Training a machine learning model to simulate images with high dose contrast agents in medical imaging applications

Abstract

A solution is proposed for training a machine learning model (420) for use in a medical imaging application. A corresponding method (700) provides (703-743, 759-763) sample sets, each of which includes a sample baseline image, a sample target image (obtained from a corresponding body part of a subject to which a particular dose of contrast agent has been administered), and a sample source dose (corresponding to a different dose of contrast agent). The machine learning model (420) is trained (744-758) to optimize its ability to generate each sample target image from the corresponding sample baseline image and sample source image. One or more sample sets are incomplete and lack a sample source image. Each incomplete sample set is completed (704-742; 759-763) by simulating a sample source image from at least the sample baseline image and the sample target image of the sample set. A computer program (500) and a computer program product for implementing the method (700) are proposed. Furthermore, a computing system (130) for executing the method (700) is proposed.

Description

[Technical field]

[0001] The present disclosure relates to the field of medical imaging applications, and more particularly, to contrast agent-based medical imaging applications. [Background technology]

[0002] The following is an introduction to the background of the present disclosure, with a description of the technologies related to the background. However, even if the description refers to documents, acts, works, etc., it does not imply or represent that the described technologies are part of the prior art or are common general knowledge in the field related to the present disclosure.

[0003] Imaging techniques are common in medical applications for the physician to examine a patient's body parts via images that provide a visual representation of them (typically in a substantially non-invasive manner, even when the body part is not directly visible). For this purpose, contrast agents are usually administered to patients undergoing a (medical) imaging procedure in order to enhance the contrast of (biological) objects of interest (e.g., lesions) and make them more visible in the images. This facilitates the physician's work in several medical applications, e.g., diagnostic applications to find / monitor lesions, therapeutic applications to delineate lesions to be treated, and surgical applications to recognize the margins of lesions to be removed.

[0004] In this context, it has also been proposed to use a reduced dose of contrast agent (e.g., at a reduced dose equal to 1 / 10 of the full dose). The reduced dose is less than the full dose of contrast agent that is standard in clinical practice. For this purpose, during the imaging procedure, a (zero-dose) image of the body part is acquired before administration of the contrast agent, and one or more (low-dose) images of the body part are acquired after administration of a low-dose contrast agent to the patient. A corresponding (full-dose) image of the body part that mimics the administration of a full-dose contrast agent to the patient is then simulated from the zero-dose image and the corresponding low-dose image by a deep learning network (DLN). The deep learning network restores the contrast enhancement from its level in the low-dose image (which is insufficient due to the reduced dose of contrast agent) to the desired level that would have been provided by the full dose of contrast agent. The deep learning network is trained by using sample sets that include zero-dose, low-dose, and full-dose images of the same type of body part, each of which is acquired before administration of contrast, after administration of low-dose, and after administration of full-dose contrast to the corresponding patient (or two or more zero-dose images acquired with different acquisition conditions, or two or more low-dose images acquired with different low-dose contrast). For example, "Enhao Gong et al., Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI, Journal of Magnetic Resonance Imaging, vol. 48, no. 2, February 13, 2018, pp. 330-340" discloses training a deep learning model for reducing gadolinium dose in contrast-enhanced brain MRI, where the model is trained on images acquired under three different conditions, namely, before contrast, after contrast with 10% dose reduction, and after contrast with 100% full dose.

[0005] The low-dose images (or zero-dose images under different acquisition conditions) required to train deep learning networks are not usually acquired in standard clinical practice. The collection of sample sets therefore requires corresponding prospective studies in which the imaging procedure is performed with a dedicated imaging protocol that deviates from the standard of care. In particular, for each desired reduction in the dose of contrast agent and for each type of body part of interest, a corresponding number of prospective studies must be performed.

[0006] However, prospective studies require a relatively complicated procedure to obtain the corresponding approval (approval of deviations from the standard of care) by the relevant health authorities. Moreover, the corresponding changes in the imaging procedure affect the clinical workflow, thereby potentially causing delays, technical difficulties, additional costs and risks for the patient (especially if the acquisition of additional images not normally required in the imaging procedure could be dangerous for the patient, e.g. if it requires exposing the patient to unnecessary radiation).

[0007] The need for prospective studies also limits the number of sample sets of available (zero-dose / low-dose / full-dose) images, and the diversity of the conditions under which they are acquired. This results in a relatively small amount and diversity of sample sets, which reduces the quality of the training of the deep learning network. The reduced quality of the training of the deep learning network negatively impacts its robustness, especially its ability to predict full-dose images. This has a negative impact on the physician's work, with corresponding risks to the patient's health (e.g., false positives / negatives or incorrect follow-up in diagnostic applications, reduced effectiveness of treatment or damage to healthy tissue in therapeutic applications, and incomplete resection of the lesion or excessive removal of normal tissue in surgical applications).

[0008] The paper "Johannes Haubold et al., Contrast agent dose reduction in computed tomography with deep learning using conditional generative adversarial network, European Radiology (2021) 31: 6087-6095" discloses the simulation of images with reduced iodine-based contrast agents (ICM) to verify the possibility of virtually enhancing ICM. For this purpose, dual-energy computed tomography (CT) images based on ICM are acquired. Separated ICM images are generated by encoding the distribution of ICM and are used to create virtual non-contrast (VCN) images. Dual-energy CT images corresponding to reduced ICM (50% or 80%) are simulated by proportional subtraction. Input and target image pairs are obtained by synthesizing the reduced and separated ICM images with the VCN image, respectively. A generative adversarial network (used to simulate and verify ICM enhancement) is trained on the input / target image pairs.

[0009] WO 2022 / 129633 (filed December 20, 2021, priority date December 18, 2020, published June 23, 2022) discloses the training of a convolutional neural network (CNN) used to generate a perfusion map from a sequence of perfusion images. For this purpose, a training base formed by a sequence of perfusion images and the associated perfusion map is provided. The training base is enriched with corresponding degraded versions of one or more sequences of perfusion images that are still associated with the corresponding perfusion map. The so enriched training base is then used to train the neural network. These degraded sequences of perfusion images are generated by simulating low doses of the corresponding contrast product. In particular, each value of the degraded sequence of perfusion images is calculated by applying a formula depending on its original value, the corresponding value over time, and the degraded coefficient of the contrast product.

[0010] WO 2022 / 129634 (filed December 20, 2021, with priority date December 18, 2020, published June 23, 2022) discloses the training of a predictive model used to predict injection parameters that provide a desired quality level when used to acquire contrast-enhanced images by administration of a contrast agent. For this purpose, training images are provided, including pre-contrast / contrast images associated with corresponding reference values ​​of the injection parameters used to acquire the training images and a manually determined reference quality level. Each pre-contrast image is applied to a predictive model for determining corresponding candidate values ​​of the injection parameters, and it is verified whether a theoretical contrast image acquired with the candidate values ​​of the injection parameters has the target quality level. If a corresponding contrast image exists, the verification is straightforward. Otherwise, the theoretical contrast image is simulated via a generator model and its quality level is determined via a classification model. [Prior art documents] [Patent documents]

[0011] [Patent Document 1] International Publication No. 2022 / 129633 [Patent Document 2] International Publication No. 2022 / 129634 [Non-patent literature]

[0012] [Non-Patent Document 1] Enhao Gong et al., Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI, Journal of Magnetic Resonance Imaging, vol. 48, no. 2, February 13, 2018, pp. 330-340 [Non-Patent Document 2] Johannes Haubold et al., Contrast agent dose reduction in computed tomography with deep learning using conditional generative adversarial network, European Radiology(2021)31:6087-6095 Summary of the Invention

[0013] A simplified summary of the disclosure is presented herein in order to provide a basic understanding of the disclosure. However, its sole purpose is to introduce some concepts of the disclosure in a simplified form as a prelude to the more detailed description below, and it should not be construed as identifying key elements or delineating the scope of the disclosure.

[0014] Generally speaking, the present disclosure is based on the idea of ​​simulating images for training machine learning (ML) models.

[0015] In particular, one embodiment provides a method for training a machine learning model for use in a medical imaging application. The method includes providing sample sets each including a sample baseline image, a sample target image (obtained from a corresponding body part of a subject administered with a particular dose of contrast agent), and a sample source dose (corresponding to a different dose of contrast agent). The machine learning model is trained to optimize its ability to generate each sample target image from the corresponding sample baseline image and sample source image. One or more sample sets are incomplete and lack a sample source image. Each incomplete sample set is completed by simulating a sample source image from at least the sample baseline image and the sample target image of the sample set.

[0016] A further aspect provides a computer program for carrying out the method.

[0017] A further aspect provides a corresponding computer program product.

[0018] A further aspect provides a computing system for implementing the method.

[0019] More specifically, one or more aspects of the present disclosure are set out in independent claims and advantageous features thereof are set out in dependent claims, including all claim language incorporated herein by reference in its entirety (wherein any advantageous feature is provided with reference to any particular aspect applying mutatis mutandis to all other aspects). [Brief description of the drawings]

[0020] The solution of the present disclosure, as well as further features and advantages thereof, will be best understood with reference to the following detailed description thereof, given purely as a non-limiting indication, read in conjunction with the accompanying drawings, in which, for simplicity, corresponding elements are indicated with equivalent or similar reference signs, their description will not be repeated and the name of each entity is generally used to indicate both its type and its attributes, such as value, content and representation, in particular: [Figure 1] FIG. 1 shows a schematic block diagram of an infrastructure that can be used to implement the solution according to an embodiment of the present disclosure. [Figure 2A] 1 illustrates different exemplary scenarios for an imaging procedure according to an embodiment of the present disclosure. [Figure 2B] 1 illustrates different exemplary scenarios for an imaging procedure according to an embodiment of the present disclosure. [Figure 2C] 1 illustrates different exemplary scenarios for an imaging procedure according to an embodiment of the present disclosure. [Figure 2D] 1 illustrates different exemplary scenarios for an imaging procedure according to an embodiment of the present disclosure. [Figure 2E] 1 illustrates different exemplary scenarios for an imaging procedure according to an embodiment of the present disclosure. [Diagram 3] 1 illustrates an example scenario for a training procedure according to one embodiment of the present disclosure. [Figure 4] 1 illustrates major software components that may be used to implement an imaging procedure according to one embodiment of the present disclosure. [Diagram 5] 1 illustrates major software components that may be used to implement a training procedure according to one embodiment of the present disclosure. [Figure 6] 1 shows an activity diagram illustrating the flow of activities for an imaging procedure according to one embodiment of the present disclosure. [Figure 7A] FIG. 1 is an activity diagram illustrating the flow of activities for a training procedure according to one embodiment of the present disclosure. [Figure 7B] FIG. 1 is an activity diagram illustrating the flow of activities for a training procedure according to one embodiment of the present disclosure. [Figure 7C] FIG. 1 is an activity diagram illustrating the flow of activities for a training procedure according to one embodiment of the present disclosure. [Figure 8A] 1 shows a representative example of experimental results for a solution according to an embodiment of the present disclosure. [Figure 8B] 1 shows a representative example of experimental results for a solution according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] With particular reference to FIG. 1, there is shown a schematic block diagram of an infrastructure 100 that may be used to implement the solution according to an embodiment of the present disclosure.

[0022] The infrastructure 100 comprises the following components:

[0023] The one or more (medical) imaging systems 105 comprise a corresponding scanner 110 and a control computing system or simply a control computer 115. Each scanner 110 is used to obtain images representing a body part of a patient during a corresponding (medical) imaging procedure based on the administration of a contrast agent to the patient for enhancing the contrast of a corresponding (biological) target such as a lesion. For example, the scanner 110 is of the Magnetic Resonance Imaging (MRI) type. In this case, not shown, the (MRI) scanner 110 has a gantry for receiving the patient, which houses a superconducting magnet (for generating a very high constant magnetic field), a set of multiple gradient coils of different axes (for adjusting the constant magnetic field), and an RF coil (with a specific structure for applying magnetic pulses to certain body parts and receiving corresponding response signals). Alternatively, the scanner 110 is of the Computed Tomography (CT) type. In this case, also not shown, the (CT) scanner 110 has a gantry for receiving the patient. The gantry houses an X-ray generator, an X-ray detector, and a motor for rotating them around the body part of the patient. A corresponding control computer 115, for example a personal computer (PC), is used to control the operation of the scanner 110. For this purpose, the control computer 115 is coupled to the scanner 110. For example, if the scanner 110 is of the MRI type, the control computer 115 is placed outside the scanner room and coupled via a cable passing through a penetration panel used to shield the scanner 110, whereas if the scanner 110 is of the CT type, the control computer 115 is placed nearby it.

[0024] The imaging systems 105 are installed in one or more medical facilities (e.g., hospitals) with a corresponding central computing system, or simply a central server 120. Each central server 120 communicates with the control computer 115 of that imaging system 105 via a network 125, e.g., a local area network (LAN) of the medical facility. The central server 120 collects information about the imaging procedures performed by the imaging systems 105, each including a (sequence of) images representative of the corresponding body part, and additional information about the imaging procedure, e.g., patient identification, results of the imaging procedure, acquisition parameters of the imaging procedure, etc.

[0025] A configuration computing device 130, or simply a configuration computer 130 (or more), is used to configure the control computer 115 of the imaging system 105. The configuration computer 130 communicates with a central server 120 of all medical facilities, for example via an Internet-based network 135. The configuration computer 130 collects (anonymously) image sequences with corresponding imaging parameters of imaging procedures performed in the medical facilities, for use in configuring the control computer 115 of the imaging system 105.

[0026] Each of the control computer 115 and the configuration computer 130 comprises several units connected between them via a bus structure 140. In particular, a microprocessor (μP) 145 or more provides the logical capabilities of the (control / configuration) computer 115, 130. A non-volatile memory (ROM) 150 stores the basic code for bootstrapping the computer 115, 130, while a volatile memory (RAM) 155 is used as working memory by the microprocessor 145. The computers 115, 130 are provided with a mass memory 160, for example a solid-state disk (SSD), for storing programs and data. Furthermore, the computers 115, 130 comprise several controllers 165 for peripherals or input / output (I / O) units. In particular, as far as the present disclosure is concerned, the peripherals include a keyboard, a mouse, a monitor, network adapters (NICs) for connecting to the corresponding networks 125, 135, drives for reading and writing to removable storage units (such as USB keys), and, for each control computer 115, a trackball and corresponding drives for the associated unit of its scanner 110.

[0027] 2A-2E, different example scenarios for an imaging procedure are shown according to one embodiment of the present disclosure.

[0028] During each imaging procedure, the corresponding scanner acquires an image sequence of (operational) acquired images representative of the patient's body part under examination. The acquired images include an (operational) baseline image (or more) and one or more (operational) administered or low-dose images. For example, the baseline image is acquired from a body part that does not contain contrast agent, and is then referred to hereinafter as an (operational) zero-dose image. The administered images are acquired from a body part of the patient to which contrast agent has been administered at an (operational) administered or low dose. A control computer associated with the scanner simulates (or synthesizes) corresponding (operational) simulated or high-dose images from the zero-dose and administered images (for example, by means of a neural network appropriately trained for this purpose, as will be explained in detail below). The simulated images mimic the administration of contrast agent to the patient at a (operational) simulated dose, or high dose, that is greater than the administered dose, i.e. with an increase factor (given by the ratio between the simulated dose and the administered dose) higher than 1. A representation of the body part based on the simulated images is then output (for example, by displaying them) to the physician in charge of the imaging procedure.

[0029] The simulated dose of contrast agent may be equal to the standard value in clinical practice. In this case, the simulated dose and the simulated image are called (operational) full dose and (operational) full dose image. The administered dose is therefore reduced for the full dose. In this case, the administered dose and the administered image are called (operational) low dose and (operational) low dose image. The simulation of the full dose image from the low dose image restores the contrast enhancement that would normally be obtained with the administration of the full dose of contrast agent. This is particularly useful in cases where the administration of the full dose of contrast agent to the patient may be dangerous (e.g. for children, pregnant women, patients suffering from certain pathologies such as renal insufficiency, etc.). In particular, by reducing the amount of contrast agent administered to the patient, the need for long-term tracking of its possible effects is eliminated. At the same time, the mimicked full dose of contrast agent keeps the contrast enhancement of the full dose image provided to the physician substantially unchanged (which would otherwise be increased by reducing the motion / aliasing artifacts that may be caused by the actual administration of the contrast agent in the full dose).

[0030] More generally, the inventors have surprisingly found that the increase factor can be applied to any administered dose to mimic a corresponding simulated dose of contrast agent, even if it is different from the one used to train the neural network.

[0031] In particular, referring to FIG. 2A, in a solution according to an embodiment of the present disclosure, the administered dose is equal to the full dose. Thus, the simulated dose is boosted with respect to the full dose. In this case, the simulated dose and the simulated image are called (operational) boosted dose and (operational) boosted dose image. The simulation of the boosted dose image from the full dose image enhances the contrast as if the boosted dose image was acquired with a higher (virtual) dose of contrast agent administered than is achievable in the current clinical practice. For example, the figure shows a zero dose image, a full dose image and two different boosted dose images (simulated with an increase factor of x2 and x10, respectively). This makes the physician's work easier. In particular, the boosted dose of contrast agent mimicked (more than the full dose) substantially increases the contrast in the boosted dose image provided to the physician (possibly reducing the movement / aliasing artifacts that may be caused instead by the actual administration of contrast agent in the boosted dose). At the same time, the full dose of contrast agent administered to the patient does not affect the standard of care and does not affect the clinical workflow. This is particularly advantageous when contrast enhancement is too low (e.g. when the target has a relatively low accumulation of contrast agent, such as in some pathologies like low-grade tumors). In any case, the proposed solution makes the target of the imaging procedure more salient and thus more easily and quickly distinguishable from other nearby (biological) features (especially when the physician has low expertise and / or is overloaded). This has a beneficial effect on the quality of the imaging procedure, for example substantially reducing the risk of false positives / negatives and incorrect follow-up in diagnostic applications, the risk of reduced efficacy of treatment or damage to normal tissue in therapeutic applications, the risk of incomplete excision of the lesion or excessive removal of normal tissue in surgical applications.

[0032] As a further improvement, the value of the boost factor can be selected, for example, from among its multiple predefined discrete values ​​(e.g., x2, x5, x10, etc.) or continuously within a predefined range (e.g., x2 to x20). This adds further flexibility. Indeed, in each imaging procedure, the physician can use the value of the boost factor that is most suitable for the specific situation (e.g., patient, body part, contrast agent, target, etc.). Furthermore, the physician can also verify the effect of various values ​​of the boost factor in real time and then select the one that provides the best contrast enhancement. This further improves the quality of the corresponding imaging procedure (e.g., the above-mentioned risks are further reduced).

[0033] Turning now to FIG. 2B, the (motion) composite image may also be generated by applying high dynamic range (HDR) techniques. Typically, HDR techniques are used in photography / videography applications to increase the contrast of an image (with or without increasing the dynamic range). For this purpose, multiple images of the same scene are acquired at different exposures. Due to the limited dynamic range of the images, distinctions are only possible within a corresponding limited luminosity range (i.e., bright features at low exposure and dark features at high exposure). Each image is then composited such that it contributes primarily in the areas that provide the optimal contrast.

[0034] In this case, the same HDR technique is used instead to generate each composite image from a zero dose image (being acquired), a full dose image (being acquired), and a corresponding boosted dose image (from which it is simulated). In general, the zero dose image has a low luminosity, the boosted dose image has a high luminosity, and the full dose image has an intermediate luminosity. Thus, the contribution to the composite image is mainly due to the zero dose image in the darkest regions, the boosted dose image in the brightest regions, and the full dose image otherwise. This makes it possible to obtain both a good contrast of the target with contrast agent (mainly due to the contribution of the boosted dose image) and a good anatomical detail of the rest of the body part without contrast agent (mainly due to the zero dose image). Thus, the target is made more prominent (thereby further improving the quality of the imaging procedure) while remaining well related to the morphology of the body part.

[0035] Moving to FIG. 2C, the contribution of the boost dose image to the composite image can also be modulated. In fact, the composite image reduces the increment of contrast enhancement. In either case, the contrast enhancement in the composite image can be increased by making the contribution of the boost dose image to the composite image more important (with respect to one of the zero dose image and the full dose image). For example, this figure shows different composite images obtained from corresponding zero dose image, full dose image and boost dose image (increase factor equal to 4), with different (relative) contributions of the boost dose image being 1.0, 1.5, 2.0 and 3.0 with respect to the contributions of the zero dose image and the full dose image. As can be seen, the contrast (between the target and the neighboring features) increases with the contribution of the boost dose image to the composite image.

[0036] With reference to FIG. 2D, a plot in arbitrary units is shown of the contrast index given by the difference between the average value of the area with tumor and the average value of the area with normal tissue in the same (zero dose / full dose / boosted dose / composite) images as above. In the zero dose image, the contrast index is almost zero (slightly negative in the example in question where the tumor appears darker than the normal tissue). In the full dose image, the contrast indicator increases (becomes positive). In the boost dose image, the contrast indicator becomes much higher according to the increase factor (x4). In the composite image, the contrast index becomes smaller relative to the boost dose image. However, the higher the contribution of the boost dose image to the composite image, the higher the corresponding contrast indicator (always exceeds the contrast indicator of the full dose image in this particular case).

[0037] Turning now to FIG. 2E, a further diagram is shown in which values ​​are plotted on the vertical axis in arbitrary units along a general line (on the horizontal axis) that intersects the areas having normal tissue in some of the above images. In particular, curve 2050 is associated with the zero dose image, and curve 205 f is related to the total dose image, and curve 205 b is related to the boost dose image, curve 205 c1 is associated with the composite image that has the lowest contribution to the boost dose image (1.0), and curve 205 c3 is associated with the composite image that has the highest contribution (3.0) to the boost dose image. As can be seen, the boost dose image (curve 205 b The spread of values ​​in the zero dose image (curve 2050) and the full dose image (curve 205 f ) is reduced relative to the spread of values ​​in the boost dose image. This means that the boost dose image contains degradation of the anatomical details of normal tissue. However, the composite image (curve 205 c1 and 205 c3 ) is independent of the contribution of the boost dose image to the composite image (1.0 to 3.0 in the example in question), and is proportional to the contribution of the zero dose / full dose image (curves 2050 and 205 f) which means that the composite image recovers the anatomical details of normal tissues even when the contribution of the boost dose image is relatively high.

[0038] Referring now to FIG. 3, an exemplary scenario for a training procedure is shown, according to one embodiment of the present disclosure.

[0039] The neural network is trained by using multiple sample sets (of sample images) representing corresponding body parts of different subjects, e.g., body parts of additional patients of the same type of body part to be imaged. Each sample set includes a (sample) baseline, image, a (sample) source image, and a (sample) target image. The baseline image is a (sample) zero dose image acquired from the corresponding body part without contrast agent. The sample target image is acquired from the corresponding body part of the subject to which contrast agent is administered at a (sample) target dose. The source image corresponds to a (sample) source dose of contrast agent that is lower than the target dose. The ratio between the source dose and the target dose is equal to (e.g., equal to) a reduction factor that corresponds to the inverse of the desired increase factor of the neural network.

[0040] The source images may also have been acquired from corresponding body parts of subjects to which contrast agent was administered at a source dose (e.g., in preclinical studies). However, in one embodiment of the present disclosure, at least some of the sample sets are received without corresponding source images (hereinafter, a sample set that already includes all of those sample images is referred to as a complete sample set, and a sample set that lacks those source images is referred to as an incomplete sample set). The source image of each incomplete sample set is instead simulated (or synthesized) (e.g., analytically) from other (acquired) sample images of the incomplete sample set, i.e., the zero dose image and the target image, so as to mimic the administration of contrast agent at a source dose to the subject.

[0041] The sample sets (either completed or already completed and received) are then used to train the neural network and optimize its ability to generate target images (ground truth) for each sample set from the zero-dose images and source images of the sample set (e.g., by using one portion of the sample set to determine the corresponding configuration of the neural network and validating it using another portion of the sample set).

[0042] For example, target images are acquired from corresponding body parts of a subject to which a full dose of contrast agent has been administered, hereafter referred to as (sample) full-dose images. The source dose is then reduced relative to the full dose. In this case, the source dose and source images are referred to as (sample) low-dose and (sample) low-dose images.

[0043] The above solution significantly facilitates training of neural networks.

[0044] In particular, this primarily requires that only zero-dose and full-dose images are actually acquired, and thus it is possible to use the zero-dose and full-dose images that are typically acquired in standard clinical practice.

[0045] As a result, the training of the neural network can be performed primarily with sample images of an incomplete sample set collected retrospectively from previously performed imaging procedures, and thus the collection of an incomplete sample set is more acceptable (e.g., by the corresponding body, such as an ethical committee or institutional review board) since it does not affect the standard of care and, consequently, poses lower risk to patients.

[0046] The acquisition of incomplete sample sets does not impact clinical workflow, thereby reducing delays, technical difficulties, additional costs and risks for the patient. In particular, this avoids (or at least substantially reduces) the acquisition of additional images that are not normally required in the imaging procedure, and is particularly important when the acquisition of these additional images may be dangerous for the patient (e.g., requiring exposure of the patient to unnecessary radiation).

[0047] Moreover, a relatively large number of incomplete sample sets, on the order of thousands, if not millions, are available (as typically collected over decades by many medical facilities). Incomplete sample sets are also typically acquired under several conditions (e.g., different scanners, patient types, body part conditions, etc.). The resulting amount and diversity of corresponding sample sets is relatively high, improving the quality of training of the neural network. The improved quality of training of the neural network has a positive impact on its robustness, especially its ability to predict (behavioral) administration images. This facilitates the physician's work in the corresponding imaging procedure (e.g., significantly reducing the above-mentioned risks).

[0048] The proposed solution allows relatively easy and fast training of the neural network for different values ​​of the augmentation factor. In fact, the required sample set (or at least a large part of them) can be generated from the same incomplete sample set by simply simulating the corresponding low-dose images for the required values ​​of the augmentation factor. This allows flexible use of the neural network with these values ​​of the augmentation factor, especially for different operating conditions (patients, body parts, lesions, etc.).

[0049] Referring now to FIG. 4, there is shown the major software components that may be used to implement an imaging procedure according to one embodiment of the present disclosure.

[0050] All software components (programs and data) are generally designated by the reference number 400. The software components 400 are typically stored in mass memory and are loaded (at least partially) into the working memory of each control computer 115 when the programs are executed, together with the operating system and other application programs not directly related to the solution of the present disclosure (and therefore omitted from the figure for simplicity). The programs are initially installed in the mass memory, for example from a removable storage device or a network. In this respect, each program may be a module, segment or part of code that includes one or more executable instructions for implementing a specified logical function.

[0051] The acquirer 405 drives the corresponding scanner components dedicated to acquire (operational) acquired images of the patient's body part during each imaging procedure, i.e. (operational) baseline images and (operational) administration images. The acquirer 405 writes the (operational) acquired image repository 410 containing the acquired images being acquired during the ongoing imaging procedure. The acquired image repository 410 has an entry for each acquired image. The entry stores a bitmap of the acquired image, which is defined by a matrix of cells (e.g. 512 rows and 512 columns) each containing the value of a voxel, i.e. an elementary image element representing a corresponding location (elementary volume) of the body part. Each voxel value defines the brightness (grayscale) of the voxel as a function of the (signal) strength of the response signal associated with the corresponding location. For example, in the case of an MRI scanner, the response signal represents the response of the location to a magnetic field applied to that location, and in the case of a CT scanner, the response signal represents the attenuation of the X-ray radiation applied to that location.

[0052] The pre-processor 415 pre-processes the acquired images (e.g. by registering them). The pre-processor 415 reads and writes to the acquired image repository 410. The (behavioral) machine learning model is used to simulate (or synthesize) (behavioral) simulation images from baseline images and corresponding administered images by applying machine learning techniques. Essentially, machine learning is used to perform a specific task (in this case, simulating a simulation image) without explicit instructions, but to automatically infer from examples how to do so (by utilizing a corresponding model learned from examples). In the particular implementation in question, deep learning techniques, which are a branch of machine learning based on neural networks, are applied. In this case, the machine learning model is a (behavioral) neural network 420. The neural network 420 is essentially a data processing system that approximates the operation of the human brain. The neural network 420 comprises basic processing elements (neurons) that perform operations based on corresponding weights. The neurons are connected via unidirectional channels (synapses) and data is transferred between the neurons. The neurons are organized into layers performing different operations, always comprising an input layer and an output layer for receiving input data of the neural network 420 and providing output data, respectively. In one embodiment of the present disclosure, the neural network 420 is a convolutional neural network (CNN), i.e. a specific type of deep neural network (with one or more hidden layers arranged successively between the input layer and the output layer along the processing direction of the neural network), one or more of whose hidden layers perform (cross)convolution operations. In particular, the neural network 420 is an autoencoder (encoder-decoder) convolutional neural network, which comprises an encoder that compresses data in a denser form (in the so-called latent space), the data thus compressed being used to perform the desired operation, and a decoder that expands the results thus obtained into the required more expanded form. More specifically, the input layer is configured to receive a baseline image and a dose image.The encoder includes three groups of three convolutional layers, followed by a corresponding max-pooling layer, and the decoder includes three groups of three convolutional layers, followed by a corresponding upsampling layer. Each convolutional layer performs a convolution operation through a convolution matrix (filter or kernel) defined by the corresponding weights, which is performed successively on a limited portion of the applied data (received field) by shifting the filter by a selected number of cells (stride) over the applied data, and also allows the filter to be applied by adding cells with zero content around the boundaries of the applied data (padding). Then, a batch normalization is applied (fixing the mean and variance of the corresponding data), followed by an activation function (introducing a nonlinear coefficient). For example, each convolutional layer applies a 3x3 filter with a padding of 1 and a stride of 1, each neuron of which applies a rectified linear unit (ReLU) activation function. Each max pooling layer is a pooling layer (that downsamples its applied data) that replaces the value of each limited portion of the applied data (window) with a single value, in this case its maximum value, by shifting the window by a selected number of cells (stride) over the applied data. For example, each max pooling layer has a 2x2 window with a stride of 1. Each upsampling layer is an unpooling layer (that inverts pooling) that expands each value into its surrounding area (window), for example using a max unpooling technique (where the value is placed in the same position as the maximum value used for downsampling and surrounded by zeros). For example, each upsampling layer has a 2x2 window. Bypass connections are added between the symmetric layers of the encoder and decoder (to avoid resolution loss), and skip connections are added within each group of convolutional layers and from the input layer to the output layer (to focus on the difference between the administered image and the baseline image).The output layer then generates a simulated image by adding the obtained results (representing the contrast enhancement at the simulated dose derived from the contrast enhancement at the administered dose that has been denoised) to the baseline image.

[0053] The neural network 420 reads an (operational) configuration repository 425 that defines one or more (operational) configurations of the neural network 420. The configuration repository 425 has an entry for each configuration of the neural network 420. The entry stores the configuration of the neural network 420 (defined by its weights) and the gain factor provided by the neural network 420 when operating according to this configuration. The neural network 420 reads an acquired image repository 410 and writes an (operational) simulation image repository 430. The simulation image repository 430 has an entry for each administered image in the acquired image repository 410. The entry stores a link to the corresponding administered image in the acquired image repository 410 and a bitmap of the corresponding simulated image, which is similarly defined by a matrix of cells (having the same size as the acquired image), each of which contains voxel values ​​for the corresponding location of the body part. The compositor 435 composites the baseline image, each administered image and the corresponding simulated image into a corresponding composite image. The compositor 435 reads the acquired image repository 410 and the simulated image repository 430 and writes the (operational) composite image repository 440. The composite image repository 440 has an entry for each composite image. The entry stores the bitmap of the composite image, which is in turn defined by a matrix of cells (having the same size as the simulated image), each containing the voxel values ​​of the corresponding location of the body part. The selector 445 exposes a user interface for selecting the value of the gain factor that the neural network 420 applies and the value of the contribution of the simulated image to the composite image. The selector 445 reads the configuration repository 425 and controls the neural network 420 and the compositor 435.

[0054] The display 450 drives the monitor of the control computer 115 to display the acquired images acquired during each imaging procedure and the generated composite image. The display 450 is fed by the acquirer 405 and reads the composite image repository 440. The imaging manager 455 manages each imaging procedure. To this end, the imaging manager 455 exposes a user interface for interacting with it. The imaging manager 455 controls the acquirer 405, the neural network 420, the combiner 435 and the display 450.

[0055] Referring now to FIG. 5, there is shown the major software components that may be used to implement a training procedure according to one embodiment of the present disclosure.

[0056] All software components (programs and data) are generally designated by the reference numeral 500. The software components 500 are typically stored in mass memory and are loaded (at least partially) into the working memory of the configuration computer 130 when the programs are executed, together with the operating system and other application programs not directly related to the solution of the present disclosure (and therefore omitted from the figure for simplicity). The programs are initially installed in the mass memory, for example from a removable storage device or a network. In this respect, each program may be a module, segment, or part of code that includes one or more executable instructions for implementing a specified logical function.

[0057] The collector 505 collects (complete / incomplete) sample sets. For example, incomplete sample sets, i.e. corresponding (sample) zero-dose images and (sample) full-dose images, are received from a central server (not shown) of a medical facility, acquired during the corresponding imaging procedure. Complete sample sets, i.e. sample sets further including corresponding (sample) low-dose images, are instead obtained in a laboratory performing preclinical studies on animals (such as rats). Indeed, the inventors have surprisingly found that operational neural networks trained on sample sets derived (at least in part) from animals provide good quality even when applied to humans. As a result, complete sample sets can be provided in a relatively simple manner. The collector 505 writes a sample set repository 510 that contains information about the sample sets. The sample set repository 510 has an entry for each sample set. The entry stores the corresponding bitmaps of the sample images of the sample set, i.e. its (acquired) zero-dose image, its (acquired) full-dose image, and its (acquired / simulated) low-dose image. As described above, the bitmap of each sample image is defined by a matrix of cells (e.g., 512 rows and 512 columns) each containing voxel values ​​for a corresponding location in the respective body part. Additionally, if the sample set was initially incomplete, the entry stores one or more acquisition parameters for the acquisition of that zero-dose / full-dose image. In particular, the acquisition parameters include one or more external parameters for the settings of the scanner used to acquire the zero-dose / full-dose image and one or more internal parameters for the corresponding body part (e.g., average values ​​of major tissues of the body part).

[0058] A pre-processor 515 pre-processes the zero-dose / full-dose images of each incomplete sample set (e.g., by co-registration, noise removal, etc.). The pre-processor 515 reads and writes to the sample set repository 510. The analysis engine 520 simulates (or synthesizes) low-dose images from the zero-dose / full-dose images of each incomplete sample set. The analysis engine 520 exposes a user interface for interacting with it. The analysis engine 520 reads and writes to the sample set repository 510. The analysis engine 520 reads a simulation equation repository 525 that stores one or more simulation equations to be used to simulate the low-dose images.

[0059] For example, in the case of an MRI scanner, when spin-echo is selected as the operating mode, the signal intensity defining each voxel value of the sample image (given by the transverse component of the magnetization of the corresponding location of the body part during relaxation of the spins of the protons of the water molecules present to return to their equilibrium state after the application of a magnetic pulse by the RF coil) is expressed by the following signal law:

number

number

[0060] Conversely, if a contrast agent is present at the location, the parameters T1 and T2 depend on the corresponding diamagnetic value plus the corresponding paramagnetic value imparted by the contrast agent, resulting in a signal intensity (M agent (distinguished as follows):

number

number

number

[0061] Therefore, in the full-dose image, the signal intensity (M full The metric is identified as:

number

number

number

[0062] From the above, the simulation formula is

number

number

number

number

[0063] The same simulation equations are obtained for other operating modes of the MRI scanner, such as gradient echo, MP-RAGE, etc.

[0064] Similarly, for a CT scanner, the signal intensity that defines each voxel value of a sample image (given by the X-ray radiation remaining after traversing the corresponding location due to attenuation) is expressed by the following signal law:

number

number

[0065] Conversely, if a contrast agent is present at the location, an additional attenuation of the X-ray radiation is caused thereby, resulting in an increase in the signal intensity (I agent The following is an example of the

number

number

[0066] Therefore, in the full-dose image, the signal intensity (I full The metric is identified as:

number

number

number

[0067] From the above, the simulation formula is

number

number

number

[0068] The proposed implementation (wherein the simulation equations are derived from the signal law linearized with respect to the local concentration of the contrast agent) is computationally very simple, and the loss of accuracy of the so obtained low-dose images (due to the linearization of the signal law) is acceptable for the purpose of training a (working) neural network.

[0069] Alternatively, the signal law is approximated to a higher order (second, third, etc.) of its Taylor series as a function of the local concentration (or density). In this case, the solution of the equation obtained for the local concentration of the contrast agent at the total dose provides a corresponding number of values ​​that need to be evaluated to discard any one of the physically meaningless values. This improves the accuracy of the simulated low-dose image (the higher the order of approximation, the higher the accuracy). As another alternative, the signal law is solved numerically (again with an evaluation of possible solutions to discard any one of the physically meaningless contrast agents) for the local concentration of the contrast agent at the total dose. This further improves the accuracy of the simulated low-dose image.

[0070] The noise corrector 530 corrects the noise of the low-dose images. In fact, the zero-dose and full-dose images of each incomplete sample set contain noise that propagates to the corresponding low-dose images according to a simulation formula. However, the noise obtained in this way (simulated noise) has a slightly different statistical distribution than the noise obtained by actually acquiring low-dose images from the corresponding body part of a patient to which a low dose of contrast agent has been administered (real noise). In particular, the noise of the zero-dose images and the noise of the full-dose images can be considered to have a normal-type statistical distribution with zero mean and corresponding standard deviation, which are in accordance with the rules of error (or uncertainty) propagation.

number

number

number

[0071] However, the inventors have found that the standard deviation σ artificial It has been found that better results are obtained by incrementing the (theoretical) value of according to an empirically determined correction factor (e.g., equal to 1.5 to 2.5, preferably 1.7 to 2.3, even more preferably 1.9 to 2.1, e.g., 2.0). The noise corrector 530 reads and writes to the sample set repository 510.

[0072] Additionally or alternatively, the low-dose images of the incomplete sample set are simulated (or synthesized) by an additional (training) machine learning model. For example, the machine learning model is a training neural network 535, in particular an autoencoder convolutional neural network as described above. The training neural network 535 reads a (training) configuration repository 540 that stores the (training) configuration of the training neural network 535 (i.e., its weights as described above). The training neural network 535 also reads and writes to the sample set repository 510.

[0073] The training engine 545 trains a copy of the operational neural network, indicated by the same reference number 420, and the training neural network 535 (if available). The training engine 545 reads the sample set repository 510. The training engine 545 writes a copy of the (operational) configuration repository, indicated by the same reference number 425, which is read by the operational neural network 420, and writes the configuration repository 540 for the training neural network 535.

[0074] Referring now to FIG. 6, an activity diagram illustrating the flow of activity associated with an imaging procedure according to one embodiment of the present disclosure is shown.

[0075] In this regard, each block may correspond to one or more executable instructions for implementing a specified logical function on a control computer. In particular, the activity diagram represents an exemplary process that may be used to image a patient's body part during each imaging procedure using method 600.

[0076] The process starts with a black start circle 603 as soon as a (new) imaging procedure is started (as indicated by a corresponding command entered by a healthcare operator, such as a doctor or radiologist, via the imaging manager user interface after the patient has reached a suitable position relative to the scanner, such as inside the gantry in case of an MRI / CT scanner). In response to this, in block 606 the acquirer starts acquiring (operational) baseline images of the body part, and the display displays them in real time on the monitor of the control computer. The baseline (zero dose) images are acquired before administering contrast agent to the patient, so that the body part does not contain contrast agent, or at least does not contain a significant amount of contrast agent (because the patient has never received contrast agent or a relatively long time has passed since the previous administration of contrast agent to the patient, ensuring that the contrast agent has been substantially removed). When the physician selects one of the zero dose images (either directly via the imaging manager user interface or by a corresponding command entered by the healthcare operator), in block 609 the acquirer stores this zero dose image in the (operational) acquired image repository (initially empty).

[0077] Then, in block 612, the indicator displays a message on the monitor of the control computer requesting administration of contrast agent (e.g., gadolinium-based for MRI applications, iodine-based for CT applications, etc.) to the patient. The contrast agent is administered in an administration dose. Typically, the administration dose is equal to the total dose of the contrast agent. The total dose has a standard value in clinical practice, required by medical authorities (i.e., organizations with jurisdiction over the application of medical care) or recommended by recognized organizations or consistent scientific publications. For example, for MRI applications, the total dose of the contrast agent is 0.1 mmol of gadolinium per kg of patient weight. For CT applications, the total dose of an iomeprol-based contrast agent with a formulation of 155-400 mg / mL, marketed under the name Iomeron by Bracco Imaging SpA™, is 20-200 mL for imaging of the head and 100-200 mL for imaging of other body part types. Alternatively, the total dose of an iopamidol-based contrast agent, marketed, for example, under the name Isovue by 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 (total iodine dose should not exceed 60 g), and 1.2-3.6 mL per kg of patient weight for a 250 mg / mL formulation or 1.0-3.0 mL per kg of patient weight for a 300 mg / mL formulation (total iodine dose should not exceed 30 g). However, the administered dose may also be less than the full dose in certain circumstances (e.g., when administration of the full dose may be dangerous).

[0078] In response, the healthcare operator administers a contrast agent to the patient. In particular, the contrast agent is adapted to reach and remain substantially fixed within a specific (biological) target, such as the tumor to be examined / resected / treated. This result can be achieved by using either a non-targeted contrast agent (adapted to accumulate at the target without specific interaction with the target, such as by passive accumulation) or a targeted contrast agent (adapted to bind to the target by specific interaction with the target, such as achieved by incorporating target-specific ligands into the formulation of the contrast agent, for example, based on chemical binding properties and / or physical structures that can interact with different tissues, vascular properties, metabolic properties, etc.). The contrast agent can be administered intravenously to the patient as a bolus (e.g., using a syringe). As a result, the contrast agent circulates within the patient's vascular system until it reaches and binds to the target. Instead, the remaining (unbound) contrast agent is removed from the patient's blood pool. After a waiting period (e.g., several minutes) that allows the contrast agent to accumulate in the (potential) target and be washed out from the rest of the patient, the imaging procedure can actually begin (e.g., as indicated by a corresponding command entered by a physician or healthcare operator via the imaging manager user interface), while the acquirer continues to acquire (operating) administered images of the body-part with a display that displays in real time on the control computer monitor.

[0079] At any time, the physician may select (directly via the user interface of the selector or by the healthcare operator) a desired (selected) value of the increase factor in block 615. In particular, in the discrete mode, the selected value of the increase factor may be selected from among those corresponding to the operational configuration of the (operational) neural network in the corresponding repository. At the same time, the physician may select (directly via the user interface of the selector or by the healthcare operator) a desired (selected) value of the (relative) contribution to the (operational) composite image of the (operational) simulation image relative to one of the zero dose / administration images. For example, the contribution of the simulation image may be set by default to be the same as one of the zero dose / administration images and increased (continuously or discretely) up to its maximum value (e.g., 5 to 10). In response, the display stops displaying the administration images on the monitor of the control computer. In block 618, the neural network configures according to the configuration of the selected value of the increase factor (retrieved from the corresponding repository). In block 621, the acquirer stores the just acquired (new) administration image in the (operational) acquired image repository. At block 624, the pre-processor pre-processes the dose image. In particular, the pre-processor co-registers the dose image with the zero-dose image (in the acquired image repository) to spatially correspond, for example, by applying a rigid transformation to the dose image. At block 627, the imaging manager provides the zero-dose image and the (pre-processed) dose image to a neural network. Moving to block 630, the neural network outputs the corresponding simulated image stored in the corresponding repository.

[0080] In block 633, the compositor combines the zero dose, dose and simulation images (obtained from the corresponding repositories) into their composite image, which is stored in the corresponding repository. For example, for this purpose, the compositor applies a modified version of an exposure blending algorithm (adapted for this different application) that implements a certain type of HDR technique that keeps the dynamic range unchanged. Specifically, the compositor calculates a (working) zero dose mask, a (working) dose mask and a (working) simulation mask from the zero dose, dose and simulation images, respectively. Each zero dose / dose / simulation mask comprises a matrix of cells (having the same size as the zero dose / dose / simulation image), each of which contains a mask value at the corresponding position. In the case of the dose mask and the simulation mask, each mask value is set to the corresponding voxel value of the dose and simulation images, respectively. In the case of the zero dose mask, each mask value is instead set to the corresponding voxel value of the zero dose image complemented to its maximum possible value. Each voxel value of the composite image is then calculated by applying the following blending formula:

number

number

number

number

[0081] In block 636, the display displays the composite image (retrieved from the corresponding repository) on the monitor of the control computer. The composite image is then displayed substantially in real time with the retrieval of the corresponding administered image (apart from a short delay due to the time required by the neural network and the compositor to generate it). In block 639, the selector verifies whether a different value of the gain factor and / or the contribution of the simulation image to the composite image has been selected. If so, the process returns to block 618 and the simulation weight w is selected according to the (new) selected value of the gain factor and / or the (new) selected contribution of the simulation image. hand then continuously repeats the same operations. Conversely, in block 642, the imaging manager verifies the status of the imaging procedure. If the imaging procedure is still in progress, the flow of activity returns to block 621 and continuously repeats the same operations. Conversely, if the imaging procedure is finished (as indicated by a corresponding command entered by a physician or healthcare operator via the imaging manager's user interface), the process ends at concentric white / black stop circles 645.

[0082] 7A-7C, an activity diagram illustrating the flow of activities associated with a training procedure according to one embodiment of the present disclosure is shown.

[0083] In this regard, each block may correspond to one or more executable instructions for implementing a specified logical function in a configuration computer. In particular, the activity diagram represents an exemplary process that may be used to train an operational neural network using method 700.

[0084] The process starts with a black start circle 701 whenever the operational neural network needs to be trained. In particular, this is done before the first delivery of the operational neural network. Moreover, this can also occur periodically in response to any significant change in the operational state of the imaging system (e.g. delivery of a new model of the corresponding scanner, variation in the patient population being imaged, etc.), in case of maintenance of the operational neural network or in case of release of a new version of the operational neural network, in order to maintain the required performance of the imaging system over time. In response to this, in block 702, the analysis engine prompts the operator (through its user interface) to input an indication of the desired increase factor by which the operational neural network must be trained, and also defines as its reciprocal the decrease factor of the (sample) low-dose images to be simulated for this purpose.

[0085] In block 703, the collector collects multiple image sequences of corresponding imaging procedures performed on different subject body parts (e.g., of one or more medical facilities for the incomplete sample set, and of a research laboratory for the complete sample set) together with corresponding acquisition parameters of the incomplete sample set. The body parts are of the same type for which the operational neural network is intended to be used. Each image sequence of the incomplete sample set includes a sequence of images acquired first without contrast agent and then with a full dose of contrast agent (e.g., actually used during the corresponding imaging procedure to provide a visual representation of the corresponding body part). Each image sequence of the complete sample set further includes a sequence of images acquired with a low dose of contrast agent (e.g., in preclinical studies). Some image sequences may also include corresponding raw data (used to generate the corresponding sample image). For example, in the case of an MRI scanner, the raw data may be acquired as k-space format (k-space) images. Each k-space image is defined by a matrix of cells whose horizontal axis corresponds to the spatial frequency, i.e., the wave number k (cycles per unit distance), and whose vertical axis corresponds to the phase of the detected response signal. Each cell includes a complex number that defines a different amplitude component of the corresponding response signal. The k-space image is complex transformed into a corresponding (complex) image by applying an inverse Fourier transform. The complex image is defined by a matrix of corresponding voxel cells. Each cell contains a complex number that represents the response signal being received from the corresponding location. Finally, the complex image is transformed into a corresponding (sampled) acquired image in magnitude form by setting each voxel value therein to the coefficient of the corresponding complex number in the complex image.

[0086] At this stage, the collector can filter the image sequences, for example to discard low quality image sequences. In either case, for each image sequence, the collector selects one of the images acquired without contrast as the (sample) zero-dose image, and one or more images acquired with contrast, up to all of them as the (sample) full-dose image and the corresponding low-dose image (if available). The collector then creates a new entry in the sample set repository for each full-dose image, adding the zero-dose image, the full-dose image, the corresponding low-dose image (if available), and a link to the corresponding acquisition parameters (if the low-dose image is not available). The sample set repository can then store a mix of incomplete and complete sample sets. For example, complete sample sets are 1-20%, preferably 5-15%, even more preferably 6-12%, e.g. 10%, of the total number of (incomplete / complete) sample sets. This further increases the quality of training of the operational neural network with limited additional effort (especially if the complete sample sets are obtained from preclinical trials).

[0087] In block 704, the analysis engine retrieves (from the corresponding repository) the simulation formula (e.g., manually selected by the operator via its user interface, defined by default, or the only one available) used to simulate the low-dose images of the incomplete sample set. Then, in block 705, a loop is entered in which the analysis engine takes into account the (current) incomplete sample set of the sample set repository (starting from the first one in any order). In block 706, the noise corrector calculates the noise of the zero-dose image (as the difference between the acquisition time and the denoising time) and then calculates its (zero-dose) standard deviation. Similarly, the noise corrector calculates the noise of the full-dose image (as the difference between the acquisition time and the denoising time) and then calculates its (full-dose) standard deviation. In both cases, the acquired (zero-dose and full-dose) images can be denoised with an autoencoder (convolutional neural network). For this purpose, the autoencoder is trained unsupervised with multiple images (such as all acquired images). In particular, the autoencoder is trained to optimize its ability to encode each image, ignore the minor data on it (due to noise), and then decode the result obtained, in order to reconstruct the same image with reduced noise. The noise corrector determines a reference standard deviation, for example equal to the average of the zero-dose standard deviation and the full-dose standard deviation. The noise corrector calculates the standard deviation of the artificial noise by applying a noise formula to the reference standard deviation and then multiplying the result obtained by a correction factor. The flow of activity branches in block 707 according to the configuration of the analysis engine (e.g., manually selected by the operator via its user interface, defined by default, or the only one available). In particular, blocks 708-726 are executed if the analysis engine is not configured to operate within the k-space block, and blocks 727-740 are executed otherwise. In both cases, the flow of activity rejoins in block 741.

[0088] Referring now to block 708 (not k-space), the pre-processor pre-processes the acquired (zero dose / full dose) images of the incomplete sample set. In particular, the pre-processor co-registers the full dose image with the zero dose image to make them spatially corresponding (e.g., by applying a rigid body transformation to the full dose image). Additionally or alternatively, the pre-processor denoises the acquired images to reduce their noise (as described above). In block 709, the analysis engine calculates a modulation coefficient for modulating the reduction factor used to apply the simulation formula. In fact, the simulation formula can introduce a higher approximation the higher the local concentration of the contrast agent. In particular, starting from a simulated value of the signal intensity (given by the simulation formula) substantially equal to the actual value of the signal intensity in the absence of the contrast agent (obtained by actually acquiring low dose images from a body part of a patient where the contrast agent was administered at a (sample) low dose), the simulated value becomes lower than the actual value as the local concentration of the contrast agent increases. To compensate for this loss of the simulated value to the actual value, it is possible to increase the value of the reduction factor used in the simulation formula (so as to limit the reduction of the simulated value to the corresponding administered value). More specifically, by solving an equation that sets the ratio between the signal law and its approximation equal to 1 for the attenuation factor, it is obtained that the value of the attenuation factor should be linearly incremented as a function of the local concentration of the contrast agent according to a proportionality factor (modulation factor) that depends on the acquisition parameters. The modulation factor is given by an analytically determined (correction) formula as a function of the acquisition parameters or by a value corresponding to the empirically determined acquisition parameters. Thus, the analysis engine obtains the acquisition parameters of the incomplete sample set from the sample set repository and then calculates the modulation factor by applying the correction formula to the acquisition parameters or by obtaining their value corresponding to the acquisition parameters from a predefined table.

[0089] The flow of activity further branches in block 710 according to the configuration of the analysis engine. In particular, if the analysis engine is configured to work with magnitude format images, a loop is entered in block 711, where the analysis engine takes into account the (current) voxel of the full dose image (starting from the first voxel of any order). In block 712, the analysis engine modulates the reduction factor used to apply the simulation formula of the voxel. For this purpose, the analysis engine calculates the contrast enhancement of the voxel as the difference between the voxel value of the full dose image and the voxel value of the zero dose image, and then calculates the modulation value of the reduction factor by multiplying the product between the modulation factor and the contrast enhancement. In block 713, the analysis engine calculates the voxel value of the low dose image by applying the simulation formula with the (modulated) reduction factor to the voxel value of the zero dose image and to the voxel value of the full dose image. Thus, in the example in question, the analysis engine subtracts the voxel value of the zero dose image from the voxel value of the full dose image, multiplies this difference by the reduction factor, and adds the obtained result to the voxel value of the zero dose image. The analysis engine then adds the voxel values ​​thus obtained to the low-dose image under construction in the sample set repository. In block 714, the analysis engine verifies whether the last voxel has been processed. If not, the activity flow returns to block 711 to repeat the same operations for the next voxel. Conversely (when all voxels have been processed), the corresponding loop is terminated by descending to block 715.

[0090] The noise corrector then injects the artificial noise in additive form into the low-dose image thus obtained. For this purpose, the noise corrector generates the artificial noise as a (noise) matrix of cells having the same size as the low-dose image. The noise matrix contains random values ​​having a normal statistical distribution with zero mean and a standard deviation equal to the artificial noise. In block 716, the noise corrector adds the noise matrix to the low-dose image for each voxel in the sample set repository. The process then continues to block 741.

[0091] Referring back to block 710, if the analysis engine is configured to operate on images in complex format, then in block 717, a branching of the flow of activity occurs according to their availability. If the zero-dose image and the full-dose image are already available in complex format, then in block 718, the analysis engine performs a phase correction by rotating the vectors representing the complex numbers of each of its cells to cancel their arguments (keep the same coefficients). This operation ensures that the result of applying the simulation formula is the same when operating on the zero-dose image and the full-dose image in complex format (because any operation applied to a corresponding complex number without an imaginary part is equivalent to the one applied to the corresponding coefficient). The process then continues to block 719. If the zero-dose image and the full-dose image are available in magnitude format, then the same point is also reached directly from block 717. In this case, the zero-dose image and the full-dose image are directly considered to be in complex format, where each voxel value (real number) is a complex number whose imaginary part is zero.

[0092] Now, similar operations as above are performed to generate the low-dose images from the zero-dose images, the full-dose images acting on them in complex form. In particular, a loop is entered in which the analysis engine takes into account the (current) voxels of the full-dose images (starting from the first voxel of any order). In block 720, the analysis engine modulates the reduction factor by calculating the contrast enhancement of the voxel (as the difference between the coefficient of the voxel value of the full-dose image and the coefficient of the voxel value of the zero-dose image), and then modulates the modulation value of the reduction factor by multiplying it by the product between the modulation coefficient and the contrast enhancement. In block 721, the analysis engine calculates the voxel values ​​of the low-dose image by applying a simulation formula with the (modulated) reduction factor to the voxel values ​​of the zero-dose image and the voxel values ​​of the full-dose image. The analysis engine then adds the voxel values ​​so obtained to the low-dose image under construction in the sample set repository. In block 722, the analysis engine verifies whether the last voxel has been processed. If not, the activity flow returns to block 719 to repeat the same operations for the next voxel. Conversely (when all voxels have been processed), the corresponding loop is terminated by descending to block 723 .

[0093] The noise corrector then injects the artificial noise in a convolutional form into the low-dose image thus obtained. For this purpose, the noise corrector generates the artificial noise as a (noise) matrix of cells having the same size as the low-dose image. The noise matrix contains random complex values ​​having a normal statistical distribution with a unitary mean and standard deviation equal to the one of the artificial noise. Then, in block 724, the noise corrector performs a convolution operation on the low-dose image in the sample set repository through the noise matrix (e.g., by shifting the noise matrix circularly over the low-dose image by one stride, the low-dose image is wrapped around in all directions). In block 725, the analysis engine converts the low-dose image thus obtained into a magnitude form. For this purpose, the analysis engine replaces each voxel value (here generally a complex number) of the low-dose image by its coefficient. The flow of activity further branches in block 726 according to the configuration of the analysis engine. In particular, if the analysis engine is configured to inject artificial noise into the low-dose image in a similarly additive form, the process proceeds to block 715 and performs the same operations as described above (and then proceeds to block 741). Conversely, the process descends directly into block 741.

[0094] Referring instead to block 727 (k-space), the analysis engine takes into account the zero-dose and full-dose images in complex form (either directly, if available, or by converting them from k-space form by applying an inverse Fourier transform). As described above, in block 728, the analysis engine performs a phase correction by rotating the vectors that represent the complex numbers of each cell of the zero-dose and full-dose images in complex form to cancel their arguments (keeping the same coefficients). In block 730, the analysis engine converts the zero-dose and full-dose images from complex form to k-space form by applying a Fourier transform.

[0095] Here, the low-dose image is generated from the zero-dose image and the full-dose image acting on them, processed in k-space format. In particular, a loop is entered in block 731, where the analysis engine takes into account the (current) cells of the full-dose image (starting from the first one in any order). In block 732, the analysis engine calculates the cell values ​​of the low-dose image by applying a simulation formula with the (original) reduction factor to the cell values ​​of the zero-dose image and to the cell values ​​of the full-dose image. The analysis engine then adds the cell values ​​so obtained to the low-dose image under construction in the sample set repository. In block 733, the analysis engine verifies whether the last cell has been processed. If not, the activity flow returns to block 731 and repeats the same operations for the next cell. Conversely (when all cells have been processed), the corresponding loop is terminated by descending to block 734.

[0096] The noise corrector then injects the artificial noise in a multiplexed form into the low-dose image thus obtained. For this purpose, the noise corrector generates the artificial noise as a (noise) matrix of cells having the same size as the low-dose image. The noise matrix contains complex random values ​​having a normal-type statistical distribution with a unitary mean and standard deviation equal to the one of the artificial noise. In block 735, the noise corrector multiplies the low-dose image cell-by-cell with the noise matrix in the sample set repository. The flow of activity further branches according to the configuration of the analysis engine in block 736. In particular, if the analysis engine is configured to inject the artificial noise into the low-dose image in an additive form as well, the process proceeds to block 737, where the noise corrector generates the artificial noise as a (further) noise matrix of cells (having the same size as the low-dose image) that now contains random complex values ​​having a normal-type statistical distribution with a null mean and a standard deviation equal to the one of the artificial noise. In block 738, the noise corrector adds the noise matrix to the cell-by-cell of the low-dose image in the sample set repository. The process then proceeds to block 739, where the same point is reached directly from block 736 if the analysis engine is not configured to inject artificial noise in additive form into the low-dose image. At this point, the analysis engine converts the low-dose image from k-space form to complex form by applying an inverse Fourier transform. In block 740, the analysis engine converts the low-dose image from complex form to magnitude form by replacing each voxel value of the low-dose image with its coefficient. The process then descends to block 741.

[0097] The above operations complete the incomplete sample set. Referring now to block 741, the analysis engine verifies whether the last incomplete sample set has been processed. If not, the flow of activity returns to block 705 to repeat the same operations for the next incomplete sample set. Conversely (when all incomplete sample sets have been processed), the corresponding loop is terminated by descending to block 742.

[0098] At this point, the flow of activity branches according to the operating mode of the configuration computer. If the sample set is used to train the operational neural network because a training neural network is not available, the training engine performs this operation directly to find optimized values ​​of the weights of the operational neural network that optimize its performance. This implementation is particularly simple and fast. At the same time, the accuracy of the simulated low-dose images is sufficient for the purpose of training the operational neural network with acceptable performance. In particular, the analysis engine of block 743 can post-process the sample (zero-dose / low-dose / full-dose) images of each sample set (either partially completed or already completed and provided as described above). For example, the analysis engine normalizes the sample images by scaling their voxel values ​​to a (common) predefined range. Furthermore, the analysis engine performs a data augmentation procedure by generating (new) sample sets from each (original) sample set to reduce overfitting in the training of the operational neural network. For example, new sample sets are generated by incrementally rotating the sample images of the original sample set by 1-5° from 0° to 90° and / or by flipping them horizontally / vertically. Additionally, if the low-dose images of the original sample set have not yet been done because they are incomplete, artificial noise is added to the low-dose images of the original / new sample set as described above. In either case, in block 744, the training engine selects multiple training sets by sampling the sample sets in the corresponding repository up to a percentage (e.g., a randomly selected 50%). In block 745, the training engine randomly initializes the weights of the operational neural network. Then, in block 746, a loop is entered in which the training engine feeds the zero-dose and low-dose images of each training set to the operational neural network. In response, in block 747, the operational neural network outputs a corresponding output image (ground truth) that must be equal to the full-dose image of the training set.In block 748, the training engine calculates a loss value based on the difference between the output image and the full dose image. For example, the loss value is given by the mean absolute error (MAE) calculated as the average of the absolute differences between corresponding voxel values ​​in the output image and the full dose image. In block 749, the training engine verifies whether the loss value is not acceptable and is still significantly improving. This operation can be performed either in an iterative mode (after processing each training set for its loss value) or in a batch mode (after processing all training sets for an accumulated value of the loss value, e.g., the average value of the loss value). If improving, in block 750, the trainer updates the weights of the operating neural network in an attempt to improve its performance. For example, a stochastic gradient descent (SGD) algorithm, such as one based on the ADAM method, is applied (the direction and amount of change are determined by the gradient of the loss function, and the loss value is given as a function of the weights, approximated with a backpropagation algorithm, according to a predefined learning rate). The process then returns to block 746 and repeats the same operations. Referring again to block 749, if the loss value becomes acceptable or if the changes in the weights do not provide a significant improvement (meaning a minimum, or at least a localized region, or flat region of the loss function has been found), the loop ends by descending to block 751. The above loop is repeated a number of times (epochs), e.g., 100-300, by adding random noise to the weights and / or starting from a different initialization of the operating neural network to find different (and possibly better) local minima and identify flat regions of the loss function.

[0099] Once a configuration of the operational neural network that provides the optimal minimum of the loss function is found, the training engine performs a validation of the performance of the operational neural network thus obtained. For this purpose, the training engine selects multiple validation sets from the sample sets in the corresponding repository (e.g., different from the training set). Then, in a loop, in block 752, the training engine feeds the zero-dose and low-dose images of the (current) validation set (starting from the first one in any order) to the operational neural network. In response, in block 753, the operational neural network outputs a corresponding output image, which must be equal to the full-dose image of the validation set. In block 754, the training engine calculates a loss value as described above based on the difference between the output image and the full-dose image. In block 755, the training engine verifies whether the last validation set has been processed. If not, the flow of activity returns to block 752 to repeat the same operations for the next validation set. Conversely (when all validation sets have been processed), the loop ends by descending to block 756. At this point, the training engine determines a global loss of the above-mentioned validation (e.g., equal to the average of the loss values ​​of all validation sets). The flow of activity branches in block 757 according to the global loss. If the global loss is (possibly strictly) higher than the tolerance value, this means that the generalization ability of the operational neural network (from its configuration learned from the training set to the validation set) is too low. In this case, the process returns to block 744 and repeats the same operation with a different training set and / or training parameters (e.g., learning rate, epochs, etc.). Conversely, if the global loss is (possibly strictly) lower than the tolerance value, this means that the generalization ability of the operational neural network is sufficient. In this case, in block 758, the training engine accepts the configuration of the operational neural network thus obtained and stores it in the corresponding repository in association with its value of the increase factor.

[0100] Referring back to block 742, if a training neural network is available to simulate low-dose images (used to train the operational neural network), in block 759 the training engine trains it by using the sample set. For example, the same operations as above may be performed with the difference that the training neural network is now optimized to generate low-dose images from the corresponding zero-dose and full-dose images. In this case, it is also possible to improve the performance of the training neural network using more complex loss functions, for example using techniques that utilize generative adversarial networks (GANs). The configuration of the training neural network thus obtained is then stored in the corresponding repository. At the same time, the analytically simulated low-dose images are deleted from the sample set repository in order to restore the corresponding incomplete sample set. Then, in block 760, a loop is entered to simulate refined versions of the low-dose images of the incomplete sample set (obtained from the sample set repository). For this purpose, the training neural network takes into account the (current) incomplete sample set (starting from the first one in any order). In block 761, the analysis engine provides the zero-dose and full-dose images of the incomplete sample set to the training neural network. Moving to block 762, the training neural network outputs the corresponding low-dose images, which are stored in the sample set repository. This completes the incomplete sample set again. In block 763, the analysis engine verifies whether the last incomplete sample set has been processed. If not, the flow of activity returns to block 760 to repeat the same operations for the next incomplete sample set. Conversely (when all incomplete sample sets have been processed), the corresponding loop ends by proceeding to block 743 to train the neural network operating as described above with the sample set (either partially completed or already completed and provided).This embodiment improves the accuracy of the low-dose images, which in turn improves the performance of the operational neural networks that are trained with them.

[0101] Referring again to block 758, the process proceeds to block 764, where the analysis engine verifies whether the configuration of the operational neural network is complete. If not, the process returns to block 702, where the same operation is repeated to configure the operational neural network for different gain coefficients. Conversely, once the configuration of the operational neural network is complete, the configuration of the operational neural network thus obtained is deployed in block 765 to a batch of instances of the control computer of the corresponding imaging system (e.g., by preloading them at the factory in case of the first delivery of the imaging system, or by uploading them over the network in case of the upgrade of the imaging system). The process then ends up to the concentric black / white stop circle 766.

[0102] 8A-8B, a representative example of experimental results for a solution according to one embodiment of the present disclosure is shown.

[0103] In particular, a dedicated preclinical study was carried out on rats bearing two brain lesions (both surgically induced): C6 glioma tumor (n=36 animals) and cerebral ischemic pathology (n=42 animals). All animals underwent a surgical procedure to induce the lesions. Animals that survived the surgical procedure and showed limited or no clinical signs during the following 2 weeks (i.e. the time window required for the pathogenesis of the lesions) were enrolled for an MRI type imaging procedure (i.e. typically 2 for each animal, 3 only in limited cases). The imaging procedure was carried out using a gadolinium-based contrast agent and a preclinical scanner spectrometer Pharmascan by Bruker Corporation™, operating at 7T and equipped with a rat head volume coil with 2 channels. The CE-MR protocol used during the acquisition was as follows: · Pre-contrast acquisition of standard T1-weighted sequences (zero-dose images); Intravenous administration of contrast agent at a low dose equal to 0.01 mmol Gd / kg; · Post-contrast acquisition of T1-weighted sequences (low-dose images); · Administer an additional intravenous dose of contrast at 0.04 mmol Gd / kg immediately after the previous contrast agent, for a total dose equal to 0.05 mmol Gd / kg; ·Post-contrast acquisition of T1-weighted sequences (full-dose images).

[0104] The study yielded 130 3D MRI volumes (i.e., 61 for rats with glioma and 69 for rats with ischemia) consisting of 24 slices each. The acquired volumes were used to construct two datasets: Acquisition data: Includes only acquired images (zero-dose, low-dose and full-dose images); · Simulated data: Includes (acquired) zero-dose and full-dose images and corresponding (simulated) low-dose images (with augmentation factor k=5).

[0105] Using the full-dose images as ground truth, a total of seven (operating) neural networks were trained by varying the mix of acquired / simulated data, i.e., by selecting all possible combinations of the following parameters: Learning rate = 0.01 Attenuation=0.001 Dataset = a mix of acquired and simulated data (noise level = 0.015 for simulated data only), i.e. 0%, 10%, 20%, 30%, 40%, 50% and 100% of acquired data and 100%, 90%, 80%, 70%, 60%, 50% and 0% of simulated data. ·Loss=MAE+ftMAE+VGG19 / 4 Relative weight of compound losses (a, b, c): a=b=c=1.

[0106] Referring now to FIG. 8A, representative examples of the original version of the full-dose images (FD) and the full-dose images as simulated by neural networks trained on different mixtures of acquired data (ACQ) and simulated data (SIM) are shown, with the same grayscale applied to all full-dose images. As can be seen, the neural network trained on 100% of the acquired data produced full-dose images (i.e., their acquired versions) that were very similar to the ground truth. It was observed that the performance of the neural network gradually and moderately deteriorated (blurring, artifacts) as the proportion of simulated data during training increased (especially when the proportion of simulated data was 60% or more). However, the addition of only 10% of the acquired data in the training set seemed sufficient to remove the major disappearance artifacts and improve the performance of the corresponding neural network.

[0107] Considering the above, the mixing of acquired / simulated images during training appeared to be a valid strategy to further improve the performance of neural networks. This consideration may be even more important when extended to datasets with lower homogeneity. Indeed, due to its inherent homogeneity (same scanner, magnetic field, coils, MRI sequences, etc.), the adopted preclinical dataset should not be optimal to gather the full potential of a mixed (acquired / simulated) training approach.

[0108] Further dedicated preclinical studies were performed on rats (n=48 animals) bearing C6 glioma tumors. All animals underwent a surgical procedure to induce lesions. Animals that survived the surgical procedure and showed limited or no clinical signs for the next 2 weeks (i.e. the time window required for pathogenesis of the lesions) were enrolled for an MRI type imaging procedure (i.e. typically 3 for each animal). The imaging procedure was performed using the commercial contrast agent ProHance by Bracco Imaging SpA™ and two preclinical scanner spectrometers, a Pharmascan by Bruker Corporation™ operating at 7T and equipped with a rat head volume coil with 2 channels, and a Biospec by Bruker Corporation™ operating at 3T and equipped with a rat head surface coil with 4 channels. The CE-MR protocol used during the acquisition was as follows: · Pre-contrast acquisition of standard T1-weighted sequences (zero-dose images); Intravenous administration of contrast agent at a total dose equal to 0.1 mmol Gd / kg; · Post-contrast acquisition of T1-weighted sequences (full-dose images); · Administer an additional intravenous dose of contrast at 0.1 mmol Gd / kg immediately following the previous dose for a total boost dose equal to 0.2 mmol Gd / kg; ·Post-contrast acquisition of T1-weighted sequences (boost-dose images).

[0109] The study yielded 122 3D MRI volumes.

[0110] A (working) neural network was trained on these (preclinical) data using the boosted dose images as ground truth with the following parameters: -Learning rate = 0.01 -Attenuation=0.001 -Loss=MAE+ftMAE+VGG19 / 4 -Relative weights of composite losses (a, b, c): a=5, b=c=1.

[0111] Another (operational) neural network was similarly trained on (clinical) data including (acquired) zero and full dose images and (simulated) boost dose images to optimize its ability to simulate boost dose images from corresponding zero and full dose images.

[0112] Once trained, the two neural networks were applied to the (clinical) zero-dose and full-dose images to predict the corresponding boosted-dose images with a boost factor k=2.

[0113] Now referring to FIG. 8B, representative examples of full-dose images and corresponding boosted-dose images that have been simulated (obtained) using neural networks trained on clinical data and neural networks trained on preclinical data are shown. As can be seen, both neural networks are successful in enhancing the contrast of the full-dose images. Surprisingly, despite the species difference (human vs. mouse), the neural networks trained on preclinical data have learned to identify locations corresponding to enhanced regions and increase such enhancement.

[0114] Considering the above, the use of preclinical data is an effective strategy for training neural networks to generate (clinical) boosted images.

[0115] Variations In order to meet local and specific requirements, those skilled in the art may apply many logical and / or physical modifications and changes to the present disclosure. More specifically, although the present disclosure has been described with a certain degree of particularity with reference to one or more embodiments thereof, it should be understood that various omissions, substitutions and changes in form and details as well as other embodiments are possible. In particular, different embodiments of the present disclosure may be practiced without specific details (such as numerical values) set forth in the preceding description to provide a more complete understanding thereof. Conversely, well-known matters may be omitted or simplified so as not to obscure the description of unnecessary matters. Furthermore, it is expressly intended that specific elements and / or method steps described in connection with any embodiment of the present disclosure may be incorporated into any other embodiment as a matter of general design choice. Furthermore, items presented in the same group and different embodiments, examples or alternatives should not be construed as being effectively equivalent to each other (they are separate and autonomous entities). In each case, each numerical value should be read as modified according to the applicable tolerances. In particular, unless otherwise indicated, terms such as "substantially," "about," "approximately," and the like, should be understood to be within 10%, preferably 5%, and even more preferably 1%. Moreover, each range of numerical values ​​should be intended to explicitly specify any possible number along a continuum within the range (including its endpoints). Order or other modifiers are merely used as labels to distinguish between elements having the same name, and do not themselves imply a priority, precedence, or order.The terms include, comprise, have, contain, involve, etc. are intended in an open, non-exhaustive sense (i.e., not limited to the listed items), the terms based, on, dependent on, according to, function of, etc. are intended as non-exclusive relationships (i.e., possible further variables are included), the term a / an is intended as one or more items (unless expressly indicated otherwise), and the term means for (or any means-plus-function expression) is intended as any structure adapted or configured to perform the relevant function.

[0116] For example, one embodiment provides a method for training a behavioral machine learning model. However, the behavioral machine learning model may be of any type (e.g., neural networks, generative models, genetic algorithms, etc.).

[0117] In one embodiment, the behavioral machine learning model is for use in a medical imaging application, however the medical imaging application may be of any type (see below).

[0118] In one embodiment, the method comprises the following steps under the control of a computing system: However, the computing system may be of any type (see below).

[0119] In one embodiment, the method includes providing (to a computing system) a plurality of sample sets. However, the sample sets may be any number and may be provided in any manner from any number and type of source (e.g., hospital, clinic, university, research lab, etc.) (e.g., downloaded via the internet, such as provided from a central server at a medical facility, retrieved automatically or manually from a corresponding imaging system, manually loaded via a corresponding LAN or via a removable storage device, manually loaded from a removable storage device copied from a central server or (separate) imaging system, etc.).

[0120] In one embodiment, the sample sets include corresponding sample baseline images, sample source images, and sample target images. However, the sample sets may be of any type (e.g., all sample sets to be completed, some sample sets already completed, sample sets constructed (completed or already completed) from image sequences, or sample sets already constructed and received, etc.). Furthermore, each sample set may include any number of sample images of any type (e.g., of any size, complexity, k-space, etc., of any dimension, size, resolution, chromaticity, bit depth, etc., related to any location of the body part, such as voxels, pixels, etc.) (e.g., sample baseline / source / target images only, one or more additional sample source images that have been acquired and / or simulated, such as corresponding to different doses of contrast agent and / or different acquisition conditions, etc.).

[0121] In one embodiment, the sample baseline / source / target images represent corresponding body parts of a subject. However, the body parts may be of any number, any type (e.g., organs, regions thereof, tissues, bones, joints, etc.) and any state (e.g., normal, pathological state with any lesion, etc.). Furthermore, the body parts may belong to any number and type of subjects (e.g., humans, animals, etc.).

[0122] In one embodiment, the sample baseline image is acquired from the corresponding body-part without contrast agent, however, the sample baseline image may be acquired in any manner (e.g., prior to administration of contrast agent, after any delay following possible pre-administration of contrast agent, etc.).

[0123] In one embodiment, the sample target image is acquired from a corresponding body part of a subject to which a contrast agent is administered in a sample target dose. However, the contrast agent may be of any type (e.g., any targeted contrast agent based on specific or non-specific interactions, any non-targeted contrast agent, etc.) and may be administered in any manner (e.g., intravenously, intramuscularly, orally, etc.). In any case, this is a (computer-implemented) data processing method that is performed independently of the acquisition of the sample image (without the need for interaction with the corresponding subject).

[0124] In one embodiment, the sample image corresponds to a sample dose of the contrast agent that is lower than the sample target dose (the ratio between the sample source dose and the sample target dose is equal to the reduction factor). However, the sample source dose and the sample target dose can have any value, either in absolute terms (e.g., the sample target dose is lower than, equal to, or higher than the full dose of the contrast agent, and the sample source dose is lower or higher than the sample target dose, etc.), and the sample image can correspond to the sample source dose in any way (e.g., fully simulated, partially acquired, etc.).

[0125] In one embodiment, the method includes training (by a computing system) a behavioral machine learning model to optimize its ability to generate sample target images for each of the sample sets from at least the sample baseline images and the sample source images of the sample sets. However, the behavioral machine learning model may be trained in any manner (e.g., by selecting any training / validation set from the sample sets using any algorithm such as stochastic gradient descent, real-time iterative learning, high-order gradient descent, extended Kalman filtering, etc., to optimize its ability to generate sample target images from one or more of the sample baseline images, the sample source images, and possibly further sample source images, by considering any loss function, such as based on mean absolute error, mean square error, perceptual loss, adversarial loss, etc., defined individually or at the level of the group of positions, any complementary information such as the state of the body part, the type of the subject, fixed gain coefficients, variable gain coefficients, etc., as parameters of the behavioral machine learning model).

[0126] In one embodiment, the step of providing a sample set includes receiving (by a computing system) one or more incomplete sample sets of the sample set, each of which lacks a sample source image, although there may be any number of incomplete sample sets (e.g., all of the sample set, only a portion thereof, etc.).

[0127] In one embodiment, the method includes completing each incomplete sample set (by a computing system) by simulating a sample source image from a sample baseline image and a sample target image of the sample set, the sample source image being simulated to represent a corresponding body part of a subject that mimics administration of a contrast agent in a sample source dose. However, the sample source image can be simulated in any manner (e.g., analytically using an analysis engine to generate a preliminary version of the sample source image, a training machine learning model trained on such a preliminary completed sample set, and a training machine learning model to generate a refined version of the sample source image, operating in any domain, such as size, complexity, k-space, etc., obtained independently from animal preclinical studies, etc., with or without pre-processing such as registration, normalization, noise removal, distortion correction, filtering of abnormal sample images, or with or without post-processing such as normalization, noise injection, etc., using a training machine learning model trained on a further sample set).

[0128] In one embodiment, the training method includes deploying (by the constituent computing systems) the operational machine learning models that have been trained for use in the medical imaging application, however, the operational machine learning models may be deployed in any manner (e.g., distributed with corresponding new imaging systems or brought online to upgrade already installed imaging systems, etc.) to any number and type of imaging systems.

[0129] In one embodiment, the motion machine learning model is used in a medical imaging application to mimic the increase in the dose of contrast agent administered to a corresponding patient according to an increase factor corresponding to the inverse of the decrease factor. However, the motion machine learning model may be used to mimic the increase in the dose of contrast agent in any manner (e.g., real-time, offline, locally, remotely, etc.). Furthermore, the increase factor may correspond to the inverse of the decrease factor in any manner (e.g., equal to, lower than, or higher than, e.g., according to a corresponding multiplication factor, etc.).

[0130] Further embodiments provide additional advantageous features that may be omitted in the basic implementation.

[0131] In one embodiment, in each sample set, the sample target images are acquired from the corresponding body parts of the subject to which the contrast agent is administered in a standard full dose. However, the full dose can be of any type (e.g., fixed for each type of medical imaging application, depending on the type of body part, depending on the type of subject, weight, age, etc.). In any case, the possibility of using a different sample target dose (e.g., lower or higher than the full dose) is not excluded.

[0132] In one embodiment, one or more of the sample sets are complete sample sets, however, there may be any number of complete sample sets (either absolutely or relative to incomplete sample sets), or none at all.

[0133] In one embodiment, each sample source image of the complete sample set is acquired from a corresponding body part of the subject to which contrast agent was administered in a sample source dose. However, the sample source images of the complete sample set may be acquired in any manner (e.g., either the same or a different manner with respect to the sample target image) from a corresponding body part of the subject to which contrast agent was administered in any manner.

[0134] In one embodiment, at least some of the subjects are animals, and the behavioral machine learning model is for use in a human medical imaging application. However, the animals can be any number and any type (e.g., rats, pigs, etc.). In any case, training of the behavioral machine learning model with a sample set obtained from the animal for use in a medical imaging application applied to humans is also possible, more generally, when all sample sets are received already completed.

[0135] In one embodiment, the subjects of the incomplete sample set are animals and the subjects of the complete sample set are humans, however, the sample sets may be collected from animals and humans in any manner (e.g., using each of them to collect a complete sample set or only a portion of it, collecting an incomplete sample set or only a portion of it, any combination thereof, etc.).

[0136] In one embodiment, the behavioral machine learning model is a behavioral neural network. However, the behavioral neural network can be of any type (e.g., an autoencoder, a multi-layer perceptron network, a recurrent network, etc., with any number of layers, connections between layers, receiving fields, strides, padding, activation functions, etc.).

[0137] In one embodiment, each of the sample baseline images, each of the sample source images, and each of the sample target images includes a plurality of sample baseline values, a plurality of sample source values, and a plurality of sample target values, respectively. However, the sample baseline / source / target values ​​may be of any number and type (e.g., voxel / pixel magnitude / complexity format, k-space format, etc.).

[0138] In one embodiment, the step of completing the incomplete sample set includes calculating (by a computing system) each of the sample source values ​​for each of the sample source images by applying a simulation equation as a function of a reduction factor, where the simulation equation may be any type (e.g., linear, quadratic, cubic, a function of corresponding sample baseline values ​​and / or sample dose values, etc.).

[0139] In one embodiment, the simulation equations are derived from a signal law describing the magnitude of the response signal of the body-part as a function of the local concentration of the contrast agent, although the signal law may be any (e.g., based on any external / internal acquisition parameters, etc.) and the simulation equations may be derived from the signal law in any manner (e.g., from any reduction of the signal law, to the actual signal law, etc.).

[0140] In one embodiment, the simulation equations are derived from a signal law that has been linearized with respect to the local concentration of the contrast agent, however, the signal law may be linearized in any way (e.g., using any series expansion, any approximation, etc.).

[0141] In one embodiment, the simulation equations are derived from the signal law by assuming a direct proportionality between the local concentration and the dose of the contrast agent. However, the simulation equations can be derived from the signal law by assuming any relationship (e.g., linear, nonlinear, etc.) between the local concentration and the dose of the contrast agent.

[0142] In one embodiment, said step of calculating each of the sample source values ​​comprises setting (by the computing system) the sample source value to the corresponding sample baseline value plus a reduction factor multiplied by the difference between the sample target dose and the sample source dose, although this does not exclude the use of other (linear / non-linear) simulation formulas.

[0143] In one embodiment, the sample baseline value, the sample source value and the sample target value represent a response signal at a corresponding location of the body part, however, the response signal may be expressed in any manner (e.g., in magnitude form, complex form, etc.).

[0144] In one embodiment, said step of calculating each of the sample source values ​​includes modulating (by the computing system) a reduction factor used to calculate the sample source value according to an indication of a local concentration of contrast agent at the corresponding location derived from the sample target image. However, the local concentration may be derived in any manner (e.g. set to the corresponding local contrast enhancement calculated from the sample target value according to a signal law, etc.) and the reduction factor may be modulated according to any linear / non-linear function thereof (so as to always remain the same).

[0145] In one embodiment, the step of modulating the reduction factor used to calculate the sample source value comprises linearly incrementing (by the computing system) the reduction factor according to the local contrast enhancement of the corresponding location depending on the difference between the corresponding sample target value and the sample baseline value. However, the reduction factor may be linearly incremented according to any modulation factor (e.g., empirically determined, calculated using average / local values ​​of any acquisition parameters, etc.).

[0146] In one embodiment, completing the incomplete sample set includes injecting (by a computing system) artificial noise into each sample source image of the incomplete sample set. However, the artificial noise may be of any type (e.g., fixed, according to a reduction factor, etc.) and may be injected into the sample source images in any manner (e.g., additive, multiplicative, convolutive, magnitude form into the sample source images, complex form, k-space form, everywhere, only where contrast agent is present, etc.) or not injected at all.

[0147] In one embodiment, the artificial noise has a statistical distribution depending on the reduction factor. However, the statistical distribution of the artificial noise may be of any type (e.g., normal, Rayleigh, Rician, etc.). Furthermore, the statistical distribution of the artificial noise may depend on the reduction factor in some way (e.g., for any linear / non-linear function of the reduction factor determined using a theoretical approach, the obtained results are heuristically corrected, etc.).

[0148] In one embodiment, injecting the artificial noise into the sample source image includes calculating (by a computing system) corresponding reference values ​​of one or more statistical parameters of the reference noise based on the noise of the corresponding sample baseline image and / or sample target image. However, the statistical parameters may be of any number and any type (e.g., standard deviation, variance, skewness, etc.) and their reference values ​​may be calculated in any manner (e.g., only from the noise of the sample baseline image, only from the noise of the sample target image, from an average of the corresponding values, only from the noise of the sample baseline image, and only from the noise of the sample target image, etc.).

[0149] In one embodiment, injecting said artificial noise into the sample source image comprises calculating (by the constituent computing system) corresponding artificial values ​​of statistical parameters of the artificial noise required to match the statistical distribution of the noise of the sample source image with the statistical distribution of the reference noise, however the artificial values ​​may be calculated according to any linear / non-linear function of the corresponding reference values.

[0150] In one embodiment, the step of injecting artificial noise into the sample source image comprises randomly generating (by a computing system) the artificial noise to have a statistical distribution with artificial values ​​of statistical parameters, however, the artificial noise may be generated in any manner (e.g., using any random or pseudorandom generator, etc.).

[0151] In one embodiment, the statistical parameter comprises a standard deviation, and said step of injecting artificial noise into the sample source image comprises setting (by a computing system) an artificial value of the standard deviation to a reference value of the standard deviation equal to the square root of two times the reduction factor multiplied by the difference between 1 and the reduction factor, however the possibility of using a different formula is not excluded.

[0152] In one embodiment, said step of calculating the artificial values ​​of the statistical parameters comprises correcting (by the computing system) the artificial values ​​of the statistical parameters according to the corresponding empirical corrections. However, the empirical corrections may be of any type (e.g., the same for all statistical parameters, different for each statistical parameter, etc.) and they may be used to correct the corresponding artificial values ​​in any way (e.g., by incrementing / decrementing them according to any linear / non-linear function, etc.).

[0153] In one embodiment, the step of injecting artificial noise into the sample source image comprises randomly generating (by a computing system) the artificial noise to have a normal statistical distribution with zero mean, although this does not exclude the possibility of artificial noise having different statistical distribution types and means.

[0154] In one embodiment, the step of injecting artificial noise into the sample source image comprises adding (by a computing system) artificial noise to the sample source image, however, the artificial noise may be injected in any manner (e.g., at the level of each cell, groups of cells of the sample source image in any form, etc.) in an additive form.

[0155] In one embodiment, the step of injecting artificial noise into the sample source image comprises randomly generating (by a computing system) the artificial noise to have a normal statistical distribution with a unitary mean, although this does not exclude the possibility of artificial noise having different statistical distribution types and means.

[0156] In one embodiment, said step of injecting artificial noise into the sample source image comprises multiplying (by a computing system) the sample source image with the artificial noise, although the artificial noise may be injected in any manner (e.g., at the level of each cell, across groups of cells of any form of sample source image, etc.) multiplexed.

[0157] In one embodiment, the step of injecting artificial noise into the sample source image includes convolving (by a computing system) the sample source image through the artificial noise, however, the artificial noise may be injected in a convolutional fashion in any manner (e.g., at the level of each cell, at groups of cells of the sample source image of any shape, circular or non-circular with any stride, padding, etc.).

[0158] In one embodiment, the training method includes denoising (by the constituent computing systems) each sample baseline image and sample target image of the incomplete sample set used to simulate the corresponding sample source image. However, the sample baseline / target images may be denoised in any manner (e.g., using an autoencoder, analysis techniques based on block matching, shrinking fields, wavelet transforms, smoothing filters, etc.) or not at all.

[0159] In one embodiment, the step of completing the incomplete sample set includes training (by a computing system) a training machine learning model to optimize its ability to generate each sample source image of the sample set from the corresponding sample baseline image and sample target image. However, the training machine learning model may be of any type and may be trained in any manner (e.g., either the same or different manner as for the operational machine learning model) by using any sample set (e.g., analytically all of the incomplete sample set after completion, analytically only the complete sample set, etc.).

[0160] In one embodiment, the step of completing the incomplete sample set includes generating (by a computing system) refined versions of each sample source image of the incomplete sample set by applying the sample baseline images and sample target images of the incomplete sample set to a training machine learning model that has been trained. However, the possibility of using the training machine learning model in a different way (e.g., refining the sample source images of the incomplete sample set and generating them directly, etc.) is not excluded.

[0161] In one embodiment, the training machine learning model is a training neural network, however, the training neural network may be of any type (e.g., either the same or different with respect to the operating neural network).

[0162] In one embodiment, the method includes repeating (by a computing system) the steps of completing an incomplete sample set and training a behavioral machine learning model for multiple values ​​of the reduction factor. However, the values ​​of the reduction factor may be any number and any type (e.g., uniformly distributed, decreasing to increasing values, having a variable pitch, etc.), and these steps may be repeated in any manner (e.g., consecutively, at different times, etc.).

[0163] In one embodiment, the method includes deploying (by a computing system) a behavioral machine learning model with corresponding configurations trained with values ​​of the decrease factor to select one or more corresponding values ​​of the increase factor in each of the medical imaging applications. However, the different configurations can be deployed in any manner (e.g., all together, added over time, etc.) and in any form (e.g., corresponding configurations of a single behavioral machine learning model, corresponding instances of a behavioral machine learning model, etc.) and can be used to select values ​​of the increase factor in any number and in any manner (e.g., discrete mode, continuous mode, the same or different ways with respect to the value of the decrease factor, etc.).

[0164] In one embodiment, the method includes extending (by the computing system) each of the sample sets by simulating one or more further sample source images from the sample baseline image and the sample target image of the sample set for a corresponding further sample source dose of contrast agent. However, the further sample source images may be any number, may correspond to any further sample source dose (or none at all), and they may be simulated in any manner (e.g., either the same or different manner with respect to the sample source images).

[0165] In one embodiment, the method includes using the motion machine learning model trained in each of the medical imaging applications to image corresponding further body parts of the patient. However, the body parts may be of any kind, in any state, and belong to any (e.g., the same or different with respect to the body parts of interest used to train the motion machine learning model) patient. Furthermore, the method can be used in any medical imaging application (e.g., diagnostic, therapeutic or surgical applications based on MRI, CT, fluoroscopy, fluorescence or ultrasound techniques, etc.). In any case, although the method may facilitate the doctor's work, it only provides intermediate results that help the doctor, and the medical activity is always strictly performed by the doctor himself.

[0166] In one embodiment, the method includes receiving (by a computing system) a motion baseline image and one or more motion administration images representing additional body parts of the patient. However, there may be any number of motion administration images, and the motion baseline / administration images may be of any type (e.g., either the same or different type with respect to the sample image). Furthermore, the motion baseline / administration images may be received in any manner (e.g., real-time, offline, locally, remotely, etc.).

[0167] In one embodiment, the operational dose image is acquired from a further body part of the patient to which the contrast agent is administered in an operational dose dose. However, the contrast agent may be administered to the patient in any manner, including non-invasive manners (e.g., orally to image the gastrointestinal tract, into the airways via a nebulizer, via topical spray application), in any case without substantial physical intervention in the patient (e.g., intramuscularly) that requires specialized medical expertise or involves health risks. Furthermore, the operational dose dose may have any value (e.g., the same or different from the sample source dose, either lower, equal or higher than the total dose of the contrast agent, etc.).

[0168] In one embodiment, the method includes simulating (by a computing system) corresponding motion simulated images from the motion baseline images and the motion administered images using a trained machine learning model. However, the motion simulated images can be simulated in any manner (e.g., in any domain of size, complexity, k-space, etc., operating in real time, offline, locally, remotely, etc.), starting from motion baseline images acquired from the body-part without contrast agent or with administration of contrast agent at a dose different from the motion administered dose and / or under different acquisition conditions.

[0169] In one embodiment, the motion simulation image represents a further body part of the patient simulating administration of contrast agent at a motion simulation dose higher than the motion administration dose (the ratio of motion simulation dose to motion administration dose corresponds to an increase factor). However, the motion simulation dose can have any value (e.g., the same or different from the sample target dose, either lower, equal or higher than the full dose of contrast agent, etc.), and the ratio between the motion simulation dose and the motion administration dose corresponds in some way to the increase factor (e.g., equal to, lower than or higher, e.g., according to a corresponding multiplication factor, etc.).

[0170] In one embodiment, the method includes outputting (by a computing system) a representation of the body part based on the motion simulation image. However, the representation of the body part may be of any type (e.g., a motion simulation image, a corresponding motion synthesis image, etc.) and may be output in any manner (e.g., displayed on any device such as a monitor, virtual reality glasses, or more generally, output in any manner such as printing, remote transmission, etc., in real time or offline).

[0171] In general, similar considerations apply if the same solution is implemented in an equivalent manner (by using more steps or similar steps having the same function of some of them, by removing some steps that are not essential, or by adding further optional steps). Furthermore, steps may be performed in different orders, simultaneously, or alternately (at least partially).

[0172] An embodiment provides a computer program, which, when executed on a computing system, is configured to cause the computing system to perform the above-mentioned method. An embodiment provides a computer program product including a computer-readable storage medium embodying a computer program, which is loadable into a working memory of the computing system and thereby configures the computing system to perform the above-mentioned method. However, the (computer) program may be executed on any computing system (see below). The program may be implemented as a stand-alone module, as a plug-in for an existing software program (e.g., a configuration application, an imaging application, etc.), or directly in the latter.

[0173] In general, similar considerations apply if the program is structured differently or if additional modules or functions are provided. Similarly, the memory structures may be of other types or may be replaced by equivalent entities (not necessarily consisting of physical storage media). The program may take any form suitable for use by a computing system, thereby configuring the computing system to perform the desired operations. In particular, the program may be in the form of external or resident software, firmware, or microcode (either object code or source code, e.g., to be compiled or interpreted). Also, the program may be provided on any computer-readable storage medium. The storage medium is any tangible medium (as opposed to the transitory signal itself) that can hold and store instructions for use by the computing system. For example, the storage medium may be of electronic, magnetic, optical, electromagnetic, infrared, or semiconductor type. Examples of such storage media are fixed disks (which may be preloaded with the program), removable disks, memory keys (e.g., USB type), etc. The program can be downloaded to the computing system from the storage medium or via a network (e.g., the Internet, a wide area network and / or a local area network including transmission cables, optical fibers, wireless connections, network devices). One or more network adapters in the computing system receive the program from the network and transfer the program for storage in one or more storage devices of the computing system. In any case, the solution according to an embodiment of the present disclosure is suitable to be implemented either in a hardware structure (e.g., by electronic circuits integrated in one or more chips of semiconductor material such as Field Programmable Gate Array (FPGA) or Application Specific Integrated Circuit (ASIC) type) or in a combination of software and appropriately programmed or otherwise configured hardware.

[0174] An embodiment provides a computing system comprising means configured to perform the steps of the above method. An embodiment provides a computing system comprising circuitry (i.e., any hardware suitably configured, for example, by software) for performing each step of the method. However, the computing system may be of any type (e.g., a constituent computing system only, a constituent computing system and a control computing system, a common computing system providing both functions, etc.) and any location (e.g., on-premise in the case of a server or virtual machine controlling multiple scanners, remote in the case of an implementation by a service provider providing corresponding services such as a cloud type or SOA type corresponding to multiple scanners, or locally in the case of separate control computers such as corresponding scanners, scanner control units, etc.).

[0175] In general, similar considerations apply when a computing system has a different structure or includes equivalent components or has other operational characteristics. In either case, each component may be separated into more elements, or two or more components may be combined into one element. Furthermore, each component may be replicated to support the execution of corresponding operations in parallel. Moreover, unless otherwise specified, the interaction between different components generally need not be sequential, but may be either direct or indirect through one or more intermediaries.

Claims

1. A method (700) for training a behavioral machine learning model (420) for use in medical imaging applications, comprising, under control of a computing system (130): providing to the computing system (130) a plurality of sample sets (703-743; 759-763) including corresponding sample baseline images, sample source images, and sample target images representing corresponding body parts of a subject, wherein the sample baseline images are acquired from the corresponding body parts of the subject that do not contain contrast agent, the sample target images are acquired from the corresponding body parts of the subject that have been administered contrast agent at a sample target dose, and the sample source images correspond to a sample dose of the contrast agent that is lower than the sample target dose, the ratio between the sample source dose and the sample target dose being equal to a reduction factor; training (744-758) the operational machine learning model (420) by the computing system (130) to optimize its ability to generate the sample target image of each of the sample set from at least the sample baseline image and the sample source image of the sample set; deploying (765) the operational machine learning model (420) trained for use in the medical imaging application to mimic an increase in the dose of the contrast agent administered to a corresponding patient according to an increase factor corresponding to the inverse of the decrease factor, by the computing system (130); Including, Providing the sample set (703-743; 759-763) receiving, by the computing system (130), one or more incomplete sample sets of the sample sets each lacking a sample source image; and completing (704-742; 759-763) each of the incomplete sample sets by simulating (704-742; 759-763) the sample source images from the sample baseline image and the sample target image of the sample set, the sample source images being simulated to represent the corresponding body part of the subject that mimics administration of the contrast agent at the sample source dose.

2. 10. The method of claim 1, wherein in each of the sample sets, the sample target images are acquired from the corresponding body-part of the subject to which the contrast agent was administered at a standard full dose.

3. 2. The method (700) of claim 1, wherein one or more of the sample sets are complete sample sets, and the sample source images of each of the complete sample sets are acquired from the corresponding body part of the subject to which the contrast agent was administered in the sample source dose.

4. 10. The method of claim 1, wherein at least a portion of the subject is an animal, and the behavioral machine learning model is for use in the medical imaging application on a human.

5. at least a portion of the subjects are animals, and the behavioral machine learning model (420) is for use in the medical imaging application on humans; 4. The method (700) of claim 3, wherein the subjects of the incomplete sample set are animals and the subjects of the complete sample set are humans.

6. The method of claim 1 , wherein the behavioral machine learning model is a behavioral neural network.

7. Each of the sample baseline images, each of the sample source images, and each of the sample target images includes a plurality of sample baseline values, a plurality of sample source values, and a plurality of sample target values, respectively, and completing the incomplete sample set (704-742; 759-763) includes:

2. The method of claim 1, further comprising: calculating, by the computing system, each of the sample source values ​​for each of the sample source images by applying a simulation equation according to the reduction factor, the simulation equation being derived from a signal law that describes a magnitude of a response signal of the body-part as a function of a local concentration of the contrast agent.

8. 8. The method (700) of claim 7, wherein the simulation equation is derived from the signal law linearized with respect to the local concentration of the contrast agent.

9. 8. The method (700) of claim 7, wherein the simulation equation is derived from the signal law by assuming a direct proportionality between the local concentration and the dose of the contrast agent.

10. Calculating each of the sample source values ​​(708-722) includes:

8. The method (700) of claim 7, comprising setting (713; 721) the sample source value to the corresponding sample baseline value plus the reduction factor multiplied by the difference between the sample target dose and the sample source dose by the computing system (130).

11. The sample baseline value, the sample source value, and the sample target value represent the response signal at corresponding locations of the body part, and calculating each of the sample source values ​​(700-722) includes:

8. The method (700) of claim 7, comprising adjusting (709; 712; 720) the reduction factor used to calculate the sample source value according to an indication of the local concentration of the contrast agent at the corresponding location derived from the sample target image, by the computing system (130).

12. Adjusting the reduction factor (709; 712; 720) used to calculate the sample source value may include:

12. The method (700) of claim 11, comprising linearly incrementing (712; 720) the reduction factor according to the local contrast enhancement of the corresponding location depending on the difference between the corresponding sample target value and the sample baseline value, by the computing system (130).

13. Completing the incomplete sample set (704-742; 759-763) 2. The method of claim 1, further comprising injecting, by the computing system, artificial noise into the sample source image of each of the incomplete sample sets, the artificial noise having a statistical distribution according to the reduction factor.

14. Injecting (706; 715-716; 723-724; 734-738) the artificial noise into the sample source image, calculating (706), by the computing system (130), corresponding reference values ​​of one or more statistical parameters of the reference noise based on the noise of the corresponding sample baseline image and / or sample target image; calculating (706) by the computing system (130) corresponding artificial values ​​of the statistical parameters of the artificial noise necessary to produce a statistical distribution of noise in the sample source image that matches the statistical distribution of the reference noise; randomly generating (715; 723; 734; 737) the artificial noise by the computing system (130) to have a statistical distribution using the artificial values ​​of the statistical parameters; 14. The method (700) of claim 13, comprising:

15. The statistical parameters include a standard deviation, and injecting (706; 715-716; 723-724; 734-738) the artificial noise into the sample source image comprises:

15. The method (700) of claim 14, comprising setting (706), by the computing system (130), the artificial value of the standard deviation to the reference value of the standard deviation equal to the square root of two times the reduction factor multiplied by the difference between 1 and the reduction factor.

16. Calculating (706) the artificial value of the statistical parameter comprises:

15. The method (700) of claim 14, comprising correcting (706), by the computing system (130), the artificial values ​​of the statistical parameters according to corresponding empirical corrections.

17. Injecting (706; 715-716; 723-724; 734-738) the artificial noise into the sample source image, Randomly generating (715; 737) the artificial noise by the computing system (130) so that the artificial noise has a normal statistical distribution with zero mean; adding (716; 738) the artificial noise to the sample source image by the computing system (130); 14. The method (700) of claim 13, comprising:

18. Injecting (706; 715-716; 723-724; 734-738) the artificial noise into the sample source image, Randomly generating (734) the artificial noise by the computing system (130) to have the statistical distribution of normal type with unitary mean; multiplying (735) the sample source image with the artificial noise by the computing system (130); 14. The method (700) of claim 13, comprising:

19. Injecting (706; 715-716; 723-724; 734-738) the artificial noise into the sample source image, Randomly generating (723) the artificial noise by the computing system (130) so that the artificial noise has the statistical distribution of normal type with unitary mean; convolving (724) the sample source image through the artificial noise by the computing system (130); 14. The method (700) of claim 13, comprising:

20. 10. The method of claim 1, further comprising: denoising, by the configuration computing system, the sample baseline image and the sample target image for each of the incomplete sample sets used to simulate the corresponding sample source image.

21. Completing the incomplete sample set (704-742; 759-763) training (759) a training machine learning model (540) by the computing system (130) to optimize its ability to generate the sample source image for each of the sample set from the corresponding sample baseline image and sample target image; generating, by the computing system (130), refined versions of the sample source images of each of the incomplete sample sets by applying the sample baseline images and the sample target images of the incomplete sample sets to the training machine learning model (540); and 10. The method of claim 1, comprising:

22. 22. The method (700) of claim 21, wherein the training machine learning model (540) is a training neural network (540).

23. Repeating (764) by the computing system (130) completing (704-742; 759-763) the incomplete sample set and training (744-758) the behavioral machine learning model (420) for multiple values ​​of the reduction factor; deploying (765) the operational machine learning model of a corresponding configuration trained with the values ​​of the decrease factor to select one or more corresponding values ​​of the increase factor in each of the medical imaging applications by the computing system (130); 10. The method of claim 1, comprising:

24. 10. The method (700) of claim 1, comprising: expanding, by the computing system (130), each of the sample sets by simulating one or more additional sample source images from the sample baseline image and the sample target image of the sample set for a corresponding additional sample source dose of the contrast agent (704-742; 759-763).

25. receiving, by the computing system (115), an operational baseline image and one or more operational administration images (609, 621) representing the further body part of the patient, the operational administration images being acquired from the further body part of the patient to which the imaging agent has been administered at an operational administration dose; simulating (624-630) by the computing system (115) corresponding motion simulation images from the motion baseline images and the motion administration images while the machine learning model (420) is trained, the motion simulated images representing the further body part of the patient simulating administration of the contrast agent at a motion simulation dose higher than the motion administration dose, with a ratio between the motion simulation dose and the motion administration dose corresponding to the increase factor; outputting, by the computing system (115), a representation of the body part based on the motion simulation image (633-636); 10. The method of claim 1, further comprising: using the machine learning model trained in each of the medical imaging applications to image a further body part of a corresponding patient by:

26. A computer program (500) configured, when executed on a computing system (130), to cause said computing system (130) to perform the method (700) of claim 1.

27. A computing system (130) comprising means (500) configured to perform the steps of the method (700) of claim 1.

Images