System and method for uniform energy conversion in computed tomography imaging

Abstract

A method includes receiving a computed tomography (CT) image (402) acquired at a first energy level; obtaining an input prior (502) including an input energy spectrum corresponding to the first energy level; obtaining a target prior (506) including a target energy spectrum corresponding to a target energy level; processing the CT image through a neural network (308) to generate at least one feature map (406); generating a feature adjustment vector (518) based on the input prior (502) and the target prior (506) using a feature attention generator (452, 456); adjusting features of the at least one feature map (406) using the feature adjustment vector (518) to generate adjusted features; and processing the adjusted features through the neural network (308) to generate a transformed CT image at the target energy level. The feature attention generator (452, 456) can provide precise control of feature adjustments based on global energy transformation requirements, thereby allowing a single neural network to perform transformations across a continuous range of energy levels.

Description

Technical Field

[0001] The embodiments of the subject matter disclosed herein relate to medical imaging, and more specifically, to energy conversion in computed tomography (CT) imaging systems. Background Technology

[0002] In a computed tomography (CT) imaging system, an X-ray source emits an X-ray beam toward a subject or object (such as a patient). After being attenuated by the subject, the X-ray beam is projected onto an array of detectors. The intensity of the attenuated beam radiation received at the detector array depends on the attenuation of the X-ray beam by the subject. Each detector element in the detector array generates a separate electrical signal, which is sent to a data processing system for analysis and generation of medical images.

[0003] CT scans at various energy levels can provide improved tissue characterization and contrast quantification / visualization. For example, lower energy level images can increase the contrast visualization of the region of interest (ROI) in a CT image. However, CT images at lower energy levels are also prone to noise and image artifacts, which can reduce the visualization of the ROI and decrease the overall image quality.

[0004] Dual-energy CT imaging systems can generate CT images at different selected energy levels by acquiring projection data at multiple peak energy levels (e.g., 80 kVp and 140 kVp). For example, virtual monochrome images can be generated at any desired energy level between 40 keV and 140 keV. However, dual-energy CT imaging systems require specialized hardware configurations (such as fast kVp switching capabilities or dual-source detector arrangements), making them more complex and expensive than single-energy CT systems.

[0005] While single-energy CT imaging systems are more prevalent, they have traditionally lacked the multi-energy imaging capabilities of dual-energy systems. Deep learning methods have been proposed to convert single-energy CT images to different energy levels. However, existing methods require separate trained models for each specific combination of energy levels (e.g., separate models for converting 70 keV to 50 keV versus 70 keV to 40 keV). Additionally, these models are sensitive to contrast agent phase variations, necessitating different models for different contrast agent phases. This necessitates multiple separate models to handle various energy conversions and contrast agent phases, increasing computational overhead and system complexity.

[0006] Furthermore, the quality of energy conversion can be significantly dependent on the magnitude of the energy level variation. Direct conversion between large energy level differences (e.g., from 120 kVp to 30 keV) can lead to unsatisfactory image quality. This complexity is further increased by the need to accurately transform both contrast-enhanced and non-contrast-enhanced tissues while maintaining consistency across different display fields of view and reconstruction kernels. Summary of the Invention

[0007] This disclosure provides systems and methods for unified energy conversion in CT imaging, which at least partially address the aforementioned problems by providing a single "energy-to-energy" neural network model capable of performing multiple energy level conversions while maintaining consistency across different contrast agent phases and display fields of view. In one embodiment, a method includes: receiving a CT image acquired at a first energy level; obtaining an input prior comprising an input energy spectrum corresponding to the first energy level; obtaining a target prior comprising a target energy spectrum corresponding to a target energy level; processing the CT image via a neural network to generate at least one feature map; generating a feature adjustment vector based on the input prior and the target prior using a feature attention generator; adjusting features of the at least one feature map using the feature adjustment vector to generate adjusted features; and processing the adjusted features via the neural network to generate a transformed CT image at the target energy level.

[0008] This method addresses the limitations of existing approaches by providing a unified energy-to-energy neural network capable of performing multiple energy level conversions, rather than requiring separate models for each energy combination. The feature attention generator dynamically adjusts the network's behavior based on the provided energy spectrum and other prior information, enabling accurate energy conversion while maintaining image quality across different contrast agent phases and display fields of view.

[0009] In another embodiment, a method includes: receiving a CT image acquired at a first energy level; obtaining an input prior, the input prior including an input energy spectrum corresponding to the first energy level; obtaining a target prior, the target prior including a target energy spectrum corresponding to a target energy level; and processing the CT image through a neural network to generate a transformed CT image at the target energy level, wherein the neural network includes: an encoder path including a plurality of encoder levels, wherein each encoder level is configured to receive an input feature map and convolve the input feature map with one or more convolutional filters to generate an output encoder feature map; and a decoder path including features corresponding to the plurality of encoder levels. Multiple decoder layers, each configured to receive an input decoder feature map including a convolved feature map from a lower decoder layer, cascaded with the output encoder feature map from the corresponding encoder layer; a feature adjustment vector is generated at each decoder layer using a feature attention generator based on the input prior and the target prior; the features of the input decoder feature map are adjusted at each decoder layer using the corresponding feature adjustment vector to generate adjusted features; the adjusted features are convolved at each decoder layer to generate an output decoder feature map; and the transformed CT image is output based on the output decoder feature map from the highest level of the decoder path.

[0010] This method utilizes a multi-level encoder-decoder architecture with feature attention at each decoder level to dynamically adjust the network's behavior based on both local and global image features. By incorporating skip connections and applying targeted feature scaling at multiple network depths, the system maintains fine anatomical detail and structural consistency while performing energy transformations, which is particularly beneficial when transforming between large energy levels, where conventional methods often suffer significant quality degradation.

[0011] In another embodiment, a system includes: a display device, one or more processors, and a memory storing instructions executable by the one or more processors to perform the following operations: receiving a computed tomography (CT) image acquired at a first energy level; obtaining an input prior, the input prior including an input energy spectrum corresponding to the first energy level; obtaining a target prior, the target prior including a target energy spectrum corresponding to a target energy level; processing the CT image via a neural network to generate a transformed CT image at the target energy level, wherein the neural network generates at least one feature map; generating a feature adjustment vector using a feature attention generator based on the input prior and the target prior; adjusting features of the at least one feature map using the feature adjustment vector to generate adjusted features; generating the transformed CT image using the adjusted features; and displaying the transformed CT image on the display device.

[0012] This system enables the integration of energy conversion capabilities into existing single-energy CT workflows. A feature attention mechanism, guided by physics-based energy spectrum and other prior information, allows for precise control of the transformation process across multiple scales, resulting in high-quality transformed images that retain diagnostic value without the need for expensive dual-energy CT hardware. This allows facilities with conventional single-energy CT systems to acquire advanced energy-specific imaging capabilities.

[0013] The above advantages, as well as other advantages and features of this specification, will become apparent from the following detailed description when considered alone or in conjunction with the accompanying drawings. It should be understood that the above summary is provided to present a series of concepts further described in the detailed description in a simplified form. This is not intended to identify key or essential features of the claimed subject matter, the scope of which is uniquely defined by the claims following the detailed description. Furthermore, the claimed subject matter is not limited to specific implementations that address any shortcomings pointed out above or in any part of this disclosure. Attached Figure Description

[0014] This patent or patent application document contains at least one color-drawn drawing. A published copy of this patent or patent application with the color drawing will be provided by the Patent Office upon request and payment of the necessary fees.

[0015] A better understanding of the various aspects of this disclosure can be achieved by reading the following detailed description and referring to the accompanying drawings, in which:

[0016] Figure 1 A drawing view of a CT imaging system including the disclosed embodiment is shown;

[0017] Figure 2 It shows Figure 1 The schematic block diagram of the system illustrated herein;

[0018] Figure 3 An example image processing system according to an embodiment of the present disclosure is shown;

[0019] Figure 4 An example neural network architecture including a feature attention generator for energy-to-energy conversion is illustrated according to an embodiment of the present disclosure;

[0020] Figure 5 An example feature attention generator architecture according to an embodiment of this disclosure is illustrated schematically;

[0021] Figure 6 This is a flowchart illustrating a method for transforming a CT image from a first energy level to a target energy level using a feature adjustment vector according to an embodiment of this disclosure;

[0022] Figure 7 This is a flowchart illustrating the details of processing CT images via an encoder path according to an embodiment of the present disclosure;

[0023] Figure 8 This is a flowchart illustrating the details of processing a feature map via a decoder path notified by a feature adjustment vector according to an embodiment of this disclosure;

[0024] Figure 9 This is a flowchart illustrating a method for generating feature adjustment vectors according to an embodiment of the present disclosure;

[0025] Figure 10 This is a flowchart illustrating a method for training a neural network to transform medical images between different energy levels according to an embodiment of the present disclosure;

[0026] Figure 11 This is a flowchart illustrating a method for training a neural network to transform a medical image between different energy levels using multiple target energy levels, according to an embodiment of the present disclosure;

[0027] Figure 12 This is a graphical representation of the X-ray spectrum according to an embodiment of the present disclosure, which shows the relationship between the spectral amplitude and the voltage at various operating voltages (kVp);

[0028] Figure 13 An example of a coded contrast agent phase timeline used as a priori information for energy-to-energy conversion according to an embodiment of this disclosure is shown;

[0029] Figure 14 This is a comparative display of CT images, illustrating input images, target images, and predicted images at different energy levels, according to an embodiment of this disclosure; and

[0030] Figure 15 This is a comparative visualization of the results of energy level conversion using an energy-to-energy neural network according to an embodiment of this disclosure. Detailed Implementation

[0031] The following description relates to transforming images from one energy level to another in CT imaging. Specifically, the following description relates to a unified framework for transforming CT images obtained at a first energy level into one or more different energy levels using a single neural network model, which incorporates feature modulation mechanisms guided by prior information associated with the input image and one or more “target” images.

[0032] Unlike single-energy CT imaging systems, dual-energy CT imaging systems can generate images at any selected energy level between 40 keV and 140 keV by acquiring projection data at multiple peak energy levels (e.g., 80 kVp and 140 kVp). For example, virtual monochrome images can be generated at any desired energy level by performing material decomposition using dual-energy projection data. Therefore, dual-energy CT imaging systems offer advantages for certain imaging tasks, such as improved contrast visualization and tissue characterization. However, dual-energy CT systems require specialized hardware configurations, making them more complex and expensive than single-energy CT systems.

[0033] While single-energy CT imaging systems are more prevalent, they have traditionally lacked the multi-energy imaging capabilities of dual-energy systems. Previous approaches have attempted to address this limitation by using deep learning to convert single-energy CT images to different energy levels. However, these methods require separate trained models for each specific combination of energy levels (e.g., separate models for converting 70 keV to 50 keV versus 70 keV to 40 keV). Additionally, these models are sensitive to contrast agent phase variations, necessitating different models for different contrast agent phases. This necessitates multiple separate models to handle various energy conversions and contrast agent phases, increasing computational overhead and system complexity.

[0034] In one implementation scheme, such as Figures 1 to 2 The CT imaging system shown can acquire CT images and use Figure 3 The image processing system shown processes these images. The image processing system includes a trained energy-to-energy neural network 308, which, through... Figure 4 The illustrated encoder-decoder architecture processes images by combining... Figure 5The feature attention generator is described in detail below. The feature attention generator dynamically adjusts the network's behavior based on prior information, including the energy spectrum stored in the energy spectrum database 316 and the temporal timeline information from the temporal timeline module 314. For example... Figure 6 As illustrated, the entire process includes: receiving CT images; obtaining input priors and target priors; processing via the encoder path; generating and applying feature modulation; and processing via the decoder path to produce a transformed image. Figure 7 The encoder path processing steps are described in detail, and Figure 8 The decoder path operation, including the application of feature adjustment vectors, is illustrated. Figure 9 An implementation scheme for generating feature adjustment vectors using input priors and target priors is shown. Figures 10 to 11 An example is given of a method for training a neural network to transform medical images between different energy levels using feature adjustment vectors, wherein... Figure 10 The basic training method is shown, and Figure 11 The training demonstrated utilizes multiple target energy levels. Figure 12 Graphical representations of X-ray spectra at various operating voltages are provided, which serve as physics-based priors for energy conversion processes. Figure 13 An example of an encoded contrast agent phase timeline used as prior information is shown, while Figures 14 to 15 The study presents a comparative visualization of input, target, and predicted images at different energy levels, demonstrating high-quality energy conversion achieved across different contrast agent phases and display fields of view.

[0035] This unified approach addresses the limitations of existing methods by providing a single neural network model capable of performing multiple energy level conversions while maintaining consistency across different contrast agent phases and display fields of view. Feature modulation mechanisms guided by physics-based energy spectra and other prior information allow for precise control of the transformation process at multiple scales, resulting in high-quality transformed images that retain diagnostic value without requiring expensive dual-energy CT hardware.

[0036] refer to Figure 1 A CT imaging system 100 is shown for acquiring and processing medical imaging data at multiple energy levels. The CT imaging system 100 is configured to: acquire single-energy CT images; and transform them into different energy levels using a feature attention-based neural network.

[0037] The CT imaging system 100 includes a gantry 102 that houses various imaging components. In one embodiment, the gantry 102 is configured to rotate about a central aperture in which a subject 112 is positioned during imaging. The gantry 102 supports an X-ray source 104 and a detector array 108, positioning them opposite each other across the aperture. In another embodiment, the gantry 102 includes mounting hardware and electrical connections for multiple X-ray sources and detector arrays to enable dual-energy imaging capability.

[0038] X-ray source 104 is configured to emit an X-ray beam toward subject 112 for imaging. In one embodiment, X-ray source 104 operates at a single-peak kilovolt (kVp) level (such as 120 kVp), producing an X-ray spectrum of approximately 70 keV monochromatic X-rays. In another embodiment, X-ray source 104 enables rapid kVp switching between multiple energy levels. X-ray source 104 is mechanically and electrically coupled to rack 102 and receives control signals from image processor unit 110 to coordinate X-ray generation with readout of detector array 108.

[0039] Detector array 108 is positioned relative to X-ray source 104 across an aperture in rack 102 to detect X-rays that have passed through subject 112. In one embodiment, detector array 108 includes multiple rows of detector elements that generate individual electrical signals proportional to the attenuated X-ray intensity at each element. In another embodiment, detector array 108 may include an energy integration detector that measures the total deposited energy. Detector array 108 is electrically coupled to image processor unit 110 to transmit detector signals for image reconstruction and processing.

[0040] Image processor unit 110 receives and processes detector signals to reconstruct CT images. In one embodiment, image processor unit 110 includes a trained energy-to-energy neural network that processes images via an encoder-decoder architecture with a feature attention generator. The neural network receives input priors (including energy spectrum and temporal timeline information) to guide the energy level conversion process. In another embodiment, image processor unit 110 includes a prior information module, a prior encoding module, and a temporal timeline module for preprocessing image data and associated prior information prior to neural network processing.

[0041] Subject 112 may be a patient, a phantom, or other object positioned on examination table 114 for CT imaging. In one embodiment, subject 112 is a patient receiving a contrast agent to enhance visualization of a specific anatomical structure, wherein the contrast agent produces different enhancement patterns across multiple contrast agent phases. In another embodiment, subject 112 is a calibration phantom composed of known materials used to validate the energy conversion capabilities of the system.

[0042] The examination table 114 supports the subject 112 and positions it within an aperture in the gantry 102. In one embodiment, the examination table 114 is electrically powered to move the subject 112 (into / out of the aperture) along the z-axis, thereby enabling helical scanning, wherein projection data is acquired as the examination table 114 continuously moves through the gantry 102. In another embodiment, the examination table 114 includes positioning controls to adjust height, tilt, and lateral position, thereby optimizing subject positioning relative to the X-ray source 104 and the detector array 108. The movement of the examination table 114 is coordinated with the image acquisition process via control signals from the image processor unit 110.

[0043] Figure 2 Examples similar to Figure 1 An exemplary imaging system 200 of a CT imaging system 100. Imaging system 200 is configured to image a subject 112. In one embodiment, imaging system 200 includes a detector array 108 (see [link to documentation]). Figure 1 The detector array 108 also includes a plurality of detector elements 202 that together sense an X-ray beam passing through the subject 112 (such as a patient) to acquire corresponding projection data. Therefore, in one embodiment, the detector array 108 is manufactured in a multi-slice configuration including multiple rows of cells or detector elements 202. In such a configuration, another row or multiple rows of detector elements 202 are arranged in parallel to acquire projection data.

[0044] In some implementations, the imaging system 200 is configured to traverse different angular positions around the subject 112 to acquire desired projection data. Therefore, the gantry 102 and the components mounted thereon can be configured to rotate about a rotation center 206 for acquiring projection data at a single energy level.

[0045] As the X-ray source 104 and detector array 108 rotate, detector array 108 collects data from the attenuated X-ray beam. The data collected by detector array 108 undergoes preprocessing and calibration to adjust the data to represent the line integral of the attenuation coefficient of the scanned subject 112. The processed data is typically referred to as the projection.

[0046] The detector element 202 of the detector array 108 may include an energy integral detector that measures the total deposited energy from the X-ray beam. The acquired projection data is used to reconstruct CT images at the acquired energy level.

[0047] Once reconstructed, the CT images reveal the internal features of the subject 112. Using the energy-to-energy neural network described in this paper, the images can be transformed to appear as if they were acquired at different energy levels. Compared to images acquired directly at lower energy levels, the transformed images provide enhanced contrast visualization while maintaining low noise and artifacts.

[0048] In one embodiment, the imaging system 200 includes a control mechanism 208 to control the movement of components, such as the rotation of the gantry 102 and the operation of the X-ray source 104. In some embodiments, the control mechanism 208 further includes an X-ray controller 210 configured to provide power and timing signals to the X-ray source 104. Additionally, the control mechanism 208 includes a gantry motor controller 212 configured to control the rotational speed and / or position of the gantry 102 based on imaging requirements.

[0049] In some embodiments, control unit 208 also includes a data acquisition system (DAS) 214 configured to sample analog data received from detector element 202 and convert the analog data into digital signals for subsequent processing. The data sampled and digitized by DAS 214 is sent to a computer or computing device 216. In one example, computing device 216 stores the data in a mass storage device or storage device 218. Storage device 218 may also store energy spectrum information, temporal timeline data, and other prior information (DFOV) used by energy-to-energy neural networks.

[0050] The computing device 216 includes a trained energy-to-energy neural network that processes images through an encoder-decoder architecture with a feature attention generator. The neural network receives input priors (including energy spectrum and temporal timeline information) to guide the energy level conversion process. The computing device 216 may also include a prior information module, a prior encoding module, and a temporal timeline module for preprocessing image data and associated prior information prior to neural network processing.

[0051] The operator console 220 allows the operator to control scanning parameters and view both the original and transformed images. The operator can select the desired target energy level for image transformation through the user interface of the operator console 220.

[0052] The imaging system 200 can be coupled to multiple displays, printers, workstations, and / or similar devices for viewing both the original and transformed images. The transformed images can be transmitted across a network to allow for remote viewing and analysis.

[0053] In one embodiment, the imaging system 200 includes or is coupled to a picture archiving and communication system (PACS) 224. The PACS 224 can store both the original image and transformed images generated at different target energy levels.

[0054] The computing device 216 operates the inspection table motor controller 226 using operator-supplied and / or system-defined commands and parameters, which controls the electric inspection table 228. The inspection table motor controller 226 moves the inspection table 228 to position the subject 112 in the rack 102 for acquiring projection data.

[0055] Image reconstructor 230 uses sampled and digitized X-ray data to perform high-speed reconstruction of CT images at the acquired energy level. The reconstructed images can then be processed by an energy-to-energy neural network to generate transformed images at different target energy levels.

[0056] The computing device 216 can send both the original reconstructed image and the transformed image to a display device 232 communicatively coupled to the computing device 216. The display device 232 can display the original image and the transformed image side by side to allow comparison of tissue contrast and detail at different energy levels.

[0057] refer to Figure 3 An image processing system 300 for processing energy-based imaging data is shown. The image processing system 300 includes an image processing device 302 configured to receive and process CT imaging data at multiple energy levels. In one embodiment, the image processing device 302 is integrated into the CT imaging system as an image processor unit or computing device. In another embodiment, at least a portion of the image processing device 302 is disposed at a separate device communicatively coupled to the CT imaging system via a wired or wireless connection.

[0058] Image processing device 302 includes processor 304 configured to execute machine-readable instructions stored in non-transitory memory 306. In one embodiment, processor 304 may be single-core or multi-core, and the program executing on it may be configured for parallel or distributed processing. In another embodiment, processor 304 may include components distributed across multiple devices configured for coordinated processing. Processor 304 may also be virtualized and executed by a remotely accessible networked computing device in a cloud computing configuration.

[0059] Non-transitory memory 306 stores various modules and databases for implementing energy conversion in CT imaging. Memory 306 includes a trained energy-to-energy neural network 308 that processes images by incorporating various neural network architectures with a feature attention generator. In one embodiment, neural network 308 receives input priors and target priors to guide the energy level conversion process via feature adjustment vectors. In another embodiment, neural network 308 enables an encoder-decoder architecture with skip connections between corresponding encoder and decoder levels to preserve fine anatomical details during energy conversion, wherein the feature adjustment vectors adjust feature maps in both the encoder and decoder paths.

[0060] The prior information module 310 manages various types of prior information used to guide the energy conversion process. In one embodiment, the prior information module 310 processes input priors, which include energy spectra, contrast agent phase timelines, display field values, and reconstructed nucleus types. In another embodiment, the module 310 processes target priors, which include target energy spectra and matching or different contrast agent phase information.

[0061] The prior encoding module 312 transforms prior information into features that influence the training and inference processes. In one implementation, module 312 encodes the energy spectrum, temporal timeline, and other priors into feature vectors used by the feature attention generator. In another implementation, the prior encoding module 312 processes the priors through multiple MLPs to generate encoded features for feature conditioning vector generation.

[0062] The phase timeline module 314 manages the contrast agent phase timeline information used for training and inference. In one embodiment, module 314 maintains phase timeline data indicating patterns of contrast agent enhancement across different anatomical structures over time. In another embodiment, module 314 provides phase timeline priors to enable the network to maintain consistency across different contrast agent phases.

[0063] The energy spectrum database 316 stores energy spectrum information for different kVp and keV levels. In one embodiment, the energy spectrum database 316 contains X-ray spectra corresponding to common acquisition energy levels (e.g., 120 kVp) and virtual monochromatic energy levels from 40 keV to 140 keV. In another embodiment, the database 316 provides spectral priors to guide energy-specific transformations in a neural network.

[0064] Training module 318 implements the process for training an energy-to-energy neural network. In one embodiment, module 318 trains using image pairs with different energy levels generated from dual-energy projection data. In another embodiment, module 318 updates the network parameters by comparing the transformed images with a ground truth image while incorporating prior information.

[0065] The projection / image database 320 stores CT projection data and reconstructed images. In one embodiment, the projection / image database 320 contains a training dataset that includes paired images of different energy levels and contrast agent phases. In another embodiment, the database 320 stores transformed images generated by a neural network for validation and clinical use.

[0066] Display device 334 enables the visualization of raw CT images and transformed CT images. In one embodiment, display device 334 includes a computer monitor that displays a side-by-side comparison of images at different energy levels. In another embodiment, display device 334 presents transformed images with enhanced contrast for improved diagnostic visualization.

[0067] User input device 332 enables a user to interact with the image processing system. In one embodiment, user input device 332 includes a touchscreen, keyboard, and mouse for selecting an image and specifying a target energy level for transformation. In another embodiment, input device 332 allows a user to interactively adjust processing parameters and view results.

[0068] refer to Figure 4 This paper illustrates an image processing architecture 400 that combines a feature attention mechanism with multi-scale processing for transforming CT images from a first energy level to a target energy level. The architecture implements an encoder-decoder framework with multiple encoding and decoding levels, where a feature attention generator is incorporated at each decoder level to dynamically adjust features based on the input and target energy spectra, as well as other prior information. This architecture enables accurate energy conversion while preserving anatomical detail and image quality across different contrast agent phases and display fields of view through the coordinated use of multi-scale processing and the feature attention mechanism. This unified approach eliminates the need for separate models for each energy combination and contrast agent phase, thereby reducing computational overhead and system complexity while maintaining consistent image quality. The feature attention generator, guided by physically based energy spectra and other prior information, allows for precise control of the transformation process at multiple scales, resulting in high-quality transformed images that retain diagnostic value without requiring expensive dual-energy CT hardware.

[0069] The encoder path begins with a first encoding level that receives an input CT image 402 acquired at a first energy level (e.g., 70 keV or 120 kVp). The input image 402 can be a 2D slice image, a 3D volume rendering, or a maximum intensity projection (MIP) image obtained from projection data. The input image 402 is processed by a convolutional layer 404, which applies multiple learned convolutional filters and nonlinear activation functions to generate an initial feature map 406. This initial feature map captures low-level image features, such as the edges, textures, and intensity gradients of different tissue types and contrast-enhanced regions. For example, the convolutional layer 404 may apply 3×3 or 5×5 convolutional kernels followed by a ReLU activation function to extract meaningful features while preserving spatial relationships. In some implementations, a pooling layer 408 performs spatial downsampling on the feature map 406 using a max-pooling operation with a 2×2 window and a stride of 2 to reduce the spatial dimensionality while preserving key features such as high-contrast boundaries and enhanced vascular structures. Pooling operations help achieve translation invariance and reduce computational complexity in deeper network layers. The feature attention generator 452 processes prior information 450 to generate a feature adjustment vector 454, which includes the input energy spectrum characterizing the X-ray attenuation properties at a first energy level, the target energy spectrum defining the desired attenuation properties, a contrast agent phase timeline indicating the time-contrast agent dynamics, and display field-of-view parameters specifying the image reconstruction settings. The feature adjustment vector 454 is used to adjust features at the corresponding highest decoder level, enabling the network to selectively modify different feature channels based on specific energy conversion requirements and imaging conditions. The prior information 450 may also include the reconstructed nuclear type, anatomical region indicators, and other acquisition parameters affecting the energy conversion process.

[0070] The second encoding layer receives the downsampled output from pooling layer 408 as a feature map 410, which preserves structural and contrast information while operating at reduced spatial resolution. Feature map 410 is processed by convolutional layer 412, which applies an additional set of learned filters, in some implementations including multiple 3×3 and 5×5 convolutional kernels, followed by a nonlinear activation function (such as ReLU or LeakyReLU) to generate a feature map 414 that captures intermediate layer image features and energy-specific properties. These intermediate features may include tissue boundaries, contrast-enhanced vascular structures, organ textures, and local energy-related attenuation patterns in different anatomical regions. Pooling layer 416 typically uses max pooling with a 2×2 window and a stride of 2 to perform further spatial downsampling to reduce computational complexity while preserving a rich feature set that retains the main activation patterns. Feature attention generator 456 processes prior information 450 to generate a feature adjustment vector 458, which dynamically adjusts the features at the corresponding decoder layer based on desired energy transformation parameters. The characteristic adjustment vector 458 can be calculated by using a dedicated MLP to process the energy spectrum and temporal information, and by generating channel-by-channel adjustment factors that selectively modify different characteristic channels based on specific energy conversion requirements.

[0071] The encoder path continues through additional encoding layers (as indicated by the ellipsis between pooling layers 416 and 418), each following a similar pattern of feature map generation, convolution, and pooling operations to progressively capture higher-level image features at different scales. These additional layers (numbered between 3 and 5 in some implementations depending on the input image resolution and the desired feature hierarchy) enable the network to learn increasingly abstract representations that combine both local and global image features. Each layer reduces the spatial dimensionality while increasing the number of feature channels (in some implementations, from doubling from 64 to 128 to 256, etc.), allowing the network to capture more complex patterns and relationships. The bottom encoding layer generates feature map 420, which is processed by convolutional layer 422 (using a dedicated kernel designed to capture global contextual information) to produce feature map 424. The final feature map 424 may typically contain a more abstract representation of the input image and energy-specific properties compared to previous encoding layers, including anatomical context, global contrast distribution patterns, and high-level energy-related imaging properties that guide subsequent decoding processes. In some implementations, the lowest layers of the energy-to-energy neural network architecture 400 operate at 1 / 16 or 1 / 32 of the original image resolution, with 512 or 1024 feature channels, to maintain rich representational power despite the reduction in spatial dimensionality.

[0072] The decoder path begins at the lowest level, where feature map 424 is up-convolved through up-convolutional layer 426 to increase spatial dimension. In some implementations, the up-convolution operation uses a transposed convolution with a learnable filter of size 2×2 or 4×4 and a stride of 2, effectively doubling the spatial dimension while reducing the number of feature channels. This is followed by additional decoder levels (as indicated by ellipses), each corresponding to an encoder level and incorporating skip connections by combining the output encoder feature maps from the corresponding encoder level to preserve fine spatial details. In some implementations, the number of decoder levels matches the number of encoder levels and can vary between 4 and 6 levels depending on the input image resolution. The penultimate decoder level receives the up-convolved feature map via up-convolutional layer 428, which may undergo a similar transposed convolution operation but with a different number of output channels (typically half the number of the previous level). The convolutional feature map 430 is generated by cascading the convolutional feature map with the feature map 414 from the corresponding encoder layer via skip connections, thereby enabling direct access to earlier spatial information that may have been lost during downsampling. The concatenated features of the feature map 430 are adjusted using a feature adjustment vector 458 generated by the feature attention generator 456 to selectively modify features related to the target energy level. In some implementations, the feature adjustment vector 458 includes a channel-wise adjustment factor computed based on global energy transformation requirements. The adjusted features are then processed by a convolutional layer 432 (which may employ multiple 3×3 convolutional filters followed by batch normalization and ReLU activation) to produce a feature map 434 with refined local details while preserving energy-specific properties.

[0073] Feature map 434 is up-convolved to the highest decoder level via up-convolutional layer 436 to recover full spatial resolution. In some implementations, this final up-convolution employs a matched number of output channels to maintain consistency with the corresponding encoder level features. At this level, the up-convolved features are concatenated with feature map 406 from the first encoding level via skip connections to produce feature map 438. The feature channels of feature map 438 are then adjusted using feature adjustment vector 454 from feature attention generator 452, which modifies the feature weights based on input energy characteristics and target energy characteristics. In some implementations, multiple MLPs are used to process the energy spectrum information to compute feature adjustment vector 454. The adjusted features of feature map 438 are then processed by convolutional layer 440, which may employ a final set of convolutional filters (e.g., 32 or 64 3×3 filters), followed by 1×1 convolutions to generate the final target image 442 at the desired target energy level (e.g., 40 keV, 50 keV, or 60 keV). In some implementations, additional post-processing operations, such as residual connection or adaptive normalization, can be applied to enhance the quality of the final output.

[0074] Figure 4 The energy-to-energy neural network architecture 400, described in detail, offers several technical advantages over existing methods. A multi-level encoder-decoder framework with feature attention at each decoder level enables dynamic adjustment of the network's behavior based on both local and global image features. By incorporating skip connections between corresponding encoder and decoder levels and applying targeted feature tuning via feature tuning vectors at multiple network depths, the system maintains fine anatomical detail and structural consistency while performing energy conversions. This is particularly beneficial when converting between vastly different energy levels, where conventional methods often exhibit significant quality degradation. A feature attention generator guided by physically based energy spectra and other prior information allows for precise control of the transformation process at multiple scales. Additionally, the unified approach eliminates the need for separate models for each energy combination and contrast agent phase, thereby reducing computational overhead and system complexity while maintaining consistent image quality across different display fields. The architecture's ability to process both the input and target energy spectra as priors enables accurate energy-specific transformations that preserve diagnostic value without requiring expensive dual-energy CT hardware.

[0075] While the architecture described above primarily illustrates feature adjustment vectors for adjusting features within the decoder path, in some implementations, feature adjustment vectors can also be used to adjust feature maps continuing through the encoder path, allowing the remainder of the encoder path to process the adjusted feature maps. Furthermore, although the encoder-decoder neural network architecture has been described in detail, the mechanism of using feature adjustment vectors to adjust feature maps can be applied to virtually any neural network architecture, including but not limited to U-Net, Residual Dense Networks (RDN), transformer-based architectures, and other convolutional neural network designs.

[0076] refer to Figure 5 The diagram illustrates a feature attention generator architecture 500 for generating feature conditioning vectors. This architecture includes two parallel processing paths that converge to generate feature conditioning vectors for dynamically adjusting features in the energy-to-energy conversion of CT images. In one implementation, the feature attention generator architecture 500 processes an input prior (which includes energy spectrum and temporal information) and a target prior (which includes a target energy spectrum) to generate feature conditioning vectors that control feature conditioning at the decoder level.

[0077] The first processing path begins with an input prior 502, which characterizes the properties of the input CT image. In one embodiment, the input prior 502 includes an input energy spectrum corresponding to a first energy level of the input CT image, thereby providing X-ray attenuation values ​​across the energy range. In another embodiment, the input prior 502 also includes an input contrast agent phase timeline, a display field of view (DFOV) value, and a reconstruction kernel type for reconstructing the input CT image, indicating the dynamics of the time-contrast agent. The input prior 502 is processed by a first multilayer perceptron 504 to generate a first processed prior. The first multilayer perceptron 504 may include multiple fully connected neural network layers with nonlinear activation functions to learn representations of input image properties that influence the energy conversion process.

[0078] The second processing path begins with a target prior 506, which specifies the desired characteristics of the transformed CT image. In one embodiment, the target prior 506 includes a target energy spectrum corresponding to a target energy level, thereby providing the desired X-ray attenuation characteristics. In another embodiment, the target prior 506 also includes a target contrast agent phase timeline matched to the input timeline and a target DFOV matched to the input DFOV. The target prior 506 is processed by a second multilayer perceptron 508 to generate a second processed prior. Similar to the first multilayer perceptron 504, the second multilayer perceptron 508 learns a representation of the target image characteristics to guide the transformation.

[0079] The outputs from the two processing paths, the first processed prior from multilayer perceptron 504, and the second processed prior from multilayer perceptron 508 are combined to form a combined processed feature. In one embodiment, the outputs are concatenated to preserve the distinct information from each path. The combined processed feature is then fed into a third multilayer perceptron 516, which learns the relationship between the input features and the target features. The third multilayer perceptron 516 processes these relationships to generate a feature conditioning vector 518.

[0080] The feature adjustment vector 518 has a dimension corresponding to the number of channels in the feature map from the corresponding decoder layer. In one embodiment, the feature adjustment vector 518 includes a channel-wise adjustment factor ranging from 0 to 1, which is used to selectively adjust different feature channels through various operations. In another embodiment, the adjustment factor in the feature adjustment vector 518 is generated using a sigmoid activation function in the final layer of the multilayer perceptron 516 to ensure an appropriate range. The feature adjustment vector 518 achieves precise control over the transformation process by dynamically adjusting the features based on specific energy conversion requirements defined by the input prior and the target prior.

[0081] This feature attention generator architecture 500 can be replicated at multiple decoder layers in an energy-to-energy neural network, with each instance processing the same input and target priors. Multiple feature conditioning vectors generated by these replicated architectures work together to perform energy transformations between different energy levels and across different contrast agent phases while maintaining fine anatomical details and structural consistency.

[0082] The technical advantage of the feature attention generator architecture 500 is that it enables precise control over feature tuning at multiple network depths based on global energy transformation requirements. By incorporating an input prior 502 (which includes continuous X-ray energy spectra rather than discrete energy categories), the architecture can efficiently interpolate or extrapolate to generate transformed images at any target energy level, including those not explicitly represented in the training data. Multilayer perceptrons 504, 508, and 516 learn the complex nonlinear relationship between the input and target energy spectra, allowing the network to dynamically adjust its behavior based on specific energy transformation requirements. Additionally, by processing the input and target priors using dedicated paths before combining them, the architecture maintains awareness of both source and desired target characteristics throughout the transformation process. The generated tuning vector 518 provides factors that adjust different feature channels based on their correlation with the desired energy transformation, enabling the network to maintain fine anatomical detail and structural consistency even when transforming between vastly different energy levels. Physics-based guidance from energy spectrum priors allows the network to learn generalizable relationships between energy levels, rather than memorizing discrete mappings, thus making the architecture more robust and adaptable to novel energy conversion scenarios.

[0083] refer to Figure 6 A flowchart of a method 600 for transforming CT images between energy levels using a neural network with feature attention is shown. Method 600 enables the transformation of a CT image acquired at one energy level to appear as if it were acquired at different energy levels while preserving image quality and anatomical detail. Method 600 can be developed by an image processing system (such as...) Figure 3 The image processing system 300 is executed.

[0084] At operation 602, the image processing system receives CT images acquired at a first energy level. In one embodiment, CT images may be acquired during an imaging protocol at 120 kVp, which produces an X-ray spectrum approximately equal to 70 keV monochromatic X-rays. In another embodiment, the CT images may be maximum intensity projection (MIP) images reconstructed based on helical scan projection data acquired during a contrast-enhanced imaging protocol.

[0085] At operation 604, the image processing system obtains an input prior, which includes an input energy spectrum corresponding to a first energy level. In one embodiment, the input energy spectrum is selected from a pre-recorded library of empirically measured energy spectra stored in the energy spectrum database 316 by matching the acquisition parameters of the input CT image (including kVp settings, X-ray tube target material, and beam filtering) with corresponding parameters associated with each stored energy spectrum. In another embodiment, the input energy spectrum is determined by analyzing image metadata, including scan protocol information, to identify the acquisition energy level and the corresponding spectrum. The input prior also includes an input contrast agent phase timeline indicating the time-contrast dynamics derived from the radiologist's phase markers, the display field of view (DFOV) value used during image reconstruction, and a reconstruction kernel type indicating the filter applied during image reconstruction. In some embodiments, the input prior additionally includes scan-specific parameters, such as tube current, exposure time, and detector calibration data used during image acquisition. In other embodiments, the input prior may include geometric factors affecting spatial resolution characteristics, such as source-to-detector distance and detector element size.

[0086] At operation 606, the image processing system obtains a target prior, which includes a target energy spectrum corresponding to the target energy level. In one embodiment, the target energy spectrum is selected from a pre-recorded library of virtual single-chromatograms stored in the energy spectrum database 316, where each spectrum represents a narrowband X-ray energy of 40 keV to 140 keV generated by physical-based modeling of the interaction between X-rays and matter. In another embodiment, the target energy spectrum is derived by empirically measuring beam attenuation at different energy levels using a dual-energy CT system with the aid of a calibration phantom. The target prior also includes a target contrast agent phase timeline matched to the input timeline to maintain temporal consistency, a target DFOV matched to the input DFOV to preserve spatial characteristics, and a reconstruction kernel type used for the input image to maintain consistent filtering effects. In some embodiments, the target prior additionally includes noise and resolution characteristics expected at the target energy level based on system modeling. In other embodiments, the target prior may include an ideal material decomposition basis representing the expected attenuation behavior of different tissue types at the target energy level.

[0087] At operation 608, the image processing system processes the CT image through the encoder path of the neural network to generate at least one feature map. In one embodiment, the encoder path includes multiple encoding levels operating at progressively decreasing spatial resolution, each level having an associated feature attention generator to enable precise control over feature emphasis at each scale of the transformation process. In another embodiment, feature attention generators are selectively combined at specific encoder levels based on the magnitude of the energy level transformation being performed, where larger energy level changes (e.g., 70 keV to 40 keV) utilize more feature attention generators to maintain transformation accuracy, while smaller energy level changes may require fewer feature attention generators to balance computational efficiency. In yet another embodiment, feature attention generators are primarily combined at intermediate encoder levels where spatial resolution and feature channel dimensions provide a balance between computational cost and transformation control, where earlier encoder levels focus on extracting general image features, and later encoder levels maintain the global context without requiring additional attention mechanisms. The encoder path can comprise 3 to 6 encoding levels, depending on the input image resolution and the desired depth of the feature hierarchy. Each level applies convolutional operations followed by pooling to reduce spatial dimensions while increasing feature channels. This allows the network to perform two operations: capturing fine anatomical details at higher resolution and capturing global contextual information at lower resolution. The specific configuration of the feature attention generator across encoder levels can be determined during network architecture design based on empirical evaluations of the transform quality versus computational cost for different types of energy level transitions and contrast agent phases.

[0088] At operation 610, the image processing system uses a feature attention generator to generate feature modulation vectors based on input and target priors. In one implementation, the feature attention generator processes the input and target energy spectra using a parallel multilayer perceptron (MLP) with three hidden layers having 128, 256, and 128 nodes, respectively, and generates channel-wise modulation factors via the final MLP. In another implementation, the feature attention generator incorporates anatomical priors via an additional branch that processes segmentation masks of key anatomical structures such as bone, soft tissue, and contrast-enhanced blood vessels. In yet another implementation, the feature attention generator includes separate processing paths for contrast-enhanced phase timelines and DFOV information, where a dedicated MLP with two to four hidden layers learns the relationship between these parameters and the desired feature modulation. In some implementations, the feature attention generator first processes the input energy spectrum through an MLP with layer sizes [64, 128, 64], processes the target energy spectrum through a parallel MLP with layer sizes [64, 128, 64], concatenates these outputs, and processes them through a final MLP with layer sizes [512, 256, C], where C is the number of feature channels. In addition to the final sigmoid activation, batch normalization and ReLU activation are performed between the layers.

[0089] At operation 612, the image processing system adjusts the features of the feature map using a feature adjustment vector. In one implementation, the adjustment vector contains channel-specific factors between 0 and 1, generated using a sigmoid activation function, to selectively modify different feature channels based on the correlation between different feature channels and the target energy level transformation. In another implementation, the adjustment factors can be spatially variable maps that allow independent adjustment of different regions based on local image characteristics such as tissue type, contrast enhancement level, and anatomical location. In yet another implementation, feature adjustment can be implemented as a combination of channel-wise adjustment and spatial adjustment, where each channel is first globally modified by the adjustment vector, followed by element-wise operations using a spatial adjustment map that modulates the features based on local context. In one implementation, the spatial adjustment map is generated by processing the local 3×3 neighborhood of the feature map using 1×1 convolutions followed by sigmoid activation, resulting in pixel-wise adjustment factors that can modify features based on both the channel identifier and spatial context. Spatially aware adjustment allows the network to adjust its feature transformation behavior based on local anatomical structures and their specific energy-related attenuation characteristics.

[0090] At operation 614, the image processing system processes the adjusted features through the decoder path of the neural network to generate a transformed CT image. In one embodiment, the decoder path includes multiple decoder levels corresponding to the number of encoder levels, where each decoder level uses transposed convolution operations to progressively upsample features while reducing channels, and skip connections from the corresponding encoder level enable direct access to spatial details that may have been lost during encoding. In another embodiment, the number of decoder levels can be dynamically determined based on the input image resolution and the desired feature reconstruction hierarchy, echoing the dynamic level determination of the encoder path. In yet another embodiment, each decoder level applies a residual block with dual 3×3 convolutions to refine features before upsampling to the next resolution level, where a specific number and configuration of convolution operations are matched to maintain symmetry with the corresponding encoder level. The decoder path may incorporate a feature attention generator at some or all levels to enable selective feature tuning during reconstruction, where the inclusion of attention mechanisms at specific levels is determined based on empirical evaluations of reconstruction quality versus computational cost for different types of energy conversion. In another implementation, the decoder layers gradually reduce the number of feature channels (e.g., from 512 to 256 to 128 to 64) while increasing the spatial dimension through transposed convolutions. These channel variations match the corresponding encoder layers to maintain consistent feature representation across the network. The decoder path can also incorporate additional skip connections within each layer to facilitate gradient flow during training and enable the learning of residual features associated with energy-specific transformations, echoing the residual connections in the encoder path.

[0091] At operation 616, the image processing system displays and / or saves the transformed CT image at the target energy level. In one embodiment, the transformed image is displayed side-by-side with the original image to allow direct comparison of tissue contrast and anatomical detail between the energy levels. In another embodiment, the transformed image is saved in DICOM format, with appropriate labels indicating the composite nature of the image and the energy level transformation performed.

[0092] The technical advantages of Method 600 are as follows: By coordinating the use of physics-based priors and learned feature modulation mechanisms, it enables accurate energy conversion of CT images while preserving anatomical details. Multi-scale processing through selective feature modification of dynamically generated modulation vectors at each decoder level allows the network to maintain high-frequency details while appropriately transforming contrast and tissue properties according to known X-ray physics. By combining both the input and target energy spectra, along with the contrast agent phase timeline and reconstruction parameters, as priors, this method enables precise control over energy conversion across different contrast agent phases and display fields. A feature attention generator (which can be selectively incorporated at specific encoder-decoder levels based on computational efficiency considerations) provides channel-wise modulation factors and potentially spatially varying modulation factors that adjust the transformation behavior based on both local and global image characteristics. This unified approach eliminates the need for separate models for each energy combination and contrast agent phase while ensuring consistent quality across different anatomical regions, contrast agent enhancement patterns, and acquisition protocols.

[0093] refer to Figure 7 A flowchart of a method 700 for processing CT image data through multiple encoder levels is shown. Method 700 can be manufactured by an image processing system (such as...) Figure 3 The image processing system 300 shown executes according to instructions stored in memory and executed by one or more processors. Method 700 can be executed as part of operation 608 of method 600 to process CT images via an encoder path. Method 700 enables processing of CT images via a multi-level encoder path, which incorporates feature conditioning mechanisms at each level to dynamically adjust features based on input and target priors. In one embodiment, the number of encoder levels can vary between 3 and 6 depending on the input image resolution and the desired feature hierarchy depth, wherein each level applies a convolutional operation followed by pooling to reduce spatial dimensionality while increasing feature channels.

[0094] At operation 702, the image processing system processes the CT image through a first encoder layer, which includes several coordinated sub-operations working together to extract meaningful features from the input image and process those meaningful features. At operation 704, the image processing system applies convolutional filters and activation functions to the CT image. In one embodiment, the convolutional layer applies multiple learned convolutional filters of size 3×3 or 5×5, followed by the application of a nonlinear activation function (such as ReLU) to extract low-level image features, including edge, texture, and intensity gradient properties of different tissue types and contrast-enhanced regions. In another embodiment, a specialized kernel may be designed to capture tissue boundaries, contrast-enhanced vascular structures, and local energy-related attenuation patterns. At operation 706, the image processing system generates a first output encoder feature map based on the applied convolutional filters and activation functions. This feature map preserves structural and contrast information while beginning to capture energy-specific properties. At operation 708, the image processing system stores the output encoder feature map for use as a skip connection with the corresponding decoder level, thereby enabling direct access to earlier spatial information that may have been lost during downsampling.

[0095] At operation 712, the image processing system determines whether there are additional encoder layers. If yes, the method proceeds to operation 714. If no, the method proceeds to operation 726. This decision point allows the network to process the image through multiple encoding layers while maintaining precise control over feature extraction at each layer. In one implementation, the number of encoder layers can vary between 3 and 6 depending on the input image resolution and the desired depth of the feature hierarchy. The specific number of layers can be determined during network architecture design based on empirical evaluations of the transform quality versus computational cost for different types of energy level transformations and contrast agent phases. In another implementation, the decision to continue through additional encoder layers can be dynamically determined based on the magnitude of the energy level transformation being performed, where larger energy level changes (e.g., 70 keV to 40 keV) utilize more encoder layers to maintain transform accuracy.

[0096] At operation 714, the image processing system downsamples the output encoder feature map for use in the next encoder level. In one implementation, pooling operations using a 2×2 window with a stride of 2 reduce the spatial dimensionality while preserving key features such as tissue boundaries, contrast-enhanced vascular structures, and local energy-related decay patterns. Downsampling helps achieve translation invariance and reduces computational complexity in deeper network layers. In another implementation, downsampling can be achieved via strided convolutions that learn downsampling patterns during training. The number of feature channels is typically doubled at each successive encoder level (e.g., from 64 to 128 to 256) to maintain rich feature representations despite the reduction in spatial dimensionality. This gradual increase in feature channels allows the network to capture increasingly complex patterns that combine both local and global image characteristics while efficiently managing computational resources.

[0097] At operation 716, the image processing system processes the downsampled feature map through the i-th encoder level. This operation includes several sub-operations: at operation 718, the image processing system applies a convolutional filter and an activation function at the i-th encoder level. In one implementation, a kernel of size 3×3 or 5×5 is designed to capture intermediate-level features, including tissue boundaries, contrast-enhanced vascular structures, and local energy-related attenuation pattern characteristics of different anatomical regions. The convolutional filter may employ dilated convolutions with different dilation rates (e.g., 2, 4, or 8) to capture multi-scale contextual information without increasing computational overhead. In another implementation, the activation function includes LeakyReLU with a negative slope of 0.01 to prevent ReLU death during training, or PReLU with a learnable slope per channel. At operation 720, the image processing system generates an i-th output encoder feature map containing a more abstract representation than previous layers, where the number of feature channels is typically doubled from previous layers (e.g., from 128 channels to 256 channels) to maintain rich feature representation despite the reduction in spatial dimension. At operation 722, the image processing system stores the i-th output encoder feature map for use in skip connections to the corresponding decoder layer, thereby enabling direct access to intermediate features that help preserve fine anatomical details during upsampling. Skip connections may employ 1×1 convolutions to match the channel dimension before concatenation with decoder features.

[0098] At operation 726, the image processing system downsamples the Nth output encoder feature map and passes it to the first decoder level, where N represents the total number of encoder levels in the network architecture. The final encoder level operates at the lowest spatial resolution while maintaining rich feature representations through increased channel depth. In one implementation, the lowest level operates at 1 / 16 or 1 / 32 of the original image resolution, with 512 or 1024 feature channels, and is downsampled using max pooling or strided convolutions. The increased channel depth at this level enables the capture of highly abstract representations that combine both local and global image features related to energy transformation. In another implementation, the final encoder level may incorporate residual connections or dense connections to facilitate gradient flow during training, where each residual block contains dual 3×3 convolutions followed by batch normalization and ReLU activation. During architecture design, the specific configuration of the lowest level (including resolution reduction factor and number of feature channels) can be determined based on the input image resolution and the desired feature hierarchy depth, where larger images potentially benefit from additional downsampling to achieve an appropriate balance between computational efficiency and feature representation capability. After operation 726, method 700 returns to continue operation 610 of method 600.

[0099] refer to Figure 8 A flowchart of a decoder-based CT image transformation method 800 is shown. Method 800 can be generated by an image processing system (such as...) Figure 3 The image processing system 300 executes to process the feature map through the decoder path and generate a transformed CT image at the target energy level. Method 800 may be executed as part of operation 614 of method 600 to process the scaled features through the decoder path.

[0100] At operation 802, the image processing system processes the first decoder layer through several coordinated sub-operations. At operation 804, the image processing system receives the Nth encoder feature map from the lowest level of the encoder path. This Nth encoder feature map contains an abstract representation combining both local and global image features related to energy conversion. In one embodiment, the Nth encoder feature map is operated at 1 / 16 or 1 / 32 of the original image resolution, having 512 or 1024 feature channels to maintain rich representational power despite the reduction in spatial dimension. In another embodiment, the Nth encoder feature map includes residual connections to facilitate gradient flow during training, where each residual block contains dual 3×3 convolutions followed by batch normalization and ReLU activation. At operation 806, the image processing system applies convolutional filters and activation functions to the Nth encoder feature map. In one implementation, multiple learned convolutional filters of size 3×3 or 5×5 are applied, followed by a nonlinear activation function (such as ReLU) to process the feature map, where specialized kernels are designed to capture tissue boundaries, contrast-enhanced vascular structures, and energy-related attenuation patterns. In another implementation, the convolutional operation employs dilated convolutions with different dilation rates to capture multi-scale contextual information without increasing computational overhead.

[0101] At operation 808, the image processing system generates an output decoder feature map based on the applied convolutional filters and activation functions. The output decoder feature map preserves structural and contrast information while beginning to capture energy-specific characteristics through the coordinated application of learned filters and nonlinear transformations. In one implementation, the output decoder feature map maintains the same channel dimension as the input while refining the feature representation through the applied transformations. In another implementation, additional post-processing operations, such as channel attention or adaptive normalization, can be applied to enhance the quality of the output feature map.

[0102] At operation 810, the image processing system determines whether additional decoder layers are available. This decision point allows the network to process features through multiple decoder layers while maintaining precise control over feature reconstruction at each layer. The specific number of decoder layers can be determined during network architecture design based on empirical evaluations of transform quality versus computational cost for different types of energy level transformations and contrast agent phases. In one implementation, the number of decoder layers matches the number of encoder layers and can vary between 4 and 6 layers depending on the input image resolution and the desired depth of the feature hierarchy. In another implementation, the decision to continue through additional decoder layers can be dynamically determined based on the magnitude of the energy level transformation being performed, where larger energy level changes utilize more decoder layers to maintain transform accuracy.

[0103] If more decoder layers exist, at operation 812, the image processing system upconvolves the output decoder feature map for use in the next decoder layer. In one implementation, a transposed convolution operation with a learnable filter of size 2×2 or 4×4 and a stride of 2 is used to double the spatial dimension while reducing the number of feature channels, typically halving the number of channels at each layer to maintain computational efficiency while preserving salient features. In another implementation, the upconvolution may employ different interpolation methods (such as bilinear upsampling) followed by standard convolution, where a particular method is selected based on empirical evaluation of reconstruction quality and computational requirements. The upconvolution operation is designed to minimize checkerboard artifacts while maintaining smooth feature reconstruction across different spatial scales.

[0104] At operation 814, the image processing system processes the convolved feature map through the i-th decoder level using several coordinated sub-operations. At operation 816, the image processing system concatenates the convolved feature map with the output encoder feature map from the corresponding encoder level via skip connections. This allows direct access to earlier spatial information that may have been lost during encoding, thus helping to preserve fine anatomical details and structural consistency. In one implementation, a 1×1 convolution can be applied before concatenation to match the channel dimension and learn feature combinations. At operation 818, the image processing system generates a feature conditioning vector using the i-th feature attention generator, based on input priors including energy spectrum and temporal information, and target priors.

[0105] At operation 820, the image processing system adjusts the cascaded feature maps using the i-th feature adjustment vector. In one implementation, each channel of the feature map is modified by the corresponding value in the feature adjustment vector, which may involve scaling, shifting, or other transformations to selectively modify different feature channels based on their correlation with the target energy level. In another implementation, the adjustment can be implemented as a combination of channel-wise adjustment and spatial adjustment, where each channel is first globally modified by the adjustment vector, followed by element-wise operations using a spatial adjustment map that modulates the features based on the local context. At operation 822, the image processing system applies convolutional filters and activation functions to the adjusted features. In one implementation, multiple 3×3 convolutional filters are applied, followed by batch normalization and ReLU activation to process the adjusted features, where a specific number and configuration of convolutional operations are matched to maintain symmetry with the corresponding encoder level. At operation 824, the image processing system generates an i-th output decoder feature map that contains refined local details while preserving energy-specific characteristics.

[0106] If no further decoder levels are available at operation 810, method 800 proceeds to operation 826. At operation 826, the image processing system applies final convolutional filters and activation functions to produce a transformed CT image at the target energy level. In one implementation, the final set of convolutional filters (e.g., 32 or 64 3×3 filters) (followed by 1×1 convolution) generates the output image, where a specific filter configuration is designed to enhance the final reconstruction quality while maintaining computational efficiency. In another implementation, additional post-processing operations (such as residual connections or adaptive normalization) may be applied to enhance the final output quality, where specific enhancements are selected based on empirical assessments of image quality metrics and clinical requirements. By coordinating the use of multi-scale processing, feature attention mechanisms, and specialized convolutional operations throughout the decoder path, the transformed image retains both fine anatomical detail and accurate energy-specific properties.

[0107] refer to Figure 9 The diagram illustrates a flowchart of a method 900 for generating feature adjustment vectors for a decoder layer. Method 900 can be generated by an image processing system (such as...) Figure 3 The image processing system 300 executes, as part of operation 610 of method 600, to generate a feature adjustment vector that enables precise control over feature adjustment during energy-to-energy transitions of CT images. The feature attention generator enables a single unified network to handle transitions between arbitrary energy levels by combining the continuous energy spectrum as a prior of both the input and target images. This allows the network to learn a shared manifold for energy-to-energy transitions while maintaining fine anatomical detail and structural consistency across different contrast agent phases. This approach is both computationally efficient (requiring only one model instead of separate models for each energy level combination) and robust because the network can leverage a larger, combined training dataset across multiple energy levels and contrast agent phases.

[0108] At operation 908, the image processing system processes the input priors via a first MLP to generate processed input priors. In one embodiment, the input priors include an input energy spectrum corresponding to a first energy level (providing X-ray attenuation values ​​across the energy range), an input contrast agent phase timeline indicative of the contrast agent dynamics, and display field of view values ​​used during image reconstruction. The first MLP processes these priors through multiple fully connected layers to learn representations influencing the energy conversion process, thereby enabling the network to adjust its behavior based on acquisition parameters. In another embodiment, the input priors may additionally include the reconstruction kernel type and scan-specific parameters, such as tube current and exposure time, that influence image characteristics.

[0109] At operation 910, the image processing system processes the target prior via a second MLP to generate a processed target prior. In one embodiment, the target prior includes a target energy spectrum corresponding to the target energy level and a target contrast agent phase timeline matched to the input timeline to maintain temporal consistency. The second MLP has an architecture similar to the first MLP but with independent weights, allowing it to learn a specific representation of the desired target properties while maintaining architectural symmetry. In another embodiment, the target prior may include an ideal material decomposition basis representing the expected decay behavior of different tissue types at the target energy level.

[0110] At operation 912, the image processing system concatenates the outputs of the first MLP and the second MLP to form a combined processed feature. In one embodiment, the concatenation operation preserves the distinct information from each processing path by stacking the processed input prior and the processed target prior into a single feature vector, which maintains a separate representation of the input energy characteristics and the target energy requirement. The combined features enable the network to consider multiple factors simultaneously when determining appropriate feature adjustment factors. In another embodiment, the concatenated features may undergo additional normalization or dimensionality adjustment operations to ensure a balanced contribution from each information stream.

[0111] At operation 914, the image processing system processes the combined features through a third MLP to generate a feature adjustment vector for the i-th decoder layer. In one implementation, the third MLP outputs a vector with dimensions matching the number of channels in the corresponding decoder layer feature map, where values ​​are normalized between 0 and 1 using a sigmoid activation function to provide an appropriate adjustment factor for selective feature adjustment. The feature adjustment vector achieves channel-wise adjustment of the decoder features through various operations, allowing the network to selectively modify different feature channels based on specific energy conversion requirements. In another implementation, the third MLP may incorporate residual connections to facilitate gradient flow during training, where each residual block contains a double fully connected layer followed by batch normalization and ReLU activation.

[0112] refer to Figure 10 A flowchart of a method 1000 for training a neural network to transform medical images between different energy levels is shown. Method 1000 can be generated by an image processing system (such as...) Figure 3 The image processing system 300 executes to train an energy-to-energy neural network 308 to transform CT images from one energy level to another while maintaining image quality across different contrast agent phases and display fields of view. Method 1000 enables the training of a unified energy-to-energy neural network (which can perform multiple energy level transformations) instead of requiring separate models for each energy combination and contrast agent phase.

[0113] At operation 1002, the image processing system receives multiple training image pairs, each pair comprising an input image at a first energy level and a target image at a second energy level. In one embodiment, the input and target images are monochrome images generated from dual-energy projection data, wherein the input image is acquired at a first energy level (e.g., 70 keV), and the target image is acquired at a second energy level (e.g., 40 keV, 50 keV, or 60 keV). In another embodiment, the training image pairs may include images from various contrast agent phases (e.g., CTPA phases, CTA phases, arterial phases, portal venous phases, delayed phases, and delayed equilibrium phases) and a display field of view (DFOV) ranging from 29 cm to 50 cm, thereby enabling the neural network to learn robust transformations across different imaging conditions. The training image pairs may be pre-labeled by a radiologist to indicate their corresponding contrast agent phases, thereby allowing the network to learn phase-specific transformation characteristics.

[0114] At operation 1004, the image processing system obtains input priors, which include an input energy spectrum corresponding to a first energy level and an input contrast agent phase timeline. In one embodiment, an input energy spectrum is selected from a pre-recorded library of empirically measured spectra stored in energy spectrum database 316, thereby providing X-ray attenuation values ​​across an energy range corresponding to the first energy level. In another embodiment, the input priors also include an input contrast agent phase timeline indicating temporal contrast agent dynamics, a display field of view (DFOV) value used during image reconstruction, and a reconstruction kernel type indicating the filtering applied during image reconstruction. These input priors provide physically-based guidance to the neural network during training, enabling the neural network to learn to consider energy-specific transformations of both image content and acquisition parameters.

[0115] At operation 1006, the image processing system obtains target priors, which include a target energy spectrum corresponding to the second energy level and a target contrast agent phase timeline. In one embodiment, a target energy spectrum is selected from a pre-recorded library of virtual single-chromatograms stored in the energy spectrum database 316, where each spectrum represents a narrowband X-ray energy from 40 keV to 140 keV. In another embodiment, the target priors also include a target contrast agent phase timeline matched to the input timeline to maintain temporal consistency, a target DFOV matched to the input DFOV to preserve spatial characteristics, and a reconstruction kernel type for the input image. These target priors define the desired characteristics of the transformed image, thereby guiding the network to produce an output that accurately reflects the expected appearance at the target energy level.

[0116] At operation 1008, the image processing system processes the input image through an encoder path of a neural network to generate feature maps. In one embodiment, the encoder path includes multiple encoding levels operating at progressively decreasing spatial resolution, where each level applies a convolutional operation followed by pooling to reduce spatial dimensionality while increasing feature channels. In another embodiment, the encoder path may include 3 to 6 encoding levels, depending on the input image resolution and the desired depth of the feature hierarchy, where each level extracts increasingly abstract representations that combine both local and global image properties related to energy conversion. The encoder path enables the network to perform two operations: capturing fine anatomical details at higher resolution and capturing global contextual information at lower resolution.

[0117] At operation 1010, the image processing system uses a feature attention generator to generate a feature adjustment vector based on input and target priors. In one implementation, each feature attention generator processes the input and target energy spectra via a parallel MLP and generates a channel-wise adjustment factor via a final MLP. In another implementation, the feature attention generator can be combined at multiple encoder-decoder layers to enable precise control over feature adjustment at each scale of the transformation process. The feature adjustment vector allows the network to dynamically adjust its behavior based on global target energy characteristics, thus allowing it to selectively modify different feature channels based on the correlation between different feature channels and the target energy level.

[0118] At operation 1012, the image processing system adjusts the features of the feature map using a feature adjustment vector. In one implementation, the adjustment vector contains channel-specific factors between 0 and 1, generated using a sigmoid activation function, to selectively modify different feature channels. In another implementation, the adjustment can be implemented as a combination of channel-wise adjustment and spatial adjustment, where each channel is first globally modified by the adjustment vector, followed by element-wise operations using a spatial adjustment map that modulates features based on local context. This selective feature adjustment enables the network to adjust its transformation behavior based on specific energy conversion requirements.

[0119] At operation 1014, the image processing system processes the adjusted features through the decoder path to generate a transformed image at a second energy level. In one embodiment, the decoder path includes multiple decoder levels corresponding to the number of encoder levels, where each decoder level uses transposed convolution operations to progressively upsample the features while reducing channels, and skip connections from the corresponding encoder levels enable direct access to spatial details that may have been lost during encoding. In another embodiment, each decoder level applies residual blocks with double 3×3 convolutions to refine the features before upsampling to the next resolution level, where a specific number and configuration of convolution operations are matched to maintain symmetry with the corresponding encoder levels. The decoder path reconstructs the transformed image at the target energy level while preserving fine anatomical details and structural consistency.

[0120] At operation 1016, the image processing system compares the transformed image with the target image and adjusts the neural network parameters via backpropagation. In one embodiment, a loss is determined based on the difference between the transformed image and the target image, where a loss function is used to update the parameters of the neural network via backpropagation. In another embodiment, the loss function may include multiple components, such as per-pixel mean squared error, a perceptual loss term based on feature differences in a pre-trained network, and a structural similarity term to ensure preservation of anatomical structures. In one embodiment, the loss function may specifically incorporate an energy-related weighting factor that emphasizes areas of most pronounced energy-specific contrast differences, such as contrast-enhanced blood vessel and organ boundaries. In yet another embodiment, the loss function may include a cycle consistency term that ensures the transformed image can be accurately converted back to the original energy level via inverse transformation, thereby preserving structural integrity and preventing cross-voxel tissue or contrast mixing. This forward-backward consistency is particularly important for maintaining diagnostic quality across different energy levels and contrast agent phases. The loss function may also include a phase-specific component adjusted based on the contrast agent phase timeline provided as prior information, where higher weights are assigned to anatomical regions expected to show significant enhancement during a specific contrast agent phase. For example, during arterial phase transitions, the loss function may place greater emphasis on accurately preserving contrast differences in arterial structures, while during portal vein phase transitions, it may place greater emphasis on the liver parenchyma and portal vein vessels. In some implementations, the loss function may also incorporate physically-based constraints derived from the input and target energy spectra to ensure that the transformed image reflects the expected attenuation changes based on known X-ray physics.

[0121] The backpropagation process enables the network to iteratively improve its transformation accuracy by adjusting weights and biases to minimize the difference between the transformed image and the ground truth target image. For the feature attention generator, the backpropagation process specifically updates the parameters of multiple MLPs that process spectral and temporal information, allowing the network to learn feature tuning factors for different combinations of energy levels and contrast agent phases. This targeted parameter tuning ensures that the network can generalize to combinations of energy levels and contrast agent phases not explicitly represented in the training data, enabling a single unified model to handle multiple energy conversion tasks that would otherwise require separate dedicated models.

[0122] By combining the input and target energy spectra along with the contrast agent phase timeline and reconstruction parameters as priors during training, Method 1000 enables the neural network to learn physical relationships between energy levels, rather than memorizing discrete mappings. The feature tuning mechanism trained through Operations 1010 to 1012 provides channel-specific and potentially spatially variable tuning factors that adjust transformation behavior based on global image characteristics. The backpropagation process at Operation 1016 specifically updates the parameters of multiple MLPs processing energy spectrum and phase information, allowing the network to learn feature tuning factors for different combinations of energy levels and contrast agent phases. This targeted parameter tuning ensures the network can generalize to combinations of energy levels and contrast agent phases not explicitly represented in the training data, enabling a single unified model to handle multiple energy conversion tasks that would otherwise require separate dedicated models. Additionally, the cycle consistency loss component incorporated in Operation 1016 ensures that the transformed image can be accurately converted back to the original energy level via inverse transformation, preserving structural integrity and preventing cross-voxel tissue or contrast blending. The training method of Method 1000 significantly reduces computational overhead and system complexity while ensuring consistent quality across different anatomical regions, contrast-enhanced patterns, and acquisition protocols.

[0123] refer to Figure 11 A flowchart of a method 1100 for training a neural network to transform medical images between different energy levels is shown. Method 1100 can be generated by an image processing system (such as...) Figure 3 The image processing system 300 executes to train the energy-to-energy neural network 308 to transform CT images from one energy level to multiple target energy levels while maintaining image quality across different contrast agent phases and display fields of view. Method 1100 enables the training of a unified energy-to-energy neural network (which can perform multiple energy level transformations) instead of requiring separate models for each energy combination and contrast agent phase.

[0124] At operation 1102, the image processing system receives a training image set comprising an input image at a first energy level and multiple target images at different energy levels, the input image and the multiple target images being generated from dual-energy projection data. In one embodiment, the input image and target images are monochrome images generated from dual-energy projection data, wherein the input image is acquired at a first energy level (e.g., 70 keV), and the target images are acquired at different target energy levels (e.g., 40 keV, 50 keV, and 60 keV). In another embodiment, the training image set may include images from various contrast agent phases (e.g., CTPA phases, CTA phases, arterial phases, portal venous phases, delayed phases, and delayed equilibrium phases) and a display field of view (DFOV) ranging from 29 cm to 50 cm, thereby enabling the neural network to learn robust transformations across different imaging conditions. The training image set pairs may be pre-labeled by a radiologist to indicate their corresponding contrast agent phases, thereby allowing the network to learn phase-specific transformation characteristics.

[0125] At operation 1104, the image processing system obtains input priors, which include an input energy spectrum corresponding to a first energy level and an input contrast agent phase timeline. In one embodiment, an input energy spectrum is selected from a pre-recorded library of empirically measured spectra stored in energy spectrum database 316, thereby providing X-ray attenuation values ​​across an energy range corresponding to the first energy level. In another embodiment, the input priors also include display field of view (DFOV) values ​​used during image reconstruction and a reconstruction kernel type indicating the filtering applied during image reconstruction. These input priors provide physically-based guidance to the neural network during training, enabling the neural network to learn to consider energy-specific transformations of both image content and acquisition parameters.

[0126] At operation 1106, the image processing system obtains target priors, which include target energy spectra corresponding to target energy levels and target contrast agent phase timelines for each target energy level. In one embodiment, target energy spectra are selected from a pre-recorded library of virtual single-chromatograms stored in the energy spectrum database 316, where each spectrum represents a narrowband X-ray energy from 40 keV to 140 keV. In another embodiment, the target priors also include target contrast agent phase timelines matched to the input timelines to maintain temporal consistency, target DFOVs matched to the input DFOVs to preserve spatial characteristics, and a reconstruction kernel type for the input image. These target priors define the desired characteristics of the transformed image, thereby guiding the network to produce outputs that accurately reflect the expected appearance at each target energy level.

[0127] At operation 1108, the image processing system processes the input image using input and target priors via a neural network to map the input image to each target energy level, including an identity transformation at the input energy level. In one implementation, the neural network processes the input image via an encoder path to generate a feature map, generates a feature adjustment vector based on the input and target priors using a feature attention generator, adjusts the features of the feature map using the feature adjustment vector, and processes the adjusted features via a decoder path to generate a transformed image at each target energy level. The identity transformation at the input energy level is used as an additional constraint during training to ensure that the network learns to preserve the original image characteristics without introducing artifacts or distortion when the target energy level matches the input energy level.

[0128] At operation 1110, the image processing system compares each transformed image with its corresponding target image at the corresponding target energy level, including comparing the identity-transformed image with the input image. In one implementation, the comparison includes calculating multiple similarity measures between each transformed image and its corresponding target image based on feature differences in a pre-trained network, such as mean squared error, structural similarity index, and perceptual loss. In another implementation, the comparison may incorporate an energy-related weighting factor that emphasizes regions with the most pronounced energy-specific contrast differences, such as contrast-enhanced blood vessel and organ boundaries. The comparison of the identity-transformed image with the input image serves as a self-consistency check, thereby ensuring that the network maintains image fidelity even when energy transformation is not required.

[0129] At operation 1112, the image processing system aggregates the losses from all comparisons and adjusts the neural network parameters via backpropagation. In one implementation, the aggregated loss function combines the individual losses from each energy level transformation with different weighting factors based on the magnitude of the energy level change. In some implementations, larger energy level transformations receive higher weights to ensure transformation fidelity across larger energy differences between the input and the target. In another implementation, the loss function may include a cycle consistency term that ensures the transformed image can be accurately converted back to the original energy level via inverse transformation, thereby preserving structural integrity and preventing tissue or contrast blending across voxels. This multi-objective training approach enables the network to learn a shared manifold for energy-to-energy transformations while maintaining fine anatomical detail and structural consistency across different contrast agent phases.

[0130] At operation 1114, the image processing system stores a trained neural network model. In one embodiment, the trained model is stored as an energy-to-energy neural network 308 in non-transitory memory 306 for subsequent use in transforming clinical CT images between different energy levels. In another embodiment, the stored model includes not only network weights and architecture but also a feature attention generator and a priori encoding module, which enable the network to process spectral and temporal timeline information during inference. The stored model provides a unified solution for energy transformation across different contrast agent phases and display fields of view, thereby eliminating the need for separate models for each energy combination and contrast agent phase.

[0131] The technical advantage of Method 1100 is that it enables the training of a single unified neural network that can perform multiple energy level transformations while maintaining consistency across different contrast agent phases and display fields of view. By training simultaneously with multiple target energy levels and combining the input and target energy spectra along with the contrast agent phase timeline as priors, this method allows the network to learn physical relationships between energy levels, rather than memorizing discrete mappings. The multi-target training method with aggregation loss significantly reduces computational overhead and system complexity while ensuring consistent quality across different anatomical regions, contrast agent enhancement patterns, and acquisition protocols. Including identity transformations during training further enhances the network's ability to maintain image fidelity across the entire energy transformation spectrum.

[0132] refer to Figure 12 The diagram illustrates a graphical representation of an X-ray spectrum 1200, showing the relationship between spectral intensity and voltage at various operating voltages. The X-ray spectrum 1200 includes a voltage axis 1202 ranging from 0 kV to 150 kV and a spectral amplitude axis 1204 ranging from 0 to 14 × 10⁶ units. The X-ray spectrum 1200 illustrates how X-ray intensity varies with energy, which serves as the physical basis for the input and target priors used by the energy-to-energy neural network described herein.

[0133] The X-ray spectrum 1200 displays multiple spectral curves corresponding to different peak kilovolt (kVp) settings commonly used in CT imaging, including 70 kVp curve 1206, 80 kVp curve 1208, 90 kVp curve 1210, 100 kVp curve 1212, 110 kVp curve 1214, 120 kVp curve 1216, 130 kVp curve 1218, and 140 kVp curve 1220. In one embodiment, these spectral curves represent data stored in… Figure 3 The energy spectrum in the energy spectrum database 316 is provided as both the input prior and the target prior. Figure 4Feature attention generators 452 and 456 are shown. In another embodiment, the spectral curve represents X-ray intensity values ​​across an energy range encoded by the prior coding module 312 prior to processing by the feature attention generator.

[0134] Each spectral curve exhibits a primary characteristic peak 1222 and a secondary characteristic peak 1224, which are distinguishing features generated by the X-ray tube target material. In one embodiment, the primary characteristic peak 1222 corresponds to the K-α emission line of the target material, while the secondary characteristic peak 1224 corresponds to the K-β emission line. In another embodiment, these characteristic peaks provide unique spectral features that enable the feature attention generator to distinguish different energy levels when transforming the image. The position and amplitude of the primary characteristic peak 1222 and the secondary characteristic peak 1224 vary with different kVp settings, thereby providing energy-specific information to guide the transformation process. This energy-specific information enables the energy-to-energy neural network 308 to accurately transform the image between different energy levels.

[0135] The spectral curves illustrate that X-ray intensity typically increases across the energy spectrum as kVp increases, with peak intensity shifting to higher energies. This physical relationship forms the basis for the contrast differences observed when images are transformed between different energy levels (e.g., from 70 keV to 50 keV or 40 keV). By providing these energy spectra as input and target priors to a feature attention generator, an energy-to-energy neural network can accurately model the energy-related properties of different tissues and contrast agents, enabling precise energy conversion while maintaining image quality across different contrast agent phases and display fields of view.

[0136] refer to Figure 13 An example of an encoded time-phase contrast agent timeline is shown in time series diagram 1300, illustrating how contrast agent phase information can be encoded into a vector for an energy-to-energy neural network. Time series diagram 1300 includes multiple sequences representing different contrast agent phases, each sequence being encoded as a vector of predetermined length, wherein a subset of consecutive elements is set to a non-zero value (denoted as one) corresponding to the time period of the indicated contrast agent phase, while the remaining elements are set to zero.

[0137] Sequence 1302 shows activation of the delayed equilibrium phase, where non-zero values ​​appear at the rightmost part of the vector, indicating that the contrast agent has reached delayed equilibrium, at which point it leaves the body. Sequence 1304 shows activation of the delayed phase, where non-zero values ​​are located earlier than in sequence 1302 but later than in the other sequences, representing the time period when the contrast agent has diffused into the interstitial space. Sequence 1306 shows activation of the portal vein phase, where non-zero values ​​are located in the middle part of the vector, corresponding to the time when the contrast agent enhances the portal vein and hepatic vein. Sequence 1308 shows activation of the arterial phase, where non-zero values ​​are located even earlier in the timeline, representing the time period when the contrast agent primarily enhances arterial structures.

[0138] The timeline phases are arranged from left to right, starting with the non-contrast agent phase (before contrast agent injection), followed by the CTPA (pulmonary artery phase), aortic (CTA) phase, late arterial (arterial) phase, portal vein phase, delayed phase, and delayed equilibrium phase. This progression corresponds to the physiological timeline of the contrast agent flowing through the body after injection.

[0139] The phase timeline module 314 maintains the phase timeline data indicating the pattern of contrast agent enhancement across different anatomical structures over time, thereby providing a phase timeline prior so that the network can maintain consistency across different contrast agent phases during energy conversion.

[0140] These encoded contrast agent phase timelines can be incorporated into prior information 450, which guides a feature attention generator to dynamically adjust features based on specific contrast agent phases of the input image, thereby enabling accurate energy conversion across the entire contrast agent phase cycle.

[0141] refer to Figure 14 The accompanying comparative display shows input, target, and predicted images at different energy levels (70 keV, 60 keV, and 50 keV). The display includes a top row of reference images and a bottom row of predicted images generated using the unified energy-to-energy neural network described herein. This comparative display enables a visual evaluation of the ability of the energy-to-energy neural network disclosed herein to transform CT images across different energy levels while preserving anatomical detail and image quality across different contrast agent phases and display fields of view.

[0142] The top row includes an input 70 keV image 1402, which in one embodiment is a raw CT image acquired at 70 keV (approximately equal to 120 kVp) during the portal vein contrast phase. In another embodiment, the input 70 keV image 1402 may be a virtual monochrome image generated from dual-energy projection data. The input 70 keV image 1402 serves as the source image for energy conversion and provides baseline anatomical information for comparison with the transformed image. Adjacent to the input image is a target 60 keV image 1404, which in one embodiment is a reference ground truth image acquired at 60 keV using dual-energy CT imaging technology. In another embodiment, the target 60 keV image 1404 may be a reference image generated by material decomposition of dual-energy projection data. The target 60 keV image 1404 demonstrates the expected appearance and contrast characteristics at the 60 keV energy level. Further to the right is the target 50 keV image 1406, which in one embodiment is a baseline ground truth image acquired at 50 keV using dual-energy CT imaging technology. In another embodiment, the target 50 keV image 1406 can be a reference image generated by material decomposition of dual-energy projection data. The target 50 keV image 1406 demonstrates the expected appearance and contrast characteristics at the 50 keV energy level, showing enhanced contrast compared to 60 keV and 70 keV images.

[0143] The bottom row includes a predicted 70 keV image 1408, which in one embodiment is generated by an energy-to-energy neural network when provided with an input 70 keV image 1402 and both an input prior and a target prior specifying 70 keV as the input energy level and the target energy level, respectively. In another embodiment, the predicted 70 keV image 1408 may be generated as part of a cycle consistency verification process to verify the network's ability to maintain image quality when the target energy matches the input energy. The predicted 70 keV image 1408 should closely match the input 70 keV image 1402, thus demonstrating the network's ability to preserve image characteristics when no energy transformation is required. The neighboring image 1408 is a predicted 60 keV image 1410, which in one embodiment is generated by an energy-to-energy neural network when provided with an input 70 keV image 1402, an input prior specifying 70 keV, and a target prior specifying 60 keV. The predicted 60 keV image 1410 closely matches the target 60 keV image 1404, demonstrating the network's ability to accurately transform images to intermediate energy levels. Further to the right is the predicted 50 keV image 1412, which in one implementation is generated by an energy-to-energy neural network when provided with an input 70 keV image 1402, an input prior specifying 70 keV, and a target prior specifying 50 keV. The predicted 50 keV image 1412 closely matches the target 50 keV image 1406, demonstrating the network's ability to accurately transform images to lower energy levels using enhanced contrast visualization.

[0144] refer to Figure 15 A comparative visualization of the results of energy level conversion using the energy-to-energy neural network disclosed in this paper is shown. Figure 15 It includes three panels that show the transformation from a higher energy (70 keV) input image to a lower energy (50 keV) predicted image, as well as the actual target image used for comparison.

[0145] exist Figure 15The left panel shows the input 70 keV image 1502. In one embodiment, the input 70 keV image 1502 is a raw CT image acquired at 70 keV during the portal vein contrast phase. In another embodiment, the input 70 keV image 1502 may be a virtual monochrome image generated from dual-energy projection data. The input 70 keV image 1502 serves as the source image for energy conversion and provides baseline anatomical information for comparison with the transformed image. The input 70 keV image 1502 exhibits typical contrast characteristics at this energy level, with moderate tissue discrimination and contrast enhancement in vascular structures.

[0146] exist Figure 15 The predicted 50 keV image 1504 is shown in the middle panel. In one embodiment, the predicted 50 keV image 1504 is generated by an energy-to-energy neural network when provided with an input 70 keV image 1502, an input prior specifying the 70 keV energy spectrum, and a target prior specifying the 50 keV energy spectrum. Compared to the input 70 keV image 1502, the predicted 50 keV image 1504 exhibits enhanced contrast, particularly in contrast-enhanced regions.

[0147] exist Figure 15 The right panel shows a target 50 keV image 1506. In one embodiment, the target 50 keV image 1506 is a baseline ground truth image acquired at 50 keV using dual-energy CT imaging. In another embodiment, the target 50 keV image 1506 can be a reference image generated by material decomposition of dual-energy projection data. The target 50 keV image 1506 demonstrates the expected appearance and contrast characteristics at the 50 keV energy level, showing enhanced contrast compared to a 70 keV image. The close visual similarity between the predicted 50 keV image 1504 and the target 50 keV image 1506 demonstrates the accuracy of the energy-to-energy neural network in performing energy conversion while preserving anatomical details and enhancing contrast visualization.

[0148] This disclosure also provides support for a method comprising: receiving a computed tomography (CT) image acquired at a first energy level; obtaining an input prior, the input prior including an input energy spectrum corresponding to the first energy level; obtaining a target prior, the target prior including a target energy spectrum corresponding to a target energy level; processing the CT image via a neural network to generate at least one feature map; generating a feature adjustment vector based on the input prior and the target prior using a feature attention generator; adjusting features of the at least one feature map using the feature adjustment vector to generate adjusted features; and processing the adjusted features via the neural network to generate a transformed CT image at the target energy level. In a first example of the method, the input prior further includes: an input contrast agent phase timeline and an input display field of view (DFOV). In a second example of the method (optionally including the first example), generating the feature adjustment vector includes: processing the input prior through a first neural network to generate a first processed prior; processing the target prior through a second neural network to generate a second processed prior; concatenating the first processed prior and the second processed prior to form a combined processed feature; and processing the combined processed feature through a third neural network to generate the feature adjustment vector, wherein the feature adjustment vector has a dimension corresponding to the number of channels in the at least one feature map. In a third example of the method (optionally including one or both of the first and second examples), adjusting the features of the at least one feature map using the feature adjustment vector includes: modifying the at least one feature map by one or more operations based on the value of the feature adjustment vector. In a fourth example of the method (optionally including one or more of the first to third examples), the one or more operations include: multiplying each channel of the at least one feature map by the corresponding scalar value in the feature adjustment vector.

[0149] This disclosure also provides support for a method comprising: receiving a computed tomography (CT) image acquired at a first energy level; obtaining an input prior, the input prior including an input energy spectrum corresponding to the first energy level; obtaining a target prior, the target prior including a target energy spectrum corresponding to a target energy level; and processing the CT image through a neural network to generate a transformed CT image at the target energy level, wherein the neural network comprises: an encoder path including a plurality of encoder levels, wherein each encoder level is configured to receive an input feature map and convolve the input feature map with one or more convolutional filters to generate an output encoder feature map; and a decoder path including a decoder path with the plurality of encoder levels. The method comprises multiple decoder layers corresponding to encoder layers, each decoder layer being configured to receive an input decoder feature map, the input decoder feature map including a concatenated feature map from a lower decoder layer concatenated with the output encoder feature map from the corresponding encoder layer; generating a feature adjustment vector at each decoder layer using a feature attention generator based on the input prior and the target prior; adjusting the features of the input decoder feature map at each decoder layer using the corresponding feature adjustment vector to generate adjusted features; convolving the adjusted features at each decoder layer to generate an output decoder feature map; and outputting the transformed CT image based on the output decoder feature map from the highest level of the decoder path. In a first example of the method, generating the feature adjustment vector at each decoder layer includes: processing the input prior through a first MLP to generate a processed input prior; processing the target prior through a second MLP to generate a processed target prior; combining the processed input prior and the processed target prior to generate a combined feature; and processing the combined feature through a third MLP to generate the feature adjustment vector. In a second example of the method (optionally including the first example), combining the processed input prior and the processed target prior includes concatenating the processed input prior and the processed target prior. In a third example of the method (optionally including one or both of the first and second examples), adjusting the features of the input decoder feature map includes multiplying each channel of the input decoder feature map by the corresponding scalar value in the feature adjustment vector. In a fourth example of the method (optionally including one or more or each of the first to third examples), the input prior further includes a contrast agent phase timeline indicating the contrast agent phase of the CT image, and wherein the target prior further includes a target contrast agent phase timeline.In a fifth example of the method (optionally including one or more, or each, of the first to fourth examples), the contrast agent phase timeline comprises a one-dimensional curve indicating one or more of the following: pulmonary artery phase, aortic phase, late arterial phase, portal vein phase, delayed phase, or delayed equilibrium phase. In a sixth example of the method (optionally including one or more, or each, of the first to fifth examples), the contrast agent phase timeline is encoded as a vector of predetermined length, wherein a subset of consecutive elements in the vector is set to non-zero values ​​corresponding to the time period of the indicated contrast agent phase, while the remaining elements of the vector are set to zero. In a seventh example of the method (optionally including one or more, or each, of the first to sixth examples), the input prior further includes display field of view (DFOV) values ​​and a reconstruction kernel type for reconstructing the CT image. In an eighth example of the method (optionally including one or more or each of the first to seventh examples), the target energy level corresponds to an energy level within a continuous range between 40 keV and 140 keV. In a ninth example of the method (optionally including one or more or each of the first to eighth examples), the input energy spectrum includes X-ray attenuation values ​​across an energy range corresponding to the first energy level, and wherein the target energy spectrum includes X-ray attenuation values ​​across an energy range corresponding to the target energy level. In a tenth example of the method (optionally including one or more or each of the first to ninth examples), the neural network is trained using multiple pairs of training images, each pair including: a first training image acquired at the first energy level; a second training image acquired at the target energy level, wherein the first training image and the second training image are monochromatic images generated from dual-energy projection data; and an input prior, the input prior including: an input energy spectrum corresponding to the first energy level, an input contrast agent phase timeline indicating the contrast agent phase of the first training image, and a display for reconstructing the first training image. The training includes: a field of view (DFOV) value and a reconstruction kernel type for reconstructing the first training image; and a target prior, the target prior including: a target energy spectrum corresponding to the target energy level, a target contrast agent phase timeline matching the input contrast agent phase timeline, the DFOV value, and the reconstruction kernel type, wherein the training includes: processing the first training image and the input prior through the neural network to generate a transformed training image; determining a loss based on comparing the transformed training image with a second training image; and updating the parameters of the neural network based on the loss to train the feature attention generator.

[0150] This disclosure also provides support for a system comprising: a display device, one or more processors, and a memory storing instructions executable by the one or more processors to perform the following operations: receiving a computed tomography (CT) image acquired at a first energy level; obtaining an input prior, the input prior including an input energy spectrum corresponding to the first energy level; obtaining a target prior, the target prior including a target energy spectrum corresponding to a target energy level; processing the CT image via a neural network to generate a transformed CT image at the target energy level, wherein the neural network generates at least one feature map; generating a feature adjustment vector based on the input prior and the target prior using a feature attention generator; adjusting features of the at least one feature map using the feature adjustment vector to generate adjusted features; generating the transformed CT image using the adjusted features; and displaying the transformed CT image on the display device. In a first example of the system, the input prior further includes an input contrast phase timeline and an input display field of view (DFOV) indicating the contrast phase of the CT image, and wherein the target prior further includes a target DFOV matching the input DFOV. In a second example of the system (optionally including the first example), the instructions for generating the feature adjustment vector include instructions for: processing the input prior through a first multilayer perceptron (MLP) to generate a first processed prior; processing the target prior through a second MLP to generate a second processed prior; concatenating the first processed prior and the second processed prior to form a combined processed feature; and processing the combined processed feature through a third MLP to generate the feature adjustment vector, wherein the feature adjustment vector has a dimension corresponding to the number of channels in the at least one feature map. In a third example of the system (optionally including one or both of the first and second examples), the instructions for adjusting the features of the at least one feature map using the feature adjustment vector include instructions for: multiplying each channel of the at least one feature map by a corresponding scalar value in the feature adjustment vector to generate a weighted feature channel.

[0151] When describing elements of various embodiments of this disclosure, the articles “a,” “an,” and “the” are intended to indicate the presence of one or more such elements. The terms “first,” “second,” etc., do not indicate any order, quantity, or importance, but are used to distinguish one element from another. The terms “comprising,” “including,” and “having” are intended to be inclusive and indicate that additional elements may exist in addition to the listed elements. As used herein, the terms “connected to,” “coupled to,” etc., indicate that an object (e.g., a material, element, structure, component, etc.) may be connected to or coupled to another object, regardless of whether the one object is directly connected to or coupled to the other object, or whether one or more intervening objects exist between the one object and the other object. Furthermore, it should be understood that references to “an embodiment” or “an embodiment” of this disclosure are not intended to be construed as excluding the existence of additional embodiments also incorporating the referenced features.

[0152] In addition to any modifications previously indicated, those skilled in the art can devise many other variations and alternative arrangements without departing from the spirit and scope of this specification, and the appended claims are intended to cover such modifications and arrangements. Therefore, although the information has been described in particular and detail above in conjunction with what is now considered to be the most practical and preferred aspects, it will be apparent to those skilled in the art that many modifications, including but not limited to those in form, function, mode of operation, and purpose, can be made without departing from the principles and concepts set forth herein. Likewise, as used herein, embodiments and implementations are intended to be illustrative in all respects and should not be construed as limiting in any way.

Claims

1. A method, the method comprising: Receive computed tomography (CT) images acquired at a first energy level (402); Obtain the input prior (502), the input prior including the input energy spectrum corresponding to the first energy level; Obtain the target prior (506), the target prior including the target energy spectrum corresponding to the target energy level; The CT image (402) is processed by a neural network (308) to generate at least one feature map (406). The feature attention generator (452, 456) is used to generate a feature adjustment vector (518) based on the input prior (502) and the target prior (506). The features of the at least one feature map (406) are adjusted using the feature adjustment vector (518) to generate adjusted features; and The adjusted features are processed by the neural network (308) to generate a transformed CT image at the target energy level.

2. The method according to claim 1, wherein the input prior (502) further comprises: Enter the contrast agent phase timeline; and Input Display Field of View (DFOV).

3. The method according to claim 1, wherein generating the feature adjustment vector (518) comprises: The input prior (502) is processed by a first neural network (504) to generate a first processed prior; The target prior (506) is processed by a second neural network (508) to generate a second processed prior; The first processed prior and the second processed prior are concatenated to form a combined processed feature; and The combined processed features are processed by a third neural network (516) to generate the feature adjustment vector (518), wherein the feature adjustment vector (518) has a dimension corresponding to the number of channels in the at least one feature map (406).

4. The method according to claim 1, wherein adjusting the features of the at least one feature map (406) using the feature adjustment vector (518) comprises: The at least one feature map (406) is modified by one or more operations based on the value of the feature adjustment vector (518).

5. The method of claim 4, wherein the one or more operations include: Each channel of the at least one feature map (406) is multiplied by the corresponding scalar value in the feature adjustment vector (518).

6. A method, the method comprising: Receive computed tomography (CT) images acquired at a first energy level (402); Obtain the input prior (502), the input prior including the input energy spectrum corresponding to the first energy level; Obtain the target prior (506), the target prior including the target energy spectrum corresponding to the target energy level; The CT image (402) is processed by a neural network (308) to generate a transformed CT image at the target energy level, wherein the neural network (308) comprises: An encoder path comprising multiple encoder levels, wherein each encoder level is configured to receive an input feature map (410) and convolve the input feature map (410) with one or more convolutional filters (404, 412) to generate an output encoder feature map (406, 414); and The decoder path includes multiple decoder levels corresponding to the plurality of encoder levels, wherein each decoder level is configured to receive an input decoder feature map, the input decoder feature map including an over-convolved feature map from a lower decoder level concatenated with the output encoder feature map (406, 414) from the corresponding encoder level. The feature attention generator (452, 456) generates a feature adjustment vector (518) at each decoder level based on the input prior (502) and the target prior (506). The features of the input decoder feature map are adjusted at each decoder level using the corresponding feature adjustment vector (518) to generate adjusted features; At each decoder level, the adjusted features are convolved to generate an output decoder feature map; and The transformed CT image is output based on the output decoder feature map from the highest level of the decoder path.

7. The method of claim 6, wherein generating the feature adjustment vector (518) at each decoder layer comprises: The input prior (502) is processed by the first MLP (504) to generate the processed input prior; The target prior (506) is processed by the second MLP (508) to generate the processed target prior; The processed input priors and the processed target priors are combined to generate combined features; and The features of the combination are processed by a third MLP (516) to generate the feature adjustment vector (518).

8. The method of claim 7, wherein combining the processed input prior and the processed target prior comprises: The processed input priors and the processed target priors are concatenated.

9. The method of claim 7, wherein adjusting the features of the input decoder feature map comprises: Each channel of the input decoder feature map is multiplied by the corresponding scalar value in the feature adjustment vector (518).

10. The method of claim 6, wherein the input prior (502) further includes a contrast agent phase timeline indicating the contrast agent phase of the CT image (402), and wherein the target prior (506) further includes a target contrast agent phase timeline.

11. The method of claim 10, wherein the contrast agent phase timeline comprises a one-dimensional curve indicating one or more of the following: pulmonary artery phase, aortic phase, late arterial phase, portal vein phase, delayed phase, or delayed equilibrium phase.

12. The method of claim 11, wherein the contrast agent phase timeline is encoded as a vector of predetermined length, wherein a subset of consecutive elements in the vector is set to non-zero values ​​corresponding to the time period of the indicated contrast agent phase, while the remaining elements of the vector are set to zero.

13. The method of claim 6, wherein the input prior (502) further includes a display field of view (DFOV) value and a reconstruction kernel type for reconstructing the CT image (402).

14. The method of claim 6, wherein the target energy level corresponds to an energy level in a continuous range between 40 keV and 140 keV.

15. The method of claim 6, wherein the input energy spectrum comprises x-ray intensity values ​​across an energy range corresponding to the first energy level, and wherein the target energy spectrum comprises x-ray intensity values ​​across an energy range corresponding to the target energy level.

Images