Simulating images with higher contrast enhancement in medical applications based on inverse problems

Abstract

A solution for imaging a body part of a patient in a medical application is presented. A corresponding imaging method (400) comprises simulating (424-487) one or more simulated high-contrast images corresponding to an acquired low-contrast image; the simulated high-contrast images simulate an increase in contrast enhancement relative to the acquired low-contrast image. Each simulated high-contrast image is simulated by optimizing an objective function comprising a fidelity term based on a comparison between the acquired low-contrast image and a simulated low-contrast image; the simulated low-contrast image is generated from the simulated high-contrast image to simulate a corresponding decrease in contrast enhancement. A corresponding computer program (300) and computer program product for implementing the imaging method (400) are presented. Moreover, a computing system (110) for performing the imaging method (400) and an imaging system (100) comprising the computing system (110) and a scanner (105) are presented. In addition, a corresponding medical method is presented.

Description

Technical Field

[0001] This disclosure relates to the field of medical applications. More specifically, this disclosure relates to contrast agent-based imaging procedures in medical applications. Background Technology

[0002] The background of this disclosure is introduced below through a discussion of the techniques in relation to its context. However, even when the discussion involves documents, actions, artifacts, etc., it does not imply or represent that the techniques discussed are part of the prior art or common general knowledge in the field related to this disclosure.

[0003] Imaging procedures are common in medical applications, allowing physicians to examine a patient's body parts by providing images that offer a visual representation of the body parts (typically in a largely non-invasive manner, even if the body parts are not directly visible). For this purpose, a contrast agent is typically administered to the patient undergoing each imaging procedure to provide contrast enhancement to the (biological) target of interest (e.g., a lesion), making it more prominent in the image. This facilitates the physician's task in several medical applications, such as diagnostic applications for detecting / monitoring lesions, therapeutic applications for identifying lesions to be treated, and surgical applications for identifying lesions to be resected.

[0004] In this context, the practice of simulating increases in contrast agent dosage from low to high (e.g., from a lower reduced dose to the full dose standard in clinical practice, or from the full dose to a higher enhancement dose) has also been proposed. For this purpose, during each imaging procedure, one or more (low-dose) images of the body site are acquired after the administration of a low-dose contrast agent to the patient, in addition to possible zero-dose images of the body site acquired before contrast agent administration. Corresponding (high-dose) images of the body site are then simulated from the low-dose images (and possible zero-dose images). Therefore, in the case of administering a reduced dose and simulating a full dose, this allows for the administration of a smaller amount of contrast agent when it may pose a risk to the patient, while restoring contrast enhancement to the desired level that would otherwise be provided by a full dose of contrast agent. Alternatively, in the case of administering a full dose and simulating an enhancement dose, this allows for obtaining higher contrast enhancement than is actually achievable in practice without compromising standards of care.

[0005] Furthermore, WO-A-2023 / 135056 describes the use of contrast agents for magnetic resonance imaging (MRI) applications in computed tomography (CT) imaging procedures. These contrast agents are widely available for many targets, whereas comparable contrast agents are relatively rare for CT imaging procedures. As described above, during each imaging procedure, low-dose images are acquired after administering a low-dose contrast agent (such as gadolinium-based) to the patient for MRI procedures; then, corresponding high-dose images of the body parts are simulated from the low-dose images, simulating the administration of the corresponding high-dose contrast agent used in the CT imaging procedure to the patient.

[0006] Deep neural networks (DNNs) are generally used to simulate high-dose images from low-dose images. This (deep) neural network is trained using sample sets, each containing possible zero-dose, reduced-dose, and full-dose images of body parts of a subject (such as a human or animal) corresponding to no contrast agent, reduced-dose, and full-dose contrast agents, respectively; all (zero-dose / reduced-dose / full-dose) images in the sample set have been obtained from the subject's body parts, or some of them have been simulated from other images in the sample set.

[0007] However, this approach has inherent limitations, primarily because the problem of determining the correspondence between a high-dose image and a low-dose image is ill-posed (because it lacks a unique solution that varies continuously with the low-dose image). In particular, it is impossible to simultaneously achieve optimal stability and accuracy. Thus, if neural networks are configured to provide high accuracy, they will overfit and subsequently exhibit variance errors (amplifying small perturbations in the low-dose image), or conversely, if neural networks are configured to provide high stability, they will underfit and subsequently exhibit bias errors (i.e., low consistency between the simulated high-dose image and the corresponding high-dose image that would actually be obtained with the administration of a high-dose contrast agent).

[0008] This can negatively impact the quality of high-dose images simulated by neural networks. In fact, in real-world scenarios, low-dose images are subject to small, unavoidable perturbations, primarily due to the presence of unquantifiable noise. Consequently, high-dose images may exhibit artifacts and may fail to capture small structural changes (e.g., tumors in their early stages of development) (especially if only low noise was considered during the training phase of the neural network).

[0009] All of the above make it difficult to distinguish the target from other nearby (biological) features. This adversely affects the physician's task and poses corresponding risks to the patient's health (e.g., false positives / negatives or incorrect follow-up in diagnostic applications, reduced effectiveness of therapies or damage to healthy tissues in therapeutic applications, and incomplete resection of lesions or excessive removal of healthy tissues in surgical applications). Summary of the Invention

[0010] This document presents a simplified summary of the invention in order to provide a basic understanding thereof; however, the sole purpose of this summary is to introduce some concepts of the disclosure in a simplified form as a preamble to its more detailed description thereafter, and it should not be construed as identifying its key elements or defining its scope.

[0011] Generally speaking, this disclosure is based on the idea of ​​making full use of the inverse problem.

[0012] Specifically, one aspect provides an imaging method for imaging body parts of a patient in a medical application. This imaging method includes simulating one or more simulated high-contrast images corresponding to an acquired low-contrast image; the simulated high-contrast images simulate an increase in contrast enhancement relative to the acquired low-contrast image. Each simulated high-contrast image is simulated by optimizing an objective function including a fidelity term based on a comparison between the acquired low-contrast image and the simulated low-contrast image; the simulated low-contrast image is generated from the simulated high-contrast image to simulate a corresponding decrease in contrast enhancement.

[0013] A further aspect provides a computer program for implementing this imaging method.

[0014] A further aspect provides a corresponding computer program product.

[0015] A further aspect provides a computational system for implementing this imaging method.

[0016] A further aspect provides an imaging system that includes a computing system and a scanner.

[0017] A further aspect provides a corresponding medical approach.

[0018] More specifically, one or more aspects of this disclosure are set forth in the independent claims, and their advantageous features are set forth in the dependent claims, the wording of all claims being incorporated herein by reference verbatim (with reference and necessary modifications to apply to any particular aspect provided by each of the other aspects). Attached Figure Description

[0019] The solutions of this disclosure, as well as its further features and advantages, will be best understood by referring to the following detailed description, which is given only by non-limiting indication and whose explanations are intended to be applied by analogy to each aspect therein (regardless of the context in which they appear); this description should be read in conjunction with the accompanying drawings (where, for simplicity, corresponding elements are indicated by the same or similar reference numerals, and their explanations are not repeated, and the name of each entity is generally used to indicate its type and attributes, such as value, content, and representation). In particular: Figure 1 A schematic block diagram of an imaging system that can be used to implement solutions according to embodiments of the present disclosure is shown. Figure 2 The general principles of a solution according to embodiments of this disclosure are illustrated. Figure 3 The main software components that can be used to implement the solutions according to embodiments of this disclosure are shown. Figures 4A-4B An activity diagram illustrating the flow of activities related to an imaging procedure according to an embodiment of the present disclosure is shown, and Figures 5A-5G Representative examples of experimental results related to solutions according to embodiments of this disclosure are shown. Detailed Implementation

[0020] For specific references Figure 1 A schematic block diagram of an imaging system 100 that can be used to implement solutions according to embodiments of the present disclosure is shown.

[0021] The imaging system 100 includes the following components.

[0022] Scanner 105 is used to acquire images of a patient's body part during a corresponding imaging procedure based on the administration of a contrast agent to enhance the contrast of the corresponding (biological) target (such as a lesion). For example, scanner 105 is of the MRI type and is used in an MRI procedure. In this case (not shown in detail in the figure), MRI scanner 105 has a gantry for receiving the patient; the gantry houses a superconducting magnet (for generating a very high static magnetic field), multiple sets of gradient coils for different axes (for adjusting the static magnetic field), and RF coils (with specific structures for applying magnetic pulses to a type of body part and for receiving corresponding response signals). Alternatively, scanner 105 is of the CT type and is used in a CT (imaging) procedure. In this case (again not shown in detail in the figure), CT scanner 105 has a gantry for receiving the patient; the gantry houses an X-ray generator, an X-ray detector, and motors for rotating them around the patient's body part.

[0023] A computing system 110, such as a (personal) computer, is used to control the operation of scanner 105. For this purpose, computer 110 is coupled to scanner 105. For example, in the case of MRI scanner 105, computer 110 is located outside the scanner chamber used to shield scanner 105 and is coupled to scanner 105 via a cable passing through a perforated panel; while in the case of CT scanner 105, computer 110 is located close to it and coupled to it via a short cable. Computer 110 includes several units connected between them via bus structure 115. In particular, one or more microprocessors ( Microprocessor 120 provides the logical capabilities of computer 110. Non-volatile memory (ROM) 125 stores the basic code for booting computer 110, and volatile memory (RAM) 130 is used as working memory by microprocessor 120. Computer 110 is provided with mass storage 135 for storing programs and data, such as a solid-state drive (SSD). Furthermore, computer 110 includes multiple controllers 140 for peripheral devices, or input / output (I / O) units. Specifically, in relation to this disclosure, peripheral devices include corresponding drivers for associated units of scanner 105, a monitor for displaying images, a keyboard for entering information / commands, a trackball for moving a pointer on the monitor, drives for reading / writing removable storage units (such as USB flash drives), and so on.

[0024] Now for reference Figure 2 This illustrates the general principles of a solution based on embodiments of the present disclosure.

[0025] During each imaging procedure, a computer (not shown in the figure) receives one or more (acquired low-contrast) images that provide a contrast-enhanced representation of the body part of the patient being examined. The scanner (not shown in the figure) has acquired low-contrast images to provide a target contrast enhancement (given by a contrast agent already administered to the patient) with a low value, referred to below as low contrast. For example, the acquired low-contrast image is an acquired low-dose image of a body part of the patient to which a low dose (e.g., less than or equal to the full dose standard in clinical practice) of contrast agent has been administered. Additionally or alternatively, the acquired low-contrast image is an acquired low-efficiency image of a body part of the patient to which an inefficient (inefficient) contrast agent (e.g., a contrast agent specific to a different medical procedure or body part) has been administered in the ongoing medical application. Furthermore, the scanner may also acquire one or more (acquired baseline) images of the body part. For example, the acquired baseline image is a body part without contrast enhancement (acquired zero-contrast) image (i.e., without contrast agent applied to the patient) and / or represents a body part after an acquisition delay following the application of contrast agent, less than the acquired low-contrast image (acquired low-delay image). (Hereinafter, for simplicity, reference will always be made to the acquired zero-contrast image, and it should be understood that the same considerations apply to any other(one or more) acquired baseline image). A computer-simulated or synthesized one or more (simulated high-contrast) images corresponding to the acquired low-contrast image; the simulated high-contrast image represents a body part of the patient whose simulated (or mimicked) target contrast enhancement increases from low contrast to a value higher than it, hereinafter referred to as high contrast. In particular, in the case of the acquired low-dose image, the simulated high-contrast image is a simulated high-dose image that simulates the application of contrast agent to the patient at a high dose (e.g., equal to or higher than the full dose) higher than the low dose. Furthermore, in the case of acquired low-contrast images, the simulated high-contrast images are simulated high-contrast images that simulate the administration of a high-efficiency contrast agent (e.g., a contrast agent specific to the same medical procedure and body site) to the patient in the ongoing medical application, providing a higher efficiency than the low-contrast image. The computer then outputs information related to the body site based on the simulated high-contrast image to the physician responsible for the imaging procedure (e.g., by displaying the simulated high-contrast image on a computer monitor).

[0026] In the solution according to embodiments of this disclosure, the simulated high-contrast image is not generated directly from a corresponding acquired low-contrast image. Instead, each simulated high-contrast image is determined by fully utilizing the inverse problem of simulating a corresponding (simulated low-contrast) image from the simulated high-contrast image; the simulated low-contrast image represents a body part of the patient, whose simulated contrast enhancement decreases from high contrast to low contrast. The simulated low-contrast image is generated directly from the corresponding simulated high-contrast image and a possible acquired zero-contrast image (e.g., by means of a decreasing neural network appropriately trained for this purpose). In particular, the simulated high-contrast image is simulated by optimizing an objective function that includes a fidelity term based on a comparison between the acquired low-contrast image and the simulated low-contrast image (e.g., as a function that minimizes their difference).

[0027] In the following text, for completeness, reference will always be made to generating a simulated low-contrast image from a simulated high-contrast image and an acquired zero-contrast image; however, it should be understood that the same considerations apply when generating a simulated low-contrast image from a simulated high-contrast image and any other (one or more) acquired baseline image, or when generating a simulated low-contrast image from a simulated high-contrast image alone (i.e., without any acquired baseline image).

[0028] The solution mentioned above allows for the utilization of the fact that simulating a decrease in contrast enhancement is easier than simulating an increase in contrast enhancement. Indeed, images can change from high to low contrast in a way that can be predicted with relatively high confidence (especially when a zero-contrast image is available), but this is not the case when changing from low to high contrast.

[0029] Therefore, it is possible to achieve high stability without sacrificing accuracy; thus, the simulated high-contrast images are in good agreement with the high-contrast images that would actually be acquired under corresponding conditions (i.e., with high-dose administration of contrast agents and / or administration of different contrast agents that provide high efficiency for medical applications); simultaneously, the simulated high-contrast images are relatively insensitive to small perturbations in the acquired low-contrast images. Therefore, the solution mentioned above allows for both low bias error and low variance error, thereby reducing both underfitting and overfitting. All of the above have a beneficial effect on the quality of the simulated high-contrast images, regardless of any (unquantifiable) noise unavoidable in the acquired low-contrast images. In particular, the simulated high-contrast images can exhibit fewer artifacts and can even capture small structural changes (e.g., tumors in their early stages of development).

[0030] The solutions mentioned above make it easier to distinguish a target from other nearby (biological) features. This facilitates the physician's task, significantly reducing the risk of false positives / negatives and erroneous follow-ups in diagnostic applications, reduced efficacy or damage to healthy tissue in therapeutic applications, and incomplete resection of lesions or excessive removal of healthy tissue in surgical applications.

[0031] Optimizing the fidelity term (which determines the required high-contrast image for each simulation) involves a small delay from acquiring the corresponding low-contrast image. However, this delay is relatively low (e.g., about a few seconds), making it acceptable in most practical cases, i.e., always for offline applications and generally for real-time applications as well.

[0032] As a further improvement, each simulated high-contrast image is simulated by optimizing an objective function that includes a fidelity term and an appended regularization term. The regularization term increases the cost, which ensures that the optimization considers not only the matching of the simulated low-contrast image to the acquired low-contrast image, but also the specific characteristics of the simulated high-contrast image (e.g., penalties for low variability or priors to its first estimate). This stabilizes the process of determining the simulated high-contrast image (thanks to the fact that the process relies on two sources of information, where the fidelity term prioritizes the consistency between the simulated high-contrast image and the acquired low-contrast image, and the regularization term prioritizes insensitivity to fluctuations in the acquired low-contrast image), thereby significantly improving its quality. In fact, in this way, the problem of determining the simulated high-contrast image becomes well-posed (i.e., has a unique solution that changes continuously with the acquired low-contrast image), stable, and converges as the noise level of the acquired low-contrast image approaches zero.

[0033] For example, in a specific implementation, each simulated high-contrast image is determined by solving a minimization problem (in variational or weak form, with conditions satisfied in an average sense but not necessarily at every point, in order to simplify its solution). This involves determining the minimization ( min A high-contrast image of the simulated loss function given by the fidelity term plus a regularization term:

[0034] in Fid() It is a fidelity item. Ia L It is a low-contrast image that has been acquired. Is L It is a simulated low-contrast image. It is a regularization parameter. Reg() It is a regularization term. For example, a mapping function generated by applying a decreasing neural network is used to generate simulated low-contrast images. Is L , Is H It is a simulated high-contrast image and It is possible to obtain a zero-contrast image; regularization parameter ( ≥0) determines the degree of regularization applied. The lower the regularization parameter, the lower the regularization (thus prioritizing consistency between the simulated high-contrast image and the acquired low-contrast image, and then accuracy); the higher the regularization parameter, the higher the regularization (thus prioritizing insensitivity to fluctuations in the acquired low-contrast image, and then stability). For example, the regularization parameter can preferably be set to 0.05-0.50, more preferably 0.10-0.40, and even more preferably 0.15-0.30, such as 0.2.

[0035] Fidelity Fid() It can be defined as the difference between a measured low-contrast image and a simulated low-contrast image. For example, the fidelity term. Fid() It is equal to the squared Euclidean norm (or 2-norm) of the vector difference between the acquired low-contrast image and the simulated low-contrast image:

[0036] More specifically, each image (the acquired low-contrast image, the simulated low-contrast image, the simulated high-contrast image, and possibly the acquired zero-contrast image) consists of images having... N lines and M The bitmap definition of the matrix of cells in the column (e.g., having N = 256-1024 and M=256-1024 Each cell contains the value of a basic image element (e.g., a voxel representing a basic volume) representing the corresponding location of a body part; each voxel value defines the voxel's brightness (in grayscale) as a function of the intensity of the response signal associated with the corresponding location; for example, in the case of an MRI scanner, the response signal represents the location's response to a magnetic field applied to it, while in the case of a CT scanner, the response signal represents the attenuation of X-ray radiation applied to that location. For example, fidelity terms... Fid() This is equivalent to the difference between the acquired low-contrast image and the simulated low-contrast image in terms of their corresponding... N×M Square Euclidean distance in dimensional Euclidean space:

[0037] in and These are the first two images, the acquired low-contrast image and the simulated low-contrast image, respectively. i row and number j The column's voxel values, and It is used for high-contrast images in simulation. Is H (vector-valued) mapping function The corresponding number i and the j Each component.

[0038] Regularization term Reg() It can be defined in different ways. For example, regularization terms. Reg() Defined by the total variation of a regularized image (as mentioned above, consisting of elements having...) N lines and M The regularization image is defined as a bitmap of a matrix of cells (each cell containing voxel values ​​for the corresponding location of a body part). The regularization image is based on a simulated high-contrast image. Total variation measures the variability / uniformity of the regularized image, and then (directly or indirectly) measures the variability / uniformity of the simulated high-contrast image. For example, the regularization term... Reg() Gradient based on voxel values ​​of a regularized image ( ,in i= 1..N and j=1…M Specifically, for each voxel, the squared (horizontal) gradient along the corresponding row is calculated. and the square (vertical) gradient along the corresponding column :

[0039] (when hour ,when hour Total variation can be isotropic. In this case, the local variation of the voxel... Calculated as:

[0040] in It is introduced for numerical stability and has very low values ​​(e.g., Such as = 10 -8 The (bias) constant is used because the square root is not differentiable at 0. The regularized image is then computed by accumulating the local variations of its voxels (e.g., by summing them). Ir Total variation of ) TV() ):

[0041] Alternatively, the total variation can be anisotropic. In this case, the total variation is given by:

[0042] For example, a regularized image is equivalent to a simulated high-contrast image:

[0043] This implementation reduces artifacts in simulated high-contrast images (emphasizing stability over accuracy).

[0044] Alternatively, a preliminary version of the simulated high-contrast image, or a preliminary high-contrast image, is generated directly from the acquired low-contrast image and possibly the acquired zero-contrast image (e.g., by means of an incremental neural network appropriately trained for this purpose), thus providing a first estimate of it. The image is then regularized to the difference between the simulated high-contrast image and the preliminary high-contrast image.

[0045] in It is a preliminary high-contrast image and For example, it is generated by applying an incremental neural network to produce an initial high-contrast image. and the acquired zero-contrast image The mapping function is used. In the following text, for completeness, the initial high-contrast image will always be generated from the acquired zero-contrast image and the acquired low-contrast image; however, it should be understood that the same considerations apply when generating the initial high-contrast image from any other (or more) acquired baseline image(s) and the acquired low-contrast image, or only from the acquired low-contrast image (i.e., without any acquired baseline image). The difference between the simulated high-contrast image and the initial high-contrast image is calculated by voxel-by-voxel subtraction. This implementation reduces the time required to solve the minimization problem and thus obtain the simulated high-contrast image (emphasizing accuracy over stability).

[0046] Now for reference Figure 3 This illustrates the main software components that can be used to implement solutions according to embodiments of this disclosure.

[0047] All software components (programs and data) are collectively indicated by reference numeral 300. Software component 300 is typically stored in mass storage and, when the program is running, is loaded into the working memory of computer 110 along with the operating system and other applications not directly related to the solution disclosed herein (and therefore omitted in the figures for simplicity). Programs are initially installed into mass storage, for example, from removable storage or from a network. In this respect, each program may be a module, fragment, or portion of code, comprising one or more executable instructions for implementing a specified logical function.

[0048] Acquirer 305 is a component of the scanner specifically designed to acquire a sequence of possible (operational) zero-contrast images and one or more (operational) low-contrast images of the patient's body parts during each imaging procedure. Acquirer 305 exposes a user interface for interacting with to control the imaging procedure. Acquirer 305 writes to (operational) acquired image storage 310, which contains the acquired (zero-contrast / low-contrast) images being acquired during the ongoing imaging procedure. Acquired image storage 310 has an entry for storing the bitmap of each acquired image.

[0049] After generating the acquired images from the corresponding raw signals, the preprocessor 315 optionally preprocesses the acquired images (e.g., by registering them, denoising them, etc.). The preprocessor 315 reads / writes to the acquired image storage 310. The optimizer 320 (by optimizing an objective function including fidelity terms and possibly added regularization terms) simulates (operationally) high-contrast images corresponding to the acquired low-contrast images. The optimizer 320 reads the acquired image storage 310. Moreover, the optimizer 320 writes to (operationally) simulated image storage 325, which contains a sequence of one or more simulated high-contrast images being simulated during the ongoing imaging procedure. The simulated image storage 325 has an entry for storing the bitmap of each simulated high-contrast image.

[0050] Optimizer 320 may optionally use an incremental machine learning model, such as an incremental neural network 330, configured to determine a corresponding preliminary high-contrast image from each acquired low-contrast image and possibly acquired zero-contrast image by applying machine learning techniques. The incremental neural network 330 reads the acquired image repository 310. In any case, optimizer 320 uses a decremental machine learning model, such as a decremental neural network 335, configured to again determine a corresponding simulated low-contrast image from each simulated high-contrast image and possibly acquired zero-contrast image by applying machine learning techniques. The decremental neural network 335 reads the simulated image repository 325 and may read the acquired image repository 310. The incremental neural network 330 and the decremental neural network 335 are provided in one or more configurations respectively for corresponding increment and decrement factors (wherein each pair of increment and decrement factors corresponds to the reciprocal of the other, e.g., is equal to it). An increment factor quantifies the increase in contrast enhancement from low contrast (of the acquired low-contrast image) to high contrast (of the initial high-contrast image), and vice versa; a decrement factor quantifies the decrease in contrast enhancement from high contrast (of the simulated high-contrast image) to low contrast (of the acquired low-contrast image). For example, the increment factor can range from 1.1 to 10.0, preferably from 1.5 to 8.0, and more preferably from 2.0 to 5.0 (such as 2.5), and the decrement factor can range from 1 / 1.1 to 1 / 10.0, preferably from 1 / 1.5 to 1 / 8.0, and more preferably from 1 / 2.0 to 1 / 5.0 (such as 1 / 2.5). Specifically, in the case of different doses of contrast agent, the increment factor corresponds to the ratio between high and low doses, and the decrement factor corresponds to the ratio between low and high doses. Furthermore, in the case of contrast agents with different efficiencies, the increment factor corresponds to the ratio between high-efficiency indications and low-efficiency indications, and the decrement factor corresponds to the ratio between low-efficiency indications and high-efficiency indications (e.g., defined by cumulative values ​​such as average, total, etc., of the voxel values ​​of the corresponding image).

[0051] During each imaging procedure, display 340 drives the monitor of computer 110 to display information based on the (acquired / simulated) images. Display 340 reads the acquired image library 310 and the simulated image library 325.

[0052] Essentially, the machine learning techniques applied by the ascending neural network 330 and the descending neural network 335 are used to perform a specific task (in this case, generating a preliminary high-contrast image and a simulated low-contrast image, respectively), without using explicit instructions, but automatically inferring how to proceed from examples (by utilizing corresponding models already learned from them). In the specific embodiment discussed, deep learning techniques are applied, a branch of machine learning based on deep neural networks, i.e., neural networks (each comprising one or more layers of neurons performing operations based on corresponding weights, followed by activation functions introducing nonlinear factors, these neurons being connected via synapses that unidirectionally transmit data between them), having one or more hidden layers arranged between the input and output layers (in the case of the ascending neural network 330, for receiving the acquired low-contrast image and possibly the acquired zero-contrast image and for providing a preliminary high-contrast image, or in the case of the descending neural network 335, for receiving the simulated high-contrast image and possibly the acquired zero-contrast image and for providing a simulated low-contrast image).

[0053] Each (incrementing / decrementing) neural network 330, 335 can be implemented using any architecture available in the field of image processing for medical applications, and it is possible to empirically customize the architecture to facilitate its training and improve its performance. For example, neural networks 330, 335 are convolutional neural networks (CNNs), i.e., a specific type of deep neural network in which one or more hidden layers in its hidden layers perform (cross)convolution operations by shifting kernels sequentially to a finite portion of their input with a selected stride. More specifically, neural networks 330, 335 are residual neural networks including skip connections, each skip connection adding (through an identity mapping) the input of a layer to the output of the next layer that follows it (to facilitate training as the depth defined by the hidden layers increases), such as based on the ResV-Net architecture; the ResV-Net architecture is characterized by two symmetrical paths (similar to the standard U-Net architecture, but with pooling operations replaced by convolution operations to facilitate training): an encoder (which reduces spatial information while increasing feature information) followed by a decoder (which extracts feature information and spatially expands it). Different implementations of the ResV-Net architecture are available. For example, neural networks 330 and 335 are variations of the neural network described in “A. Bone, S. Ammari, J.-P. Lamarque, M. Elhaik, E. Chouzenoux, F. Nicolas, P. Robert, C. Balleyguier, N. Lassau and M.-M. Roh’e – 'Contrast-enhanced brain MRI synthesis with deep learning: key input modalities and asymptotic performance' - 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI) - IEEE, 2021, pp. 1159–1163”, which in turn is derived from “F. Milletari, N. Navab, and S.A. Ahmadi – 'V-net: Fully convolutional neural networks for volumetricmedical image segmentation' - 2016 fourth international conference on 3Dvision (3DV)”. The neural network tuning described in IEEE, 2016, pp. 565-571.Specifically, the encoder comprises four blocks, each consisting of a residual convolutional module followed by a downsampling module. The residual convolutional module is formed by multiple convolutional layers (1 for the first block, 2 for the second block, 3 for the third block, and 3 for the fourth block), each convolutional layer applying a 3x3 kernel with a stride of 1; each convolutional layer is followed by a 0.2 leaky rectified linear unit (ReLU) activation function (for the fourth block). x≥0 , LeakyReLU(x)=x ,for x<0 , LeakyReLU(x) = 0.2 x Each residual convolutional module doubles the number of channels in the data (for each cell's value), for example, starting with 12. The input to each residual convolutional module is saved and then component-wise added to the output of a leaky ReLU activation function of 0.2. Downsampling modules are formed by convolutional layers with 2x2 kernels applied at a stride of 2. The decoder consists of four blocks (symmetric to those in the encoder) followed by the final residual convolutional modules. Each block consists of a residual convolutional module (like its counterpart in the encoder, but now with half the number of channels) followed by an upsampling module. The upsampling module is formed by convolutional layers with a 2x2 transposed kernel applied at a stride of 2. The input to each residual convolutional module in the decoder is concatenated with the output of the corresponding residual convolutional module in the encoder. The corresponding block connecting the encoder and decoder (often called the bottleneck) consists of a single residual convolutional module formed by three convolutional layers, each with a 3x3 kernel applied at a stride of 1; each convolutional layer is followed by a leaky ReLU activation function of 0.2. The final residual convolutional module consists of three convolutional layers. The first two apply a 3x3 kernel with a stride of 1 (reducing the number of channels to 1), and the third applies a 1x1 kernel. The first two convolutional layers are followed by a leaky ReLU activation function of 0.2, and the last convolutional layer is followed by a standard ReLU activation function (for...). x≥0 , ReLU(x) = x ,for x<0 , ReLU(x) = 0 ).

[0054] Neural networks 330 and 335 can be trained (incrementally / decrementally) as usual using multiple sample sets, each sample set defining the inputs and corresponding desired outputs of the neural networks 330 and 335 under their supervision. Each sample set includes optional zero-contrast images, low-contrast images, and high-contrast images; the zero-contrast images correspond to no contrast enhancement, the low-contrast images correspond to low contrast of the contrast agent, and the high-contrast images correspond to high contrast of the contrast agent, with corresponding increases / decreases equal to the increase / decrease in contrast enhancement to be simulated. For example, in the case of different doses of contrast agent, the low dose and high dose are equal to the reduced dose and full dose of the contrast agent, respectively. The full dose has a standard value in clinical practice, which is required by healthcare institutions (i.e., institutions with jurisdiction over healthcare applications, such as the European Medicines Agency (EMA) in Europe or the U.S. Food and Drug Administration (FDA) in the United States), or recommended by accredited bodies or consistent scientific publications; thus, the reduced dose is below this standard value. For example, in MRI procedures, the full dose of contrast agent is 0.1 mmol of gadolinium per kilogram of patient body weight. In CT protocol, the full dose of iomeprol-based contrast agents (e.g., commercially available under the name Iomeron (trademark) of Bracco Imaging SpA, formulated at 155-400 mg / mL) is 20-200 mL for imaging the head and 100-200 mL for imaging other body parts; alternatively, the full dose of iopamidol-based contrast agents (e.g., commercially available under the name Isovue (trademark) of Bracco Imaging SpA) is 100-230 mL for a 250 mg / mL formulation or 100-200 mL for a 300 mg / mL formulation in adults (where the total dose of iodine should not exceed 60 g), and 1.2-3.6 mL per kg of patient body weight for a 250 mg / mL formulation or 1.0-3.0 mL per kg of patient body weight for a 300 mg / mL formulation (where the total dose of iodine should not exceed 30 g). g). Reduce the dose below the full dose (e.g., equal to 0.1-0.5 times it). As another example, in the case of contrast agents with different efficiencies, the low-efficiency contrast agent is the one that can be used for certain medical applications (i.e., the type of imaging procedure and / or body part), while the high-efficiency contrast agent simulates the corresponding contrast agent for different medical applications in which neural networks 330, 335 will be used. For example, the low-efficiency contrast agent is used for MRI procedures (such as gadolinium-based), and the high-efficiency contrast agent simulates the corresponding contrast agent used for CT procedures.

[0055] The sample set has been acquired, at least in part, from the corresponding body parts of the subjects (of the same type as the body part to be imaged). The subjects in the sample set are humans and / or animals (for clinical applications or testing studies), e.g., rats (for preclinical studies). In the case of incomplete sample sets, the (missing) images for each sample set have been simulated from other (available) images in the sample set, such as analytically. For example, all (possible) zero-dose, reduced-dose, and full-dose images for each sample set may have been acquired from humans (in prospective studies involving non-standard imaging protocols) or from animals; additionally or alternatively, the (possible) zero-dose and full-dose images for each sample set may have been acquired from humans (backwards from previously performed standard imaging protocols), and the reduced-dose images for the sample set have been simulated from them. As another example, all (possible) zero-contrast, low-efficiency, and high-efficiency images for each sample set may have been acquired from animals (when both low-efficiency and high-efficiency contrast agents are available and can be used). Additionally or alternatively, the (possible) zero-contrast and low-efficiency images of each sample set may have been obtained from humans (back to standard imaging procedures performed in the past) or from animals, while the high-efficiency images of the sample set have been simulated from them.

[0056] Then, as usual, the sample set is used to train each neural network 330, 335 to optimize its ability to generate (target) images (high-contrast images of the ascending neural network 330 or low-contrast images of the descending neural network 330) that define the ground truth for each sample set from other (source) images in the sample set (possible zero-contrast and low-contrast images of the ascending neural network 330, or possible zero-contrast and high-contrast images of the descending neural network 330). Specifically, a portion of the sample set is used as a training subset. The source images of each sample set are applied to neural networks 330, 335 (whose weights are randomly initialized) to obtain the corresponding output images. Until the loss value based on the difference between the output image and the target image in the sample set (e.g., defined by the Structural Similarity Index (SSIM)) is acceptable, the weights of neural networks 330 and 335 are updated, for example, by applying an Adaptive Moments Estimation (ADAM) method based on the Stochastic Gradient Descent (SGD) algorithm (where, given the loss value as a function of the weights approximated by backpropagation, the SGD algorithm determines the direction and amount of the update based on the gradient of the loss function, and the ADAM method automatically determines the corresponding learning rate). This operation can be performed in batch mode (after processing a batch of sample sets for the cumulative value of their loss values, such as their average). This process is repeated multiple times (periods). After the configuration of neural networks 330 and 335 that provides the optimal minimum loss value has been found, neural networks 330 and 335 are validated. In particular, the remainder of the sample set is used as a validation subset. The source image of each sample set is applied to neural networks 330 and 335 to obtain the corresponding output image, and its loss value relative to the target image of the sample set is calculated as described above. If the cumulative value of all loss values ​​(such as their average) is acceptable, then this means that the generalization ability of neural networks 330 and 335 (from the configuration they learn from the training subset to the validation subset) is satisfactory; otherwise, repeat the training of neural networks 330 and 335 with different training subsets and / or training parameters (such as learning rate, epochs, etc.).

[0057] Now for reference Figures 4A-4B This illustrates an activity diagram describing the activity flow associated with an imaging procedure according to an embodiment of the present disclosure.

[0058] In this respect, each box may correspond to one or more executable instructions for implementing a specified logical function on a computer. In particular, the activity diagram representation can be used as an exemplary process for imaging a patient's body parts during an imaging procedure using method 400.

[0059] After the imaging protocol is initiated, the process begins at the black starting circle 403 (as instructed by a physician or healthcare operator such as a radiologist via a corresponding command entered through the acquirer's user interface after the patient has been positioned appropriately relative to the scanner (e.g., inside the gantry in the case of an MRI / CT scanner). In response, the acquirer begins acquiring zero-contrast images of the body part at box 406, which are then displayed in real-time on a computer monitor. At this stage, the body part contains no contrast agent, or at least not a significant amount of contrast agent (because the patient has never been given any contrast agent, or a relatively long period of time has elapsed since any contrast agent was previously administered, ensuring that the contrast agent has been substantially removed). Once the physician has selected one of the zero-contrast images to acquire (via a corresponding command entered directly by the physician or by a healthcare operator through the acquirer's user interface), the acquirer saves this acquired zero-contrast image to the acquired image library (initially empty) at box 409.

[0060] The display then shows a message prompting the administration of a contrast agent on the computer monitor at box 412. In response, the healthcare operator administers the contrast agent to the patient. For example, the contrast agent may be of the targeted type, i.e., adapted to reach a specific target, such as a lesion, like a tumor to be examined / resected / treated, and adapted to remain substantially fixed therein. This result can be achieved by using either a non-target-specific contrast agent (adapted to accumulate in the target without any specific interaction with it, such as through passive accumulation) or a target-specific contrast agent (adapted to attach to the target by means of a specific interaction with it, such as by incorporating a target-specific ligand into the formulation of the contrast agent, for example, based on chemical bonding properties and / or physical structure capable of interacting with different tissues, vascular properties, metabolic characteristics, etc.). The contrast agent may be administered to the patient as a bolus intravenous injection (e.g., using a syringe). Thus, the contrast agent circulates within the patient's vascular system until it reaches and bonds to the target; any remaining (unbonded) contrast agent is instead removed from the patient's blood pool. The contrast agent is administered in low doses. A low dose can be equal to or less than the full dose of contrast agent. Additionally or alternatively, an imaging procedure may be performed using a scanner for one medical application (imaging protocol and / or body part), while the contrast agent is used for another; for example, an imaging protocol may be performed using a CT scanner with a contrast agent used for an MRI protocol. After a waiting period (e.g., several minutes) allowing the contrast agent to accumulate in the (potential) target and be flushed away from the rest of the patient, the imaging protocol may actually begin at box 415, during which time the acquirer continues to acquire low-contrast images of the body part and the display shows them in real time on a computer monitor. At this point, or at any later time, the physician may decide to increase the contrast enhancement (e.g., as indicated by a corresponding command entered by the physician or healthcare operator via the acquirer's user interface); this increase may be unspecified or based on a desired increment factor, which can be manually selected in a discrete / continuous manner (default definition or only available).

[0061] In response, the display stops showing the acquired low-contrast image on the computer monitor. The optimizer configures the increasing neural network (if used) and decreasing neural network (if necessary) at box 418 according to the required increasing and decreasing factors. The acquirer retrieves the (new) acquired low-contrast image (not displayed) from the acquired image repository at box 421. The preprocessor optionally preprocesses the acquired low-contrast image at box 424. For example, the preprocessor co-registers the acquired low-contrast image with the acquired zero-contrast image (to achieve spatial correspondence, such as by applying a rigid transformation to the acquired low-contrast image), denoises the acquired low-contrast image (to reduce its noise, such as by means of an autoencoder trained in an unsupervised manner to optimize its encoded image, ignore unimportant data (due to noise) on it, and then decodes the resulting image to reconstruct the same image with reduced noise), and so on.

[0062] Depending on the optimizer's configuration (manually selected, default defined, or uniquely available), the activity flow branches at box 427. If the regularization terms (i.e., the type and / or regularization parameters) can be dynamically defined, the optimizer estimates the noise of the acquired low-contrast image (if not yet completed) at box 430. For this purpose, the optimizer computes a noise matrix defined by a matrix of cells (with the same size as the acquired low-contrast image), each cell containing a value representing the noise intensity in the corresponding voxel of the acquired low-contrast image; for example, the noise matrix is ​​obtained by subtracting the denoised acquired low-contrast image from the initially acquired low-contrast image.

[0063] Moving to box 433, if the type of regularization term can be dynamically selected (based on the optimizer's configuration), the process descends to box 436. The optimizer now selects the type of regularization term based on the noise in the acquired low-contrast image. For example, the optimizer calculates the standard deviation of the noise matrix values; if the standard deviation is (possibly strictly) below a threshold, such as 0.02–0.04, the type of regularization term is set to the total variation of the difference between the simulated high-contrast image and the initial high-contrast image; otherwise, the type of regularization term is set to the total variation of the simulated high-contrast image. If the type of regularization term is statically set (configurable or fixed), the process continues from box 436 or directly from box 433 to box 439.

[0064] At this point, if the regularization parameter can be dynamically selected (based on the optimizer's configuration), the process descends to box 442. The optimizer now sets the regularization parameter based on the acquired low-contrast image, either its entirety or its noise level. For example, the regularization parameter can be set based on the entirety of the acquired low-contrast image by applying the L-curve method. For this purpose, the regularization parameter is calculated by using the Euclidean norm (|||) of a simulated high-contrast image. Is H ||, its minimized loss function And therefore depends on the regularization parameter. ) relative to the Euclidean norm of the residual term ( The L-curve is determined by plotting; the L-curve typically has a general "L" shape, where the vertical portion is dominated by the consistency (high accuracy) between the simulated high-contrast image and the acquired low-contrast image, while the horizontal portion is dominated by regularization (high stability). The regularization parameter is set to a value corresponding to the maximum curvature of the L-curve. In this way, the regularization parameter thus obtained provides a trade-off between the opposing requirements of accuracy and stability of the simulated high-contrast image. The maximum curvature of the L-curve can be determined using spline or triangle methods. As a further example, the regularization parameter is set based on the noise level of the acquired low-contrast image by applying the discrepancy principle. For this purpose, the optimizer calculates the noise level as equal to the norm (such as the Euclidean norm) of the noise matrix value. The optimizer then solves for the regularization parameter... The difference equation is used to determine the regularization parameter:

[0065] in It is a fidelity term (depending on the simulated high-contrast image). Is H And thus it depends on the regularization parameter. ), It is a control parameter (e.g., equal to 1.01). It is a noise matrix, and It is the squared Euclidean norm of the noise matrix (defining the noise level of the acquired low-contrast image). In this way, given an estimated noise level, the obtained regularization parameter provides a simulated high-contrast image that reproduces the acquired low-contrast image as closely as possible. The difference equation can be solved using any root-finding algorithm (which provides an approximation of the regularization parameter) (e.g., based on Newton's method). The process continues from box 442 to box 445, and if the regularization parameter is statically set (configurable or fixed), then proceeds directly from box 439 to box 445; or if the regularization term is completely statically set (or configurable or fixed), then proceeds directly from box 427 to box 445.

[0066] The optimization loop now iteratively determines a simulated high-contrast image corresponding to the acquired low-contrast image. Specifically, the optimizer initializes the simulated high-contrast image with the acquired low-contrast image. Is H =Ia L Even when a preliminary high-contrast image is available, the choice to always start with the acquired low-contrast image limits the dependence on it, which has a beneficial impact on quality. The optimizer applies the (possibly) acquired zero-contrast image and the simulated high-contrast image to the decreasing neural network at box 448 to obtain the corresponding simulated low-contrast image. The optimizer calculates the fidelity term at box 451 as a measure of the difference between the acquired low-contrast image and the simulated low-contrast image (e.g., ).

[0067] The activity flow branches at box 454 based on the number of iterations of the optimization loop. In the first iteration of the optimization loop, or if the simulated high-contrast image is directly determined (without any iterations of the optimization loop), the process continues to box 457. At this point, the activity flow further branches based on the type of regularization term. If the regularization term is not based on any initial high-contrast image, the optimizer computes the regularization term as the total variation of the simulated high-contrast image at box 460. Instead, the optimizer applies the (possibly) acquired zero-contrast image and the acquired low-contrast image to the incremental neural network at box 463 to obtain an initial high-contrast image. The optimizer calculates the regularization term at box 466 as a total variation of the difference between the simulated high-contrast image and the initial high-contrast image. Returning to reference box 454, in any subsequent iteration of the optimization loop, the process instead continues to box 469, where the optimizer computes the regularization term as a total variation of the difference between the simulated high-contrast image (as set for the next iteration of the optimization loop) and the simulated high-contrast image obtained in the previous iteration of the optimization loop (immediately following), using superscript... k-1 To distinguish ( The process now continues from box 460, from box 466, or from box 469 to box 472. At this point, the optimizer calculates the loss function by adding the regularization term to the fidelity term after multiplying the regularization term by the regularization parameter. ,in It is the first k The regularization parameter for the next iteration, and the parameter for the first iteration. (As initialized as described above). The optimizer checks at box 475 whether the minimum of the loss function has been reached (e.g., if the loss function has not improved significantly in two or more iterations). If not, the optimizer updates the simulated high-contrast image at box 478, attempting to reduce the loss function (e.g., using a deterministic gradient descent (GD) algorithm). The process then returns to box 448, repeating the same operation with the (updated) simulated high-contrast image.

[0068] Conversely, after reaching the minimum of the loss function, the process moves from box 475 to box 481. The optimizer now verifies the termination condition of the optimization loop. For example, the termination condition is based on a comparison of the convergence index (a measure of how well the simulated high-contrast image converges toward the desired result) with a threshold value. This termination condition could also be based on the difference principle; in this case, the termination condition is based on an equality... The convergence index is equal to The threshold value is compared. Additionally or alternatively, the termination condition is based on the number of iterations of the optimization loop (…). k From the first iteration of the optimization loop k=0 (start) and maximum value ( Max Comparisons such as 5-10). If the termination condition is not met, then the process proceeds to box 484. Specifically, in the example discussed, this occurs when the convergence exponent (strictly) is above a threshold and the number of iterations of the optimization loop (strictly) is below the maximum value: and

[0069] In this case, the optimizer performs further iterations of the optimization loop. For this purpose, the optimizer saves the simulated high-contrast image thus obtained to a temporary variable to utilize in the next iteration of the optimization loop (as described above). Furthermore, the optimization loop can iterate in a non-stationary manner by changing the regularization parameter in each iteration; for example, the optimizer updates the regularization parameter at box 487 according to a geometric progression:

[0070] in k It is the number of iterations. It is the first k The regularization parameter for the next iteration This is the initial value of the regularization parameter (for the first iteration). k=0 And q is a constant (having 0<q<1 The process then returns to box 445 to repeat the same operation for the next iteration of the optimization loop. In this way, the quality of the simulated high-contrast image generally improves with each iteration of the optimization loop. In any case, the quality of the final simulated high-contrast image depends significantly less on the choice of the regularization parameter (which can also be set to a predefined value without any noticeable quality degradation). Returning to reference box 481, after the termination condition is met, the optimization loop is exited by saving the thus obtained simulated high-contrast image to the corresponding repository, and then descends to box 490. Specifically, in the example discussed, the convergence exponent is lower than a threshold value, the number of iterations of the optimization loop is equal to the maximum value, or both (when no iterations are performed or Max=1 This occurs when (when the time is always true): or

[0071] The above implementation provides a good trade-off between the opposing requirements of high accuracy and low computation time. In fact, while the iteration of the optimization loop is not guaranteed to converge from a strictly theoretical point of view, it provides acceptable quality in most practical cases within a relatively short time.

[0072] At this point, the display shows information related to the body part based on a simulated high-contrast image obtained in this way (manually selected, default defined, or the only available type); for example, the simulated high-contrast image is displayed alone, superimposed on an acquired zero-contrast image, combined with an acquired zero-contrast image and an acquired low-contrast image by applying high dynamic range (HDR) technology, and so on. In cases of different contrast agent doses, when the high dose equals the full dose, the low dose is a reduced dose (below the full dose). Therefore, the simulated full-dose image restores the contrast enhancement that would normally be obtained when the contrast agent is administered at the full dose (if it is not increased by reducing motion / aliasing artifacts that might be caused by the actual administration of the full dose of contrast agent); at the same time, the reduced dose of contrast agent administered to the patient limits his / her health risks, especially for children, pregnant women, or patients with specific conditions such as renal insufficiency. Alternatively, when the low dose is the full dose, the high dose is an enhanced dose (above the full dose). Therefore, the simulated enhanced-dose images increase contrast enhancement as if they were acquired with an enhanced-dose image obtained by administering a higher (virtual) enhanced dose of contrast agent than is possible in current clinical practice (further reducing motion / aliasing artifacts that may otherwise result from actual administration of contrast agent at an enhanced dose, where possible); meanwhile, the full dose of contrast agent administered to the patient does not affect standards of care or clinical workflow. In the case of contrast agents with different efficiencies, low-contrast images are low-efficiency images (acquired with low-efficiency contrast agents for different medical applications) and high-contrast images are high-efficiency images (simulating the administration of high-efficiency contrast agents for the medical application in question). Therefore, the simulated high-efficiency images increase contrast enhancement as if they were acquired by administering (virtual) high-efficiency contrast agents for the medical application in question (i.e., for the same imaging protocol performed with the same type of scanner and for the same body part type). This allows for the desired contrast enhancement to also be achieved in medical applications where a corresponding contrast agent is unavailable or cannot be used due to its (actual or potential) side effects (such as toxicity, short- / long-term damage, risks, etc.). The acquirer verifies at box 493 whether different values ​​of the increment factor have been selected (where possible) via its user interface. If so, the process returns to box 418 to update the configurations of the increment neural network (if used) and the decrement neural network based on the different values ​​of the increment factor and the corresponding decrement factor, respectively, and then repeats the same operation continuously. In this way, the physician can verify (also in real-time) the effects of different values ​​of the increment factor and then select the one that provides the best contrast enhancement. Conversely, the process descends from box 493 to box 496, where the acquirer verifies the status of the imaging protocol.If the imaging procedure is still in progress, the activity flow returns to box 421 to repeat the same operation on the next low-contrast image acquired (corresponding to the storage). Conversely, if the imaging procedure has ended (as indicated by a corresponding command entered by a physician or healthcare operator via the user interface of the acquirer), the process ends at concentric white / black stop circle 499.

[0073] Now for reference Figures 5A-5G The illustration shows a representative example of experimental results related to the solution according to embodiments of the present disclosure.

[0074] For this purpose, a dataset of images acquired from 45 experimental rats (with tumors) undergoing corresponding imaging procedures of the MRI type in preclinical studies was used. Specifically, 24 slices of each rat were imaged during three imaging procedures: before contrast agent administration, after administration of a low dose equal to 20% of the full dose, and after administration of a high dose equal to the full dose, resulting in corresponding zero-dose, low-dose, and high-dose images, each with 256x256 pixels, for a total of 1080 acquired (zero / low / high-dose) images for each (zero / low / high) dose of contrast agent. The acquired high-dose images were preprocessed using block matching and 3D filtering (BM3D) algorithms to reduce their noise. For this purpose, the standard deviation of the noise for each acquired high-dose image was automatically calculated from its background-representing portion (where pixel values ​​should be zero in the absence of noise). The average value obtained for all acquired high-dose images has been calculated, and this value (0.006) has been used to apply the BM3D algorithm to remove corresponding amounts of noise. The dataset has been organized into two complementary subsets: a training subset containing 840 acquired images for each contrast dose, and a validation subset containing 240 acquired images for each contrast dose. The incremental and decremental neural networks (implemented using the TensorFlow library and Keras API) have been trained independently as described above (using the ADAM algorithm with a fixed learning rate of 10). -3 (with a batch size of 16 and a period number of 80), the first time was performed using a sample set that included the corresponding low-dose image and the acquired high-dose image, and the second time was performed using a sample set that included the corresponding zero-dose image, the acquired low-dose image and the acquired high-dose image.

[0075] Simulated high-dose images corresponding to the acquired low-dose images are generated in the following different modes (the loss function of the inverse problem is achieved by using a fixed step size of 10). -2 (And minimize using the GD algorithm with a fixed number of iterations equal to 150). Each simulated high-dose image is generated from the corresponding acquired low-dose images by using an incremental neural network trained without acquiring zero-dose images (NN-L2H mode). Each simulated high-dose image is generated from the corresponding acquired zero-dose image and acquired low-dose image using an incremental neural network trained with the acquired zero-dose image (NN-0L2H mode). A decreasing neural network, trained without any acquired zero-dose images, is used to minimize a loss function using the inverse problem described above. This generates each simulated high-dose image from the corresponding acquired low-dose images. The regularization term is equal to the total variation of the simulated high-dose images, and the regularization parameter... (IP-L2H-TV(H) mode); Each simulated high-dose image is generated from the corresponding acquired zero-dose image and the acquired low-dose image by minimizing the loss function using the inverse problem as described above, through a decreasing neural network trained with the acquired zero-dose image. The regularization term is equal to the total variation of the simulated high-dose image, and the regularization parameter is... (IP-0L2H-TV(H) mode); Each simulated high-dose image is generated from the corresponding acquired zero-dose image and the acquired low-dose image by minimizing the loss function using the inverse problem as described above with a decreasing neural network trained on the acquired zero-dose image. The regularization term is equal to the total variation of the difference between the simulated high-dose image and the initial high-dose image generated by the increasing neural network trained on the acquired zero-dose image, and the regularization parameter is... (IP-0L2H-TV(HP) mode).

[0076] Special Reference Figure 5A and Figure 5B Two exemplary high-dose images (numbered 39 and 141 in the dataset, which are significantly different and characterized by various anatomical objects and contrast enhancements) are shown, along with corresponding simulated high-dose images generated using the aforementioned pattern.

[0077] For each simulated high-dose image, its SSIM value is calculated with respect to the corresponding acquired high-dose image (measuring its matching degree in the range from -1 for perfect anticorrelation, through 0 for no similarity, to 1 for perfect similarity):

[0078] (The high-dose images acquired are different from the corresponding low-dose images, and their SSIM values ​​are 0.6626 and 0.6330, respectively.)

[0079] In all modes, the simulated high-dose image showed a significant increase in SSIM value compared to the acquired high-dose image, compared to the acquired low-dose image. However, the highest SSIM value was consistently provided by the IP-0L2H-TV (HP) mode.

[0080] Move to Figure 5C The image shows the same acquired high-dose image (number 39) and corresponding simulated high-dose images in NN-OL2H and IP-OL2H-TV(HP) modes, with background noise essentially removed (by resetting every pixel value below a threshold of 0.14 to zero (black)). As can be seen, the performance improvement of the IP-OL2H-TV(HP) mode relative to the NN-OL2H mode becomes clear. Indeed, as indicated by the added corresponding white circles, the simulated high-dose image in the NN-OL2H mode generates unwanted black holes (not present in the acquired high-dose image), while the simulated high-dose image in the IP-OL2H-TV(HP) mode maintains a perfectly smooth pattern.

[0081] Move to Figure 5D To verify the performance in the presence of unknown noise, additive white Gaussian noise (normally distributed with zero mean) was measured at different standard deviations. Injected into the same acquired low-dose image (number 39), particularly with N =10 A value that linearly increases from 0 to 0.1 (i.e., , i=0…N-1 For each (noisy) acquired low-dose image thus obtained, a corresponding simulated high-dose image was then generated in NN-0L2H mode, IP-L2H-TV(H) mode, IP-0L2H-TV(H) mode, and IP-0L2H-TV(HP) mode, and their SSIM values ​​relative to the corresponding acquired high-dose image were calculated. The figure shows the SSIM values ​​for different modes plotted on the vertical axis and the standard deviation of additive (Gaussian white) noise plotted on the horizontal axis. The icon.

[0082] In this case, due to the presence of additive noise, the regularization parameters in the IP-L2H-TV(H) mode, IP-0L2H-TV(H) mode, and IP-0L2H-TV(HP) mode ( The update was performed based on additive noise, specifically using the following linear formula:

[0083] in The initial values ​​of the regularization parameters are those without additive noise (for IP-L2H-TV(H) mode). For IP-0L2H-TV(H) mode And for IP-0L2H-TV (HP) mode Therefore, when there is no additive noise (for = 0 , i=0 When ), the regularization parameter is from Increase, when additive noise is at its maximum (for = 0.1 , i=N-1 When ), increase to .

[0084] Generally, the SSIM value decreases with increasing additive noise. As observed in the table above, when additive noise is absent or small (left part of the graph), the NN-0L2H, IP-0L2H-TV(H), and IP-0L2H-TV(HP) modes provide very high SSIM values, while the IP-L2H-TV(H) mode provides a lower SSIM value. However, when additive noise is high (right part of the graph), the IP-0L2H-TV(H) and IP-0L2H-TV(H) modes provide significantly higher SSIM values ​​than the SSIM value provided by the NN-0L2H mode, which reasonably utilizes the value of the increasing regularization parameter, and significantly higher than the SSIM value provided by the IP-0L2H-TV(HP) mode, which reasonably utilizes the decoupling from the initial high-dose image generated by the incremental neural network (which suffers from the typical instability of neural networks when a given input is affected by high noise).

[0085] Move to Figure 5E Simulated high-dose images are shown generated from the same low-dose image (number 39) with minimal additive noise in NN-OL2H mode, IP-OL2H-TV(H) mode, and IP-OL2H-TV(HP) mode, specifically with standard deviation. = 0 (left column), = 0.011 (Middle column) and = 0.022 (Right column).

[0086] In all cases, the simulated high-dose images remained rich in detail, but noisy patterns also appeared. However, in IP-0L2H-TV(HP) mode, fine details were smoothed out (thanks to the regularization term being strongly influenced by the prior value defined by the initial high-dose image).

[0087] Move to Figure 5F The image shows a magnified view of a portion of the tumor representation within the same simulated high-dose image (in...). Figure 5E (Indicated by the white square within one of them).

[0088] As can be seen, in NN-0L2H mode, the shape of the tumor is altered after the addition of noise, while in IP-0L2H-TV(H) and IP-0L2H-TV(HP) modes, it remains essentially unchanged.

[0089] Move to Figure 5G The following models were generated using NN-0L2H mode, IP-L2H-TV(H) mode, IP-0L2H-TV(H) mode, and IP-0L2H-TV(HP) mode, with no additive noise and standard deviation. = 0.022 Simulated high-dose images were generated from a subset of 80 (randomly selected) acquired low-dose images with additive noise, and their SSIM values ​​relative to the corresponding acquired high-dose images were calculated. The figure shows a box plot for each different mode representation with two whiskers detached from it (horizontally extending). The boxes are drawn from the first quartile to the third quartile and are separated by a line indicating the median of the corresponding SSIM value; the (lower) whisker and (upper) whisker are defined by the minimum and maximum values ​​of the corresponding SSIM values, respectively.

[0090] As can be seen, the minimum, maximum, and median SSIM values ​​in the NN-0L2H mode change significantly after the injection of additive noise, while they remain essentially unchanged in the IP-0L2H-TV(H) and IP-0L2H-TV(HP) modes.

[0091] All of the above demonstrates that the solutions according to the embodiments of this disclosure (especially when regularization terms are added) provide both good accuracy (comparable to, or even better than, the accuracy provided by techniques known in the art) and good stability.

[0092] Revise

[0093] To meet specific and particular requirements, those skilled in the art may apply numerous logical and / or physical modifications and alterations to this disclosure, provided that it remains within the scope of the claims. More specifically, while this disclosure has been described with a degree of particularity with reference to one or more embodiments thereof, it should be understood that various omissions, substitutions, and changes in form and detail, as well as in other embodiments, are possible. In particular, different embodiments of this disclosure may even be practiced without the specific details (such as numerical values) set forth in the foregoing description, in order to provide a more thorough understanding thereof; rather, well-known features may have been omitted or simplified to avoid unnecessary detail that obscures the description. Moreover, it is expressly contemplated that specific elements and / or method steps described in connection with any embodiment of this disclosure may be incorporated in any other embodiment as a matter of general design choice. Furthermore, items presented in the same group as well as in different embodiments, examples, or alternatives should not be construed as being substantially equivalent to each other (rather than being separate and independent entities). In any case, each numerical value should be modified according to applicable tolerances; in particular, unless otherwise indicated, the terms “substantially,” “approximately,” “about,” etc., should be understood to be within 10%, preferably within 5%, and more preferably within 1%. Furthermore, each numerical range should be intended to explicitly specify any possible numbers along the range (including its endpoints). Ordinal numbers or other qualifiers are used only as labels to distinguish elements with the same name, but they do not in themselves indicate any priority, precedence, or order. Terms such as “containing,” “including,” “having,” “comprising,” “involving,” etc., should be intended to have an open, non-exhaustive meaning (i.e., not limited to the listed items), terms such as “based on,” “depending on,” “according to,” “a function of,” etc., should be intended to indicate a non-exclusive relationship (i.e., involving possible other variables), terms such as “a / an” should indicate one or more items (unless otherwise explicitly indicated), and terms such as “component for,” (or any component plus function), should be intended to indicate any structure suitable for or configured to perform the relevant function.

[0094] For example, an embodiment provides an imaging method for imaging a patient's body part in a medical application. However, the body part can be of any type (e.g., an organ, its region, tissue, bone, joint, etc.) and in any condition (e.g., healthy, pathologically affected, etc.), and can belong to any patient (e.g., human, animal, etc.); moreover, the imaging method can be used in any medical application (e.g., diagnostic, therapeutic, or surgical applications based on MRI, CT, fluoroscopy, fluorescence, or ultrasound techniques, etc.). In any case, while the imaging method can facilitate the physician's task, it only provides intermediate results that can help him / her, and the strictly medical activity is always performed by the physician himself / herself.

[0095] In this embodiment, the imaging method includes the following steps under the control of a computing system. However, the computing system can be of any type (see below).

[0096] In an embodiment, the imaging method includes receiving one or more acquired low-contrast images (by a computing system). However, the acquired low-contrast images can be any number and of any type (e.g., providing a spatial / frequency representation, in 2D / 3D form, having any size and resolution, pixel / voxel values ​​with any chromaticity and bit depth, etc.), and they can be received in any manner (e.g., in real time, offline, locally, remotely, etc.).

[0097] In this embodiment, the acquired low-contrast image provides a contrast-enhanced representation of the body part with targeted contrast enhancement. However, contrast enhancement can be achieved using any contrast agent already administered to the patient in any manner (e.g., any targeted contrast agent, such as those based on specific or non-specific interactions, any non-targeted contrast agent, etc.), including in a non-invasive manner (e.g., orally for gastrointestinal imaging, via nebulizer into the airway, via local spray application, etc.), and in no case requiring any substantial physical intervention on the patient that would require specialized medical knowledge or pose any health risk to him / her (e.g., intramuscular); moreover, the target can be of any type (e.g., any type of body part to which the contrast agent is fixed (e.g., for detecting lesions), a vascular system in which the contrast agent circulates (e.g., for determining hemodynamic characteristics, etc.)). In any case, this is a data processing method that only includes steps performed by a computing system (these steps can even be performed independently of the acquisition of the low-contrast image, thus requiring no interaction with the patient).

[0098] In this embodiment, the contrast enhancement of the low-contrast image has low contrast. However, low contrast can be of any type (e.g., quantitatively defined by any value, qualitatively defined by the dose of the contrast agent, the type of contrast agent, both of them, etc.).

[0099] In an embodiment, the imaging method includes (by a computing system) simulating one or more simulated high-contrast images corresponding to the acquired low-contrast images. However, the simulated high-contrast images can be simulated in any way (e.g., in real time, offline, locally, remotely, for all or only some of the acquired low-contrast images, for the entire content of the acquired low-contrast images or for a manually or automatically selected region of interest, etc.).

[0100] In an embodiment, the simulated high-contrast image provides a representation of contrast enhancement of a body part, which simulates an increase in contrast enhancement of the target from low contrast to high contrast above that low contrast. However, the increase in contrast enhancement can be of any type (e.g., simulating a higher dose of contrast agent, using a more efficient contrast agent, both, etc.) and to any degree (e.g., quantitatively, defined in any way based on the content of the corresponding low-contrast / high-contrast image, based on the corresponding low / high dose of contrast agent, etc., qualitatively, fixedly, or selectable in a discrete / continuous mode, etc.).

[0101] In this embodiment, each simulated high-contrast image is simulated by optimizing an objective function. However, the objective function can be of any type (e.g., with or without any regularization, to be minimized / maximized, etc.) and it can be optimized in any way (e.g., in discrete / continuous mode, in weak / strong form, with any initialization of the simulated high-contrast image, in a direct / iterative manner, etc.).

[0102] In an embodiment, the objective function includes a fidelity term based on a comparison between the corresponding acquired low-contrast image and a simulated low-contrast image. However, the fidelity term can be any type of measure based on the difference or similarity between the acquired low-contrast image and the simulated low-contrast image, such as taxi norm, Euclidean norm, p-norm, etc., defined by any distance between their values ​​(such as Euclidean distance, Wasserstein distance, Chebyshev distance, Manhattan distance, Minkowski distance, etc.).

[0103] In one embodiment, a simulated low-contrast image is generated from a simulated high-contrast image to simulate the contrast enhancement from high contrast to low contrast. However, the simulated low-contrast image can be generated in any manner (e.g., analytically, using or not using additional information (such as information relating to the patient, body part, clinical condition), etc., solely from the simulated high-contrast image or together with any acquired baseline image(s), utilizing any machine learning model (such as neural networks, generative models, genetic algorithms, etc.).

[0104] In an embodiment, the imaging method includes (by a computing system) outputting information related to a body part based on a simulated high-contrast image. However, the information can be of any type (e.g., providing a representation of the body part (such as being represented solely by a simulated high-contrast image, superimposed on an acquired zero-contrast image, combined in any way with an acquired zero-contrast image and an acquired low-contrast image, etc.), providing one or more parameters indicating the condition of the body part (such as local / global parameters based solely on values ​​of a simulated high-contrast image in a region of interest (ROI) or combined with a region representing healthy tissue), etc.), and it can be output in any manner (e.g., displayed on any device (such as a monitor, virtual reality glasses, etc.), or more generally output in any way, either in real time or offline, such as printing, remote transmission, etc.).

[0105] Additional embodiments provide extra advantageous features, but these features may be omitted entirely in the basic implementation. In this regard, it is explicitly stated that features of each of the following embodiments may be combined with the features described above, either individually or in combination with features of any number of other subsequent embodiments.

[0106] In this embodiment, the low-contrast image is a corresponding low-dose image representing a body part of a patient to which a low dose of contrast agent has been administered. However, the low dose can have any value (e.g., less than, equal to, or greater than the full dose of contrast agent).

[0107] In this embodiment, the simulated high-contrast image is a corresponding simulated high-dose image representing a body part of the patient, which simulates the administration of contrast agent at a high dose higher than the low dose. However, the high dose can have any value (e.g., less than, equal to, or higher than the full dose of contrast agent).

[0108] In this embodiment, the low dose is the full dose that serves as the standard in medical applications, and the high dose is an enhanced dose that is higher than the full dose. However, the full dose can be of any type (e.g., depending on the type of imaging procedure, depending on the type of body part, depending on the patient type, weight, age, etc., fixed, etc.), and the enhanced dose can have any value (e.g., in an absolute sense or in a relative sense relative to the full dose).

[0109] In this embodiment, a low dose is a reduced dose below the full dose, and a high dose is the full dose. However, the reduced dose can have any value (e.g., in an absolute sense or in a relative sense relative to the full dose).

[0110] In this embodiment, the low-contrast image is a corresponding low-efficiency image representing a body part of a patient to which a low-efficiency contrast agent, which provides low efficiency in a medical application, has been applied. However, the low-efficiency contrast agent can be of any type (e.g., for different medical applications, body parts, targets, etc.).

[0111] In this embodiment, the simulated high-contrast image is a corresponding simulated high-efficiency image representing a part of the patient's body, which simulates the application of a high-efficiency contrast agent that provides higher efficiency than low efficiency in medical applications. However, the high-efficiency contrast agent can be of any type (e.g., a non-existent contrast agent, a contrast agent that cannot be used due to its side effects, a contrast agent that is currently unavailable, a contrast agent that requires too long a waiting time, etc.).

[0112] In an embodiment, the imaging method further includes (by a computing system) receiving at least one acquired baseline image. However, the acquired baseline images can be any number and can be of any type (e.g., acquired without contrast agent, acquired in advance relative to an acquired low-contrast image, acquired with a different dose of contrast agent, acquired with different contrast agents, acquired under different acquisition conditions, etc.), up to none.

[0113] In an embodiment, the imaging method includes (by a computing system) simulating each simulated high-contrast image by optimizing an objective function that includes a fidelity term based on a comparison between a corresponding acquired low-contrast image and a corresponding simulated low-contrast image generated from an acquired baseline image and the simulated high-contrast image. However, the possibility of generating simulated low-contrast images solely from the simulated high-contrast images is not excluded.

[0114] In this embodiment, the acquired baseline image includes a zero-contrast image representing a body part without contrast enhancement. However, the acquired zero-contrast image can be of any type (e.g., acquired with any advance relative to the administration of the contrast agent, acquired together with the administration of the contrast agent, etc.).

[0115] In an embodiment, the acquired baseline image includes an acquired low-latency image representing a body part after an acquisition delay following the application of the contrast agent, which is lower than the acquisition delay of the acquired low-contrast image. However, the acquisition delay can have any value (e.g., in an absolute or relative sense).

[0116] In an embodiment, for each simulated high-contrast image, the simulation step includes (by the computing system) generating a simulated low-contrast image using a decrementing neural network. However, the decrementing neural network can be of any type (e.g., residual networks, multilayer perceptron networks, recurrent networks, etc., having any number of layers, connections between layers, receptive fields, strides, padding, activation functions, etc.).

[0117] In this embodiment, the decrementing neural network has been trained to optimize its ability to simulate contrast enhancement from high contrast to low contrast. However, the decrementing neural network can be trained in any way (e.g., with a sample set including images from humans / animals (living or non-living), from artificial objects mimicking corresponding body parts, using a sample set that is at least partially simulated, with or without zero-contrast images, using any algorithm (such as stochastic gradient descent, real-time recurrent learning, higher-order gradient descent, extended Kalman filtering, etc.), any loss function (such as based on mean absolute error, mean squared error, perceptual loss), etc.).

[0118] In an embodiment, the objective function includes a regularization term that provides regularization for the optimized objective function. However, the regularization term can be of any type (e.g., penalty, prior, one or more constraints such as p-norm, total variation, with or without a universal differential operator, with or without a threshold operator, with or without any initial guess of a simulated high-contrast image, etc.).

[0119] In an embodiment, the regularization term includes at least one regularization parameter that defines the degree of regularization. However, the regularization parameters can be any number (e.g., general parameters for the entire regularization term, specific parameters for its components, etc.) and can be of any type (e.g., static / dynamic, additive / multiplicative / exponential, etc.).

[0120] In an embodiment, for each simulated high-contrast image, the simulation includes (by a computing system) calculating a noise indicator based on the noise of at least the acquired low-contrast image. However, the noise indicator can be calculated in any way (e.g., based on any data source (such as a noise matrix determined in any way from the acquired low-contrast image, a background region of the acquired low-contrast image, etc.), and any technique can be employed for each data source (e.g., based on any corresponding statistical parameter, norm, etc.), and it can be based in any way on the acquired low-contrast image (e.g., equal to its noise indicator, equal to any combination of noise indicators of a set of acquired images including it (such as their average, maximum, etc.), etc.).

[0121] In an embodiment, for each simulated high-contrast image, the simulation includes (by a computing system) selecting the type of regularization term from a plurality of candidates based on a noise indicator. However, the number of candidates can be any and they can be of any type (see above).

[0122] In an embodiment, for each simulated high-contrast image, the simulation includes (by the computing system) setting a regularization parameter based on the acquired low-contrast image. However, the regularization parameter can be based on the acquired low-contrast image in any way (e.g., based on any feature (such as all content or noise indicators), based solely on features of the acquired low-contrast image, based on any indication of features of a set of acquired low-contrast images including it (such as their average), etc.), and it can be set accordingly in any way (e.g., by directly calculating the regularization parameter using any technique (such as the difference principle, L-curve method, cross-validation, generalized cross-validation, heuristic hypothesis, prior selection rule, posterior selection rule, etc.), by applying any linear / nonlinear function (such as a logarithmic function) to a predefined value, etc.).

[0123] In an embodiment, for each simulated high-contrast image, the simulation includes (by a computing system) calculating an additional noise indicator based on the noise of at least the acquired low-contrast image. However, the additional noise indicator can be of any type (e.g., the same as or different from the noise indicator).

[0124] In an embodiment, for each simulated high-contrast image, the simulation includes setting a regularization parameter (by the computing system) based on an additional noise indicator. However, the regularization parameter can be set based on the additional noise indicator in any way (e.g., by calculating it directly from the additional noise indicator in any way, by applying any linear-nonlinear function (such as a logarithmic function) that depends on the additional noise indicator to a predefined value (such as increasing the regularization parameter as the additional noise indicator increases), etc.).

[0125] In an embodiment, for each simulated high-contrast image, the simulation step includes (by the computing system) calculating a regularization term that indicates the total variation of the regularized image. However, for any regularized image (e.g., equal to the simulated high-contrast image, equal to any difference between the simulated high-contrast image and any initial high-contrast image, etc.), the total variation of the regularized image can be calculated in any way (e.g., isotropic / anisotropic type, any gradient based on its value (such as along any number of directions, forward, backward, or both), considering all directions or only the most significant direction, accumulating gradients in any way (such as by calculating their sum, mean, median, or any norm), adding any bias term until none, etc.).

[0126] In this embodiment, the regularized image is equal to a simulated high-contrast image. However, the regularized image can be set to a simulated high-contrast image in any way (e.g., as is, after any post-processing such as applying a smoothing filter to it, etc.).

[0127] In this embodiment, the regularized image is equal to the difference between the simulated high-contrast image and the initial high-contrast image. However, the difference between the simulated high-contrast image and the initial high-contrast image can be defined in any way (e.g., by simply subtracting their values, by downsampling the image, subtracting their values, and then upsampling the result, etc.).

[0128] In one embodiment, a preliminary high-contrast image is generated from an acquired low-contrast image to simulate contrast enhancement from low to high contrast. However, the preliminary high-contrast image can be generated in any manner (e.g., the same as or different from the simulated low-contrast image).

[0129] In an embodiment, for each simulated high-contrast image, the simulation step includes generating an initial high-contrast image (by a computing system) using an incremental neural network. However, the incremental neural network can be of any type (e.g., the same as or different from a decremental neural network).

[0130] In this embodiment, the incremental neural network has been trained to optimize its ability to simulate contrast enhancement from low to high contrast. However, the incremental neural network can be trained in any way (e.g., the same as or different from the decremental neural network).

[0131] In an embodiment, for each simulated high-contrast image, the simulation step includes (by the computing system) initializing the simulated high-contrast image to an acquired low-contrast image. However, the simulated high-contrast image can be initialized in any way (e.g., initialized as is to the acquired low-contrast image, multiplied by an increment factor or a percentage thereof, etc.), or even in any other way (e.g., initialized to a preliminary high-contrast image, etc.).

[0132] In an embodiment, for each simulated high-contrast image, the simulation step includes (by the computing system) repeating multiple iterations of the optimization objective function. However, the iterations can be any number and of any type (e.g., Tikhonov iteration, Landwebber iteration, Arnoldi iteration, Bregman iteration, with or without Krylov projection, non-stationary or stationary with the same regularization parameter, etc., depending on how the regularization parameter changes).

[0133] In this embodiment, the iterations are repeated until a termination condition is met. However, the termination condition can be of any type (e.g., based on any convergence exponent, any maximum number of iterations, both, etc.).

[0134] In an embodiment, each subsequent iteration, unlike the first iteration, utilizes a simulated high-contrast image from the previous iteration preceding that iteration. However, the simulated high-contrast image from the previous iteration can be utilized in any way (e.g., to define regularization terms, to initialize the simulated high-contrast image, etc.).

[0135] In an embodiment, the imaging method includes (by a computing system) calculating a regularization term for each subsequent iteration to indicate the total variation of a regularized image, the total variation of which is equal to the difference between the simulated high-contrast image of the next iteration and the simulated high-contrast image of the previous iteration. However, the difference between the simulated high-contrast images can be defined in any way (e.g., by subtracting their values ​​as is, with different weights, etc.).

[0136] In an embodiment, the imaging method includes (by a computing system) setting the regularization parameter for each subsequent iteration by applying a variation law to an initial value of the regularization parameter. However, the variation law can be of any type (e.g., exponential, logarithmic, linear, etc.) and can be applied to any initial value (e.g., based on an acquired low-contrast image, predefined, etc., as described above).

[0137] Generally speaking, if the same solution is achieved by an equivalent method, then similar considerations apply, provided that it is still within the scope of the claims (by using similar steps with more steps or parts thereof that have the same function, or by removing some unnecessary steps or adding other optional steps); moreover, the steps may be performed in different orders, concurrently, or (at least partially) in an interleaved manner.

[0138] An embodiment provides a computer program configured to cause the computing system to perform the imaging method mentioned above when executed on a computing system. An embodiment also provides a computer program product including one or more computer-readable storage media implementing the computer program, which can be loaded into the working memory of a computing system, thereby configuring the computing system to perform the same imaging method. However, the program can be executed on any computing system (see below). The program can be implemented as a standalone module, as a plug-in to existing software programs (e.g., imaging applications), or even directly within the latter.

[0139] Generally, similar considerations apply if the program is structured differently, or if additional modules or functions are provided; similarly, the memory structure can be of other types or can be replaced by an equivalent entity (not necessarily composed of physical storage media). The program can take any form suitable for use by any computing system, thereby configuring the computing system to perform the desired operation; in particular, the program can be in the form of external or resident software, firmware, or microcode (either object code or source code, for example, to be compiled or interpreted). Moreover, it is possible to provide the program on any computer-readable storage medium. A storage medium is any tangible medium (essentially different from transient signals) that can retain and store instructions for use by the computing system. For example, a storage medium can be of electronic, magnetic, optical, electromagnetic, infrared, or semiconductor types; examples of such storage media are fixed disks (where programs can be preloaded), removable disks, memory cards (e.g., USB type), etc. The program can be downloaded from the storage medium or via a network (e.g., the Internet, wide area network, and / or a local area network including transmission cables, fiber optics, wireless connections, network devices); one or more network adapters in the computing system receive the program from the network and forward it to one or more storage devices of the computing system for storage. In any case, the solutions according to embodiments of this disclosure can be implemented even with hardware structures (e.g., electronic circuits integrated on one or more chips of semiconductor material, such as field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs)), or with a combination of software and hardware that are properly programmed or otherwise configured.

[0140] An embodiment provides a computing system including components configured to perform the steps of the imaging method described above. Another embodiment provides a computing system including corresponding circuitry (i.e., hardware of any suitable configuration (e.g., software)) for performing the steps of the same imaging method. However, the computing system can be of any type (e.g., a computer, a controller, etc.).

[0141] The embodiments provide an imaging system for imaging body parts of a patient in a medical application. However, the imaging system can be used in any medical application to image body parts of any type, any condition, and any patient (see above).

[0142] In this embodiment, the imaging system includes the aforementioned computing system and a scanner for acquiring the acquired low-contrast images. However, the scanner can be of any type (e.g., MRI, CT, fluoroscopy, fluorescence, ultrasound, PET, etc.).

[0143] In one embodiment, the computing system is coupled to the scanner to receive the acquired low-contrast image from it. However, the computing system and the scanner can be coupled in any way (e.g., locally / remotely via any type of wired and / or wireless connection, etc.).

[0144] Generally, similar considerations apply if the computing system and the imaging system each have different structures, or include equivalent components, or have other operational features, provided that they remain within the scope of the claims. In any case, each component may be separated into multiple elements, or two or more components may be combined into a single element; moreover, each component may be replicated to support the parallel execution of corresponding operations. Moreover, unless otherwise specified, any interaction between different components generally does not need to be sequential, and it may be either direct or indirect through one or more intermediaries.

[0145] The embodiment provides a medical method applied to a patient's body part. However, this medical method can be applied to any body part of any patient (see above).

[0146] In one embodiment, the medical method includes administering a contrast agent to a patient, the contrast agent providing contrast enhancement to a target with low contrast. However, the contrast agent can be administered in any manner (e.g., with a syringe, an infusion pump, in advance, shortly before image acquisition, continuously during acquisition, etc.), and it can provide low contrast in any manner (see above).

[0147] In an embodiment, the medical method includes acquiring one or more acquired low-contrast images in response to the administration of a contrast agent, the one or more acquired low-contrast images providing a contrast-enhanced representation of a body part having a target with low contrast enhancement. However, the acquired low-contrast images can be any number, any type, and acquired in any manner (see above).

[0148] In this embodiment, the corresponding simulated high-contrast image is modeled from the acquired low-contrast image, and information related to the body part based on the simulated high-contrast image is output according to the imaging method described above. However, these operations can be performed in any manner (e.g., outputting information substantially simultaneously with the acquisition of the acquired low-contrast image or with any delay, etc.).

[0149] In one embodiment, the medical method includes performing a medical procedure related to a body part based on corresponding information. However, the medical procedure can be of any type (e.g., a diagnostic procedure, a treatment procedure, a surgical procedure, etc.).

[0150] In this embodiment, the medical method includes a diagnostic method that assesses the health status of a body part based on corresponding information. However, the proposed method can be applied in any kind of diagnostic application in the broadest sense of the term (e.g., to detect new lesions, monitor known lesions, etc.).

[0151] In this embodiment, the medical method includes a treatment method that treats body parts according to corresponding information. However, the proposed method can be applied to any kind of treatment method in the broadest sense of the term (e.g., aimed at curing a pathological condition, preventing its progression, preventing the occurrence of a pathological condition, or simply improving patient comfort).

[0152] In this embodiment, the medical method includes surgical procedures performed on body parts based on corresponding information. However, the proposed method can be applied to any kind of surgical procedure in the broadest sense of the term (e.g., for therapeutic purposes, for preventative purposes, for aesthetic purposes, etc.).

Claims

1. An imaging method (400) for imaging a patient's body parts in a medical application, wherein the imaging method (400) includes, under the control of a computing system (110): The computing system (110) receives (421) one or more acquired low-contrast images, the one or more acquired low-contrast images providing a contrast-enhanced representation of the body part, the contrast-enhanced representation having contrast enhancement for a target with low contrast. The computing system (110) simulates (424-487) one or more simulated high-contrast images corresponding to the acquired low-contrast images. These simulated high-contrast images provide a representation of contrast enhancement for the body part, simulating an increase in contrast enhancement from low contrast to higher contrast than the low contrast. Each simulated high-contrast image is simulated by optimizing an objective function (445-487) that includes a fidelity term based on a comparison between the corresponding acquired low-contrast image and a simulated low-contrast image generated from the simulated high-contrast image to simulate a decrease in contrast enhancement from high to low. The computing system (110) outputs (490) information related to body parts based on simulated high-contrast images.

2. The imaging method (400) according to claim 1, wherein the low-contrast image is a corresponding low-dose image representing a body part of the patient to which a low dose of contrast agent has been administered, and the simulated high-contrast image is a corresponding simulated high-dose image representing a body part of the patient to which a high dose of contrast agent higher than the low dose has been administered, and specifically, the low dose is the full dose as standard in medical applications and the high dose is an enhanced dose higher than the full dose, or the low dose is a reduced dose lower than the full dose and the high dose is the full dose.

3. The imaging method (400) according to claim 1, wherein the low-contrast image is a corresponding low-efficiency image representing a body part of the patient to which a low-efficiency contrast agent has been applied in a medical application to provide low efficiency, and the simulated high-contrast image is a corresponding simulated high-efficiency image representing a body part of the patient to which a high-efficiency contrast agent has been applied in a medical application to provide higher efficiency than low efficiency.

4. The imaging method (400) according to any one of claims 1 to 3, wherein the imaging method (400) further comprises: The computing system (110) receives (409) at least one acquired baseline image, and The computing system (110) simulates each simulated high-contrast image by optimizing an objective function (445-487), which includes a fidelity term based on a comparison between a corresponding acquired low-contrast image and a corresponding simulated low-contrast image generated from an acquired baseline image and a simulated high-contrast image.

5. The imaging method (400) according to claim 4, wherein the acquired baseline image comprises: The acquired zero-contrast image represents body parts without contrast enhancement, and / or The acquired low-latency image represents a body part acquired after a delay following the application of the contrast agent, which is lower than the acquisition delay of the acquired low-contrast image.

6. The imaging method (400) according to any one of claims 1 to 5, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computing system (110) generates (448) a simulated low-contrast image using a decreasing neural network (335), which is trained to optimize its ability to simulate contrast enhancement from high contrast to low contrast.

7. The imaging method (400) according to any one of claims 1 to 6, wherein the objective function includes a regularization term that provides a regularization of the optimized (445-487) objective function, the regularization term including at least one regularization parameter defining the degree of regularization.

8. The imaging method (400) according to claim 7, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computing system (110) calculates (430) a noise indicator based on the noise of at least the acquired low-contrast image, and The computing system (110) selects the type of regularization term (433-436) from among its multiple candidates based on the noise indicator.

9. The imaging method (400) according to claim 7 or 8, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computing system (110) sets the regularization parameters (430, 439-442) based on the acquired low-contrast image.

10. The imaging method (400) according to claim 9, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computing system (110) calculates (430) additional noise indicators based on the noise of at least the acquired low-contrast image, and The computing system (110) sets (439-442) regularization parameters according to the noise indicator.

11. The imaging method (400) according to any one of claims 7 to 10, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computing system (110) calculates (460) a regularization term to indicate the total variation of a regularized image equal to the simulated high-contrast image.

12. The imaging method (400) according to any one of claims 7 to 10, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computing system (110) calculates (463-466) regularization terms to indicate the total variation of the regularized image, which is equal to the difference between the simulated high-contrast image and the preliminary high-contrast image, which is generated from the acquired low-contrast image to simulate contrast enhancement from low to high contrast.

13. The imaging method (400) according to claim 12, wherein for each simulated high-contrast image, the simulation (424-487) comprises: A preliminary high-contrast image is generated (463) by a computing system (110) using an incremental neural network (330), which is trained to optimize its ability to simulate contrast enhancement from low to high contrast.

14. The imaging method (400) according to any one of claims 1 to 13, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The simulation high-contrast image is initialized (445) by the computing system (110) into the acquired low-contrast image.

15. The imaging method (400) according to any one of claims 1 to 14, wherein for each simulated high-contrast image, the simulation (424-487) comprises: The computational system (110) repeats (481-487) multiple iterations of the optimization (445-478) objective function until the termination condition is met, and each subsequent iteration, unlike the first iteration in the iteration, utilizes a simulated high-contrast image of the previous iteration prior to the next iteration.

16. The imaging method (400) according to claim 15, when directly or indirectly dependent on claim 7, wherein the imaging method (400) comprises: The computing system (110) calculates (469) a regularization term for each next iteration to indicate the total variation of the regularized image equal to the difference between the simulated high-contrast image of the next iteration and the simulated high-contrast image of the previous iteration.

17. The imaging method (400) according to claim 15 or 16, when directly or indirectly dependent on claim 7, wherein the imaging method (400) comprises: The computing system (110) sets (487) the regularization parameter for each subsequent iteration by applying a variation law to the initial value of the regularization parameter used for the first iteration.

18. A computer program (300) configured to cause the computing system (110) to perform the imaging method (400) according to any one of claims 1 to 17 when the computer program is executed on a computing system (110).

19. A computer program product comprising one or more computer-readable storage media implementing a computer program, said computer program being loadable into the working memory of a computing system, thereby configuring the computing system to perform the imaging method according to any one of claims 1 to 17.

20. A computing system (110) includes a component (135) configured to perform the steps of the imaging method (400) according to any one of claims 1 to 17.

21. A computing system comprising corresponding circuitry for performing steps of the imaging method according to any one of claims 1 to 17.

22. An imaging system (100) for imaging body parts of a patient in a medical application, wherein the imaging system (100) includes a computing system (110) according to claim 20 or 21 and a scanner (105) for acquiring the acquired low-contrast image, the computing system (110) being coupled to the scanner (105) to receive the acquired low-contrast image therefrom.

23. A medical method applied to a patient's body part, wherein the medical method comprises: A contrast agent is administered to the patient, which provides contrast enhancement to a target with low contrast. In response to the application of the contrast agent, one or more acquired low-contrast images are acquired, the one or more acquired low-contrast images providing a contrast-enhanced representation of a body part, the contrast-enhanced representation having contrast enhancement for a target with low contrast, the imaging method according to any one of claims 1 to 16 simulating a corresponding simulated high-contrast image from the acquired low-contrast images and outputting body part-related information based on the simulated high-contrast image, and Perform the medical procedures related to the body part based on the corresponding information.

24. The medical method of claim 23, wherein the medical method is a diagnostic method, comprising assessing the health status of a body part based on corresponding information.

25. The medical method of claim 23, wherein the medical method is a treatment method comprising treating a body part according to corresponding information.

26. The medical method of claim 23, wherein the medical method is a surgical method, comprising performing surgery on a body part based on corresponding information.

Images