (EIDs)). In a two-step conversion process, absorbed X-rays are first converted into visible light in a scintillation crystal. This light is then converted into an electrical signal by a photodiode attached to the back of each detector unit. PCCT utilizes direct-conversion X-ray detectors, where the energy of incident X-ray photons is directly recorded as an electronic signal. Photon counting detectors (PCDs) directly convert the deposited X-ray energy into an electronic signal because a large voltage is applied across the semiconductor, creating electron-hole pairs when photons strike the detector. By using energy-resolved detectors instead of EIDs, PCCT systems are able to count individually incident X-ray photons and measure their energy. This energy information can then be used to generate images and for other tasks, such as material decomposition. For material decomposition, energy-selective images are generated based on the number of counts registered in each energy bin. Based on these images, a set of material concentration maps is generated using a data processing method called material decomposition. Material concentration maps can be used to generate images, but they also help to enhance simulated ultrasound data using fine tissue data.

[0031] In current clinical EID (Enhanced Electronic Dosing), the pixel size at isogonal points is approximately 0.4-0.6 mm, thus limiting their resolution. In fact, the smaller detector pixel size design, compared to the detector area, leads to an increase in the relative area of ​​the septum and a decrease in geometric dose efficiency. In PCD (Procedure for Diagnosis), due to the absence of mechanical separation, there are no technical limitations on pixel spacing, and it can reach 0.15-0.225 mm at isogonal points. PCCT data acquired by PCD is transmitted to image processing unit 130, which generates PCCT image 120 and / or PCCT image data.

[0032] PET data is acquired using a PET scan performed by a PET / PCCT scanner. The spatial resolution of PET images is typically 4-5 mm. The inherent resolution of PET images is significantly higher than that of CT / PCCT images. The resolution of PET images can be improved by using CT / PCCT images from the same patient. PET scans provide data related to the metabolic or biochemical function of the tissues and organs of the imaging patient 225. PET scans can use radiopharmaceuticals known as tracers to visualize both typical and atypical metabolic activity. The PET system 210 includes multiple detectors, such as crystals or other photon detectors. For example, the detector is a scintillation crystal coupled to an avalanche photodiode. In other embodiments, the scintillation crystal is coupled to a photomultiplier tube. The scintillation crystal is bismuth germanium oxide, gadolinium oxysilicate, or lutetium oxysilicate crystals, but other crystals may also be used. Solid-state or semiconductor detectors may be used.

[0033] Detectors 215 are arranged individually or in groups. Detector blocks or groups are arranged around the hole in any pattern, such as rings. Detector rings 215 are spaced apart but placed adjacent to or adjacent to each other. Any gaps may be provided between blocks within a ring, between detectors within a block, and / or between rings. Any number of detectors in blocks, detector blocks in rings, and / or rings may be used. Rings may extend completely or only partially around the hole.

[0034] PET system 210 is a nuclear imaging system. Detectors detect gamma rays indirectly emitted by positron emission tracers. Pairs of gamma rays generated by the same positron can be detected using a detector ring. The pairs of gamma rays travel approximately 180 degrees apart. If their directions of travel intersect the detector arrangement at two locations, an overlapping pair can be detected. To distinguish specific pairs, the consistency of the detected gamma rays is determined. Receiver timing is used to pair the detected gamma rays. As cue data, timing can also indicate the time of flight and generally provides information about the location where the emission occurred along the response line. Each individual detection output from the detector includes energy, position, and timing information. Alternatively, the detector outputs energy information, and the receiver processor determines the timing and position (e.g., based on port assignment or connection). Timing is used by a consistency processor to determine the detection consistency by different detectors, as well as the general position along the emission response line. Gamma ray pairs associated with the same positron emission are identified. Based on the detected event, given the detector involved in detecting that event, the response line is determined. The detected event is transmitted to image processing unit 130, which generates PET image data.

[0035] Alternatively, the systems and methods described herein can be applied to any PET image 115 from the same patient 225 with paired high-resolution CT images acquired using a PET-PCCT system, a PET-CT system, and / or a PET and CT scanner.

[0036] Image processing unit 130 / controller may include an image processor that uses machine learning network 110 (machine learning model 110) to generate fused images and / or higher resolution images. The image processor is a general-purpose processor, digital signal processor, 3D data processor, graphics processing unit, application-specific integrated circuit, field-programmable gate array, artificial intelligence processor, digital circuit, analog circuit, a combination thereof, or another device now known or later developed for image generation. The image processor can be a single device, multiple devices, or a network. For more than one device, parallel or sequential processing partitioning can be used. Different devices constituting the image processor can perform different functions. In one embodiment, the image processor is also a control processor or other processor of a PET / PCCT imaging device. Other image processors of the PET / PCCT imaging device or image processors external to the PET / PCCT imaging device can be used. The image processor is configured by software, firmware, and / or hardware to process data acquired by the PET / PCCT imaging device and output one or more images. The image processor can reconstruct intermediate images based on data from the PET / PCCT imaging device and then use machine learning network 110 to fuse or upgrade the images to provide higher resolution / more detailed images. Instructions for implementing the processes, methods, and / or techniques discussed herein are provided on a non-transitory computer-readable storage medium or memory, such as a cache, buffer, RAM, removable media, hard disk drive, or other computer-readable storage medium. The instructions may be executed by a processor or another processor. Computer-readable storage media include various types of volatile and non-volatile storage media. The functions, actions, or tasks illustrated in the figures or described herein are performed in response to one or more sets of instructions stored in or on a computer-readable storage medium. The functions, actions, or tasks are independent of the instruction set, storage medium, processor, or processing strategy, and may be performed individually or in combination by software, hardware, integrated circuits, firmware, microcode, etc. In one embodiment, the instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, the instructions are stored at a remote location for transfer via a computer network. In still other embodiments, the instructions are stored in a given computer, CPU, GPU, or system. Because some of the system components and method steps depicted in the accompanying figures can be implemented in software, the actual connections between system components (or process steps) may vary depending on how this embodiment is programmed.

[0037] Generally, trained machine learning networks mimic human cognitive functions associated with other human thought processes. Specifically, through training on training data, machine learning networks can adapt to new environments and detect and infer patterns. Another term for a “trained machine learning network” is a “trained function.” Generally, the parameters of a machine learning network can be adapted through training. Specifically, supervised training, semi-supervised training, unsupervised training, reinforcement learning, and / or active learning can be used. Furthermore, representation learning (an alternative term is “feature learning”) can be used. Specifically, the parameters of a machine learning network can be iteratively adapted through several training steps. Specifically, during training, a certain cost function can be minimized. Specifically, in the training of neural networks, backpropagation algorithms can be used. Specifically, machine learning networks can include neural networks, support vector machines, decision trees, and / or Bayesian networks, and / or machine learning networks can be based on k-means clustering, Q-learning, genetic algorithms, and / or association rules. Specifically, neural networks can be deep neural networks, convolutional neural networks, or convolutional deep neural networks. Furthermore, neural networks can be adversarial networks, deep adversarial networks, and / or generative adversarial networks.

[0038] In an embodiment, the machine learning network 110 may be provided or implemented by a neural network trained using deep learning. Each trained network may be defined as multiple sequential feature units or layers. The sequence is used to indicate the general flow of output feature values ​​from the input of one layer to the next. Information from that next layer is fed into the next layer, and so on, until the final output. These layers may be feedforward only, or they may be bidirectional, including some feedback to the previous layer. Nodes in each layer or unit may be connected to all nodes in the previous and / or subsequent layer or unit, or only to a subset of nodes. Skip connections may be used, such as passing the layer output to the next sequential layer and other layers. Unlike pre-programming features and attempting to associate features with attributes, deep architectures are defined as learning features at different levels of abstraction from the input data. These features are learned to reconstruct lower-level features (i.e., more abstract or compressed levels of features). For example, features are learned to generate fused images or higher-resolution images. For the next unit, features are learned to reconstruct the features of the previous unit, thus providing more abstraction. Each node in a unit represents a feature. Different units are provided to learn different features.

[0039] Various units or layers can be used, such as convolutions, pooling (e.g., max pooling), deconvolutions, fully connected layers, or other types of layers. Within a unit or layer, any number of nodes can be provided. For example, 100 nodes can be provided. Later or subsequent units can have more, fewer, or the same number of nodes. Generally, for convolutions, subsequent units have a higher level of abstraction. For example, the first unit provides features from the image, such as a node or feature being a line found in the image. The next unit combines the lines, making one of the nodes an angle. The next unit can combine features from the previous unit (e.g., the length of the angle and the line), making the nodes provide shape indications. For transposed convolutions, the level of abstraction is reversed. Each unit or layer reduces or compresses the level of abstraction.

[0040] Unlike conventional methods that primarily rely on mathematical models and hardware improvements to enhance resolution, this embodiment leverages advanced machine learning techniques to learn directly from the data how to improve image quality. This approach can be applied to existing PET / PCCT datasets without requiring additional hardware. Trained using real high-resolution PCCT 120 and PET images 115, the generated high-resolution PET images are both accurate and realistic, potentially reducing artifacts common in conventional methods. Compared to traditional reconstruction techniques, this process may also be more efficient in terms of both time and computational resources, as it utilizes the image processing capabilities of AI. In this embodiment, the use of GANs for PET / CT image enhancement represents an innovative application of machine learning in medical imaging. Unlike the static mathematical models used in traditional methods, this approach can be continuously improved as more data becomes available and the model is further refined.

[0041] Figure 3 An embodiment of an artificial neural network 500 according to one or more embodiments is shown. Alternative terms for "artificial neural network" are "neural network," "artificial neural network," or "neural network." The artificial neural network 500 can be used in part in one or more machine learning-based networks, such as those utilized in GANs, autoencoders, CNNs, unfolded networks, etc.

[0042] The artificial neural network 500 includes nodes 502-522 and edges 532, 534, ..., 536, where each edge 532, 534, ..., 536 is a directed connection from a first node 502-522 to a second node 502-522. Generally, the first node 502-522 and the second node 502-522 are different nodes 502-522, but they can also be the same. For example, in... Figure 5In the diagram, edge 532 is a directed connection from node 502 to node 506, and edge 534 is a directed connection from node 504 to node 506. Edges 532, 534, ..., 536 from the first node 502-522 to the second node 502-522 are also labeled as "incoming edges" of the second node 502-522 and "outgoing edges" of the first node 502-522.

[0043] In this embodiment, nodes 502-522 of the artificial neural network 500 can be arranged in layers 524-530, wherein these layers can include an inherent order introduced by edges 532, 534, ..., 536 between nodes 502-522. In particular, edges 532, 534, ..., 536 may exist only between adjacent node layers. Figure 5 In the illustrated embodiment, there is an input layer 524 consisting only of nodes 502 and 504 without incoming edges, an output layer 530 consisting only of node 522 without output edges, and hidden layers 526 and 528 located between the input layer 524 and the output layer 530. Generally, the number of hidden layers 526 and 528 can be arbitrarily chosen. The number of nodes 502 and 504 in the input layer 524 is often related to the number of input values ​​of the neural network 500, and the number of nodes 522 in the output layer 530 is often related to the number of output values ​​of the neural network 500.

[0044] In particular, (real) numbers can be assigned as values ​​to each node 502-522 of the neural network 500. Here, x (n) i This indicates the value of the i-th node 502-522 in the n-th layer 524-530. The values ​​of nodes 502-522 in the input layer 524 are equivalent to the input values ​​of the neural network 500, and the value of node 522 in the output layer 530 is equivalent to the output value of the neural network 500. Furthermore, each edge 532, 534, ..., 536 may include real-valued weights, specifically, the weights are real numbers within the interval [-1, 1] or the interval [0, 1]. Here, w... (m,n) i,j This indicates the weight of the edge between the i-th node 502-522 in layer m (524-530) and the j-th node 502-522 in layer n (524-530). Further, the abbreviation is w. (n) i,j Defined as weight w (n,n+1) i,j .

[0045] In particular, to calculate the output value of neural network 500, the input value is propagated through the neural network. Specifically, the values ​​of nodes 502-522 in the (n+1)th layer 524-530 can be calculated based on the values ​​of nodes 502-522 in the nth layer 524-530 as follows:

[0046]

[0047] Here, the function f is the transfer function (another term is the "activation function"). Known transfer functions are step functions, sigmoid functions (e.g., logic functions, generalized logic functions, hyperbolic tangent functions, arctangent functions, error functions, smooth step functions), or rectifier functions. Transfer functions are primarily used for standardization purposes.

[0048] In particular, these values ​​are propagated layer by layer through the neural network, where the value of the input layer 524 is given by the input of the neural network 500, the value of the first hidden layer 526 can be calculated based on the value of the input layer 524 of the neural network, the value of the second hidden layer 528 can be calculated based on the value of the first hidden layer 526, and so on.

[0049] To set the value w of the edge (m,n) i,j Training data is required to train the neural network 500. Specifically, the training data includes training input data and training output data (denoted as t). i For the training step, neural network 500 is applied to the training input data to generate computed output data. Specifically, the training data and the computed output data include multiple values, the number of which is equal to the number of nodes in the output layer.

[0050] In particular, the weights within the neural network 500 are recursively adapted using a comparison between the calculated output data and the training data (backpropagation algorithm). Specifically, the weights are changed according to the following...

[0051]

[0052] Where γ is the learning rate, and the numerical δ (n) j can be recursively calculated as

[0053]

[0054] Based on δ (n+1) j, if the (n+1)th layer is not an output layer, and

[0055]

[0056] If the (n+1)th layer is the output layer 530, where f' is the first derivative of the activation function, and y (n+1) j It is the comparison training value of the j-th node of the output layer 530.

[0057] Figure 4A convolutional neural network 600 according to one or more embodiments is shown. Machine learning networks described herein, such as GANs, autoencoders, unfolded networks, etc., can be implemented using the convolutional neural network 600.

[0058] exist Figure 4 In the illustrated embodiment, the convolutional neural network 600 includes an input layer 602, a convolutional layer 604, a pooling layer 606, a fully connected layer 608, and an output layer 610. Alternatively, the convolutional neural network 600 may include several convolutional layers 604, several pooling layers 606, and several fully connected layers 608, as well as other types of layers. The order of the layers can be arbitrarily chosen; often, the fully connected layer 608 is used as the last layer before the output layer 610.

[0059] Specifically, within the convolutional neural network 600, nodes 612-620 of layer 602-610 can be viewed as arranged as a d-dimensional matrix or a d-dimensional image. In particular, in the two-dimensional case, the values ​​of nodes 612-620 in the nth layer 602-610, indexed by i and j, can be denoted as x. (n) [i,j] However, the arrangement of nodes 612-620 in a layer 602-610 has no effect on the computations performed within the convolutional neural network 600 itself, since these computations are given only by the structure and weights of the edges.

[0060] In particular, the convolutional layer 604 is characterized by the structure and weights of its incoming edges forming a convolution operation based on a certain number of kernels. Specifically, the structure and weights of the incoming edges are selected such that the value x of node 614 of the convolutional layer 604 is... (n) k Calculated as the value x of node 612 based on the previous layer 602. (n-1) convolution x (n) k =K k *x (n-1) Where convolution* is defined in the two-dimensional case as:

[0061]

[0062] Here, the k-th core K kIt is a d-dimensional matrix (a two-dimensional matrix in this embodiment), which is typically small compared to the number of nodes 612-618 (e.g., a 3x3 or 5x5 matrix). Specifically, this implies that the weights of the incoming edges are not independent but are chosen such that they produce the convolution equation. In particular, for a 3x3 kernel, regardless of the number of nodes 612-620 in the corresponding layers 602-610, there are only 9 independent weights (one independent weight per entry in the kernel matrix). Specifically, for convolutional layer 604, the number of nodes 614 in the convolutional layer is equal to the number of nodes 612 in the previous layer 602 multiplied by the number of kernels.

[0063] If the nodes 612 of the previous layer 602 are arranged as a d-dimensional matrix, then using multiple kernels can be interpreted as adding another dimension (denoted as the "depth" dimension), such that the nodes 614 of the convolutional layer 604 are arranged as a (d+1)-dimensional matrix. If the nodes 612 of the previous layer 602 have already been arranged as a (d+1)-dimensional matrix including the depth dimension, then using multiple kernels can be interpreted as extending along the depth dimension, such that the nodes 614 of the convolutional layer 604 are also arranged as a (d+1)-dimensional matrix, where the (d+1)-dimensional matrix is ​​larger in the depth dimension than the previous layer 602 by the number of kernels.

[0064] The advantage of using convolutional layer 604 is that it can take advantage of the spatial local correlation of the input data by enforcing local connectivity patterns between nodes of adjacent layers, in particular by having each node connect only to a small region of the node in the previous layer.

[0065] exist Figure 6 In the illustrated embodiment, the input layer 602 comprises 36 nodes 612 arranged in a two-dimensional 6x6 matrix. The convolutional layer 604 comprises 72 nodes 614 arranged in two two-dimensional 6x6 matrices, each of which is the result of convolving the input layer values ​​with the kernel. Equivalently, the nodes 614 of the convolutional layer 604 can be interpreted as arranged in a three-dimensional 6x6x2 matrix, where the last dimension is the depth dimension.

[0066] The pooling layer 606 can be characterized by the structure and weights of its incoming edges, as well as the activation function of its nodes 616, which forms a pooling operation based on a nonlinear pooling function f. For example, in the two-dimensional case, the value x of node 616 in pooling layer 606... (n) It can be based on the value x of node 614 of the previous layer 604. (n-1) Calculated as

[0067] x (n) [i,j]=f(x (n-l) [id1,jd2],...,x (n-1) [id1+d1-1,jd2+d2-1])

[0068] In other words, by using pooling layer 606, the number of nodes 614 and 616 can be reduced by replacing the number d1·d2 of adjacent nodes 614 in the previous layer 604 with a single node 616, which is calculated as a function of the number of adjacent nodes in the pooling layer. Specifically, the pooling function f can be a maximum function, an average function, or an L2 norm function. Furthermore, for pooling layer 606, the weights of the incoming edges are fixed and are not modified during training.

[0069] The advantage of using pooling layer 606 is that it reduces the number of nodes 614 and 616 and the number of parameters. This results in a reduction in the computational cost of the network and controls overfitting.

[0070] exist Figure 6 In the illustrated embodiment, pooling layer 606 is max pooling, replacing four adjacent nodes with only one node, where the value is the maximum of the four adjacent node values. Max pooling is applied to each d-dimensional matrix of the previous layer; in this embodiment, max pooling is applied to each of the two 2D matrices, reducing the number of nodes from 72 to 18.

[0071] The characteristic of the fully connected layer 608 is that there are most, in particular all, edges between the node 616 of the previous layer 606 and the node 618 of the fully connected layer 608, and the weight of each edge can be adjusted individually.

[0072] In this embodiment, the nodes 616 of the layer preceding the fully connected layer 606 are displayed both as a two-dimensional matrix and additionally as unrelated nodes (indicated as a row of nodes, where the number of nodes is reduced for better presentation). In this embodiment, the number of nodes 618 in the fully connected layer 608 is equal to the number of nodes 616 in the layer preceding the fully connected layer 606. Alternatively, the number of nodes 616 and 618 can be different.

[0073] The convolutional neural network 600 may also include ReLU (Rectified Linear Unit) layers or activation layers with nonlinear transfer functions. Specifically, the number and structure of nodes in the ReLU layer are identical to those in the previous layer. In particular, the value of each node in the ReLU layer is computed by applying the rectification function to the value of the corresponding node in the previous layer.

[0074] The inputs and outputs of different convolutional neural network blocks can be wired using summation (residual / dense neural networks), element-wise multiplication (attention), or other differentiable operators. Therefore, if the entire pipeline is differentiable, the convolutional neural network architecture can be nested rather than sequential.

[0075] In particular, the convolutional neural network 600 can be trained based on the backpropagation algorithm. To prevent overfitting, regularization methods can be used, such as dropping nodes 612-620, random pooling, using artificial data, weight decay based on L1 or L2 norm, or maximum norm constraints. Different loss functions can be combined to train the same neural network to reflect the joint training objective. A subset of neural network parameters can be excluded from the optimization to retain weights pre-trained on another dataset.

[0076] In this embodiment, the machine learning network 110 is trained using a generative adversarial process. Figure 5 The diagram shows a data flow diagram according to an embodiment for creating high-resolution images using a generative adversarial network (GAN) based on input data indistinguishable from real output data. The GAN includes a generator function 320 and a discriminator function 310, whereby the generator function 320 creates synthetic data and the discriminator function 310 distinguishes between synthetic and real data. By training the generator function 320 and / or the discriminator function 310, on the one hand, the generator function 320 is configured to create synthetic data that is incorrectly classified as real by the discriminator function 320; on the other hand, the discriminator function 310 is configured to distinguish between real data and synthetic data generated by the generator function. In game theory terms, the GAN can be interpreted as a zero-sum game. The training of the generator function and / or the discriminator function is specifically based on minimizing a cost function. The GAN is trained on a dataset of paired input images 115, 120 and their corresponding fused or higher-resolution images, which serve as ground truth. The generator network 310 is trained to produce a fused image similar to a real benchmark or a higher resolution image, while the discriminator network 320 is trained to distinguish between real and fake images. The loss function used during training can be a combination of adversarial loss and content loss to ensure that the fused or higher resolution image is realistic and retains the most relevant information from the input image.

[0077] In an embodiment, the machine learning network 110 is or includes a cycle consistent adversarial network, also known as CycleGAN. Figure 6 An example of CycleGAN is depicted. In Figure 6In the CycleGAN architecture, there are two generators G and F, and two discriminators X and Y. Generator G transforms the distribution from one input X to another input Y such that discriminator Y cannot distinguish the transformed output Y = G(X) from the original input Y. One problem with this system is that the generator may transform all images in the same way, so it can only generate a single example Y, which is not the desired task. To mitigate this problem, the CycleGAN architecture uses constraints by defining another generator F, which acts as the inverse transformation of G. This guarantees that the transformation of X will not be reduced to a single example. A new loss function characterizing the cycle consistency loss is added during training, thereby encouraging transformations to verify the properties FGX≈X and G(FY)≈Y. There are two discriminators: discriminator Y, which distinguishes the generated Y from Y, and discriminator X, which distinguishes the generated X from the real X. The cycle consistency loss is divided into two distinct parts: the first part (1) corresponds to the loss between the elements of X and its reconstruction, and the second part (2) corresponds to the loss between the elements of Y and its reconstruction. During training, the generator and discriminator optimize the same overall loss function, which consists of two loss sub-functions associated with the generator (adversarial loss): the discriminator maximizes it (so that they can distinguish between generated and real data), and the generator minimizes it so that they can create examples that are increasingly difficult to distinguish from real data.

[0078] During training, additional loss values, parameters, or step sizes are used. In the example, CycleGAN, as described above, includes an objective function to minimize the loss. The total loss can be divided into three parts: adversarial loss (each corresponding to one of the two domains) and cycle consistency loss. The adversarial loss function used in CycleGAN is similar to the adversarial loss function in a typical GAN. It involves setting an objective for the generator G to produce an image G(x) that is visually similar to an image from domain Y, while the discriminator Dy aims to distinguish between the generated sample G(x) and the real sample y. The goal is to minimize the objective of G, while the adversary D attempts to maximize it. Additionally, a similar loss function is introduced for the mapping function F: Y->X and its discriminator Dx. For each image x from domain X, the image transformation cycle should be able to bring x back to the original image (forward cycle consistency), i.e., x→G(x)→F(G(x))≈x. Similarly, for each image y from domain Y, G and F should also satisfy backward cycle consistency: y→F(y)→G(F(y))≈y. Add additional loss values / functions to the adversarial loss in either domain.

[0079] In this embodiment, the machine learning network 110 is or includes a CNN 800, such as those described above. Figure 4 As described in [the text]. Figure 7A and Figure 7BAn example CCN 800 that can be used to generate fused images is depicted. Figure 7A In the first example, the CNN 800 determines the weights of how to fuse the image features decomposed from the input image. These fused image features are then used to generate the fused image. Figure 7B In the second example, image features are extracted by the corresponding CNN 800 and then fused to generate a fused image. The CNN 800 automatically learns spatial features from the input images, which can be used to create fused or higher-resolution representations by extracting relevant information from the source images. The CNN 800 architecture is trained on datasets containing pairs of input images and their corresponding fused or higher-resolution images. A loss function is used during training to ensure that the fused or higher-resolution image retains the most relevant information from the input images. Once the CNN 800 has been trained, it can fuse new pairs of input images or generate new higher-resolution images. Input images are fed through the CNN 800, and the resulting feature maps from each input modality are combined to create fused or higher-resolution representations.

[0080] In an embodiment, the machine learning network 110 is or includes an autoencoder. Figure 8 An example of a machine learning network 110 including an autoencoder is depicted. The autoencoder can be used to perform feature-level fusion, where input images are encoded into a lower-dimensional feature space and then decoded to create a fused representation. The autoencoder is trained on a dataset of paired input images and their corresponding fused images. The loss function used during training typically combines mean squared error (MSE) and a structural similarity index (SSIM) to ensure that the fused image retains the most relevant information from the input images.

[0081] In this embodiment, the machine learning network 110 is or includes a transformer. An attention mechanism is used for image fusion to enable the model to focus on the most relevant features from each input modality. Figure 9An example of a machine learning network110 using a transformer is depicted. A CNN is used to extract features from the input image and then feed them into the transformer to determine which features are most important for reconstructing the fused image. Different attention mechanisms can be used for the fusion task. Self-attention mechanisms compute attention scores for each feature map within a single modality. This allows the model to focus on the most informative regions in each modality and can improve the overall quality of the fused image. Cross-attention mechanisms compute attention scores between different modalities. This allows the model to identify the most informative features from each modality and combine them in the fused image. Cross-attention can be particularly useful when fusing images of different resolutions or sensor types. Multi-level attention mechanisms can be used to compute attention scores at multiple levels of abstraction. For example, the model can compute self-attention scores at the pixel level and higher levels of abstraction, such as object or scene levels. This allows the model to capture fine-grained details and global contextual information in the fused image. Per-channel attention mechanisms compute attention scores for each channel within a single modality or between different modalities. This allows the model to identify the most informative channels and weight them accordingly when fusing the input images.

[0082] In an embodiment, the machine learning network 110 includes, or alternately includes, gradient updates and regularized iterative reconstruction, wherein the machine learning network is provided for regularization via an iterative sequence. Each given iteration in the unfolded network or via repeated reconstruction operations includes at least regularization. Gradients or comparisons that associate image objects with measurements can be used. Gradient updates use a scaling factor determined based on scan settings or operator input. The scaling factor can be related to the gradient step size. In one embodiment, the scaling factor is determined based on the sampling pattern. In another embodiment, the scaling factor is determined based on the noise level in the scan data. Regularization is provided in one, some, or all iterations and may include the application of a machine learning network, such as a convolutional neural network (CNN).

[0083] The outputs of processes and methods can be output for further processing or displayed to the operator. The image processing system 130 includes an operator interface 15 formed by inputs and outputs. Inputs can be interfaces, such as interfaces connecting to computer networks, memory, databases, medical image storage, or other input data sources. Inputs can be user input devices, such as mice, trackpads, keyboards, ball bearings, touchpads, touchscreens, or other devices for receiving user input. Inputs can receive scanning protocols, imaging protocols, or scanning parameters. Individuals can select inputs, such as manually or physically entering values. Previously used values ​​or parameters can be entered from the interface. Default, institutional, facility, or group setting levels can be entered, such as from memory to the interface.

[0084] The output can be a display device or any other type of interface. For example, it can display an image output by the method. For example, it can display an image of a region of patient 225. The generated images of the selected model and simulated scan are presented on the display of operator interface 115. Analysis / interpretation can also be displayed on the display device. Image processing system 130 can be configured to generate reports / evaluations of the images displayed on the display device. The display is a CRT, LCD, plasma, projector, printer, or other display device. The display is configured by loading the image onto a display plane or buffer. The display is configured to display a fused / higher resolution image of a region of patient 225. The operator interface may include a graphical user interface (GUI) that allows the user to interact with image processing system 130 and make modifications substantially in real time.

[0085] Figure 10 An example method for generating high-resolution PET-CT images using high-resolution CT images, and particularly PCCT images 120, is described. This method leverages the high-resolution capability of photon-counting computed tomography (PCCT) within a PET / PCCT scanner setup to enhance the spatial resolution of PET image 115. The method uses high-resolution spatial data from PCCT to significantly enhance the image quality of PET image 115. In an embodiment, a generative adversarial network (GAN) is used. The GAN uses existing high-resolution CT data from PCCT to enhance the resolution of PET image 115, making this application of the GAN specifically tailored for utilizing real, applicable medical imaging data.

[0086] At action A110, medical imaging device 105 acquires PET images 115 of the area of ​​patient 225. Positron emission tomography (PET) scans capture images based on emissions from radioactive material delivered to patient 225.

[0087] At action A120, medical imaging device 105 acquires a PCCT image 120 of a region of patient 225. PCCT uses crystalline semiconductor materials instead of ceramic scintillators to directly generate charges, thereby achieving high resolution while eliminating electronic noise and reducing radiation dose. Detected photons are counted individually, providing a more accurate image signal, and classified according to their energy levels, enabling spectral discrimination at the detector level. Medical imaging device 105 is configured to acquire both PET image 115 and PCCT image 120 in a single imaging session of patient 225. In an embodiment, medical imaging device 105 acquires reconstructed data to generate PCCT image 120 and PET image 115.

[0088] At action A130, PET image 115 and PCCT image 120 are fed into a machine learning network, which is trained to output a high-resolution PET-CT image. The machine learning network 110 is trained using real high-resolution PCCT and PET images 115. The resulting high-resolution PET image is both accurate and realistic, potentially reducing artifacts common in conventional methods.

[0089] In this embodiment, the machine learning network 110 includes or is based on paired PCCT and PET images 115 to generate an image using a recurrent consistent adversarial network (Recurrent Generative Network). When the underlying structures are similar, the Recurrent GAN framework effectively transforms the image between the source and target domains. Additionally, the Recurrent GAN enforces an inverse transformation. This recurrent consistency allows for a higher level of accuracy than other machine learning-based methods because the model is subject to dual constraints. The Recurrent GAN model relies on the continuous improvement of both the generator and discriminator networks. The accuracy of these two networks directly depends on the design of their corresponding loss functions. In this embodiment, a two-part loss function consisting of an adversarial loss and a recurrent consistency loss is used.

[0090] Alternative models or networks, such as CNNs, autoencoders, unfolded networks, or other generative networks, can be used.

[0091] At action A140, the machine learning network 110 outputs a high-resolution PET-CT image of the patient's region 225. The image can be displayed to the user or further processed / analyzed.

[0092] Figure 11 A method for generating high-resolution fused PET-CT images using high-resolution CT images, particularly PCCT images 120, is described. This method leverages the high-resolution capability of photon counting computed tomography (PCCT) within the PET / PCCT scanner setup to enhance the spatial resolution of PET image 115. The method uses high-resolution spatial data from PCCT to significantly enhance the image quality of PET image 115. In an embodiment, a generative adversarial network (GAN) is used. The GAN uses existing high-resolution CT data from PCCT to enhance the resolution of PET image 115, making this application of the GAN specifically tailored for utilizing real, applicable medical imaging data.

[0093] At action A210, medical imaging device 105 acquires a PET image 115 of a region of patient 225. PET imaging is limited by poor spatial resolution (often 4-5 mm) and noise. By integrating photon counting CT (PCCT), which provides superior resolution (up to 0.11 mm) and sensitivity, this embodiment improves the resolution of PET image 115, thereby enhancing disease detection, staging, and treatment monitoring. At action A220, medical imaging device 105 acquires a PCCT image 120 of a region of patient 225. Medical imaging device 105 is configured to acquire both PET image 115 and PCCT image 120 in a single imaging session of patient 225. In this embodiment, medical imaging device 105 acquires reconstructed data to generate PCCT image 120 and PET image 115.

[0094] At action A230, PET image 115 and PCCT image 120 are fed into a machine learning network 110, which is trained to output a high-resolution fused PET-CT image. Machine learning network 110 may contain or include models or networks such as CNNs, autoencoders, unfolding networks, GANs, or other generative networks.

[0095] At action A240, machine learning network 110 outputs a high-resolution fused PET-CT image of the patient's region 225.

[0096] While the invention has been described above with reference to various embodiments, many changes and modifications can be made without departing from the scope of the invention. Therefore, the detailed description above is intended to be illustrative rather than restrictive, and it should be understood that the following claims, including all equivalents, are intended to define the spirit and scope of the invention.

[0097] The following is a list of non-limiting illustrative embodiments disclosed herein:

[0098] Illustrative Example 1. A method for generating a high-resolution PET image, the method comprising: acquiring a PET image of a patient region using a medical imaging device; acquiring a PCCT image of the patient region using the medical imaging device; inputting the PET image and the PCCT image into a machine learning network, the machine learning network being trained to output a high-resolution PET image including a higher resolution than the input PET image; and outputting a high-resolution PET image of the patient region.

[0099] Illustrative Example 2. The method according to Illustrative Example 1 further includes: displaying a high-resolution PET image.

[0100] Illustrative Example 3. The method described in the preceding illustrative examples, wherein the machine learning network is trained using a generative adversarial process.

[0101] Illustrative Example 4. The method according to Illustrative Example 3, wherein the machine learning network includes CycleGAN.

[0102] Illustrative Example 5. The method described in the preceding illustrative examples, wherein the machine learning network is trained using real PET and PCCT images.

[0103] Illustrative Example 6. The method according to the preceding illustrative examples, wherein the medical imaging apparatus includes a combined PET / PCCT imaging system configured to acquire PET images and PCCT images during a single imaging session.

[0104] Illustrative Example 7. The method described in the preceding illustrative examples, wherein the high-resolution PET image includes a resolution of less than 1 mm.

[0105] Illustrative Example 8. A method for generating a fused high-resolution PET-CT image from a pair of CT and PCCT images, the method comprising: acquiring a PET image of a patient region via a medical imaging device; acquiring a PCCT image of the patient region via the medical imaging device; inputting the PET image and the PCCT image into a machine learning network, the machine learning network being trained to output a high-resolution PET-CT image including a higher resolution than the input PET image; and outputting a high-resolution fused PET-CT image of the patient region.

[0106] Illustrative Example 9. The method according to the preceding illustrative examples further includes: displaying high-resolution fused PET-CT images.

[0107] Illustrative Example 10. The method according to the preceding illustrative examples, wherein the machine learning network is trained using a generative adversarial process.

[0108] Illustrative Example 11. The method described in the preceding illustrative examples, wherein the machine learning network includes CycleGAN.

[0109] Illustrative Example 12. The method described in the preceding illustrative examples, wherein the machine learning network is trained using real PET and PCCT images.

[0110] Illustrative Example 13. The method according to the foregoing illustrative examples, wherein the medical imaging device includes a combined PET / PCCT imaging system configured to acquire PET images and PCCT images during a single imaging session.

[0111] Illustrative Example 14. The method described in the preceding illustrative examples, wherein the high-resolution fused PET-CT image includes a resolution of less than 1 mm.

[0112] Illustrative Example 15. A system for generating high-resolution PET images from a pair of CT and PCCT images, the system comprising: a medical imaging apparatus configured to acquire PET images and PCCT images of a patient region; an image processing unit configured to input the PET images and PCCT images into a machine learning network trained to generate a high-resolution PET-CT image including a higher resolution than the input PET images; and a display configured to display the high-resolution PET-CT image generated from the PET images and PCCT images.

[0113] Illustrative Example 16. The system according to the preceding illustrative examples, wherein the machine learning network is trained using a generative adversarial process.

[0114] Illustrative Example 17. The system according to the preceding illustrative examples, wherein the machine learning network includes CycleGAN.

[0115] Illustrative Example 18. The system according to the preceding illustrative examples, wherein the machine learning network is trained using real PET and PCCT images.

[0116] Illustrative Example 19. According to the system described in the preceding illustrative examples, the medical imaging apparatus includes a combined PET / PCCT imaging system configured to acquire PET and PCCT images during a single imaging session.

[0117] Illustrative Example 20. The system according to the foregoing illustrative examples includes high-resolution PET-CT images with a resolution of less than 1 mm.

Claims

1. A method for generating high-resolution PET images, the method comprising: PET images of the patient's area are acquired using medical imaging equipment; PCCT images of the patient region are acquired using the medical imaging device. The PET image and the PCCT image are input into a machine learning network, which is trained to output a high-resolution PET image that has a higher resolution than the input PET image; and Output high-resolution PET images of the patient area.

2. The method according to claim 1, further comprising: The high-resolution PET image is displayed.

3. The method of claim 1, wherein the machine learning network is trained using a generative adversarial process.

4. The method of claim 3, wherein the machine learning network comprises CycleGAN.

5. The method of claim 1, wherein the machine learning network is trained using real PET and PCCT images.

6. The method of claim 1, wherein the medical imaging device comprises a combined PET / PCCT imaging system configured to acquire the PET image and the PCCT image during a single imaging session.

7. The method of claim 1, wherein the high-resolution PET image comprises a resolution of less than 1 mm.

8. A method for generating a fused high-resolution PET-CT image from paired CT and PCCT images, the method comprising: PET images of the patient's area are acquired using medical imaging equipment; PCCT images of the patient region are acquired using the medical imaging device. The PET image and the PCCT image are input into a machine learning network, which is trained to output a high-resolution PET-CT image that has a higher resolution than the input PET image. as well as Output a high-resolution fused PET-CT image of the patient region.

9. The method of claim 8, further comprising: The high-resolution fused PET-CT image is displayed.

10. The method of claim 8, wherein the machine learning network is trained using a generative adversarial process.

11. The method of claim 9, wherein the machine learning network comprises CycleGAN.

12. The method of claim 8, wherein the machine learning network is trained using real PET and PCCT images.

13. The method of claim 8, wherein the medical imaging device comprises a combined PET / PCCT imaging system configured to acquire the PET image and the PCCT image during a single imaging session.

14. The method of claim 8, wherein the high-resolution fused PET-CT image comprises a resolution of less than 1 mm.

15. A system for generating high-resolution PET images from paired CT and PCCT images, the system comprising: A medical imaging device configured to acquire PET images of a patient region and PCCT images of the patient region; An image processing unit is configured to input the PET image and the PCCT image into a machine learning network, the machine learning network being trained to generate a high-resolution PET-CT image that includes a higher resolution than the input PET image; as well as A display is configured to display the high-resolution PET-CT image generated from the PET image and the PCCT image.

16. The system of claim 15, wherein the machine learning network is trained using a generative adversarial process.

17. The system of claim 16, wherein the machine learning network comprises CycleGAN.

18. The system of claim 15, wherein the machine learning network is trained using real PET and PCCT images.

19. The system of claim 15, wherein the medical imaging device comprises a combined PET / PCCT imaging system configured to acquire the PET image and the PCCT image during a single imaging session.

20. The system of claim 15, wherein the high-resolution PET-CT image includes a resolution of less than 1 mm.

Images