Deep-learning based monoenergetic imaging at different energies

Abstract

Virtual monoenergetic images (VMIs) are generated from single-energy CT images using a machine learning model. The machine learning model is trained on training data to generate a VMI at a target energy level from an input single-energy CT image acquired with a CT system having an associated tube potential. CT scan protocol parameters used when acquiring the single-energy CT image are applied as an additional input to the machine learning model.

Description

DEEP-LEARNING BASED MONOENERGETIC IMAGING AT DIFFERENT ENERGIESCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Patent Application Serial No. 63 / 674,777, filed on July 23, 2024, and entitled “DEEP-LEARNING BASED MONOENERGETIC IMAGING AT DIFFERENT ENERGIES," which is herein incorporated by reference in its entirety.STATEMENT OF FEDERALLY SPONSORED RESEARCH

[0001] This invention was made with government support under EB028590 awarded by the National Institutes of Health. The government has certain rights in the invention.BACKGROUND

[0002] Multi-energy computed tomography (CT), including conventional dual-energy CT with energy-integrating-detectors (EID) and photon-counting detector (PCD) CT, enables the generation of virtual monoenergetic images (VMIs) from the acquired multi-energy data. These VMIs, obtained at various photon energies (i.e., keVs), exhibit distinct image properties suitable for diverse clinical applications. For instance, VMIs at lower keV levels enhance iodine and bone contrasts, but often exhibit blooming artifacts from highly attenuating materials such as calcium and stents. On the other hand. VMIs at higher keV levels effectively reduce beam hardening, calcium blooming, and metal artifacts, though they result in reduced contrast, especially for organs and tissues with iodine. Thus, there are intrinsic trade-offs between VMIs at different keV levels.SUMMARY OF THE DISCLOSURE

[0003] According to an aspect of the present disclosure, a method for generating a virtual monoenergetic image (VMI) at a target energy level from single-energy computed tomography (CT) image data is provided. The method includes accessing single-energy CT image data with a computer system, wherein the single-energy CT image data comprise one or more single-energy' CT images acquired with a CT system using a tube potential. The method also includes accessing CT scan protocol data with the computer system, wherein the CT scan protocol data comprise CT scan protocol parameters used by the CT system when acquiringthe single-energy CT image data. The method further includes accessing a machine learning model with the computer system, wherein the machine learning model has been trained on training data to synthesize a VMI at a target energy level from a single-energy CT image acquired at a particular tube potential. The method additionally includes inputting the singleenergy CT image data and the CT scan protocol data to the machine learning model using the computer system, generating a VMI at the target energy level as an output. Finally, the method includes outputting the VMI with the computer system. Other embodiments of this aspect include corresponding systems (e.g., computer systems), programs, algorithms, and / or modules, each configured to perform the steps of the methods.

[0004] According to another aspect of the present disclosure, a method for training a neural network to generate a virtual monoenergetic image (VMI) at a target energy level from an input image having a different energy level is provided. The method includes accessing training data with a computer system, wherein the training data comprises: first VMI data comprising VMIs associated with a first energy level corresponding to a tube potential of a CT system; second VMI data comprising VMIs associated with a second energy level corresponding to the target energy level, wherein the first energy level is different from the second energy level; and CT scan protocol data associated with CT scan protocol parameters corresponding to the first VMI data and the second VMI data. The method also includes accessing a neural network with the computer system, training the neural network on the training data, and storing the trained neural network with the computer system. Other embodiments of this aspect include corresponding systems (e g., computer systems), programs, algorithms, and / or modules, each configured to perform the steps of the methods.

[0005] According to another aspect of the present disclosure, a method for generating a computed tomography (CT) image is provided. The method includes accessing a first CT image with a computer system, wherein the first CT image was acquired with a CT system using a first kV setting. The method also includes accessing a machine learning model with the computer system, wherein the machine learning model has been trained on training data to generate a new CT image corresponding to a second kV setting from a CT image acquired at the first kV setting. The method further includes inputting the first CT image to the machine learning model using the computer system, generating a second CT image as an output, wherein the second CT image corresponds to a CT image having been acquired using the second kV setting. Finally, the method includes outputting the second CT image with the computer system. Other embodiments of this aspect include corresponding systems (e.g., computersystems), programs, algorithms, and / or modules, each configured to perform the steps of the methods.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] FIG. 1 is a schematic overview of an example machine learning model framework (e.g., a DIAMOND framework) for generating VMIs at a target energy level from a single-energy CT image input. The framework includes two stages. Stage 1: Network Training on VMI Data. Stage 2: DIAMOND Inference on single energy CT with subsequent clinical analysis.

[0007] FIG. 2 is a schematic overview of an example machine learning model framework (e.g., a DIAMOND+ framework) for generating VMIs at a target energy level from a single-energy CT image input. The DIAMOND+ additionally incorporates CT protocol parameters together with original single-energy’ CT images to enable model generalizability for different imaging conditions.

[0008] FIG. 3 is a flowchart of an example method for generating VMIs at a target energy’ level from single-energy CT images using a suitably trained machine learning model.

[0009] FIG. 4 is a flowchart of an example method for training a machine learning model to generate a VMI at a target energy level from single-energy CT images acquired with a CT system having a particular tube potential.

[0010] FIG. 5 illustrates an example of a DIAMOND model transforming an input 70 keV VMI from an example study to a 50 keV VMI, obtaining CT numbers that matched those of the true 50 keV VMI.

[0011] FIG. 6 illustrates a diagram of one-sided stenosis phantoms and the results of stenosis assessment of the SECT-Original and DIAMOND images. The stenosis percent from the software was shown on the top left of each figure. Image display window (WL / WW) for SECT: 200 / 300 HU, for DIAMOND: 100 / 180 HU.

[0012] FIG. 7 illustrates sample cCTA images from EID-CT, DIAMOND-enhanced. and PCD-CT of the same patient, scanned on the same day. DIAMOND substantially reduced blooming artifacts compared to EID-CT, leading to a decrease in percent diameter stenosis (values shown in the lower comer), aligning the visual and quantitative outcomes more closely with PCD-CT.

[0013] FIG. 8 illustrates a percent diameter stenosis quantification among original EID- CT, DIAMOND, and high-resolution PCD-CT images demonstrate that DIAMONDsignificantly lowers percent diameter stenosis compared to original EID-CT, yielding results similar to those from PCD-CT.

[0014] FIG. 9 is a flowchart of an example method for converting a first CT image acquired using a first kV setting to a second CT image having properties as if the second CT image was acquired using a second kV setting, or alternatively to a VMI having an effective energy level corresponding to the second kV setting.

[0015] FIG. 10 is a flowchart of an example method for training a machine learning model to convert a first CT image acquired with a first kV setting to a second CT image and / or VMI associated with a second kV setting.

[0016] FIG. 11 shows another example DIAMOND framework. EID = energy integrating detector, VMI = virtual monoenergetic image.

[0017] FIG. 12 is a block diagram of an example system for generating VMIs at a target energy level from single-energy7CT images according to some embodiments described in the present disclosure.

[0018] FIG. 13 is a block diagram of example components that can implement the system of FIG. 12.

[0019] FIGS. 14A and 14B illustrate an example CT sy stem that can be implemented in accordance with some embodiments described in the present disclosure.DETAILED DESCRIPTION

[0020] Described here are systems and methods for generating virtual monoenergetic images (VMIs) from CT images acquired with a single-energy CT (SECT) system. For instance, the disclosed systems and methods can produce VMIs for SECT to enhance its spectral capability without incurring additional upgrade costs and without sacrificing temporal resolution. The methods implement an artificial intelligence (Al) based framework that may be referred to as deep-leaming based monoenergetic imaging at different energies (DIAMOND).

[0021] It is an aspect of the present disclosure that the disclosed systems and methods can be used with images acquired using an EID-CT system, a PCD-CT system, a cone-beam CT system, or other suitable CT scanner system. Although most PCD-CT scanners are intrinsically multi-energy and may not need an extra method to generate VMIs, certain applications may not have the capability of multi-energy imaging available due to various reasons (e.g., image transfer limitations in certain high-resolution applications, or differentresolution at various energy bins). In these situations, the PCD-CT scanner may only output single-energy images.

[0022] A common problem in Al or deep learning is generalizability. For instance, Al algorithms trained on one system may not generalize well to other systems. This creates issues with generalizability and commonly requires retraining the same algorithms each time a new system is introduced, which is labor-intensive and costly. It is another advantage of the disclosed systems and methods that a machine learning model (e.g., a neural network) is fully trained using VMIs at different effective energies as the input and using target energy VMIs as the label. The VMIs used for training can be collected from any MECT scanner, either EID- CT or PCD-CT. The, for inference, the only information needed from the single energy CT is its effective energy, which can be obtained through HVL or phantom measurements as an example. This allows the same training dataset to be used for inference with any single-energy CT scanner model from any vendor, without the need to acquire new training data, by selecting the model with the same effective energy. In addition, the machine learning model can incorporate generalization to other CT protocol settings. This can be achieved by introducing CT scanning and reconstruction parameters (e.g., tube potential, reconstruction field of view, and kernel) into the model.

[0023] It is another aspect of the present disclosure that the disclosed systems and methods can operate directly on CT images without the need for projection data, raw data, or other proprietary information from the CT scanner. In this way, the disclosed systems and methods ensure compatibility across a broad range of scanner types and models.

[0024] The disclosed systems and methods provide a framework to create VMIs from single-energy CT images using a deep learning model, or other machine learning model. The framework, which in some instances may be referred to as a DIAMOND framework as noted above, generally includes two stages: a training stage in which a machine learning model is trained on training data that includes VMIs with two different energy levels (e.g., keVs); and an inference stage in which a single-energy CT image (e.g., a EID-CT image) is input to the trained machine learning model to generate a VMI at a target energy level. The trained machine learning model is selected to match the particular tube potential on which the model was trained with the tube potential of the input single-energy CT image.

[0025] A high-level example of this framework is illustrated in FIG. 1. In the illustrated example, during the training stage the input to the DIAMOND model is provided by VMI at 70 keV. This choice of VMI energy level (70 keV) aligns with the effective energy of a 120kV beam, which is equivalent or otherwise comparable to the single-energy CT images obtained from EID-CT scans conducted at 120 kV. In other implementations, however, the VMI energy level can be different from 70 keV when training the model for scanners with other kV or effective energy. The training focuses on learning monoenergetic imaging with the target energy level. In the illustrated example, the target energy level is a 100 keV VMI. As another non-limiting example, the target energy level may be 50 keV. In the inference stage, the trained DIAMOND model processes the single-energy CT images scanned at the tube potential, which in the illustrated example is 120 kV.

[0026] In a non-limiting example, the DIAMOND model may implement a neural network. The neural network may implement a U-Net architecture with nine modules, as an example. Each module involves convolution, batch normalization (BN), and exponential linear unit (eLU) activation operations sequentially. The max pooling layer and convolution transpose operator can also be applied. Concatenation can be added to the network to preserve the similarity between the input and output.

[0027] Additionally or alternatively, the DIAMOND model can be constructed to incorporate additional data to be generalizable and / or adaptable to various CT protocol settings. This can be achieved by introducing CT scanning and reconstruction parameters (e.g., tube potential (kV), image field-of-view (FOV), and reconstruction kernel) into the model. An example of this enhanced model framework is illustrated in FIG. 2. These enhancements can be encoded into the model, influencing its activation layers at multiple depths to accurately represent symmetric image characteristics across different feature levels. A tailored incremental training regimen can be adopted to progressively intensify the complexify of learning tasks to ensure thorough adaptability to varied imaging scenarios. This model framework aims to markedly enhance the accuracy and adaptability of multi-energy CT image generation. In some implementations, expert assessments of intermediate outcomes derived from developmental data can be used to inform the fine-tuning of neural network hyperparameters, such as network dimensions and depth, leveraging expert insights to refine the model’s performance further.

[0028] The disclosed DIAMOND framework can transfer the learned monoenergetic imaging knowledge from training data containing VMIs at different energy levels to singleenergy CT images (e.g., those acquired with EID-CT, PCD-CT operating in a single-energy mode, cone-beam CT, etc.) to enable multi-energy imaging capabilities. As an advantageous application, the disclosed framework can improve coronary stenosis assessment by decreasingblooming artifacts when compared to the original single-energy CT images, which can improve patient management. Additionally, using the VMIs generated using the DIAMOND framework, a fully automated stenosis quantification framework can be implemented, which can leverage the full potential of multi-energy features aided by advanced deep-learning algorithms to improve clinical diagnosis.

[0029] As will be described in more detail below, the systems and methods described in the present disclosure provide a technical solution for generating VMIs from single-energy CT image data using machine learning. This approach enables the creation of VMIs without requiring specialized multi-energy CT hardware, thereby improving the capabilities of existing single-energy CT systems.

[0030] In some embodiments, machine learning models (e.g., neural networks) are integrated with CT imaging protocols to synthesize VMIs at target energy levels from singleenergy CT images or VMIs with different energy levels. By incorporating CT scan protocol data as additional input to the machine learning model, the technique achieves improved generalizability’ across different CT systems and scanning parameters. This allows the same trained model to be applied across various CT scanner models and manufacturers, enhancing the practical applicability- of the solution.

[0031] The disclosed systems and methods also provide an approach for converting CT images acquired at one kV setting to equivalent images at a different kV setting. This capability addresses challenges in quantitative imaging and radiomics applications where CT numbers can vary based on acquisition parameters. By enabling consistent CT number representation across different kV settings, the systems and methods facilitate more robust quantitative analysis and comparison of CT images from diverse sources.

[0032] By leveraging existing CT hardware and integrating machine learning techniques, the systems and methods described in the present disclosure improve the field of CT imaging. For instance, the systems and methods expand the capabilities of single-energy CT systems to approach those of more advanced multi-energy systems, while maintaining broad compatibility’ across different CT platforms. This practical application of artificialintelligence in medical imaging enhances diagnostic accuracy, enables new quantitative analysis capabilities, and improves the overall utility of CT imaging in clinical practice.

[0033] Referring now to FIG. 3, a flowchart is illustrated as setting forth the steps of an example method for generating a VMI from single-energy CT data using a suitably trained neural network or other machine learning algorithm. As will be described, the neural network or other machine learning algorithm takes a single-energy CT image acquired at a tube potential as input data and generates a VMI at a target energy level as output data. In some implementations, the machine learning model additional takes CT scan protocol data associated with the CT scan protocol used to acquire the single-energy CT image data as an input.

[0034] The method includes accessing single-energy CT image data with a computer system, as indicated at step 302. Accessing the single-energy CT image data may include retrieving such data from a memory or other suitable data storage device or medium. Additionally or alternatively, accessing the single-energy CT image data may include acquiring such data with a CT system and transferring or otherwise communicating the data to the computer system, which may be a part of the CT system.

[0035] The tube potential of the CT system used to acquire the single-energy CT image data is then determined, as indicated at step 304. The tube potential may be determined by the user inputting or otherwise selecting the tube potential as a known parameter. Alternatively, the tube potential can be determined from HVL or phantom measurements.

[0036] CT scan protocol data are also accessed with the computer system, as indicated at step 306. The CT scan protocol data are associated with the scanning protocol used when acquiring the single-energy CT image data. Examples of parameters or other data contained in the CT scan protocol data can include tube potential (kV), image field-of-view (FOV), and reconstruction kernel.

[0037] A trained machine learning model (e.g., a neural network or other suitable machine learning model) is then accessed with the computer system, as indicated at step 308. In general, the machine learning model is trained, or has been trained, on training data in order to generate a VMI at a target energy level from a single-energy CT image acquired with a tube potential associated with an input energy level that is different from the target energy level. Accordingly, the machine learning model can be selected from a plurality of trained machine learning models each associated with a different particular tube potential. The machine learning model to be used can be selected from the plurality of machine learning models by matching the determined tube potential with the respective particular tube potential of the model.

[0038] Accessing the trained machine learning model may include accessing model parameters (e.g.. weights, biases, or both) that have been optimized or otherwise estimated by training the machine learning model on training data. In some instances, retrieving the machine learning model can also include retrieving, constructing, or otherwise accessing the particular model architecture to be implemented. For instance, data pertaining to the layers in a neural network architecture (e.g.. number of layers, type of layers, ordering of layers, connections between layers, hyperparameters for layers) may be retrieved, selected, constructed, or otherwise accessed.

[0039] An artificial neural network generally includes an input layer, one or more hidden layers (or nodes), and an output layer. Typically, the input layer includes as many nodes as inputs provided to the artificial neural network. The number (and the type) of inputs provided to the artificial neural network may vary based on the particular task for the artificial neural network.

[0040] The input layer connects to one or more hidden layers. The number of hidden layers varies and may depend on the particular task for the artificial neural network. Additionally, each hidden layer may have a different number of nodes and may be connected to the next layer differently. For example, each node of the input layer may be connected to each node of the first hidden layer. The connection between each node of the input layer and each node of the first hidden layer may be assigned a weight parameter. Additionally, each node of the neural network may also be assigned a bias value. In some configurations, each node of the first hidden layer may not be connected to each node of the second hidden layer. That is, there may be some nodes of the first hidden layer that are not connected to all of the nodes of the second hidden layer. The connections between the nodes of the first hidden layers and the second hidden layers are each assigned different weight parameters. Each node of the hidden layer is generally associated with an activation function. The activation function defines how the hidden layer is to process the input received from the input layer or from a previous input or hidden layer. These activation functions may vary and be based on the type of task associated with the artificial neural network and also on the specific type of hidden layer implemented.

[0041] Each hidden layer may perform a different function. For example, some hidden layers can be convolutional hidden layers which can, in some instances, reduce the dimensionality of the inputs. Other hidden layers can perform statistical functions such as max pooling, which may reduce a group of inputs to the maximum value; an averaging layer; batchnormalization; and other such functions. In some of the hidden layers each node is connected to each node of the next hidden layer, which may be referred to then as dense layers. Some neural networks including more than, for example, three hidden layers may be considered deep neural networks.

[0042] The last hidden layer in the artificial neural network is connected to the output layer. Similar to the input layer, the output layer typically has the same number of nodes as the possible outputs. For example, the output layer can include an output for a VMI at the target energy level.

[0043] The single-energy CT image data are then input to the trained machine learning model associated with the tube potential used when acquiring the single-energy CT image data, generating a VMI at the target energy level as an output, as indicated at step 310.

[0044] Additionally or alternatively, other data can be input to the trained machine learning model. For example, the CT scan protocol data can be input to the machine learning model as an additional input. In some implementations, the CT scan protocol data can first be encoded using a protocol encoder and the encoded CT scan protocol data can then be input to the machine learning model.

[0045] The VMI at the target energy level can then be displayed to a user, stored for later use or further processing, or both, as indicated at step 312. As one example, the target energy VMI can be further processed to provide a stenosis assessment for the subject. For instance, the VMI can be processed to quantify the percent diameter stenosis. The processed VMI can additionally be classified or otherwise analyzed to assign each identified lesion in the processed VMI with a stenosis severity category, which can be determined by evaluating the measured stenosis percentage, resulting in distinct classifications: absence of stenosis (0%), minimal stenosis (1-24%), mild stenosis (25-49%), moderate stenosis (50-69%), severe stenosis (70-99%), and total occlusion (100%) according to the CAD-RADS scoring system.

[0046] Referring now to FIG. 4, a flowchart is illustrated as setting forth the steps of an example method for training one or more machine learning models (e.g., neural networks or other suitable machine learning models) on training data, such that the one or more machine learning models are trained to receive single-energy CT image data as input data in order to generate VMIs at a target energy level as output data.

[0047] In general, the machine learning model may include a neural network. The neural network(s) can implement any number of different neural network architectures. For instance, the neural network(s) could implement a convolutional neural network, a residualneural network, or the like. As one non-limiting example, the machine learning model may be a neural network implemented with a U-Net architecture with nine modules. Each module can involve convolution, batch normalization (BN), and exponential linear unit (eLU) activation operations sequentially. Alternatively, activation functions other than eLU activation may be used. A max pooling layer and convolution transpose operator can also be applied. Concatenation can also be added to the network to preserve the similarity between the input and output.

[0048] Alternatively, the neural network(s) could be replaced with other suitable machine learning or artificial intelligence algorithms or models, such as those based on supervised learning, unsupervised learning, deep learning, ensemble learning, adversarial learning (e.g., as may be used in a generative adversarial network), dimensionality reduction, and so on.

[0049] The method includes accessing training data with a computer system, as indicated at step 402. Accessing the training data may include retrieving such data from a memory or other suitable data storage device or medium. Alternatively, accessing the training data may include acquiring such data with one or more CT systems and transferring or otherwise communicating the data to the computer system.

[0050] In general, the training data can include VMIs generated for various energylevels. In general, the training data can include VMIs for at least two different energy levels. The first energy level may be an input energy level associated with a tube potential of a CT system. For example, the first energy level may be 70 keV, which is associated with a tube potential of 120 kV for a CT system. The second energy level may be a target energy level, which is the energy level of the VMI that will be generated by the trained machine learning model.

[0051] Additionally, the training data may include other data, such as CT scan protocol data containing CT scan parameters or other data associated with the scanning protocols used when acquiring the VMIs or related CT images. In some embodiments, the training data may include VMIs or other relevant data that have been labeled (e.g.. labeled as containing patterns, features, or characteristics indicative of different energy levels; and the like).

[0052] The method can include assembling training data from VMIs and other relevant data using a computer system. This step may include assembling the VMIs and other relevant data into an appropriate data structure on which the neural network or other machine learning algorithm can be trained. Assembling the training data may include assembling VMIs,segmented VMIs, and other relevant data. For instance, assembling the training data may include generating labeled data and including the labeled data in the training data. Labeled data may include VMIs, segmented VMIs, or other relevant data that have been labeled as belonging to, or otherwise being associated with, one or more different classifications or categories. For instance, labeled data may include VMIs and / or segmented VMIs that have been labeled as being associated with different energy levels, effective beam energies, or the like.

[0053] One or more machine learning models are trained on the training data, as indicated at step 404. In general, the machine learning model can be trained by optimizing model parameters (e.g., weights, biases, or both) based on minimizing a loss function. As one non-limiting example, the loss function may be a mean squared error loss function.

[0054] When the machine learning model includes a neural network, training a neural network may include initializing the neural network, such as by computing, estimating, or otherwise selecting initial network parameters (e.g., weights, biases, or both). During training, an artificial neural network receives the inputs for a training example and generates an output using the bias for each node, and the connections between each node and the corresponding weights. For instance, training data can be input to the initialized neural network, generating output as a VMI at the target energy level. The artificial neural netw ork then compares the generated output with the actual output of the training example in order to evaluate the quality of the target energy level VMI. For instance, the target energy level VMI can be passed to a loss function to compute an error. The current neural network can then be updated based on the calculated error (e.g., using backpropagation methods based on the calculated error). For instance, the current neural network can be updated by updating the network parameters (e.g., weights, biases, or both) in order to minimize the loss according to the loss function. The training continues until a training condition is met. The training condition may correspond to, for example, a predetermined number of training examples being used, a minimum accuracy threshold being reached during training and validation, a predetermined number of validation iterations being completed, and the like. When the training condition has been met (e.g., by determining whether an error threshold or other stopping criterion has been satisfied), the current neural network and its associated network parameters represent the trained neural network. Different types of training processes can be used to adjust the bias values and the weights of the node connections based on the training examples. The training processes may include, for example, gradient descent, Newton’s method, conjugate gradient, quasi-Newton, Levenberg-Marquardt. among others.

[0055] The artificial neural network, or other machine learning model, can be constructed or otherwise trained based on training data using one or more different learning techniques, such as supervised learning, unsupervised learning, reinforcement learning, ensemble learning, active learning, transfer learning, adversarial learning, or other suitable learning techniques for neural networks or other machine learning models. As an example, supervised learning involves presenting a computer system with example inputs and their actual outputs (e.g., categorizations). In these instances, the artificial neural network is configured to learn a general rule or model that maps the inputs to the outputs based on the provided example input-output pairs.

[0056] In some implementations, the machine learning model can additionally be trained using CT scan protocol data and / or encoded CT scan protocol data, as described above. In these instances, the CT scan protocol data can be encoded into the model, influencing its activation layers at multiple depths to accurately represent symmetric image characteristics across different feature levels. When additionally training the machine learning model on CT scan protocol data and / or encoded CT scan protocol data, a tailored incremental training regimen can be used to progressively intensify the complexity of learning tasks to ensure thorough adaptability to varied imaging scenarios. In some implementations, intermediate outcomes derived from developmental data can be used to inform fine-tuning of machine learning model hyperparameters, such as network dimensions and depth.

[0057] The one or more machine learning models are then stored for later use, as indicated at step 406. Storing the machine learning model(s) may include storing model parameters (e.g., weights, biases, or both), which have been computed or otherwise estimated by training the machine learning model(s) on the training data. Storing the machine learning model(s) may also include storing the particular model architecture to be implemented. For instance, data pertaining to the layers in a neural network architecture (e.g., number of layers, type of layers, ordering of layers, connections between layers, hyperparameters for layers) may be stored.

[0058] In an example study, the systems and methods described in the present disclosure were evaluated by generating VMIs at a target energy level from single-energy CT image data using a suitably trained neural network. The neural network was trained on a retrospectively collected patient cohort that included 10 cases acquired from a dual-source PCD-CT system in multi-energy mode. Cases were randomly split into training sets (n=6) and testing sets (n=4). In each case, VMIs at 50, 70, and 100 keV were reconstructed following astandard clinical protocol, with an iterative reconstruction (IR) algorithm at strength 4, a Bv60 kernel, 0.6 mm slice thickness, 1024x 1024 matrix, and 200x200 mm2field of view. For the network training, 70,960 paired patches with the size of 128 x 128 pixels were randomly extracted from the 70 VMIs as the input and from the 50 keV or 100 keV VMIs as the target. The initial learning rate was set as 0.001 with a scheduled descent to 0.00001. The Adam optimizer was chosen to minimize the mean-squared-error (MSE) loss function.

[0059] To evaluate the inference stage of the trained neural network, an experiment using PCD-CT, which provides a true VMI, was conducted. The neural network was trained to produce 50 and 100 keV VMIs (target energy7levels), which were then compared to the actual 50 and 100 keV VMIs reconstructed directly from the scanner. As shown in FIG. 5, the 50 keV VMI created by the trained neural network from the 70 keV VMI input was visually similar to the true 50 keV VMI. Moreover, quantitative measurements across different regions-of-interest (ROIs), representing different materials, revealed similar CT numbers between the true VMIs and the neural network output, indicating that the trained neural network accurately learned the spectral information from the single-energy equivalent input.

[0060] In another experiment, the trained neural network was applied to EID-CT scans performed using third-generation dual-source scanners following a clinical cCTA protocol. The scanning parameters used were automatic tube potential selection with a reference tube potential of 120 kV, automatic exposure control with a quality reference tube-current-time product of 120 mAs / rotation, 0.25-second rotation time, and 192 x 0.6 mm collimation (with z flying focal spot physical collimation was 96 x 0.6 mm). Images were reconstructed using 0.6 / 0.3 mm slice thickness / increment, the Bv40 kernel, a 512 matrix size, and iterative reconstruction (ADMIRE) at strength 4.

[0061] FIG. 6 shows the phantom results from the neural network (100 keV) together with the original SECT. The results demonstrate that the neural network effectively mitigated blooming artifacts, leading to enhanced visualization of the lumen in contrast to the original SECT images. Moreover, the analysis conducted through software revealed a significant decrease in stenosis percent area when utilizing the trained neural network. Notably, the stenosis percentage area reductions in Phantom- A, Phantom-B, and Phantom-C are 13%, 8%, and 6%, respectively, approaching values that closely align with the ground truth measurements.

[0062] As illustrated in FIG. 7, the trained neural network effectively reduced blooming artifacts caused by calcium, resulting in better lumen visualization compared to theEID-CT inputs, bringing the images closer to those produced by PCD-CT. FIG. 8 shows that across all ten patients, the percent diameter stenosis identified using VMIs output by the trained neural network was reduced compared to their original EID-CT images. Specifically, the average percent diameter stenosis from the original EID-CT was 43.3%, known from phantom experiments to overestimate stenosis severity. This figure dropped to 30.5% when processed with the trained neural network, aligning closely with the 28.6% seen in UHR PCD-CT, which is considered a more accurate representation of true stenosis. When using UHR PCD-CT as the reference, the VMIs generated using the trained neural netw ork overestimated stenosis by only 6.6%, a significant improvement over the 51% overestimation in the original EID-CT data.

[0063] In some implementations, the models illustrated in FIGS. 1 and 2 can be adapted to convert CT images acquired using a first kV setting to CT images as if they had been acquired using a second kV seting that is different from the first kV seting. Advantageously, this implementation can be clinically useful because patients are often scanned with the kV seting adjusted specific to the patient (e.g., adjusted to the patient’s size, exam ty pe, etc.). For example, a lower kV setting may’ be used for smaller-sized patients, contrast-enhanced exams, etc.

[0064] Advantageously, implementing the DIAMOND model framework to convert CT images between different kV setings enables comparisons between quantitative imaging metrics. Applications in quantitative imaging, radiomics. and Al rely on quantitative CT numbers. The same patient / organ scanned with different kV setings will have different CT numbers, which impose a challenge for quantitative applications such as radiomics and Al. Using the DIAMOND model framework to convert CT images between kV settings provides an effective data harmonization method that allows these quantitative applications to use patient data scanned with any kV seting. This can be particularly beneficial for studies that rely on large data sets, many of which are retrospective data sets where the kV settings used cannot be retrospectively changed.

[0065] Referring now to FIG. 9, a flowchart is illustrated as seting forth the steps of an example method for using a suitably trained machine learning model to generate a CT image having properties as if it w ere acquired using a kV seting that is different from the kV seting used to acquire the CT image input to the machine learning model. As will be described, the neural netw ork or other machine learning algorithm takes a CT image acquired using a first kV seting (e.g., tube potential) as input data and generates a second CT image as output data, where the second CT image corresponds to a CT image having been acquired at a second kVsetting that is different than the first kV setting. Additionally or alternatively, the output image may be a VMI having an effective energy level corresponding to the second kV setting (e.g., 70 keV for a 120 kV setting). In some implementations, the machine learning model additionally receives as an input CT scan protocol data associated with the CT scan protocol used to acquire the CT image data.

[0066] Additionally or alternatively, the machine learning model framework can be used to receive a first CT image acquired using a first kV setting as an input, and to generate a second CT image as an output, where the second CT image corresponds to a CT image as it would have been acquired using a second kV setting that is different from the first kV setting.

[0067] The method includes accessing CT image data with a computer system, as indicated at step 902. Accessing the CT image data may include retrieving such data from a memory or other suitable data storage device or medium. Additionally or alternatively, accessing the CT image data may include acquiring such data with a CT system and transferring or otherwise communicating the data to the computer system, which may be a part of the CT system.

[0068] Additionally, the tube potential or kV setting of the CT system used to acquire the CT image data may be determined. The tube potential or kV setting may be determined by the user inputting or otherwise selecting the tube potential or kV setting as a known parameter. Alternatively, the tube potential can be determined from HVL or phantom measurements. Additionally or alternatively, CT scan protocol data can also be accessed with the computer system. The CT scan protocol data are associated with the scanning protocol used when acquiring the CT image data. Examples of parameters or other data contained in the CT scan protocol data can include image field-of-view (FOV), reconstruction kernel, and the like.

[0069] A trained machine learning model (e.g., a neural network or other suitable machine learning model) is then accessed with the computer system, as indicated at step 904. In general, the machine learning model is trained, or has been trained, on training data in order to generate a CT image at a second kV setting from a CT image acquired using a first kV setting that is different from the first kV setting. Accordingly, the machine learning model can be selected from a plurality of trained machine learning models each associated with different kV settings. The machine learning model to be used can be selected from the plurality of machine learning models by matching the input and target output kV settings.

[0070] Accessing the trained machine learning model may include accessing model parameters (e.g.. weights, biases, or both) that have been optimized or otherwise estimated bytraining the machine learning model on training data. In some instances, retrieving the machine learning model can also include retrieving, constructing, or otherwise accessing the particular model architecture to be implemented. For instance, data pertaining to the layers in a neural network architecture (e.g., number of layers, type of layers, ordering of layers, connections between layers, hyperparameters for layers) may be retrieved, selected, constructed, or otherwise accessed.

[0071] The CT image data are then input to the trained machine learning model, generating second CT image data at the target kV setting as an output, as indicated at step 906. Additionally or alternatively, other data can be input to the trained machine learning model. For example, the CT scan protocol data can be input to the machine learning model as an additional input. In some implementations, the CT scan protocol data can first be encoded using a protocol encoder and the encoded CT scan protocol data can then be input to the machine learning model.

[0072] The second CT image data can then be displayed to a user, stored for later use or further processing, or both, as indicated at step 908.

[0073] Referring now to FIG. 10, a flowchart is illustrated as setting forth the steps of an example method for training one or more machine learning models (e.g., neural networks or other suitable machine learning models) on training data, such that the one or more machine learning models are trained to receive CT image data acquired using a first kV setting as input data in order to generate second CT image data as if they were acquired using a second kV setting that is different from the first kV setting as output data, or alternatively a VMI having an effective energy level corresponding to the second kV setting.

[0074] In general, the machine learning model may include a neural network. The neural network(s) can implement any number of different neural network architectures. For instance, the neural network(s) could implement a convolutional neural network, a residual neural network, or the like. As one non-limiting example, the machine learning model may be a neural network implemented with a U-Net architecture with nine modules. Each module can involve convolution, batch normalization (BN), and exponential linear unit (eLU) activation operations sequentially. Alternatively, activation functions other than eLU activation may be used. A max pooling layer and convolution transpose operator can also be applied. Concatenation can also be added to the network to preserve the similarity between the input and output.

[0075] Alternatively, the neural network(s) could be replaced with other suitable machine learning or artificial intelligence algorithms or models, such as those based on supervised learning, unsupervised learning, deep learning, ensemble learning, adversarial learning (e.g., as may be used in a generative adversarial network), dimensionality reduction, and so on.

[0076] The method includes accessing training data with a computer system, as indicated at step 1002. Accessing the training data may include retrieving such data from a memory or other suitable data storage device or medium. Alternatively, accessing the training data may include acquiring such data with one or more CT systems and transferring or otherwise communicating the data to the computer system.

[0077] In general, the training data can include CT images acquired with different kV settings. In general, the training data can include CT images acquired using at least two different kV settings. The first kV setting may be an input kV setting corresponding to a tube potential of a CT system. For example, the first kV setting may be 80 kV. The second kV setting may be a target kV setting.

[0078] Additionally, the training data may include other data, such as CT scan protocol data containing CT scan parameters or other data associated with the scanning protocols used when acquiring the CT images. In some embodiments, the training data may include CT images or other relevant data that have been labeled (e.g., labeled as containing patterns, features, or characteristics indicative of different energy levels and / or kV settings; and the like).

[0079] The method can include assembling training data from CT images and other relevant data using a computer system. This step may include assembling the CT images and other relevant data into an appropriate data structure on which the neural network or other machine learning algorithm can be trained. Assembling the training data may include assembling CT images, segmented CT images, and other relevant data. For instance, assembling the training data may include generating labeled data and including the labeled data in the training data. Labeled data may include CT images, segmented CT images, or other relevant data that have been labeled as belonging to, or otherwise being associated with, one or more different classifications or categories. For instance, labeled data may include CT images and / or segmented CT images that have been labeled as being associated with different kV settings, energy' levels, effective beam energies, or the like.

[0080] One or more machine learning models are trained on the training data, as indicated at step 1004. In general, the machine learning model can be trained by optimizingmodel parameters (e.g., weights, biases, or both) based on minimizing a loss function. As one non-limiting example, the loss function may be a mean squared error loss function.

[0081] When the machine learning model includes a neural network, training a neural network may include initializing the neural network, such as by computing, estimating, or otherwise selecting initial network parameters (e.g., weights, biases, or both). During training, an artificial neural network receives the inputs for a training example and generates an output using the bias for each node, and the connections betw een each node and the corresponding weights. For instance, training data can be input to the initialized neural netw ork, generating output as a new CT image as a CT image as if it w ere acquired at the target kV setting. The artificial neural network then compares the generated output w ith the actual output of the training example in order to evaluate the quality of the new CT image. For instance, the new CT image can be passed to a loss function to compute an error. The current neural netw ork can then be updated based on the calculated error (e.g., using backpropagation methods based on the calculated error). For instance, the current neural network can be updated by updating the network parameters (e.g., weights, biases, or both) in order to minimize the loss according to the loss function. The training continues until a training condition is met. The training condition may correspond to, for example, a predetermined number of training examples being used, a minimum accuracy threshold being reached during training and validation, a predetermined number of validation iterations being completed, and the like. When the training condition has been met (e.g.. by determining whether an error threshold or other stopping criterion has been satisfied), the current neural network and its associated network parameters represent the trained neural netw ork. Different types of training processes can be used to adjust the bias values and the weights of the node connections based on the training examples. The training processes may include, for example, gradient descent, Newton's method, conjugate gradient, quasi-Newton, Levenberg-Marquardt, among others.

[0082] The artificial neural network, or other machine learning model, can be constructed or otherwise trained based on training data using one or more different learning techniques, such as supervised learning, unsupervised learning, reinforcement learning, ensemble learning, active learning, transfer learning, adversarial learning, or other suitable learning techniques for neural networks or other machine learning models. As an example, supervised learning involves presenting a computer system with example inputs and their actual outputs (e.g.. categorizations). In these instances, the artificial neural network isconfigured to learn a general rule or model that maps the inputs to the outputs based on the provided example input-output pairs.

[0083] In some implementations, the machine learning model can additionally be trained using CT scan protocol data and / or encoded CT scan protocol data, as described above. In these instances, the CT scan protocol data can be encoded into the model, influencing its activation layers at multiple depths to accurately represent symmetric image characteristics across different feature levels. When additionally training the machine learning model on CT scan protocol data and / or encoded CT scan protocol data, a tailored incremental training regimen can be used to progressively intensify the complexity of learning tasks to ensure thorough adaptability to varied imaging scenarios. In some implementations, intermediate outcomes denved from developmental data can be used to inform fine-tuning of machine learning model hyperparameters, such as network dimensions and depth.

[0084] The one or more machine learning models are then stored for later use, as indicated at step 1006. Storing the machine learning model(s) may include storing model parameters (e.g., weights, biases, or both), which have been computed or otherwise estimated by training the machine learning model(s) on the training data. Storing the machine learning model(s) may also include storing the particular model architecture to be implemented. For instance, data pertaining to the layers in a neural network architecture (e.g., number of layers, type of layers, ordering of layers, connections between layers, hyperparameters for layers) may be stored.

[0085] In another example study, the deep learning-based neural networks described in the present disclosure were evaluated for providing virtual monoenergetic imaging capabilities, reducing blooming artifacts, improving stenosis assessments, and achieving results comparable to those of UHR PCD-CT.

[0086] The trained convolutional neural network, DIAMOND, used in this example study reduced blooming artifacts and decreased the average percent diameter stenosis from 35.65% in EID-CT to 25.19%, closely approaching the 24.27% measured in ultra-high resolution (UHR) PCD-CT (p<0.05). Strong agreement was noted between DIAMOND and PCD-CT, consistently showing lower stenosis measurements than EID-CT (p<0.05). The disclosed methods enable virtual monoenergetic imaging on single-energy cardiac EID-CT systems without costly upgrades and loss of temporal resolution, significantly enhancing stenosis quantification.

[0087] In the example study, the systems and methods described in the present disclosure were used to generate VMIs for SECT to provide spectral capabilities without incurring additional upgrade costs and without sacrificing temporal resolution. The DIAMOND model was trained to map SECT images to high-keV VMIs using paired PCD-CT data, with input keV matched to the effective energy of the SECT. Operating directly on CT images, DIAMOND required no projection data or proprietary information, enabling broad applicability across scanner models.

[0088] The example study included a retrospective study of subject with suspected CAD enrolled in a prospective cCTA cohort. All subjects underwent clinical EID-CT followed by same day PCD-CT research scan. Thirty-two patients with dense coronary artery calcifications and a phantom with known stenosis values (50% / 25% / l 5%) were included. Ten retrospective cases (6 training, 4 validation) were used to train the DIAMOND model with PCD-CT VMI data. The remaining 22 prospective cases and the phantom were assessed using EID-CT, PCD-CT, and DIAMOND-processed images.

[0089] Ten patients scanned with dual-source PCD-CT in standard mode (144 x 0.4 mm collimation) were used for the training dataset. VMIs at 70 and 100 keV were reconstructed with iterative reconstruction (IR, strength 4), medium-sharp vascular kernel (Bv60), 0.6 mm slice thickness, 1024 x 1024 matrix, and 200 x 200 mm2field of view. The 70 keV VMI, matching the effective energy of the 120 kV SECT, was input, while the 100 keV VMI served as label, given its ability to reduce blooming artifacts.

[0090] EID-CT testing datasets included scans acquired with a dual-source EID-CT system with the following cCTA protocol: automatic tube potential selection (CARE kV, 120 kV reference), automatic exposure control (CAREDose 4D) with a 120 mAs / rotation quality reference, 0.25-second rotation time, 192 x 0.6 mm collimation (z-flying focal spot; physical collimation: 96 x 0.6 mm). Images were reconstructed at 0.6 mm thickness, 0.3 mm increment, Bv40 kernel, 512 x 512 matrix, and IR (ADMIRE, strength 4).

[0091] Same-day PCD-CT research scans were performed after EID-CT, using retrospective ECG-gated spiral acquisition in ultra-high resolution (UHR)-only mode (120 x 0.2 mm collimation, 0.25-second rotation). VMIs were unavailable in UHR-only mode. Taskbased CARE keV automatic selection was applied: 120 kV / 50 CARE keV image quality (IQ) for patients >90 kg, and 90 kV / 100 CARE keV IQ for patients <90 kg. Sharp reconstruction kernels maximized UHR detector benefits but increased image noise, particularly in larger patients. To compensate, slice thicknesses (0.2-0.6 mm) were adjusted by weight (i.e.. thickerslices were used for larger patients). Phantom reconstructions used a 0.6 mm thickness for both PCD-CT and EID-CT.

[0092] FIG. 11 illustrates example pipelines for DIAMOND inference and training. SECT images (e.g., 120 kV) were mapped to 100 keV VMI-like outputs, reducing blooming and enhancing lumen visibility for stenosis quantification. The network training used 70 keV VMIs from PCD-CT as inputs, matching the effective energy of a 120 kV SECT, while 100 keV VMIs acted as labels. A simplified U-Net architecture including nine modules was used. Each module sequentially performed convolution, batch normalization, and exponential linear unit activation operations. The architecture included max pooling layers, convolution transpose operators, and concatenation to maintain input-output similarity. The mean-squared-error (MSE) loss function was optimized during training. The final training data included 70.965 paired patches with the size of 128 x 128 pixels from the 70 and 100 keV VMIs as the input and label, respectively, from the training set (n=6) and 7,885 from validation set (n=4), at a 9: 1 ratio. Training began with an initial learning rate of 0.001, progressively reduced to 0.00001, using the Adam optimizer to minimize MSE loss function. The training was set for 100 epochs to ensure model convergence.

[0093] A 4.5 mm phantom (FIG. 6) with 50%, 25%, and 15% diameter stenosis (375 mg / ccm CaHA, iodine-filled lumens, -300 HU at 120 kV) was scanned in 35-cm anthropomorphic thorax phantom using EID-CT and UHR PCD-CT, following the previously described protocols. Stenosis was quantified using Syngo.Via software (VB70. Siemens Healthineers): For each lesion, a central marker was placed at the most severe stenosis point, with two additional markers in healthy segments — one proximal and one distal to the lesion — to determine the average lumen area. The percent diameter stenosis was automatically calculated by the software.

[0094] For each patient, the most severe stenosis caused by dense calcification was identified in each coronary segment: left main (LM), left anterior descending (LAD), right coronary artery' (RCA), and circumflex (CX). Arteries affected by severe motion artifacts or long-segment stenosis that prevented accurate placement of healthy reference markers were excluded from analysis.

[0095] The software automatically extracted coronary artery' centerlines and performed lumen segmentation. Once a lesion w as identified, a central marker was placed at the most severe point of stenosis, and two reference markers were placed in adjacent healthy segments (one proximal and one distal). These reference points were used to estimate the average luminaldiameter of the normal vessel. All segmentations were manually reviewed and adjusted when necessary. The ratio of the effective diameters was calculated as:

[0096] where Dceritral markeris the diameter at the central stenosis and Dreference{sis the average diameter of the two healthy reference segments. The percentage diameter stenosis (PDS) was computed using:PDS = 100%x (l - 7?)

[0097] Visual matching was used to ensure that the same lesion was measured across EID-CT (with and without DIAMOND) and PCD-CT images.

[0098] For patient data, 22 cases underwent stenosis quantification in Syngo.Via with visual matching across EID-CT. DIAMOND, and PCD-C. Stenosis severity was categorized using CAD-RADS criteria: absence (0%), minimal (1-24%), mild (25-49%). moderate (50- 69%), severe (70-99%), or total occlusion (100%). Differences in CAD-RADS scores betw een modalities w ere summarized.

[0099] Statistical analyses used SciPy (vl.7.3) and Scikit-leam (vl.0.2). Agreement among EID-CT, DIAMOND, and PCD-CT for percent diameter stenosis was assessed with Bland- Altman analysis, reporting median bias and non-parametric limits of agreement (2.5th- 97.5th percentiles). Systematic and proportional biases w ere evaluated via linear regression of residuals, with slopes > 0.1 indicating significant proportional bias.

[0100] Intraclass Correlation Coefficient (ICC) with 95% confidence intervals (Cis) quantified agreement: < 0.40 (poor). 0.40-0.59 (moderate), 0.60-0.74 (good), and > 0.75 (excellent). Residual plots w ere used to assess measurement variation, and linear regression on residuals quantified bias (slope) and proportionality7(R2). Significance was set at p<0.05.

[0101] DIAMOND significantly reduced blooming artifacts and enhanced lumen visualization relative to EID-CT in the phantom. Quantitative software analysis showed a reduction in percent diameter stenosis with DIAMOND: 69% to 56% (Phantom-A), 36% to 28% (Phantom-B), and 21% to 15% (Phantom-C). The full width at half maximum (FWHM) values of attenuation line profiles across the calcified plaques were consistently lower with DIAMOND than with EID-CT: 1.37 mm vs. 1.56 mm (Phantom A), 1.17 mm vs. 1.37 mm (Phantom B), and 0.98 mm vs. 1.17 mm (Phantom C).

[0102] Twenty-six coronary artery segments were assessed across 22 patients: left main (LM. n = 1), left anterior descending (LAD. n = 10), right coronary artery (RCA, n = 10), and circumflex (CX, n = 5). Mean percent diameter stenosis decreased from 35.65% (standard deviation [SD] : 15.10%) with conventional EID single-energy CT to 25.19% (SD: 13.24%, p < 0.05) with DIAMON, comparable to the PCD-CT (24.27%, SD: 12.48%, p < 0.05). CAD- RADS reclassification occurred in 11 / 26 lesions (42%): 4 downgraded from moderate to mild and 7 from mild to minimal.

[0103] Bland-Altman analysis comparing PCD-CT with EID-CT demonstrated a proportional bias, with amedian difference of -10.00% (95% bootstrap CI: [-14.46%, -6.65%]), slope of -0.220, and R2of 0.070. ICC was 0.707, indicating EID-CT overestimates stenosis measurements relative to PCD-CT. Applying DIAMOND to EID-CT data significantly improved alignment with PCD-CT results. Comparison between DIAMOND and PCD-CT demonstrated strong agreement (median difference: -1.50%, 95% bootstrap CI: [-3.19%, 1.96%]), minimal proportional bias (slope: 0.063, R2: 0.012), and narrow LoA. ICC was 0.847, confirming strong agreement. Both DIAMOND and PCD-CT provided lower, more accurate stenosis measurements than EID-CT.

[0104] DIAMOND allowed EID-CT images to be transformed into VMIs, reducing blooming artifacts and improving stenosis assessment to levels comparable to UHR PCD-CT. While 100 keV VMIs were used here, DIAMOND can generate VMIs at any keV by adjusting the training labels, enabling flexible multi-energy imaging.

[0105] While EID-CT systems can generate VMIs using dual-energy CT, adoption in cCTA is limited by availability, cost, and reduced temporal resolution in certain systems (e.g., dual-source scanners and scanners with fast kV switching). Prior deep learning methods for generating multi-energy images from SECT often lack generalizability due to reliance on scanner-specific data. By contrast, DIAMOND is trained on VMIs and maps SECT images to target keV output, enabling robust generalization across various EID-CT systems. During inference, the only scanner-specific input required is effective energy', which may be determined by half-value layer (HVL), phantom measurements, scanner metadata, and / or standardized lookup tables based on common acquisition protocols. This allows the use of a single patient-derived training dataset by matching the VMIs to the EID-CT scanner’s effective energy. A unique advantage of VMIs is that the CT number at a given energy is a physical property independent of scanner manufacturer or model, unlike single-energy images, which depend on x-ray spectra and hardware.

[0106] While this example study used PCD-CT to generate VMIs, they can also be produced by other dual- or multi-energy CT systems, including fast kV-switching, dual-source, and dual-layer detectors. VMIs at the same keV yield consistent CT numbers across different systems. Because DIAMOND was trained with energy-defined VMI pairs rather than scannerspecific features, it can be applied across platforms by matching the input keV to the scanner’s effective energy. DIAMOND processing takes -0.21 seconds per axial slice (NVIDIA RTX 3090 GPU). While processing time may vary with computational resources, it remains comparable to standard cCTA post-processing and can be seamlessly integrated as a plug-in within workstations or cloud-based workflows, without disrupting routine clinical operations.

[0107] Based on this example study, it was observed that the DIAMOND model significantly enhanced EID-CT. reducing blooming artifacts and improving the accuracy of coronary stenosis assessments without the need for costly upgrades. This led to stenosis severity measurements that closely approximated those observed with UHR PCD-CT.

[0108] FIG. 12 shows an example of a system 1200 for generating VMIs at a target energy level from single-energy CT images in accordance with some embodiments described in the present disclosure. Additionally or alternatively, the system 12 can be used to generate new CT images corresponding to CT images having been acquired at different kV settings from input CT images. As shown in FIG. 12, a computing device 1250 can receive one or more types of data (e.g., single-energy CT image data or other CT image data, kV setting data or other CT scan protocol data) from data source 1202. In some embodiments, computing device 1250 can execute at least a portion of a DIAMOND image generation system 1204 to generate images (e.g., VMIs, CT images) from data received from the data source 1202.

[0109] Additionally or alternatively, in some embodiments, the computing device 1250 can communicate information about data received from the data source 1202 to a server 1252 over a communication network 1254, which can execute at least a portion of the DIAMOND image generation system 1204. In such embodiments, the server 1252 can return information to the computing device 1250 (and / or any other suitable computing device) indicative of an output of the DIAMOND image generation system 1204.

[0110] In some embodiments, computing device 1250 and / or server 1252 can be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smartphone, a tablet computer, a wearable computer, a server computer, a virtual machine being executed by a physical computing device, and so on. The computing device 1250 and / or server 1252 can also reconstruct images from the data.

[0111] In some embodiments, data source 1202 can be any suitable source of data (e.g., measurement data, images reconstructed from measurement data, processed image data, CT scan protocol data), such as a CT system, another computing device (e g., a server storing measurement data, images reconstructed from measurement data, processed image data, CT scan protocol data), and so on. In some embodiments, data source 1202 can be local to computing device 1250. For example, data source 1202 can be incorporated with computing device 1250 (e.g., computing device 1250 can be configured as part of a device for measuring, recording, estimating, acquiring, or otherwise collecting or storing data). As another example, data source 1202 can be connected to computing device 1250 by a cable, a direct wireless link, and so on. Additionally or alternatively, in some embodiments, data source 1202 can be located locally and / or remotely from computing device 1250, and can communicate data to computing device 1250 (and / or server 1252) via a communication network (e g., communication network 1254).

[0112] In some embodiments, communication network 1254 can be any suitable communication network or combination of communication networks. For example, communication network 1254 can include a Wi-Fi network (which can include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e g., a 3G network, a 4G network, etc., complying with any suitable standard, such as CDMA, GSM, LTE, LTE Advanced. WiMAX, etc.), other types of wireless network, a wired network, and so on. In some embodiments, communication network 1254 can be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi -private netw ork (e.g., a corporate or university intranet), any other suitable type of network, or any suitable combination of networks. Communications links shown in FIG. 12 can each be any suitable communications link or combination of communications links, such as wired links, fiber optic links, Wi-Fi links, Bluetooth links, cellular links, and so on.

[0113] Referring now to FIG. 13, an example of hardw are 1300 that can be used to implement data source 1202, computing device 1250, and server 1252 in accordance with some embodiments of the systems and methods described in the present disclosure is shown.

[0114] As shown in FIG. 13, in some embodiments, computing device 1250 can include a processor 1302, a display 1304, one or more inputs 1306, one or more communication systems 1308, and / or memory71310. In some embodiments, processor 1302 can be any suitable hardware processor or combination of processors, such as a central processing unit (“CPU"’), a graphics processing unit (“GPU’7), and so on. In some embodiments, display 1304 can includeany suitable display devices, such as a liquid crystal display (“LCD’") screen, a light-emitting diode (“LED7’) display, an organic LED (“OLED”) display, an electrophoretic display (e.g., an “e-ink” display), a computer monitor, a touchscreen, a television, and so on. In some embodiments, inputs 1306 can include any suitable input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, and so on.

[0115] In some embodiments, communications systems 1308 can include any suitable hardware, firmware, and / or software for communicating information over communication network 1254 and / or any other suitable communication networks. For example, communications systems 1308 can include one or more transceivers, one or more communication chips and / or chip sets, and so on. In a more particular example, communications systems 1308 can include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and so on.

[0116] In some embodiments, memory’ 1310 can include any suitable storage device or devices that can be used to store instructions, values, data, or the like, that can be used, for example, by processor 1302 to present content using display 1304, to communicate with server 1252 via communications system(s) 1308, and so on. Memory 1310 can include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1310 can include random-access memory (“RAM”), read-only memory (“ROM”), electrically programmable ROM (“EPROM”), electrically erasable ROM (“EEPROM”), other forms of volatile memory7, other forms of non-volatile memory7, one or more forms of semi-volatile memory7, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, and so on. In some embodiments, memory 1310 can have encoded thereon, or otherwise stored therein, a computer program for controlling operation of computing device 1250. In such embodiments, processor 1302 can execute at least a portion of the computer program to present content (e g., images, user interfaces, graphics, tables), receive content from server 1252, transmit information to server 1252, and so on. For example, the processor 1302 and the memory 1310 can be configured to perform the methods described herein (e.g., the method of FIG. 3, the method of FIG. 4, the method of FIG. 9, the method of FIG. 10).

[0117] In some embodiments, server 1252 can include a processor 1312, a display 1314, one or more inputs 1316. one or more communications systems 1318, and / or memory1320. In some embodiments, processor 1312 can be any suitable hardware processor or combination of processors, such as a CPU, a GPU, and so on. In some embodiments, display 1314 can include any suitable display devices, such as an LCD screen, LED display, OLED display, electrophoretic display, a computer monitor, a touchscreen, a television, and so on. In some embodiments, inputs 1316 can include any suitable input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, and so on.

[0118] In some embodiments, communications systems 1318 can include any suitable hardware, firmware, and / or software for communicating information over communication network 1254 and / or any other suitable communication networks. For example, communications systems 1318 can include one or more transceivers, one or more communication chips and / or chip sets, and so on. In a more particular example, communications systems 1318 can include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and so on.

[0119] In some embodiments, memory 1320 can include any suitable storage device or devices that can be used to store instructions, values, data, or the like, that can be used, for example, by processor 1312 to present content using display 1314, to communicate with one or more computing devices 1250, and so on. Memory 1320 can include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1320 can include RAM, ROM, EPROM, EEPROM, other types of volatile memory, other types of non-volatile memory, one or more types of semi-volatile memory, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, and so on. In some embodiments, memory 1320 can have encoded thereon a server program for controlling operation of server 1252. In such embodiments, processor 1312 can execute at least a portion of the server program to transmit information and / or content (e.g., data, images, a user interface) to one or more computing devices 1250, receive information and / or content from one or more computing devices 1250, receive instructions from one or more devices (e.g., a personal computer, a laptop computer, a tablet computer, a smartphone), and so on.

[0120] In some embodiments, the server 1252 is configured to perform the methods described in the present disclosure. For example, the processor 1312 and memory' 1320 can be configured to perform the methods described herein (e.g., the method of FIG. 3, the method of FIG. 4. the method of FIG. 9, the method of FIG. 10).

[0121] In some embodiments, data source 1202 can include a processor 1322. one or more data acquisition systems 1324, one or more communications systems 1326. and / or memory 1328. In some embodiments, processor 1322 can be any suitable hardware processor or combination of processors, such as a CPU, a GPU, and so on. In some embodiments, the one or more data acquisition systems 1324 are generally configured to acquire data, images, or both, and can include a CT system. Additionally or alternatively, in some embodiments, the one or more data acquisition systems 1324 can include any suitable hardware, firmware, and / or software for coupling to and / or controlling operations of the CT system. In some embodiments, one or more portions of the data acquisition system(s) 1324 can be removable and / or replaceable.

[0122] Note that, although not shown, data source 1202 can include any suitable inputs and / or outputs. For example, data source 1202 can include input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, a trackpad, a trackball, and so on. As another example, data source 1202 can include any suitable display devices, such as an LCD screen, an LED display, an OLED display, an electrophoretic display, a computer monitor, a touchscreen, a television, etc., one or more speakers, and so on.

[0123] In some embodiments, communications systems 1326 can include any suitable hardware, firmware, and / or software for communicating information to computing device 1250 (and, in some embodiments, over communication network 1254 and / or any other suitable communication networks). For example, communications systems 1326 can include one or more transceivers, one or more communication chips and / or chip sets, and so on. In a more particular example, communications systems 1326 can include hardware, firmware, and / or software that can be used to establish a wired connection using any suitable port and / or communication standard (e.g., VGA. DVI video, USB, RS-232, etc.), Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and so on.

[0124] In some embodiments, memory' 1328 can include any suitable storage device or devices that can be used to store instructions, values, data, or the like, that can be used, for example, by processor 1322 to control the one or more data acquisition systems 1324, and / or receive data from the one or more data acquisition systems 1324; to generate images from data; present content (e.g., data, images, a user interface) using a display; communicate with one or more computing devices 1250; and so on. Memory 1328 can include any suitable volatile memory, non-volatile memory’, storage, or any suitable combination thereof. For example, memory 1328 can include RAM, ROM. EPROM. EEPROM, other types of volatile memory.other types of non-volatile memory, one or more types of semi-volatile memory, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, and so on. In some embodiments, memory 1328 can have encoded thereon, or otherwise stored therein, a program for controlling operation of data source 1202. In such embodiments, processor 1322 can execute at least a portion of the program to generate images, transmit information and / or content (e.g., data, images, a user interface) to one or more computing devices 1250, receive information and / or content from one or more computing devices 1250, receive instructions from one or more devices (e.g., a personal computer, a laptop computer, a tablet computer, a smartphone, etc.), and so on.

[0125] In some embodiments, any suitable computer-readable media can be used for storing instructions for performing the functions and / or processes described herein. For example, in some embodiments, computer-readable media can be transitory or non-transitory. For example, non-transitory' computer-readable media can include media such as magnetic media (e.g., hard disks, floppy disks), optical media (e.g., compact discs, digital video discs, Blu-ray discs), semiconductor media (e.g., RAM, flash memory, EPROM. EEPROM), any suitable media that is not fleeting or devoid of any semblance of permanence during transmission, and / or any suitable tangible media. As another example, transitory7computer- readable media can include signals on networks, in wires, conductors, optical fibers, circuits, or any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and / or any suitable intangible media.

[0126] As used herein in the context of computer implementation, unless otherwise specified or limited, the terms “component,” “system,” “module,” “framework,” and the like are intended to encompass part or all of computer-related systems that include hardware, software, a combination of hardware and software, or software in execution. For example, a component may be, but is not limited to being, a processor device, a process being executed (or executable) by a processor device, an object, an executable, a thread of execution, a computer program, or a computer. By way of illustration, both an application running on a computer and the computer can be a component. One or more components (or system, module, and so on) may reside within a process or thread of execution, may be localized on one computer, may7be distributed between two or more computers or other processor devices, or may be included within another component (or system, module, and so on).

[0127] In some implementations, devices or systems disclosed herein can be utilized or installed using methods embodying aspects of the disclosure. Correspondingly, descriptionherein of particular features, capabilities, or intended purposes of a device or system is generally intended to inherently include disclosure of a method of using such features for the intended purposes, a method of implementing such capabilities, and a method of installing disclosed (or otherwise known) components to support these purposes or capabilities. Similarly, unless otherwise indicated or limited, discussion herein of any method of manufacturing or using a particular device or system, including installing the device or system, is intended to inherently include disclosure, as embodiments of the disclosure, of the utilized features and implemented capabilities of such device or system.

[0128] Referring particularly now to FIGS. 14A and 14B, an example of an x-ray computed tomography (“CT”) imaging system 1400 is illustrated. The CT system includes a gantry 1402, to which at least one x-ray source 1404 is coupled. The x-ray source 1404 projects an x-ray beam 1406, which may be a fan-beam or cone-beam of x-rays, towards a detector array 1408 on the opposite side of the gantry 1402. The detector array 1408 includes a number of x-ray detector elements 1410. Together, the x-ray detector elements 1410 sense the projected x-rays 1406 that pass through a subject 1412, such as amedical patient or an object undergoing examination, that is positioned in the CT system 1400. Each x-ray detector element 1410 produces an electrical signal that may represent the intensity of an impinging x-ray beam and, hence, the attenuation of the beam as it passes through the subject 1412. In some configurations, each x-ray detector 1410 is capable of counting the number of x-ray photons that impinge upon the detector 1410. During a scan to acquire x-ray projection data, the gantry 1402 and the components mounted thereon rotate about a center of rotation 1414 located within the CT system 1400.

[0129] The CT system 1400 also includes an operator workstation 1416, which typically includes a display 1418: one or more input devices 1420, such as a keyboard and mouse; and a computer processor 1422. The computer processor 1422 may include a commercially available programmable machine running a commercially available operating system. The operator workstation 1416 provides the operator interface that enables scanning control parameters to be entered into the CT system 1400. In general, the operator workstation 1416 is in communication with a data store server 1424 and an image reconstruction system 1426. By way of example, the operator workstation 1416, data store sever 1424, and image reconstruction system 1426 may be connected via a communication system 1428, which may include any suitable network connection, whether wired, wireless, or a combination of both.As an example, the communication system 1428 may include both proprietary or dedicated networks, as well as open networks, such as the internet.

[0130] The operator workstation 1416 is also in communication with a control system 1430 that controls operation of the CT system 1400. The control system 1430 generally includes an x-ray controller 1432, a table controller 1434, a gantry controller 1436, and a data acquisition system 1438. The x-ray controller 1432 provides power and timing signals to the x-ray source 1404 and the gantry controller 1436 controls the rotational speed and position of the gantry 1402. The table controller 1434 controls a table 1440 to position the subject 1412 in the gantry' 1402 of the CT system 1400.

[0131] The DAS 1438 samples data from the detector elements 1410 and converts the data to digital signals for subsequent processing. For instance, digitized x-ray data is communicated from the DAS 1438 to the data store server 1424. The image reconstruction system 1426 then retrieves the x-ray data from the data store server 1424 and reconstructs an image therefrom. The image reconstruction system 1426 may include a commercially available computer processor, or may be a highly parallel computer architecture, such as a system that includes multiple-core processors and massively parallel, high-density computing devices. Optionally, image reconstruction can also be performed on the processor 1422 in the operator workstation 1416. Reconstructed images can then be communicated back to the data store server 1424 for storage or to the operator workstation 1416 to be displayed to the operator or clinician.

[0132] The CT system 1400 may also include one or more networked workstations 1442. By way of example, a networked workstation 1442 may include a display 1444; one or more input devices 1446, such as a keyboard and mouse; and a processor 1448. The networked workstation 1442 may be located within the same facility as the operator workstation 1416, or in a different facility, such as a different healthcare institution or clinic.

[0133] The networked workstation 1442, whether within the same facility or in a different facility as the operator workstation 1416, may gain remote access to the data store server 1424 and / or the image reconstruction system 1426 via the communication system 1428. Accordingly, multiple networked workstations 1442 may have access to the data store server 1424 and / or image reconstruction system 1426. In this manner, x-ray data, reconstructed images, or other data may be exchanged between the data store server 1424, the image reconstruction system 1426, and the networked workstations 1442, such that the data or images may be remotely processed by a networked workstation 1442. This data may be exchanged inany suitable format, such as in accordance with the transmission control protocol (“TCP"’), the internet protocol (“IP”), or other known or suitable protocols.

[0134] The present disclosure has described one or more preferred embodiments, and it should be appreciated that many equivalents, alternatives, variations, and modifications, aside from those expressly stated, are possible and within the scope of the invention.

Claims

CLAIMS1. A method for generating a virtual monoenergetic image (VMI) at a target energy level from single-energy computed tomography (CT) image data, the method comprising: accessing single-energy7CT image data with a computer system, wherein the singleenergy CT image data comprise one or more single-energy CT images acquired with a CT system using a tube potential; accessing CT scan protocol data with the computer system, wherein the CT scan protocol data comprise CT scan protocol parameters used by the CT system when acquiring the single-energy CT image data; accessing a machine learning model with the computer system, wherein the machine learning model has been trained on training data to synthesize a VMI at a target energy level from a single-energy CT image acquired at a particular tube potential; inputting the single-energy CT image data and the CT scan protocol data to the machine learning model using the computer system, generating a VMI at the target energy level as an output; and outputting the VMI with the computer system.

2. The method of claim 1. wherein the machine learning model comprises a neural network.

3. The method of claim 2, wherein the neural network implements a U-Net architecture.

4. The method of claim 1, wherein accessing the machine learning model includes determining an effective energy corresponding to the tube potential used for acquiring the single-energy- CT image data, and selecting the machine learning model from a plurality of trained machine learning models by matching the determined effective energy to a respective target VMI keV level, wherein each of the plurality of trained machine learning models is associated yvith a different target VMI keV level.

5. The method of claim 4. wherein determining the effective energy comprises estimating the effective energy based on at least one of a half value layer (HVL) or phantom measurement data associated with the tube potential.

6. The method of claim 1 , wherein the CT scan protocol data comprise a tube potential used when acquiring the single-energy CT image data.

7. The method of claim 1 or 6, wherein the CT scan protocol data comprise an image field-of-view used when acquiring the single-energy CT image data.

8. The method of any one of claims 1, 6, or 7, wherein the CT scan protocol data comprise a reconstruction kernel used when generating the single-energy CT images of the single-energy' CT image data.

9. The method of claim 1. wherein the tube potential is 120 kV.

10. The method of claim 1 or 9, wherein the target energy' level is 50 keV.

11. The method of claim 1 or 9, wherein the target energy level is 100 keV.

12. A method for training a neural network to generate a virtual monoenergetic image (VMI) at a target energy' level from an input image having a different energy' level, the method comprising: accessing training data with a computer system, wherein the training data comprises: first VMI data comprising VMIs associated with a first energy level corresponding to a tube potential of a CT system; second VMI data comprising VMIs associated with a second energy level corresponding to the target energy level, wherein the first energy level is different from the second energy level;CT scan protocol data associated with CT scan protocol parameters corresponding to the first VMI data and the second VMI data; accessing a neural network with the computer system; training the neural network on the training data; andstoring the trained neural network with the computer system.

13. The method of claim 12, wherein training the neural network comprises generating encoded CT scan protocol data by encoding the CT scan protocol data and training the neural network on the first VMI data, second VMI data, and encoded CT scan protocol data.

14. The method of claim 12, wherein the tube potential is 120 kV and the first energy level is 70 keV.

15. The method of claim 14, wherein the second energy level is 50 keV.

16. The method of claim 14, wherein the second energy7level is 100 keV.

17. A method for generating a computed tomography (CT) image, the method comprising: accessing a first CT image with a computer system, wherein the first CT image was acquired with a CT system using a first kV setting; accessing a machine learning model with the computer system, wherein the machine learning model has been trained on training data to generate a new CT image corresponding to a second kV setting from a CT image acquired at the first kV setting; inputting the first CT image to the machine learning model using the computer system, generating a second CT image as an output, wherein the second CT image corresponds to a CT image having been acquired using the second kV setting; and outputting the second CT image with the computer system.

18. The method of claim 17, wherein the machine learning model comprises a neural network.

19. The method of claim 18, wherein the neural network implements a U-Net architecture.

20. The method of claim 17, further comprising accessing CT scan protocol data with the computer system and inputting the CT scan protocol data as an additional input to the machine learning model, wherein the CT scan protocol data comprise CT scan protocol parameters used by the CT system when acquiring the first CT image.

21. The method of claim 20, wherein the CT scan protocol data comprise a tube potential corresponding to the first kV setting used when acquiring the first CT image.

22. The method of claim 20 or 21, wherein the CT scan protocol data comprise an image field-of-view used when acquiring the first CT image.

23. The method of any one of claims 20-22, wherein the CT scan protocol data comprise a reconstruction kernel used when generating the first CT image.

24. The method of claim 17, wherein the first kV setting is 80 kV.

25. The method of claim 17 or 24, wherein the second kV setting is 120 kV.

Images