Training machine learning models for use in medical imaging applications based on combinations of incomplete sample sets and sample images simulated therefrom

Abstract

A solution for training a machine learning model (305) for use in medical imaging applications is presented. A corresponding method (400) includes providing (403-410) an incomplete sample set for each imaging process, each incomplete sample set including one (or more) sample target images and one (or more) sample baseline images for a sample target dose of contrast agent. One (411-445; 450-456) sample source images are simulated (411-445; 450-456) from each of the incomplete sample sets (or a portion thereof) to mimic contrast agent at a sample source dose that is lower than the sample target dose. One or more complete sample sets are generated (446; 457) for each imaging process by combining the incomplete sample sets with sample source images that have been simulated from other incomplete sample sets. The machine learning model (305) is then trained (458-473) using the complete set of samples. Further, a method of using a machine learning model in medical imaging applications is presented. A computer program (300) and a computer program product for implementing a method (400) are presented. Furthermore, a computing system (130) for executing the method (400) is presented. A corresponding medical method is also presented.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of machine learning. More specifically, the present disclosure relates to machine learning for medical imaging applications. BACKGROUND

[0002] The background of the present disclosure is introduced hereinafter by discussion of the techniques associated therewith. However, even if this discussion is made with regard to documents, acts, materials, etc., it does not imply, or in any way represent, that the discussed techniques are part of the prior art or represent common general knowledge related to the field of interest of the present disclosure.

[0003] Imaging techniques are common in medical applications, where a physician examines a body part of a patient by providing an image that visually represents it (typically, even if the body part cannot be seen directly, in an essentially non-invasive way). To this end, a contrast agent is usually administered to the patient undergoing the imaging procedure, in order to enhance the contrast of the (bio)target of interest (e.g., a lesion), making it more clearly visible in the image.

[0004] In this context, it has also been proposed to use a reduced dose of contrast agent; this reduced dose is lower than the full dose of contrast agent that is standard in clinical practice. To this end, during the imaging procedure, a zero-dose image of the body part is acquired before the contrast agent is administered, and one or more reduced-dose images of the body part are acquired after the reduced dose of contrast agent is administered to the patient. Then, by means of a deep learning network (DLN), a corresponding full-dose image of the body part is simulated from the zero-dose image and the corresponding reduced-dose image, thereby mimicking the administration of the full dose of contrast agent to the patient; the deep learning network recovers the contrast enhancement from the level of this enhancement in the reduced-dose image (which is insufficient due to the reduced dose of contrast agent) to the desired level that would be provided by the contrast agent under the full dose. The deep learning network is trained by using a set of samples, each of which comprises a zero-dose image, a reduced-dose image and a full-dose image of the same type of body part acquired before administration of the contrast agent to the corresponding patient, after administration of the reduced dose of contrast agent and after administration of the full dose of contrast agent (or two or more zero-dose images acquired under different acquisition conditions, or two or more reduced-dose images acquired with different reduced doses of contrast agent).

[0005] However, the number of sample sets of (zero-dose / reduced-dose / full-dose) images available for training deep learning networks is relatively low. This is mainly due to the fact that reduced-dose images are not typically acquired in standard clinical practice (as they would require corresponding modifications to the imaging procedure in order to administer contrast agent at reduced dose and full dose multiple times). The low number of sample sets reduces the quality of the training of deep learning networks. This has a negative impact on the robustness of deep learning networks, and in particular on their ability to predict full-dose images.

[0006] Simulations of reduced-dose images have also been proposed.

[0007] For example, WO-A-2021 / 061710 discloses the inclusion of an enhanced dataset obtained from simulations to the input data used for training deep learning networks for improving the quality of images acquired with reduced dose of contrast agent. To this end, low quality image data mimicking image data acquired with reduced dose of contrast agent can be generated using image data from a clinical database; this result is achieved by adding artefacts to the original image data.

[0008] The document “Johannes Haubold et al., Contrast agent dose reduction in computed tomography with deep learning using a conditional generative adversarial network, European Radiology (2021) 31:6087-6095” discloses instead simulations of images with reduced iodine-based contrast medium (ICM) to verify the possibility of virtual enhancement of ICM. To this end, dual-energy computed tomography (CT) images based on ICM are acquired. Isolated ICM images encoding the distribution of ICM are generated and used to create virtual non-contrast (VCN) images. Dual-energy CT images corresponding to reduced ICM (reduced by 50% or 80%) are simulated via proportional subtraction. By combining the reduced ICM images and the isolated ICM images with the VCN images, respectively, pairs of input images and target images are obtained. The pairs of input / target images are used to train a generative adversarial network (to be used for simulating ICM enhancement to verify it).

[0009] However, the simulation of full-dose images from zero-dose images and corresponding reduced-dose images by means of deep learning networks is still challenging. In particular, the results obtained can suffer from artefacts, texture inhomogeneity and / or signal intensity changes. SUMMARY

[0010] The summary of the disclosure is presented here in order to provide a basic understanding of the disclosure; however, the only purpose of this summary is to introduce some concepts of the disclosure in a simplified form as a prelude to the more detailed description of it that follows later, and it should not be interpreted as identifying key elements or delineating its scope.

[0011] Generally speaking, the disclosure is based on the idea of decorrelating images used to train a machine learning model.

[0012] In particular, one aspect provides a method for training a machine learning model for use in a medical imaging application. The method comprises providing, for each imaging procedure, an incomplete sample set comprising one (or more) sample target image(s) at a sample target dose of a contrast agent and one (or more) sample baseline image(s). From each incomplete sample set (or a portion thereof), one (or more) sample source image(s) is simulated to mimic the contrast agent at a sample source dose lower than the sample target dose. One or more complete sample sets are generated for each imaging procedure by combining the incomplete sample sets with sample source images that have been simulated from other incomplete sample sets. The machine learning model is then trained using the complete sample sets.

[0013] Another aspect provides a method of using a machine learning model in a medical imaging application.

[0014] Another aspect provides a computer program for implementing each method.

[0015] Another aspect provides a corresponding computer program product.

[0016] Another aspect provides a computing system for implementing each method.

[0017] Another aspect provides a corresponding medical method.

[0018] More specifically, one or more aspects of the disclosure are set out in the independent claims, and advantageous features are set out in the dependent claims, the wording of all claims being incorporated herein by reference (with any particular aspect providing any advantageous feature provided by the reference being adapted as necessary to apply to any other aspect). The features of the disclosure will be better understood by reading the following detailed description of the specific embodiments, described by way of non-limiting indication only, and made in connection with the attached drawings, in which: BRIEF DESCRIPTION OF DRAWINGS

[0019] The solution of the disclosure and its further features and advantages will be best understood by reading the following detailed description of the specific embodiments, described by way of non-limiting indication only, and made in connection with the attached drawings, in which, for simplicity, corresponding elements are denoted with the same or similar reference signs, and their explanation is not repeated, and the name of each entity is generally used to indicate its type and properties, such as values, contents and representations. In particular:

[0020] Figure 1 a schematic block diagram of an infrastructure that can be used to practice the solution according to embodiments of the disclosure is shown,

[0021] Figure 2 the general principles of the solution according to embodiments of the disclosure are shown,

[0022] Figure 3 the main software components that can be used to implement the solution according to embodiments of the disclosure are shown,

[0023] Figures 4A-4D an activity diagram describing the flow of activities related to the implementation of the solution according to embodiments of the disclosure is shown, and

[0024] Figures 5A-5C representative examples of experimental results related to the solution according to embodiments of the disclosure are shown. DETAILED DESCRIPTION

[0025] With specific reference Figure 1 a schematic block diagram of an infrastructure 100 that can be used to practice the solution according to embodiments of the disclosure is shown.

[0026] The infrastructure 100 comprises the following components.

[0027] One or more (medical) imaging systems 105 comprise a corresponding scanner 110 and a control computing system (or just control computer 115). Each scanner 110 is used to acquire images representing a body part of a subject during a corresponding (medical) imaging procedure, which is based on administering a contrast agent to it for enhancing the contrast of a corresponding (biological) target, such as a lesion. The subject can be an animal in preclinical studies, or a human (i.e., a human being) in clinical applications. For example, the scanner 110 is of the magnetic resonance imaging (MRI) type. In this case, not represented in the figure, the (MRI) scanner 110 has a gantry for receiving the subject; the gantry houses a superconducting magnet (for generating a very high static magnetic field), multiple sets of gradient coils (for adjusting the static magnetic field) for different axes, and RF coils (with a specific structure for applying magnetic pulses to a certain type of body part and for receiving a corresponding response signal). As an alternative, the scanner 110 is of the computed tomography (CT) type. In this case, also not represented in the figure, the (CT) scanner 110 has a gantry for receiving the subject; the gantry houses an X-ray generator, an X-ray detector, and a motor for rotating them around the body part of the subject. The corresponding control computer 115 (e.g., a personal computer (PC)) is used to control the operation of the scanner 110. To this end, the control computer 115 is coupled with the scanner 110. For example, in case the scanner 110 is of the MRI type, the control computer 115 is arranged outside a scanner room for shielding the scanner 110 and is coupled with it via a cable through a penetration panel, while in case the scanner 110 is of the CT type, the control computer 110 is arranged in close proximity to the scanner.

[0028] The imaging systems 105 are installed at one or more facilities, which can be a laboratory facility (e.g., a university) in case of preclinical type imaging procedures, or a health facility (e.g., a hospital) in case of clinical type imaging procedures. Each (laboratory / health) facility comprises one or more imaging systems 105. Moreover, the facility can comprise a central computing system (or just central server 120), which communicates with the control computers 115 of its imaging systems 105 through a network 125 (e.g., a local area network (LAN) of the facility). The central server 120 gathers from its control computers 115 through the network 125 information about imaging procedures at least partially executed by the imaging systems 105; in particular, for each imaging procedure, this includes a series of images representing the corresponding body part and additional information related to the imaging procedure, e.g., the identification of the subject, the outcome of the imaging procedure, the acquisition parameters of the imaging procedure, etc.

[0029] A configuration computing device 130 (or just one (or more) configuration computers 130) is used to configure the control computers 115 of the imaging system 105. The configuration computer 130 can communicate with the central server 120 of the facility through a network 135 (e.g. based on the Internet). The configuration computer 130 collects (anonymously) from the corresponding central server 120 through the network 135 information related to imaging processes that have been at least partially performed in the facility (alternatively, not shown in the figure, the same information can be collected from the corresponding control computers 115 through the network 135, or from the central server 120 and / or the control computers 115 via removable storage units such as USB type). The information thus collected by the configuration computer 130 is used to configure at least part of the control computers 115 of the imaging system 105, either in the field in the facility after the shipment of these computers, or in the factory before the shipment of these computers to the facility (as described in detail below).

[0030] Each of the computing machines implementing the control computers 115, the central server 120 and the configuration computer 130 comprises several units connected therein through a bus structure 140. In particular, one or more microprocessors (µP) 145 provide the logical capabilities of the computing machines 115, 120, 130. A non-volatile memory (ROM) 150 stores basic code for booting the computing machines 115, 120, 130, and a volatile memory (RAM) 155 is used by the microprocessor 145 as working memory. The computing machines 115, 120, 130 are provided with a mass storage 160, e.g. a solid state disk (SSD), for storing programs and data. Moreover, the computing machines 115, 120, 130 comprise a plurality of controllers 165 for peripheral devices or input / output (I / O) units. In particular, as far as relevant to the present disclosure, the peripheral devices include a keyboard, a mouse, a monitor, a network adapter (NIC) for connecting to the corresponding network 125, 135, a drive for reading / writing removable storage units such as USB keys, and, for each control computer 115, a trackball and corresponding drivers for the related units of its scanner 110.

[0031] Reference is now made to Figure 2 , which shows the general principles of the solution according to an embodiment of the present disclosure.

[0032] This involves the training of an (operational) machine learning model that is to be used in a medical imaging application to mimic an increase in the dose of contrast agent administered to a corresponding patient. To this end, (sample) images are collected from a plurality of (sample) imaging procedures relating to a corresponding (sample) body part of a (sample) subject. In particular, a plurality of incomplete sample sets is provided for each imaging procedure. Each incomplete sample set comprises one (or more) sample target image(s) and one (or more) sample baseline image(s). The sample target image(s) represent the corresponding body part of the subject to which contrast agent has been administered at a sample target dose; the sample baseline image(s) represent the corresponding body part under different conditions (e.g. without contrast agent). For each of the incomplete sample sets (or a portion thereof) of each imaging procedure, one (or more) sample source image(s) is (e.g. analytically) simulated (or synthesized) from the other sample images of the incomplete sample set (i.e. the sample baseline image(s) and the sample target image(s)). The sample source image(s) represent the corresponding body part of the corresponding subject using contrast agent mimicking an administration at a sample source dose. The sample source dose is lower than the sample target dose, wherein the ratio between the sample source dose and the sample target dose equals a reduction factor (e.g. 1 / 10) smaller than 1.

[0033] In the solution according to embodiments of the present disclosure, one or more complete sample sets are generated for each imaging procedure. Each complete sample set comprises the sample baseline image(s) and the sample target image(s) of the incomplete sample set of that imaging procedure and the sample source image(s) simulated from another (different) incomplete sample set of that imaging procedure.

[0034] The complete sample sets are then used to train the machine learning model to optimize its ability to predict the sample target image(s) (ground truth) of each complete sample set from the sample baseline image(s) and the sample source image(s) of the complete sample set (e.g. by using a portion of the complete sample sets to form a training set of the machine learning model and using another portion of the complete sample sets to form a validation set of the machine learning model).

[0035] The machine learning model thus trained can be used in a medical imaging application in several ways. In particular, during each (operative) imaging procedure, a corresponding scanner acquires (operative) images representing an (operative) body part of an (operative) patient under examination. The operative images comprise one (or more) operative baseline images and one or more operative administration images. The operative administration images are acquired from a patient who has been administered a contrast agent at an operative administration dose; the operative baseline images are acquired from the body part under different conditions (e.g., again without contrast agent). A control computer associated with the scanner uses the machine learning model to simulate (or synthesize) each one or more operative simulation images from the operative baseline images and the corresponding operative administration images. The operative simulation images represent the body part mimicking administration of the contrast agent to the patient at an operative simulation dose. The operative simulation dose is higher than the operative administration dose, the ratio between the operative simulation dose and the operative administration dose corresponding to (e.g., being equal to) the inverse of the downscaling factor of the machine learning model. Thus, the operative administration dose and the operative administration images are also referred to as operative low dose and operative low dose images, respectively, and the operative simulation dose and the operative simulation images are also referred to as operative high dose and operative high dose images, respectively. Then, the representations of the body part based on the operative simulation images are output (e.g., by displaying them) to the physician in charge of the imaging procedure.

[0036] The above solution significantly improves the training effect of the machine learning model, having a positive impact on its robustness. This is due to the fact that the combination of the proposed sample baseline / target images and the sample source images simulated from different sample baseline / target images removes (or at least greatly reduces) their correlation. Thus, the machine learning model is prevented from learning corresponding unwanted features (such as random fluctuations) that can be transferred into the operative simulation images simulated therefrom. This reduces artifacts, uneven textures, and / or signal intensity variations in the operative simulation images.

[0037] Thus, the machine learning model is able to predict the operative simulation images with a relatively high accuracy. This has a beneficial impact on the quality of the medical imaging application, e.g., greatly reducing the risk of false positives / negatives and errors in follow-up in diagnostic applications, of reduced effectiveness of therapy or damage to healthy tissue in therapeutic applications, and of incomplete removal of a lesion or excessive removal of healthy tissue in surgical applications.

[0038] The above solution can be directly applied to sample images acquired in preclinical imaging procedures. Moreover, it can also be extended to sample images acquired in (prospective) clinical imaging procedures by acquiring only one or more further images, without affecting the standard of care (i.e., the dose and injection protocol of the contrast agent).

[0039] In particular, the sample baseline image can have been acquired from a subject who has never been administered contrast agent or who has been administered contrast agent a relatively long time ago so as to ensure that the contrast agent has been substantially cleared, and thus the corresponding body part is free of contrast agent (or at least a significant amount of contrast agent); in this case, the sample baseline image is referred to as a sample zero-dose image. The sample target dose of contrast agent can be equal to the standard value in clinical practice (the sample target image is acquired at any time after administration of contrast agent, the same as or different from the standard value in clinical practice); in this case, the sample target dose and the sample target image are referred to as a full dose and a sample full-dose image, respectively. Thus, the sample source dose of contrast agent is reduced relative to the standard value in clinical practice; in this case, the sample source dose and the sample source image are referred to as a reduced dose and a sample reduced-dose image, respectively.

[0040] For simplicity, reference will be made in the following to this implementation of sample baseline image, sample target dose, sample target dose image, sample source dose, and sample source dose image (however, with the same considerations applicable to any other implementation thereof).

[0041] For example, Figure 2 A simple scenario is shown in which two incomplete sample sets 205a and 205b are provided for the same imaging procedure: the incomplete sample set 205a includes a sample zero-dose image 210a and a sample full-dose image 215a, while the incomplete sample set 205b includes a sample zero-dose image 210b and a sample full-dose image 215b. A sample reduced-dose image 220a is simulated from the incomplete sample set 205a (sample zero-dose image 210a and sample full-dose image 215a), while a sample reduced-dose image 220b is simulated from the incomplete sample set 205b (sample zero-dose image 210b and sample full-dose image 215b). A complete sample set 225a is generated to include the sample zero-dose image 210a and the sample full-dose image 215a (of the incomplete sample set 205a) and add the sample reduced-dose image 220b (simulated from the incomplete sample set 205b), while a complete sample set 225b is generated to include the sample zero-dose image 210b and the sample full-dose image 215b (of the incomplete sample set 205b) and add the sample reduced-dose image 220a (simulated from the incomplete sample set 205a).

[0042] Instead of referring to the use of a machine learning model in medical imaging applications, the operational baseline image as above can have been acquired from a patient who has never been administered contrast agent or a relatively long time has passed since the last administration of contrast agent so as to ensure that the contrast agent has been substantially cleared, thus the body part is free of contrast agent (or at least not a significant amount of contrast agent); in this case, the operational baseline image is referred to as an operational zero-dose image.

[0043] The operational administered dose of contrast agent and the operational simulated dose can be equal to its reduced dose and full dose, respectively; in this case, the operational administered image and the operational simulated image are referred to as operational reduced-dose image and operational full-dose image, respectively. Thus, it is possible to restore the contrast enhancement that would have been obtained if the contrast agent had been administered normally at full dose (especially advantageous when this can pose a risk to the patient) if not, even the contrast enhancement can be enhanced by reducing the motion / aliasing artifacts that can be caused by actually administering the contrast agent at full dose.

[0044] The operational administered dose of contrast agent can also be equal to its full dose, thus the operational simulated dose is boosted (or increased / amplified) with respect to the full dose; in this case, the operational administered image is referred to as operational full-dose image, while the operational simulated dose and the operational simulated image are referred to as operational boosted dose and operational boosted-dose image, respectively. Thus, it is possible to increment the contrast enhancement as if the operational boosted-dose image was obtained by administering the contrast agent at a higher (virtual) dose than what is achievable in current clinical practice (especially advantageous when the contrast enhancement is too low) while not posing any impact on the standard of care. As a further improvement, the operational combined image can each be generated from the operational baseline image, the operational full-dose image and the corresponding operational boosted-dose image by applying high dynamic range (HDR) techniques (e.g., by putting more emphasis on the contribution of the boosted-dose image to it). This makes the object more salient while maintaining its good association with the morphology of the body part.

[0045] It is also possible to provide multiple versions of the machine learning model that have been trained with different values of the downscaling factor. In this case, in medical imaging applications, the value of the upscaling factor can be selected from among the corresponding values. This further improves flexibility, allowing the physician to verify in real time the effect of different values of the upscaling factor and then select the one that provides the best contrast enhancement.

[0046] Reference is now made to Figure 3 , which shows the main software components that can be used to implement the solution according to embodiments of the present disclosure.

[0047] All software components (programs and data) are represented as a whole with reference 300. The software components 300 are usually stored in mass storage and, at program runtime, at least partially loaded into the working memory of the configuration computer 130 together with the operating system and other applications not directly related to the solution of the present disclosure (thus omitted from the figure for simplicity). The programs are initially installed into the mass storage, e.g. from a removable storage unit or from a network. In this respect, each program can be a module, a fragment or a portion comprising code of one or more executable instructions for implementing a specified logical function.

[0048] The configuration computer 130 stores a copy of a machine learning model to be trained. For example, the machine learning model is implemented by an (operative) neural network 305 (to which reference will be made hereinafter, with the same considerations applying to any other implementation thereof).

[0049] Basically, machine learning is used to perform a specific task (in this case, simulating an operational simulation image) without using explicit instructions, but rather automatically inferring how to perform it from examples (by exploiting a corresponding model learned from the examples). In embodiments of the present disclosure, a deep learning technique is applied, which is a branch of machine learning based on deep neural networks. Basically, a neural network is a data processing system inspired by the workings of the human brain. A neural network comprises basic processing elements (neurons) that perform operations based on corresponding weights; the neurons are connected via unidirectional channels (synapses) that convey data between the neurons. The neurons are organized in layers, performing different operations, always comprising an input layer and an output layer, for receiving input data of the neural network and providing output data, respectively. In particular, a deep neural network has one or more hidden layers arranged successively between the input layer and the output layer along a processing direction of the deep neural network. In embodiments of the present disclosure, the neural network is a convolutional neural network (CNN), i.e. a particular type of deep neural network, in which one or more of its hidden layers perform a (cross-) convolution operation. In particular, the neural network is an autoencoder (encoder-decoder) convolutional neural network, which comprises an encoder that compresses data in a more dense form (in a so-called latent space), the data thus compressed being used to perform the desired operation, and a decoder that expands the result thus obtained to the required more expanded form. In more detail, the input layer is configured to receive two input images (i.e. a sample baseline image and a sample source image during training, and an operational baseline image and an operational application image in a medical imaging application). The encoder comprises 3 groups of 3 convolutional layers, which groups are followed by a corresponding max-pooling layer, and the decoder comprises 3 groups of 3 convolutional layers, which groups are followed by a corresponding up-sampling layer. Each convolutional layer performs a convolution operation by a convolution matrix (filter or kernel) defined by corresponding weights, which is performed successively on a finite portion (receptive field) of the applied data by shifting the filter across the applied data by a selected number of cells (stride), where cells with zero content (padding) can be added around the borders of the applied data to allow the filter to be applied therewith as well. Batch normalization (fixing the mean and variance of the corresponding data) is then applied, followed by an activation function (introducing a non-linearity factor). For example, each convolutional layer applies a 3x3 filter with padding 1 and stride 1, where its each neuron applies a rectified linear unit (ReLU) activation function. Each max-pooling layer is a pooling layer (down-sampling the data applied thereto) that replaces the values of each finite portion (window) of the applied data by a single value (in this case the maximum value) by shifting a window across the applied data by a selected number of cells (stride). For example, each max-pooling layer has a 2x2 window with stride 1.Each up-sampling layer is a de-pooling layer (inverse operation of pooling) that expands around each value in its region (window), such as using max de-pooling technique (where the value is placed in the same position as the maximum used for down-sampling and surrounded by zeros). For example, each up-sampling layer has a window of 2x2. Bypass connections are added between symmetric layers of the encoder and decoder (to avoid resolution loss), and skip connections are added within each set of convolutional layers and from the input layer to the output layer (to focus on differences between input images). Then, the output layer generates an output image corresponding to an input image (i.e., a sample target image during training and an operational simulation image in a medical imaging application), for example, by adding the obtained result (representing contrast enhancement between input images according to requirements) to the sample / operational baseline image.

[0050] The operational neural network 305 reads an operational configuration repository 310 that defines one or more (operational) configurations of the operational neural network 310. For example, the operational configuration repository 310 has an entry for each configuration of the operational neural network 305. The entry stores the configuration of the operational neural network 305 (defined by its weights) and the magnification factor provided by it when operated according to this configuration.

[0051] The sample image repository 315 contains information related to sample images to be used for training the operative neural network 305. For example, the sample image repository 315 has an entry for each (sample) imaging procedure. The entry stores a plurality of sample images (as described below) acquired or generated for the imaging procedure. Each sample image is defined by a corresponding bitmap, i.e. a matrix of cells (e.g. 512 rows by 512 columns), each cell containing a value of a basic picture element representing a corresponding position of a body part in the imaging procedure (e.g. a voxel of a basic volume in case of a 3D sample image); each voxel value (e.g. in gray scale) defines a brightness of the voxel, which is a function of the (signal) intensity of the response signal associated with the position. For example, in case of an MRI scanner, the response signal represents a response of the position to a magnetic field applied to the position, and in case of a CT scanner, the response signal represents an attenuation of X-ray radiation applied to the position. Furthermore, the entry stores one or more acquisition parameters related to the acquisition of its sample images; in particular, the acquisition parameters include one or more external parameters related to the settings of the scanner used for acquiring the sample images, and one or more internal parameters related to the corresponding body part (e.g. mean values of the main tissues of the body part). The collector 320 collects the (acquired) sample images from a central server or control computer (not shown in the figure) of the laboratory / health facility. At least part of the sample images of the imaging procedures are incomplete, as they only include one or more sample zero-dose images and one or more sample full-dose images of the corresponding body part. Optionally, the sample images of one or more of the imaging procedures are complete, as they also include one or more sample reduced-dose images of the corresponding body part. The collector 320 writes to the sample image repository 315. The preprocessor 325 optionally pre-processes the sample images of each imaging procedure (e.g. by co-registering, de-noising, etc.); furthermore, the preprocessor can also reduce and / or augment the sample images by generating one or more (generated) sample images (as described below). The preprocessor 325 reads / writes from / to the sample image repository 320.

[0052] The creator 330 creates, for each imaging procedure, a plurality of incomplete sample sets (each incomplete sample set being formed of at least one sample zero-dose image and at least one sample full-dose image), so as to provide, by combining them, an incomplete sample image (or at least part thereof). The creator 330 reads the sample image repository 320 and writes an incomplete sample set repository 335 containing information related to the incomplete sample sets. For example, the incomplete sample set repository 335 has an entry for each imaging procedure, providing an incomplete sample image. This entry stores the corresponding incomplete sample set, each incomplete sample set comprising a bitmap of its sample zero-dose image(s) and sample full-dose image(s). Moreover, this entry indicates the acquisition parameters of the corresponding imaging procedure (e.g. by a pointer to them in the sample image repository 315).

[0053] The analysis engine 340 simulates (or synthesizes) at least one sample reduced-dose image from the zero-dose / full-dose images of each incomplete sample set (or part thereof). The analysis engine 340 exposes a user interface for interacting therewith. The analysis engine 340 reads the incomplete sample set repository 335 and a simulation formula repository 345, which stores one or more simulation formulas to be used for simulating the sample reduced-dose image.

[0054] For example, in the case of an MRI scanner, when the spin echo is chosen 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 position of the body part during the spin relaxation of the protons of the water molecules present, for returning to its equilibrium condition after the magnetic pulse applied by the RF coil) is expressed by the following signal law:

[0055]

[0056] where M is the signal intensity, M0 is a parameter depending on the density of the protons, the size of the voxel, the magnetic pulse and the strength of the magnetic field, TE is the echo time (between the application of the magnetic pulse and the reception of the echo signal), T2 is the transverse relaxation time of the protons, TR is the repetition time (between the successive sequences of magnetic pulses) and T1 is the longitudinal relaxation time of the protons. In the sample zero-dose image (without contrast agent), the parameters T1 and T2 can be replaced respectively by the corresponding diamagnetic values, generally denoted T 10 and T 20 , so that the signal intensity (distinguished as M zero ) becomes:

[0057]

[0058] Conversely, when the contrast agent is present in that location, the parameters T1 and T2 depend on the corresponding diamagnetic value plus the corresponding paramagnetic value given by the contrast agent, so the signal intensity (distinguished as M agent ) becomes:

[0059]

[0060] where r2 is the transverse relaxation rate of the contrast agent, c is the local concentration of the contrast agent in that location and r1 is the longitudinal relaxation rate of the contrast agent. Linearizing this formula as a first order approximation of its Taylor series as a function of the local concentration of the contrast agent, one obtains (except for negligible errors depending on the square of the local concentration of the contrast agent):

[0061]

[0062] where

[0063] Therefore, in the sample full dose image, the signal intensity (distinguished as M full ) becomes:

[0064]

[0065] where c full is the local concentration of the contrast agent when administered at full dose. Likewise, in the sample reduced dose image, the signal intensity (distinguished as M reduced ) becomes:

[0066]

[0067] where c reduced is the local concentration of the contrast agent when administered at reduced dose. The local concentration of the contrast agent is essentially linear with the amount of contrast agent administered, so:

[0068]

[0069] where d is the reduction factor (corresponding to the magnification factor k = 1 / d).

[0070] In view of the above, the simulation formula is:

[0071]

[0072] Replacing the local concentration c full obtained from the definition of the corresponding intensity signal (M ) at full dose, the simulation formula becomes:

[0073]

[0074] Thus

[0075]

[0076] The same formula is obtained in other operating modes of the MRI scanner, such as gradient echo, MP-RAGE, etc.

[0077] Also, in the case of a CT scanner, the signal intensity defining each voxel value of the sample image (given by the X-ray radiation remaining after passing through the corresponding location due to attenuation) is expressed by the following signal law:

[0078]

[0079] where I is the signal intensity, I0is the initial intensity of the X-ray radiation, µ is the linear attenuation coefficient, p is the density and x is the thickness of the location. In the sample zero-dose image (without contrast agent), the parameters µ, p and x are parameters related to the corresponding material of the body part, denoted by µ M , p M and x M , respectively, so the signal intensity (distinguished as I zero ) becomes:

[0080]

[0081] Conversely, when the contrast agent is present in the location, it causes an additional attenuation of the X-ray radiation, so the signal intensity (distinguished as I agent ) becomes:

[0082]

[0083] where µ A is the linear attenuation coefficient, p A is the local concentration of the contrast agent and x A is the thickness of the location with this local concentration of the contrast agent. Linearizing this formula as a first order approximation of its Taylor series as a function of the local concentration of the contrast agent, it is obtained (except for negligible errors depending on the square of the local concentration of the contrast agent):

[0084]

[0085] Thus, in the sample full-dose image, the signal intensity (distinguished as I full ) becomes:

[0086]

[0087] where p fullis the local concentration of the contrast agent when administered at full dose. Likewise, in the sample reduced dose image, the signal intensity (denoted as I reduced ) becomes:

[0088]

[0089] where p reduced is the local concentration of the contrast agent when administered at reduced dose. The density of the contrast agent is essentially linear with the amount of contrast agent administered, so:

[0090]

[0091] (d is again the zoom-out factor corresponding to the zoom-in factor k = 1 / d).

[0092] In view of the above, the simulation formula is:

[0093]

[0094] Replacing the local concentration p full at full dose obtained from the definition of the corresponding intensity signal , the simulation formula becomes:

[0095]

[0096] Thus, as above

[0097]

[0098] The proposed implementation (where the simulation formula is derived from a signal law linearized with respect to the local concentration of the contrast agent) is computationally very simple, where the loss of accuracy of the sample reduced dose image thus obtained (due to the linearization of the signal law) is acceptable for the purpose of training the operational neural network.

[0099] As an alternative, the signal law is approximated as a higher order (second, third, etc.) of its Taylor series as a function of the local concentration. In this case, the solution of the equation obtained for the local concentration of the contrast agent at full dose provides a corresponding number of values that need to be evaluated to discard any of them that do not have physical meaning. This improves the accuracy of the simulated sample reduced dose image (the higher the order of the approximation, the higher the accuracy). As another alternative, the signal law is numerically solved to obtain the local concentration of the contrast agent at full dose (again, the possible solutions need to be evaluated to discard any of them that do not have physical meaning). This further improves the accuracy of the simulated sample reduced dose image.

[0100] The analysis engine 340 writes the sample reduced dose images in a sample reduced dose images repository 350, which contains bitmaps of sample reduced dose images that have been simulated, each associated (e.g. by a pointer, same location, etc.) with the incomplete sample set in the repository 335 used to simulate it. Optionally, a noise corrector 355 corrects the noise of the sample reduced dose images. The noise corrector 355 reads / writes the sample reduced dose images repository 350. Indeed, the sample zero dose image and the sample full dose image of each incomplete sample set contain noise, which is propagated to the corresponding sample reduced dose image according to the simulation formula. However, the noise thus obtained (simulation noise) has a statistical distribution that slightly differs from the statistical distribution of the noise (real noise) that would be obtained by actually acquiring a sample reduced dose image from the corresponding subject to which the reduced dose of contrast agent has been administered. In particular, the noise of the sample zero dose image and the noise of the sample full dose image can be considered as having a normal statistical distribution with zero mean and corresponding standard deviation that is propagated to the sample reduced dose image according to the error (or uncertainty) propagation rules:

[0101]

[0102] where σ reduced is the standard deviation of the noise of the sample reduced dose image, σ full is the standard deviation of the noise of the sample full dose image and σ zero is the standard deviation of the noise of the sample zero dose image (d again being the demagnification factor corresponding to the magnification factor k = 1 / d). Assuming that the noise of the sample zero dose image and the noise of the sample full dose image have the same standard deviation σ = σ zero = σ full , in order to make the noise of the sample reduced dose image have the same normal statistical distribution with zero mean and standard deviation σ, an artificial noise should be injected into the sample reduced dose image having a normal statistical distribution with zero mean and standard deviation σ artificial such that:

[0103]

[0104] (in case of artificial noise injected in additive form); the standard deviation of the artificial noise is then given by the following noise formula:

[0105]

[0106] However, the inventors have found that by incrementing the standard deviation σ artificialthe (theoretical) value of the standard deviation σ

[0107] In case of injecting artificial noise in multiplication / convolution form, the inventors have found that the standard deviation σ artificial may be set in accordance with the value obtained above for injecting artificial noise in addition form, e.g. by multiplying it with an (empirical) conversion factor equal to 0.05-2.00, preferably 0.1-1.0, more preferably 0.3-0.7, such as 0.5.

[0108] For each imaging procedure (or at least part thereof) in the imaging procedure, the combiner 360 generates one or more complete sample sets (each complete sample set being formed by at least one sample zero dose image, a sample full dose image and a sample reduced dose image). In particular, when the imaging procedure provides an incomplete sample set, the combiner 360 generates a complete sample set by combining the incomplete sample set of the imaging procedure and a reduced dose sample image (simulated from other incomplete sample sets). When the imaging procedure provides a complete sample set, in contrast, the combiner 360 directly generates a complete sample set by combining the sample images of the imaging procedure. To this end, the combiner 360 reads the incomplete sample set repository 335 and the reduced dose image repository 350 (for the imaging procedure providing the incomplete sample images) and the combiner reads the sample image repository 315 (for the imaging procedure providing the complete sample images); moreover, the combiner 360 writes the complete sample set repository 365 containing information related to the complete sample sets. For example, the complete sample set repository 365 has an entry for each imaging procedure. The entry stores the corresponding complete sample set, each complete sample set comprising a bitmap of its sample zero dose image(s), sample full dose image(s) and sample reduced dose image(s).

[0109] Additionally or alternatively, the sample reduced dose images of the incomplete sample sets are simulated (or synthesized) by an additional (training) machine learning model; for example, the training machine learning model is implemented by a training neural network 370, in particular an autoencoder convolutional neural network as above (hereinafter reference will be made to this neural network, with the same considerations applicable to any other embodiment thereof). The training neural network 370 is controlled by the analysis engine 340. The training neural network 370 reads the training configuration repository 375, which stores the (training) configuration of the training neural network 370 (i.e. its weights as above). The training neural network 370 reads the incomplete sample set repository 335 and writes the sample reduced dose image repository 350.

[0110] The training engine 380 trains the operational neural network 305 and the training neural network 370 (when available). The training engine 380 reads the full sample set repository 365. The training engine 380 writes to the operational configuration repository 310 (of the operational neural network 305); and, when the training neural network 370 is available, the training engine 380 also writes to the training configuration repository 375, the sample reduced dose image repository 350, and the full sample set repository 365 (of the training neural network 370).

[0111] Reference is now made to Figures 4A-4D , showing an activity diagram depicting the flow of activities related to the implementation of the solution according to embodiments of the present disclosure.

[0112] In particular, this activity diagram represents an exemplary process that can be used to train the operational neural network with the method 400. In this regard, each block can correspond to one or more executable instructions for implementing the specified logical function on a configuration computer.

[0113] Whenever there is a need to train the operational neural network, the process starts at the black start circle 401. In particular, this happens prior to the first delivery of the operational neural network; furthermore, this can also happen periodically, in response to any significant change in the operational conditions of the imaging system (e.g., delivery of a new model of corresponding scanner, change 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 or improve over time the required performance of the imaging system, etc. In response thereto, at block 402, the analysis engine prompts the operator (via its user interface) to enter an indication of the desired magnification factor for which the operational neural network has to be trained, also defining the corresponding demagnification factor, e.g., as the inverse of the magnification factor (unless predefined as a single fixed value).

[0114] At block 403, the collector collects the (incomplete and possibly complete) sample images that have been acquired during a plurality of (sample) imaging procedures (these sample images are saved in a corresponding repository). The imaging procedures have been performed for body parts of the same type as the body parts for which the operational neural network is intended to be used. The imaging procedures can be of a clinical type, or of a preclinical type. Indeed, the inventors have surprisingly found that an operational neural network trained with sample images (at least partially) derived from animals still provides a good performance when applied to humans. Thus, the information required to train the operational neural network can be provided in a relatively simple manner. At this stage, the pre-processor can also pre-process the sample images (in the corresponding repository). For example, the pre-processor co-registers the sample images so as to make them spatially correspondent (e.g. by applying a rigid transformation). Additionally or alternatively, the pre-processor denoises the sample images to reduce their noise. To this end, it is possible to use an autoencoder (convolutional neural network). This autoencoder has been trained in an unsupervised manner with a plurality of sample images (such as all the sample images that are being collected); in particular, the autoencoder has been trained to optimize its such ability to encode each sample image, disregarding the insignificant data therein (due to noise), and then to decode the obtained result, thereby reconstructing the same sample image with reduced noise.

[0115] The sample images can also include the corresponding raw data used to generate these images. For example, in the case of an MRI scanner, it acquires the raw data in the form of K-space images. Each K-space image is defined by a matrix of cells, with the horizontal axis corresponding to the spatial frequency or wave number k (number of cycles per unit distance), and the vertical axis corresponding to the phase of the response signal being detected; each cell contains a complex number defining a different amplitude component of the corresponding response signal. By applying an inverse Fourier transform to the K-space image, it is converted into a corresponding (complex) image in the form of complex numbers. The complex image is defined by a matrix of cells of corresponding voxels; each cell contains a complex number representing the response signal being received from the corresponding location. Finally, the complex image is converted into a sample image in the form of magnitudes by setting each voxel value of the complex image to the modulus of the corresponding complex number in the complex image.

[0116] The activity flow then branches at block 404 depending on whether there is a need to reduce the number of sample images. For example, this can be performed to ensure a minimum quality level of the sample images, select a single sample zero dose image, so that the number of sample zero dose images is equal to the number of sample full dose images (and possibly sample reduced dose images), such as by reducing sample zero dose images relative to sample full / reduced dose images, and so on. If so, then at block 406 the analysis engine discards one or more of the sample images (by deleting them from the corresponding repository). For example, a quality indicator is computed for each sample image (such as a signal-to-noise ratio for that sample image); then, it is possible to discard all sample images whose quality indicator is (possibly strictly) below an acceptable value, all sample zero dose images except the sample image with the highest quality indicator, a number of zero dose sample images that exceeds the number of full / reduced dose sample images with lower quality indicators, and so on. Processing enters from block 406 into block 407, or directly from block 405 into block 407 if there is no need to reduce the number of sample images.

[0117] At this point, the activity flow branches depending on whether there is a need to increase the number of sample images. For example, this can be performed to ensure a minimum number of sample images, have a number of sample full dose images (such as in a clinical imaging procedure where only a single sample full dose image can be available), so that the number of sample zero dose images is equal to the number of sample full dose images (and possibly reduced dose sample images), such as by increasing sample full / reduced dose images relative to sample zero dose images, and so on. If so, then at block 408 the pre-processor generates one or more (generated) sample images and adds them to the corresponding repository. For example, with reference to sample full dose images (similar considerations apply to sample zero dose images and sample reduced dose images), it is possible to generate each sample full dose image by linearly combining two or more (acquired) sample full dose images, using an autoencoder fed with the corresponding (acquired) sample full dose image, adding random noise to the corresponding (acquired) sample full dose image, and so on. Processing enters from block 408 into block 409, or directly from block 407 into block 409 if there is no need to increase the number of sample images.

[0118] Now, the activity flow branches depending on the type of the (possibly reduced / increased) sample images of the imaging procedure. In particular, blocks 410-446 are performed when the sample images are incomplete, while blocks 447 are performed when the sample images are complete. In both cases, processing goes to block 448.

[0119] With specific reference to block 410 (incomplete sample image), the creator creates two or more incomplete sample sets for the imaging procedure (each sample set being formed of at least one sample zero dose image and at least one sample full dose image extracted from the sample image repository) and saves these incomplete sample sets into a corresponding repository (with a link to their acquisition parameters in the sample image repository). For example, the creator creates an incomplete sample set by combining a single sample zero dose image with each of a plurality of sample full dose images, by combining the same number of sample zero dose images one-to-one with sample full dose images, by combining sample zero dose images with sample full dose images in any possible way, and so on.

[0120] Then enters a loop at block 411, in which the analysis engine takes into account the (current) incomplete sample set of the imaging procedure (starting from the first sample set in arbitrary order from the corresponding repository). If necessary, at block 412, the noise corrector computes the noise of the sample zero dose image (as the difference between the acquired image and the denoised image) and then its (zero dose) standard deviation; likewise, the noise corrector computes the noise of the sample full dose image (as the difference between the acquired image and the denoised image) and then its (full dose) standard deviation. In both cases, the sample (zero dose / full dose) image can be denoised with the automatic encoder described above. 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 computes the standard deviation of the artificial noise to be injected into the corresponding sample source image in additive form (for example, by applying a noise formula to the reference standard deviation and then adding a correction factor to the result obtained) and / or multiplicative / convolutional form (for example, multiplying this value by a conversion factor).

[0121] The activity flow branches at block 413 according to the configuration of the analysis engine (for example, manually selected by the operator via its user interface, defined by default or uniquely available). In particular, if the analysis engine is not configured to operate in k-space, blocks 414-431 are executed. Otherwise, blocks 432-444 are executed. In both cases, the activity flow merges again at block 445.

[0122] With particular reference to block 414 (non-k-space), the optional analysis engine computes a modulation factor, which is used to modulate the downscaling factor to be used for applying the analogical formula retrieved from the corresponding repository (e.g. manually selected by the operator via its user interface, default defined or uniquely available). In fact, the analogical formula can introduce an approximation, the higher the local concentration of the contrast agent, the higher the approximation. In particular, when the contrast agent is not present, the analogical value from the signal intensity (given by the analogical formula) substantially equals the real value of the signal intensity (obtained by actually acquiring a sample-reduced dose image from the corresponding body part of a subject to which the contrast agent has been administered at a sample-reduced dose) starting from which, as the local concentration of the contrast agent increases, the analogical value becomes increasingly lower than the real value. In order to compensate this loss of the analogical value with respect to the real value, it is possible to increment the value of the downscaling factor used in the analogical formula (i.e. to decrement its denominator) in order to limit the reduction of the analogical value with respect to the corresponding administered value. More specifically, by solving the equation that sets the ratio between the approximation of the signal law and the downscaling factor equal to one, the value of the downscaling factor should be linearly incremented as a function of the local concentration of the contrast agent according to a scaling factor (modulation factor) that depends on the acquisition parameters; the modulation factor is given by a correction formula determined analytically as a function of the acquisition parameters, or by a value corresponding to the acquisition parameters empirically determined. Therefore, the analysis engine retrieves the acquisition parameters of the incomplete sample set from the incomplete sample set repository (linked in the sample image repository to these parameters), and then computes the modulation factor by applying the correction formula to the acquisition parameters or by retrieving the value corresponding to the acquisition parameters from a pre-defined table.

[0123] The activity flow further branches at block 415 depending on the configuration of the analysis engine. In particular, if the analysis engine is configured to operate on the sample image in the amplitude form, then a loop is entered at block 416 in which the analysis engine takes into account (in arbitrary order starting from the first voxel) the (current) voxel of the sample full-dose image. At block 417, the analysis engine modulates the downscaling factor to be used to apply the simulation formula for the voxel. To this end, the analysis engine computes the contrast enhancement of the voxel as the difference between the voxel value of the sample full-dose image and the voxel value of the sample zero-dose image, and then computes the modulated value of the downscaling factor by multiplying the downscaling factor by the product of the modulation factor and the contrast enhancement. At block 418, the analysis engine computes the voxel value of the reduced-dose image by applying the simulation formula with the (modulated) downscaling factor to the voxel value of the sample zero-dose image and to the voxel value of the sample full-dose image; thus, in the example at issue, the analysis engine subtracts the voxel value of the sample zero-dose image from the voxel value of the sample full-dose image, multiplies this difference by the downscaling factor, and adds the obtained result to the voxel value of the sample zero-dose image. The analysis engine then adds the voxel value thus obtained to the sample reduced-dose image being constructed in the corresponding repository. At block 419, the analysis engine verifies whether the last voxel has been processed. If not, the activity flow returns to block 416, repeating the same operations for the next voxel. Conversely (after all the voxels have been processed), the corresponding loop is exited by proceeding into block 420.

[0124] At this point, the noise corrector injects artificial noise in additive form into the sample reduced-dose image thus obtained. To this end, the noise corrector generates the artificial noise as a (noise) matrix of cells having the same dimensions as the sample reduced-dose image; the noise matrix contains random values having a normal statistical distribution with zero mean and standard deviation equal to the standard deviation of the artificial noise. At block 421, the noise corrector adds the noise matrix voxel by voxel to the sample reduced-dose image in the corresponding repository. The processing then continues to block 445.

[0125] Referring back to block 415, if the analysis engine is configured to operate on the sample images in complex form, then the activity flow branches at block 422 depending on their availability. If the sample zero-dose image and the sample full-dose image are already available in complex form, then the analysis engine performs a phase correction at block 423 by rotating the vector representing each cell of the complex number by an angle equal to its argument (maintaining the same modulus). This operation allows obtaining the same result as applying the simulation formula even when operating on the sample zero-dose image and the sample full-dose image in complex form (as all operations applied to the corresponding complex numbers (without imaginary part) are equivalent to applying them to the corresponding modulus). Then, the processing continues to block 424. If the sample zero-dose image and the sample full-dose image are available in magnitude form, then the same point can also be reached directly from block 422; in this case, the sample zero-dose image and the sample full-dose image are directly considered in complex form, with each of their voxel values (real numbers) being complex numbers with an imaginary part equal to zero.

[0126] Now, similar operations as described above are performed for generating the sample reduced-dose image from the sample zero-dose image and the sample full-dose image by processing them in complex form. In particular, a loop is entered in which the analysis engine takes into consideration (in arbitrary order, starting from the first voxel) the (current) voxel of the sample full-dose image. At block 425, the analysis engine modulates the downscaling factor by computing a contrast enhancement of the voxel (as the difference between the modulus of the voxel value of the sample full-dose image and the modulus of the voxel value of the sample zero-dose image), and then computes a modulated value of the downscaling factor by multiplying the downscaling factor by the product of the modulation factor and the contrast enhancement. At block 426, the analysis engine computes the voxel value of the sample reduced-dose image by applying the simulation formula with the (modulated) downscaling factor to the voxel value of the sample zero-dose image and the voxel value of the sample full-dose image; then, the analysis engine adds the thus obtained voxel value to the sample reduced-dose image being constructed in the corresponding repository. At block 427, the analysis engine verifies whether the last voxel has been processed. If not, then the activity flow returns to block 424, repeating the same operations for the next voxel. Conversely (after having processed all the voxels), the corresponding loop is exited by proceeding to block 428.

[0127] At this point, the noise corrector injects artificial noise in the form of a convolution into the sample reduced dose image thus obtained. To this end, the noise corrector generates the artificial noise as a (noise) matrix of cells having the same size as the sample reduced dose image; this noise matrix contains random complex values having a normal statistical distribution with unit mean and standard deviation equal to the standard deviation of the artificial noise. Then, at block 429, the noise corrector performs a convolution operation on the sample reduced dose image in the corresponding repository by means of the noise matrix (for example, by shifting the noise matrix in a toroidal manner across the sample reduced dose image by a single step, thus wrapping around the sample reduced dose image in each direction). At block 430, the analysis engine converts the sample reduced dose image thus obtained into amplitude form; to this end, the analysis engine replaces each voxel value (now generally complex) of the sample reduced dose image with its modulus. The activity flow branches further at block 431 depending on the configuration of the analysis engine. In particular, if the analysis engine is configured to inject artificial noise also in the form of an addition to the sample reduced dose image, then the processing continues to block 420 to perform the same operations described above (then proceeding into block 445). Conversely, the processing proceeds directly into block 445.

[0128] Instead of referring to block 432 (k-space), the analysis engine takes into account the sample zero dose image and the sample full dose image in complex form (if available, directly, or by converting them from k-space form by applying to them the inverse Fourier transform). As mentioned above, the analysis engine performs a phase correction at block 433 by rotating the vector representing the complex number of each cell of the sample zero dose image and of the sample full dose image, thus cancelling its argument (maintaining the same modulus). The analysis engine converts the sample zero dose image and the sample full dose image from complex form into k-space form by applying to them the Fourier transform at block 434.

[0129] Now, the sample reduced dose image is generated from the sample zero dose image and the sample full dose image by processing them in k-space form. In particular, a loop is entered at block 435 in which the analysis engine takes into account (in arbitrary order, starting from the first cell) a (current) cell of the sample full dose image. At block 436, the analysis engine computes a cell value of the sample reduced dose image by applying to the cell value of the sample zero dose image and to the cell value of the sample full dose image the analog formula with the (original) reduction factor; then, the analysis engine adds the cell value thus obtained to the sample reduced dose image being constructed in the corresponding repository. At block 437, the analysis engine verifies whether the last cell has been processed. If not, the activity flow returns to block 435, repeating the same operations for the next cell. Conversely (after all the cells have been processed), the corresponding loop is exited by proceeding into block 438.

[0130] At this point, the noise corrector injects artificial noise into the thus obtained sample reduced dose image in multiplicative form. To this end, the noise corrector generates artificial noise as a (noise) matrix of cells having the same dimensions as the sample reduced dose image; this noise matrix contains random complex values having a normal statistical distribution with unit mean and standard deviation equal to the standard deviation of the artificial noise. At block 439, the noise corrector multiplies the sample reduced dose image cell by cell with the noise matrix in the corresponding repository. The activity flow branches further at block 440 depending on the configuration of the analysis engine. In particular, if the analysis engine is configured to inject artificial noise into the sample reduced dose image also in additive form, the processing continues to block 441, where the noise corrector generates artificial noise as a (further) noise matrix of cells (having the same dimensions as the sample reduced dose image), which now contains random complex values having a normal statistical distribution with null mean and standard deviation equal to the standard deviation of the artificial noise. At block 442, the noise corrector adds the noise matrix cell by cell to the sample reduced dose image in the corresponding repository. Then, the processing continues to block 443; if the analysis engine is not configured to inject artificial noise into the sample reduced dose image in additive form, it is also possible to reach the same point directly from block 440. At this point, the analysis engine converts the sample reduced dose image from k-space form to complex form by applying an inverse Fourier transform thereto. At block 444, the analysis engine converts the sample reduced dose image from complex form to magnitude form by replacing each voxel value thereof with the modulus thereof. Then, the processing proceeds into block 445. Alternatively, not shown in the figure, it is possible to inject artificial noise into the sample reduced dose image in magnitude form (when required) in additive form by continuing from block 444 to block 420 for performing the same operations described above (then proceeding into block 445).

[0131] Reference is now made to block 445, in which the analysis engine verifies whether the last incomplete sample set of the imaging procedure has been processed. If not, the activity flow returns to block 411, repeating the same operations for the next incomplete sample set. Conversely (after all the incomplete sample sets, or a selected portion thereof, have been processed), the corresponding loop is exited by proceeding into block 446.

[0132] At this point, the combiner generates one or more complete sample sets (each complete sample set being formed by at least one sample zero dose image, a sample full dose image and a sample reduced dose image) for the imaging procedure (and saves them into the corresponding repository) by combining the incomplete sample set with the sample reduced dose images generated from (retrieved from the corresponding repository) other incomplete sample sets. This result can be achieved in different ways. In particular, in an embodiment, half of the incomplete sample set (in this case, there is no need to simulate the corresponding sample reduced dose image) and the sample reduced dose images simulated from the other half of the incomplete sample set are combined one-to-one. Thus, if there are N incomplete sample sets, this provides INT(N / 2) complete sample sets. For example, consider the scenario with four incomplete sample sets B1-F1, B2-F2, B3-F3, B4-F4 (formed by the corresponding sample zero dose images B1, B2, B3, B4 and sample full dose images F1, F2, F3, F4) and four sample reduced dose images R1, R2, R3, R4 that can be simulated from them. If the incomplete sample sets B1-F1, B2-F2 and the sample reduced dose images R3, R4 are considered, the complete sample sets (INT(4 / 2)=2) can be B1-F1-R3, B2-F2-R4. This embodiment provides a high diversity of complete sample sets that will be used to train the operational neural network. In another embodiment, the incomplete sample sets and the sample reduced dose images simulated from other incomplete sample sets are combined in all possible ways. Thus, if there are N incomplete sample sets and their corresponding sample reduced dose images, this provides N-(N-1) complete sample sets. Consider the same example above, the complete sample sets (4-3=12) will be B1-F1-R2, B1-F1-R3, B1-F1-R4, B2-F2-R3, B2-F2-R4, B2-F2-R1, B3-F3-R4, B3-F3-R1, B3-F3-R2, B4-F4-R1, B4-F4-R2, B4-F4-R3. This implementation provides a high number of complete sample sets that will be used to train the operational neural network. In yet another embodiment, the incomplete sample sets are combined one-to-one with the sample reduced dose images simulated from other incomplete sample sets. Thus, if there are N incomplete sample sets and their corresponding sample reduced dose images, this provides N complete sample sets. Consider the same example above, the complete sample sets (4) can be B1-F1-R2, B2-F2-R3, B3-F3-R4, B4-F4-R1. This embodiment provides a trade-off between the high diversity and the high number of complete sample sets that will be used to train the operational neural network. Then, the process proceeds into block 448.

[0133] Referring now to block 447 (full sample images), the combiner directly generates one or more full sample sets (each formed by at least one sample zero dose image, sample full dose image, and reduced dose sample image) for the imaging procedure by combining its sample images (retrieved from the corresponding repository) (and saves them to the corresponding repository). This result can also be achieved in different ways. In particular, it is possible to combine the same number of sample full dose images and sample reduced dose images one-to-one and then combine each pair of images thus obtained with the same (single) sample zero dose image, to combine the same number of sample zero dose images, sample full dose images, and sample reduced dose images one-to-one, to combine sample full dose images and sample reduced dose images in all possible ways and then combine each pair of images thus obtained with the same (single) sample zero dose image, to combine sample zero dose images, sample full dose images, and sample reduced dose images in all possible ways, and so on. The process then proceeds into block 448.

[0134] Referring now to block 448, it is verified whether the last imaging procedure has been processed. If not, the activity flow returns to block 404, repeating the same operations for the next imaging procedure. Conversely (after all imaging procedures have been processed), the corresponding loop is exited by proceeding into block 449. Thus, the full sample set repository can store a mix of (simulated) full sample sets for which the sample reduced dose images have been simulated and (acquired) full sample sets for which the sample reduced dose images have been acquired; for example, the acquired full sample sets represent 1-20%, preferably 5-15%, more preferably 6-12%, such as 10%, of the total number of (simulated / acquired) full sample sets. This can further improve the quality of the trained operational neural network, while requiring only limited additional effort (especially when the acquired full sample sets are obtained from preclinical imaging procedures).

[0135] Referring now to block 449, the activity flow branches depending on the operational mode in which the computer is configured. If a trained neural network is available for simulating sample reduced dose images (to be used for training the operational neural network), then at block 450 the training engine trains it by using the full sample sets (retrieved from the corresponding repository). For example, the same operations described below for training the operational neural network can be performed, with the difference that now the trained neural network is optimized for generating sample reduced dose images from the corresponding sample zero dose images and sample full dose images; in this case, it is also possible to use a more complex loss function to improve the performance of the trained neural network, e.g. using a method exploiting a generative adversarial network (GAN). The training engine saves the configuration of the trained neural network thus obtained into the corresponding repository; at the same time, the training engine deletes from its repository the sample reduced dose images and their corresponding full sample sets that have been already simulated. Then at block 451 the cycle for simulating refined versions of the sample reduced dose images for the imaging procedures providing the incomplete sample images is entered. To this end, the analysis engine takes into account the (current) imaging procedure providing the incomplete sample images (in arbitrary order from the first image in the sample image repository). Then, at block 452 the analysis engine takes into account the (current) incomplete sample set of the imaging procedure (from the corresponding repository in arbitrary order from the first image). At block 453 the analysis engine feeds the sample zero dose images and sample full dose images of the incomplete sample set to the trained neural network. Moving to block 454, the trained neural network outputs the corresponding sample reduced dose image, which is saved into the corresponding repository. At block 455 the analysis engine verifies whether the last incomplete sample set has been processed. If not, then the activity flow returns to block 452, repeating the same operations for the next incomplete sample set. Conversely (after all the incomplete sample sets or a selected portion thereof have been processed), the corresponding cycle is exited by passing to block 456, in which the analysis engine now verifies whether the last imaging procedure has been processed. If not, then the activity flow returns to block 451, repeating the same operations for the next imaging procedure. Conversely (after all the imaging procedures have been processed), the corresponding cycle is exited by proceeding to block 457. At this point, the combiner generates one or more full sample sets for the imaging procedure (and saves them into the corresponding repository) by combining the incomplete sample set with the sample reduced dose images generated from other incomplete sample sets (retrieved from the corresponding repository), as above.

[0136] Then, processing passes to block 458, whereby the thus obtained complete sample set training is used to train the operational neural network. This implementation improves the accuracy of the sample reduced dose image, which in turn improves the performance of the operational neural network trained with the corresponding complete sample set. If the operational neural network is to be trained directly with the complete sample set generated by the analysis engine, due to no trained neural network being available, then the same point is reached directly from block 449. This implementation is particularly simple and expedient; at the same time, the accuracy of the analytically simulated sample reduced dose image is sufficient for the purpose of training the operational neural network with acceptable performance.

[0137] In both cases, the training engine now performs this operation in order to find the optimal values of the operational neural network weights that optimize the performance of the operational neural network. First, the training engine can optionally post-process the sample images of each full sample set. For example, the training engine normalizes the sample images by scaling their voxel values to a (common) predefined range. Also, the training engine performs a data augmentation process by generating (new) full sample sets from each (original) full sample set, thereby reducing overfitting in the training of the operational neural network. For example, the new full sample sets are generated by rotating the sample images of the original full sample set (such as from 0° to 90° in 1-5° increments) and / or by horizontally / vertically flipping them. In any case, at block 459, the training engine selects a plurality of training sets by sampling the full sample sets in the corresponding repository by a certain percentage (e.g., 50% randomly). At block 460, the training engine randomly initializes the weights of the operational neural network. Then at block 461, a loop is entered in which the training engine feeds the sample zero-dose images and sample reduced-dose images of each training set to the operational neural network. In response thereto, at block 462, the operational neural network outputs corresponding output images that should be equal to the sample full-dose images (ground truth data) of the training set. At block 463, the training engine computes a loss value based on the difference between the output images and the sample full-dose images; for example, the loss value is given by the mean absolute error (MAE) computed as the average of the absolute differences between the corresponding voxel values of the output images and the sample full-dose images. At block 464, the training engine verifies whether the loss value is not acceptable and whether it is still significantly improving. This operation can be performed in an iterative mode (after processing the loss value for each training set) or in a batch mode (after processing the cumulative value (such as the average thereof) of the loss values for all training sets). If so, at block 465, the training engine updates the weights of the operational neural network in an attempt to improve its performance. For example, a stochastic gradient descent (SGD) algorithm such as based on the ADAM method is applied (where the direction and amount of the changes are determined by the gradient of the loss function given the loss value as a function of the weights, approximated with a backpropagation algorithm according to a predefined learning rate). Then, the processing returns to block 461 to repeat the same operation. Referring again to block 464, if the loss value has become acceptable or the changes of the weights do not provide any significant improvement (meaning that a minimum of the loss function has been found, at least a local minimum, or a flat region), then the loop is exited. The above loop is repeated a number of times (epochs), for example 100-300, by adding random noise to the weights and / or starting from different initializations of the operational neural network, in order to find different (and possibly better) local minima and to identify flat regions of the loss function.

[0138] After having found the configuration of the operational neural network that provides the optimal minimum of the loss function, the processing continues to block 466, in which the training engine performs a validation of the performance of the thus obtained operational neural network. To this end, the training engine selects a plurality of validation sets from the sample sets in the corresponding repository (e.g., sample sets different from the training sets). Then a loop is entered at block 467, in which the training engine feeds the sample zero-dose image and the sample reduced-dose images (in arbitrary order, starting from the first one) of the (current) validation set to the operational neural network. In response thereto, at block 468 the operational neural network outputs the corresponding output image, which should be equal to the sample full-dose image of the validation set. At block 469 the training engine computes the loss value as above based on the difference between the output image and the sample full-dose image. At block 470 the training engine verifies whether the last validation set has been processed. If not, then the activity flow returns to block 467 to repeat the same operations for the next validation set. Conversely (after having processed all the validation sets), the corresponding loop is exited by proceeding to block 471. At this point, the training engine determines the global loss of the above-mentioned validation (e.g., equal to the average of the loss values of all the validation sets). The activity flow branches at block 472 as a function of the global loss. If the global loss is (possibly strictly) higher than an acceptable value, this means that the generalization capabilities of the operational neural network (from its configuration learned from the training sets to the validation sets) are too poor; in this case, the processing returns to block 459 to repeat the same operations with different training sets and / or training parameters (such as learning rate, epoch, etc.). Conversely, if the global loss is (possibly strictly) lower than an acceptable value, this means that the generalization capabilities of the operational neural network are satisfactory; in this case, at block 473 the training engine accepts the configuration of the operational neural network thus obtained, and saves it to the corresponding repository in association with the value of its upscaling factor.

[0139] At block 474 the analysis engine verifies whether the configuration of the operational neural network has been completed. If not, then the processing returns to block 402 to repeat the same operations in order to configure the operational neural network for a different upscaling factor. Conversely, after having completed the configuration of the operational neural network, at block 475 the configuration of the operational neural network thus obtained is deployed 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 a first delivery of the imaging system, or by uploading them via a network or a removable storage unit in case of an upgrade of the imaging system). Then the processing ends to a concentric white / black stop circle 476.

[0140] Reference is now made to Figures 5A-5C , which shows representative examples of experimental results related to the solution according to embodiments of the present disclosure.

[0141] In particular, dedicated preclinical studies were performed on rats affected by two cerebral lesions: C6 glioma (n=36 animals) and cerebral ischemic lesion (n=42 animals). All animals underwent a surgical procedure to induce the lesion. Animals that survived the surgical procedure and showed only limited or no clinical symptoms during the following two weeks (i.e. the time window required for the pathological development of the lesion) were enrolled in the imaging procedure of the MRI type (i.e. usually 2 per animal, in limited cases only 3). The imaging procedure was performed using a gadolinium-based contrast agent and a preclinical scanner, Pharmascan® by Bruker Corporation, operating at 7T and equipped with a rat head volume coil with 2 channels. The CE-MR protocol used during each imaging procedure was the following:

[0142] • first pre-contrast acquisition of a standard Tl -weighted sequence (first sample zero-dose image);

[0143] • second pre-contrast acquisition of a standard Tl -weighted sequence (second sample zero-dose image);

[0144] • intravenous administration of the contrast agent at a sample reduced dose equal to 0.01 mmol Gd / kg;

[0145] • post-contrast acquisition of a Tl -weighted sequence (sample reduced dose image);

[0146] • intravenous administration of the contrast agent again at 0.04 mmol Gd / kg shortly after the previous intravenous administration, so that the sum reaches a sample full dose equal to 0.05 mmol Gd / kg;

[0147] • first post-contrast acquisition of a Tl -weighted sequence (first sample full dose image);

[0148] • second post-contrast acquisition of a Tl -weighted sequence (second sample full dose image).

[0149] The study resulted in the acquisition of 130 3D MRI sample images (i.e. 61 for glioma rats and 69 for ischemic rats), each composed of 24 slices. The acquired sample images were used to build two datasets:

[0150] • acquired dataset, including for each imaging procedure the acquisition of the first zero-dose image, the sample reduced dose image and the first sample full dose image;

[0151] • Simulated dataset comprising, for each imaging procedure, the acquired first sample zero dose image and first sample full dose image and the simulated sample reduced dose image (reduction factor d = 1 / 5) from the second sample zero dose image and second sample full dose image.

[0152] The operational convolutional neural network (as described in Dongwook Lee, Jaejun Yoo and Jong Chul Ye in "Deep artifact learning for compressed sensing and parallel MRI", available on https: / / arxiv.org / abs / 1703.01120) is trained with the following hyperparameters using the sample full dose image as ground truth data:

[0153] • Learning rate = 0.01

[0154] • Decay = 0.001

[0155] • Compound loss = Mean Absolute Error (MAE) + Mean Absolute Error in Fourier Transform domain (fftMAE) + Perceptual Loss (PL) with the network VGG19 pre-trained on ImageNET dataset (available on https: / / www.image-net.org / ) cropped at layer 4

[0156] • Relative weights of the compound loss: a = b = c = 1 and a = b = 0, c = 1 (where "a" is MAE, "b" is fftMAE and "c" is PL)

[0157] • Noise level (σ artificial ) injected in the sample reduced dose image being simulated: 0.01, 0.0125 and 0.015.

[0158] Different instances of the operational neural network are trained with the acquired dataset and the simulated dataset, using the corresponding full dose image as ground truth data. These instances of the neural network are then applied to the sample reduced dose images included in the acquired dataset (to verify their ability to recover the full dose contrast effect) and to operational full dose contrast images acquired in the medical imaging application with full dose contrast (to verify their ability to boost the contrast effect).

[0159] From Figure 5ATo begin, three representative examples of an acquired sample full-dose image and a corresponding sample full-dose image simulated by an operational neural network trained on an acquired dataset and a simulated dataset are shown. As can be seen, the simulated sample full-dose images are very similar to the sample full-dose images being acquired; this is correct for an operational neural network trained with either the acquired dataset or the simulated dataset. In particular, using an operational neural network trained with the simulated dataset does not introduce significant artifacts in its simulated sample full-dose images.

[0160] Moving to Figure 5B , representative examples of acquired and simulated sample full-dose images (by an operational neural network trained with a simulated dataset) are shown using different hyperparameters (i.e., relative weights a = b = c = 1 and a = b = 0, c = 1, and noise levels 0.01, 0.0125, and 0.015). As can be seen, by tuning the hyperparameters, the quality of the simulated sample full-dose images can be further improved, increasing the similarity to the acquired sample full-dose images under different criteria, such as enhancement of regions perfused with contrast agent, signal intensity of enhanced and non-enhanced regions, etc.

[0161] Moving to Figure 5C , two representative examples of an operational zero-dose image, an operational full-dose image acquired during an imaging procedure on a (human) patient, and a corresponding operational boost dose image (magnification factor k = 5) simulated by an operational neural network trained on an acquired dataset and a simulated dataset (both based on imaging procedures on animals) are shown. As can be seen, the operational boost dose image simulated by the operational neural network trained with the simulated dataset is superimposable on the operational boost dose image simulated by the operational neural network trained with the acquired dataset; in both cases, the operational boost dose image improves contrast enhancement with essentially no introduction of artifacts.

[0162] Modifying

[0163] To satisfy local and specific requirements, a person skilled in the art may apply many logical and / or physical modifications and changes to the present disclosure. More specifically, while the present disclosure has been described with some degree of particularity based on one or more embodiments of the disclosure, it should be understood that various omissions and substitutions and changes in the form and details of the methods and embodiments illustrated can be made by those skilled in the art, without departing from the spirit of the disclosure. In particular, different embodiments of the disclosure can even be practiced without the specific details (such as the numeric values) set forth in the preceding description, to provide a more thorough understanding thereof; on the contrary, well-known features can have been omitted or simplified in order not to obscure the description with unnecessary particulars. Moreover, it is expressly intended that the features described in each sentence of claims can be implemented independently of the features described in any other sentence, except where it is explicitly stated that they cannot be. In any case, the specific features described in connection with any embodiment of the disclosure can be incorporated in any other embodiment as a general design choice, in order to provide an even broader solution. Moreover, items presented in the same group and different embodiments, examples or alternatives should not be construed as facts that are in fact equivalent (but rather they are separate and autonomous entities). In any case, each numerical value should be read as modified by the applicable tolerances; in particular, the terms "substantially", "about", "approximately" and the like should be understood as within 10%, preferably within 5%, and more preferably within 1%, unless otherwise indicated. Moreover, each numerical range should be intended to explicitly specify any possible number along the continuum of the range (including its endpoints). The ordinal or other qualifiers are used merely as labels to distinguish elements having the same name but they themselves do not imply any priority, precedence or order. The terms include, contain, have, involve, etc. should be intended to have an open, non-exhaustive meaning (i.e. not limited to the items listed), the terms based on, depend on, function of, etc. should be intended as a non-exclusive relationship (i.e. involving possible other variables), the term one / a should be intended as one or more items (unless explicitly indicated otherwise), and the term means for... (or any reference to means plus function) should be intended as any structure adapted to, or configured to, perform the relevant function.

[0164] For example, embodiments provide methods for training a machine learning model. However, the machine learning model can be of any type (e.g. dense neural networks, convolutional neural networks, generative adversarial networks, linear and non-linear regression models such as polynomial regression, support vector regression, nearest neighbor regression, Gaussian process regression, etc., probabilistic models such as Bayesian networks, Markov random fields, etc., used in conjunction with optimization methods such as gradient descent-based optimization methods and other more complex ones such as genetic algorithms, etc.).

[0165] In embodiments, the machine learning model is for use in a medical imaging application. However, the medical imaging application can be any type (e.g., diagnostic, therapeutic, or surgical application based on MRI, CT, fluoroscopy, fluorescence, or ultrasound technology, etc.).

[0166] In embodiments, the method comprises the following steps under control of a computing system. However, the computing system can be any type (see below).

[0167] In embodiments, the method comprises providing (to the computing system) a plurality of incomplete sample sets for each of a plurality of imaging procedures. However, the imaging procedures can be any number and any type (e.g., preclinical, clinical, etc.), and the incomplete sample sets for each imaging procedure can be any number (the same or different between imaging procedures); and the incomplete sample sets can be provided in any manner (e.g., already formed or created from corresponding sample images, collected from any source (such as a health / laboratory facility) in any manner, such as downloaded from a central server of the facility over the Internet (where the incomplete sample sets have been automatically gathered from corresponding imaging systems over a corresponding LAN or manually gathered with the aid of removable storage units, manually loaded from the removable storage units where they have been copied from the central server or from (stand-alone) imaging systems), etc.). In any case, this is a (computer-implemented) data processing method that is performed independently of acquisition of the sample images (without requiring any interaction with the corresponding subjects).

[0168] In embodiments, the imaging procedure is related to a corresponding body part of a subject. However, the body part can be any number, any type (e.g., organ, region of an organ, tissue, bone, joint, etc., the same or different from the body part for which the machine learning model is intended in the medical imaging procedure, etc.), and in any condition (e.g., healthy, pathological state with any lesion, etc.); and the body part can belong to any number and type of subject (e.g., animal, human, etc.).

[0169] In embodiments, each of the incomplete sample sets comprises at least one sample target image. However, the sample target image can be any number (e.g., axial, coronal, and / or sagittal images corresponding to different doses of contrast agent, etc.) and any type (e.g., in any form such as magnitude, complex, k-space, etc., with any size, resolution, chroma, bit depth, etc., related to any location of the body part, such as voxels of a 3D image, pixels of a 2D image, etc.).

[0170] In embodiments, the sample target images represent corresponding body parts of the subject to which a contrast agent has been administered in the sample target dose. However, the contrast agent can be of any type (e.g., any targeted contrast agent (such as based on specific or non-specific interaction), any non-targeted contrast agent, etc.) and can be administered in any manner that ensures that it perfuses the body part in any sample target dose (e.g., equal to the full dose, lower or higher than the full dose, etc.) (e.g., intravenous injection, intramuscular injection, oral, etc. at any time prior to the imaging procedure up to immediately prior to the imaging procedure).

[0171] In embodiments, each of the incomplete sample sets includes at least one sample baseline image. However, the sample baseline images can be of any number (e.g., axial, coronal, and / or sagittal images corresponding to zero dose and / or different doses of contrast agent, etc.) and of any type (e.g., the same or different than the sample target images).

[0172] In embodiments, the sample baseline images represent images of the corresponding body parts of the subject without contrast agent. However, the sample baseline images can have been acquired in any manner that ensures that the contrast agent does not have a significant impact on their content (e.g., acquired any time in advance of administration of the contrast agent, acquired any time in delay after the previous administration of the contrast agent, etc.).

[0173] In embodiments, the sample baseline images represent corresponding body parts of the subject to which a contrast agent has been administered in a sample baseline dose lower than the sample target dose. However, the sample baseline dose can have any value (e.g., lower or higher than the sample source dose).

[0174] In an embodiment, the method includes (by the computing system) simulating a plurality of sample source images, at least one of the sample source images being simulated from each of the incomplete sample sets of each of the imaging procedures. However, the sample source images can be simulated from any number of the incomplete sample sets (up to all), from each of the incomplete sample sets in any number (e.g., for different values of the downscaling factor), and in any manner (e.g., operating in any domain, such as magnitude, complex, k-space, etc., whether or not any pre-processing (such as registration, normalization, denoising (such as with an autoencoder), analysis techniques based on block matching, shrinkage field, wavelet transform, smoothing filter, etc., distortion correction, filtering of outlier sample images, etc.), any post-processing (such as any registration, normalization, noise injection, etc.), etc. are performed). For example, in an embodiment, the sample source images are analytically generated (such as by applying a simulation formula in the case of a single sample baseline / target image, by interpolation in the case of multiple sample baseline / target images, etc.). In another embodiment, a preliminary version of the sample source images is analytically generated, a further machine learning model is trained based on the preliminary version of the complete sample set of analytically generated sample source images, and then a refined version of the sample source images is generated by the further machine learning model being trained. In yet another embodiment, the sample source images are generated by a further machine learning model trained with a further sample set acquired independently.

[0175] In an embodiment, the sample source images are simulated to represent the corresponding body part, mimicking administration of contrast agent to the corresponding subject at a sample source dose lower than a sample target dose (where the ratio between the sample source dose and the sample target dose is equal to the downscaling factor). However, the sample source dose can have any value, in absolute terms or in relative terms (e.g., the sample source dose is lower than, equal to, or higher than a full dose, the sample source dose and the sample target dose define any downscaling factor, etc.).

[0176] In an embodiment, the method includes (by the computing system) generating one or more complete sample sets for each of the imaging procedures. However, the complete sample set for each of the imaging procedures can be any number.

[0177] In embodiments, each complete sample set includes sample baseline images and sample target images for one incomplete sample set of the imaging procedure and sample source images simulated from another incomplete sample set of the imaging procedure. However, the incomplete sample sets and sample source images can be combined in any manner (e.g., each incomplete sample set combined with one or more sample source images of other incomplete sample sets, each sample source image combined with one or more other incomplete sample sets, using each of all or only a portion of the incomplete sample sets and sample source images one or more times, etc.) and each complete sample set can include any number of sample source images (simulated from one or more other incomplete sample sets).

[0178] In embodiments, the method includes training (by the computing system) a machine learning model to optimize its ability to generate sample target images for each complete sample set from sample baseline images and sample source images of the complete sample set. However, the machine learning model can be trained in any manner (e.g., by selecting any training / validation set from the complete sample sets, using any algorithm (such as stochastic gradient descent, real-time recurrent learning, higher-order gradient descent, extended Kalman filter, etc.), any loss function (such as based on mean absolute error, mean squared error, perceptual loss, adversarial loss, etc.), defined individually at a location level or in groups thereof, for a single magnification factor, for multiple magnification factors corresponding to the sample source images of each complete sample set, for a variable magnification factor as a parameter of the operational machine learning model, etc.); and the training can be based on any additional information up to no additional information (e.g., any number of complete sample sets formed from all acquired sample images (already formed provided or created from corresponding sample images), additional sample images in each of one or more complete sample sets acquired and / or simulated (such as corresponding to different doses of contrast agent and / or different acquisition conditions, health conditions of body parts, types of subjects, etc.).

[0179] In embodiments, the method includes deploying (by the computing system) the machine learning model being trained. However, the machine learning model can be deployed to any number and type of imaging systems in any manner (e.g., distributed with corresponding new imaging systems, or for upgrading installed, online imaging systems, etc.).

[0180] In embodiments, the deployed machine learning model is used in a medical imaging application to mimic an increase in a contrast agent dose administered to a corresponding patient according to a magnification factor corresponding to an inverse of a reduction factor. However, the patient can be of any type (e.g., human, animal, etc.); the machine learning model can be used to mimic an increase in a contrast agent dose in any manner (e.g., in real-time, offline, locally, remotely, etc.) where the magnification factor corresponds to (e.g., equals, is lower than, or higher than, such as according to a corresponding multiplication factor, etc.) an inverse of the reduction factor in any manner.

[0181] Additional embodiments provide additional advantageous features, although these features can be omitted entirely in basic implementations. In this regard, it is expressly stated that the features of each of the following embodiments can be combined with the above features, either individually or in combination with any number of the features of any number of the other subsequent embodiments.

[0182] In embodiments, the sample source dose is between the sample baseline dose and the sample target dose. However, the sample baseline dose can have any value (in absolute or relative terms).

[0183] In embodiments, the sample target dose is a full dose of contrast agent that is standard in clinical practice. However, the full dose can be of any type (e.g., fixed for each type of medical imaging application, dependent on a type of body part, dependent on a type, weight, age, etc. of the patient, etc.).

[0184] In embodiments, at least a portion of the subjects is an animal. However, the animal can be any percentage of the subjects (from none to all) and of any type (e.g., rat, pig, etc.).

[0185] In embodiments, the patient is a human. However, the human can be of any type (such as gender, age, health, etc.).

[0186] In embodiments, the method includes (by the computing system) receiving a plurality of sample images for each of the imaging procedures. However, the sample images can be of any number and received in any manner (see above description regarding incomplete sample sets).

[0187] In embodiments, the sample images for each of the imaging procedures include one or more of the sample baseline images and one or more of the sample target images. However, the sample images can include any number of sample baseline images and sample target images.

[0188] In embodiments, the method includes (by the computing system) creating, for each imaging procedure, an incomplete set of sample images for that imaging procedure based on the sample images received. However, the incomplete set of sample images can be created in any manner (e.g., based on sample images received or augmented / reduced, combining each sample baseline image with one or more sample target images, combining each sample target image with one or more sample baseline images, using all or only a portion of each of the sample baseline images and sample target images one or more times, etc.).

[0189] In embodiments, the method includes (by the computing system) generating, for each of one or more of the imaging procedures, one or more new sample baseline images and / or one or more new sample target images from the corresponding sample baseline images and / or sample target images received, respectively. However, new sample images can be generated for any of the imaging procedures (from none to all), in any number (for sample baseline images only, sample target images only, or both), and in any manner (e.g., each new sample image obtained by linearly / non-linearly combining two or more corresponding sample images, by adding any noise to a corresponding sample image, by using any autoencoder, etc.).

[0190] In embodiments, the method includes (by the computing system) generating, for each of the imaging procedures, new sample baseline images and / or new sample target images such that the number of sample baseline images is equal to the number of sample target images for the imaging procedure. However, new sample (baseline / target) images can be generated for any purpose (e.g., to augment with fewer numbers of sample baseline / sample target images or both, to equalize the number of both, to obtain a required number of sample baseline / target images from a single image, etc.).

[0191] In embodiments, the method includes (by the computing system) discarding, for each of one or more of the imaging procedures, one or more of the sample baseline images and / or one or more of the sample target images. However, for any of the imaging procedures (from none to all), sample images can be discarded in any number (for sample baseline images only, sample target images only, or both) and in any manner (e.g., according to their quality level, diversity, etc.).

[0192] In embodiments, the method includes (by the computing system) discarding, for each imaging procedure, a sample baseline image and / or a sample target image so that the number of sample baseline images for an imaging procedure is equal to the number of sample target images. However, sample (baseline / target) images can be discarded for any purpose (e.g., to reduce the number of sample baseline / target images or both, to equalize the number of both, to obtain a required number of sample baseline images / sample target images, to ensure a minimum quality level or diversity of sample images, etc.).

[0193] In embodiments, the method includes (by the computing system) creating an incomplete sample set for each of one or more of the imaging procedures by combining a single sample baseline image of the sample baseline images for that imaging procedure with each of the sample target images. However, this process can be applied to any number of imaging procedures (from none to all).

[0194] In embodiments, the method includes (by the computing system) creating an incomplete sample set for each of one or more of the imaging procedures by combining the same number of sample baseline images and sample target images for each of the imaging procedures one-to-one. However, this process can be applied to any number of imaging procedures (from none to all).

[0195] In embodiments, the method includes (by the computing system) creating an incomplete sample set for each of one or more of the imaging procedures by combining the sample baseline images and sample target images for each of the imaging procedures in all possible ways. However, this process can be applied to any number of imaging procedures (from none to all).

[0196] In embodiments, the method includes (by the computing system) generating a complete sample set for each of one or more of the imaging procedures by combining one half of the incomplete sample set with sample source images simulated from the other half of the incomplete sample set for that imaging procedure one-to-one. However, this process can be applied to any number of imaging procedures (from none to all).

[0197] In embodiments, the method includes (by the computing system) generating a complete sample set for each of one or more of the imaging procedures by combining the incomplete sample set with sample source images simulated from other incomplete sample sets in the incomplete sample set for that imaging procedure in all possible ways. However, this process can be applied to any number of imaging procedures (from none to all).

[0198] In an embodiment, the method includes (by the computing system) generating the complete set of samples for each of the one or more imaging procedures by combining the incomplete set of samples one-to-one with a simulated sample source image from other incomplete sets of samples of the imaging procedure. However, this process can be applied to any number of imaging procedures (from none to all).

[0199] In an embodiment, each of the sample baseline images, each of the sample target images, and each of the sample source images includes a plurality of sample baseline values, a plurality of sample target values, and a plurality of sample source values, respectively. However, the sample baseline / sample source / sample target values can be of any number and any type (e.g., real / complex, with any range, in grayscale or color, etc.).

[0200] In an embodiment, the method includes (by the computing system) calculating each of the sample source values of each of the sample source images by applying a simulation formula that depends on the reduction factor. However, the simulation formula can be of any type (e.g., linear, quadratic, cubic, a function of the corresponding sample baseline value and / or sample administered value, etc.).

[0201] In an embodiment, the simulation formula is derived from a signal law that expresses the amplitude of the response signal of the body part as a function of the local concentration of the contrast agent. However, the signal law can be of any type (e.g., based on any external / internal acquisition parameters, etc.) and the simulation formula can be derived from the signal law in any way (e.g., from any approximation of the signal law, from the actual signal law, etc.).

[0202] In an embodiment, the simulation formula is derived from a signal law that is linearized with respect to the local concentration of the contrast agent. However, the signal law can be linearized in any way (e.g., with any series expansion, any approximation, assuming any linear / non-linear relationship between the local concentration and the dose of the contrast agent, etc.).

[0203] In an embodiment, the simulation formula is derived from a signal law that assumes a direct proportionality between the local concentration and the dose of the contrast agent. However, the direct proportionality between the local concentration and the dose of the contrast agent can be based on any proportionality factor.

[0204] In an embodiment, the sample baseline values, the sample target values, and the sample source values represent the response signal of the corresponding location of the body part. However, the response signal can be represented in any way (e.g., in amplitude form, in complex form, in positive / negative form, etc.).

[0205] In an embodiment, the method includes (by the computing system) modulating a downscaling factor to be used for computing each of the sample source values in each of the sample source images according to an indication of a local concentration of the contrast agent at the corresponding location derived from the corresponding sample target value. However, the local concentration can be derived in any way (e.g., set to a corresponding local contrast enhancement, computed from the sample target value according to a signal law, etc.) and the downscaling factor can be modulated according to any linear / non-linear function thereof (e.g., an empirically determined modulation factor, computed by using an average / local value of any acquisition parameter, etc.) up to always maintaining the same.

[0206] In an embodiment, the method includes (by the computing system) injecting artificial noise into each of the sample source images. However, the artificial noise can be of any type (e.g., depending on the downscaling factor, fixed, etc.) and can be injected into the sample source images in any way (e.g., in additive form or in multiplicative form (such as at each cell, cell group level, etc.), in a cyclic or non-cyclic way in a convolutional form (such as with any stride, padding, etc.), in a magnitude form, complex form, K-space form, at every place, only at places where the contrast agent is present, etc.) up to none.

[0207] In an embodiment, the artificial noise has a statistical distribution depending on the downscaling factor. However, the statistical distribution of the artificial noise can be of any type (e.g., normal distribution with any mean, Rayleigh distribution, Rician distribution, etc.) and can be obtained in any way (e.g., by computing artificial values of one or more statistical parameters, such as standard deviation, variance, skewness, etc., by applying any linear / non-linear function to reference values of the statistical parameters, such as obtained from the noise of the sample baseline images and the noise of the sample target images, from the noise of the sample baseline images only, from the noise of the sample target images only, etc., then the artificial values of the statistical parameters can be heuristically corrected, such as in the same way for all statistical parameters or differently for each statistical parameter, by incrementing / decrementing them, according to any linear / non-linear function, etc.).

[0208] In an embodiment, the method includes (by the computing system) training a further machine learning model to optimize its ability to generate the sample source images for each of the at least part of the complete sample sets from the corresponding sample baseline images and sample target images. However, the further machine learning model can be of any type and can be trained in any way (e.g., same or different with respect to the machine learning model) by using any sample sets (e.g., all sample sets after analytically completing the incomplete sample sets, only providing completed sample sets, etc.).

[0209] In an embodiment, the method comprises (by the computing system) generating a refined version of each of the sample source images by applying to the sample baseline images and the sample target images of the corresponding incomplete sample set the further machine learning model that is being 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, generating them directly, etc.) is not excluded.

[0210] In an embodiment, the method comprises (by the computing system) repeating the steps of simulating the sample source images, generating the complete sample set and training the machine learning model for a plurality of values of the downscaling factor. However, the values of the downscaling factor can be of any number and of any type (e.g. uniformly distributed, with variable pitch such as decreasing for increasing values, etc.) and these steps can be repeated in any way (e.g. consecutively, at different times, etc.).

[0211] In an embodiment, the method comprises (by the computing system) deploying the machine learning model in a corresponding configuration, trained with the values of the downscaling factor, for selecting one or more corresponding values of the upscaling factor in each of the medical imaging applications. However, the different configurations can be deployed in any way (e.g. all together, added over time, etc.) and in any form (e.g. corresponding configurations of a single operative machine learning model, corresponding instances of operative machine learning models, etc.) and they can be used to select values of the upscaling factor in any number and in any way (e.g. discrete mode, continuous mode, same or different with respect to the values of the downscaling factor, etc.).

[0212] In an embodiment, the machine learning model is a neural network. However, the neural network can be of any type (e.g. autoencoder, multilayer perceptron network, recurrent network, generative adversarial network, etc. with any number of layers, shallow or deep, with any connections, receptive fields, strides, padding, activation functions, etc. between the layers).

[0213] An embodiment provides a method of using a machine learning model trained as above in a medical imaging application for imaging a body part of a patient. However, the body part can be of any type, in any condition and can belong to any patient (see above). In any case, while the method can facilitate the tasks of a physician, it only provides intermediate results that can be helpful to the physician, while the strict medical activity is always performed by the physician himself.

[0214] In an embodiment, the method comprises performing the following steps under the control of a computing system. However, the computing system can be of any type (see below).

[0215] In embodiments, the method includes receiving (by the computing system) one or more operational administration images representative of a body part of a patient to which contrast agent has been administered in an operational administration dose. However, the operational administration images can be of any number and of any type (e.g., the same as or different from the sample images) and can be received in any manner (e.g., in real-time, offline, locally, remotely, etc.); and the operational administration dose can have any value (e.g., the same as or different from the sample source dose, lower than, equal to, or higher than a full dose of contrast agent, etc.). In any case, the contrast agent can be administered to the patient in any manner, including in a non-invasive manner (e.g., orally for imaging the gastrointestinal tract, into the respiratory tract via a nebulizer, via topical spray application), and in any case without requiring any substantial physical intervention of the patient that requires professional medical expertise or poses a health risk to the patient (e.g., intramuscular injection).

[0216] In embodiments, the method includes receiving (by the computing system) at least one operational baseline image representative of a body part of a patient to which contrast agent has not been administered or to which contrast agent has been administered in an operational baseline dose lower than the operational administration dose. However, the operational baseline images can be acquired in any manner that ensures that the contrast agent does not significantly affect their content, or after administration of contrast agent in any operational administration dose (e.g., the same as or different from the source baseline image).

[0217] In embodiments, the method includes simulating (by the computing system) a corresponding operational simulation image from the operational baseline image and the operational administration image using the machine learning model. However, the operational simulation image can be simulated in any manner (e.g., operating in any domain, such as magnitude, complex, k-space, etc., in real-time, offline, locally, remotely, etc.).

[0218] In embodiments, the operational simulation image is representative of the body part of the patient to which contrast agent is simulated to have been administered in an operational simulation dose higher than the operational administration dose (where a ratio between the operational simulation dose and the operational administration dose is equal to a magnification factor corresponding to an inverse of a reduction factor of the machine learning model). However, the operational simulation dose can have any value (e.g., the same as or different from the sample target dose, lower than, equal to, or higher than a full dose, etc.).

[0219] In embodiments, the method includes outputting (by the computing system) a representation of the body part based on the operational simulation image. However, the representation of the additional body part can be of any type (e.g., the operational simulation image, a corresponding operational combined image, etc.) and can be output in any manner (e.g., displayed on any device, such as a display, virtual reality glasses, etc., or more generally output in any manner in real-time or offline, such as printing, remote transmission, etc.).

[0220] In general, similar considerations apply if the same solution is achieved using equivalent means, provided that it still falls within the scope of the claims (by using similar steps with more steps or parts of them having the same functionality, removing some non-essential steps or adding additional optional steps); moreover, steps can be executed in a different order, concurrently or (at least partially) in an interleaved way.

[0221] Embodiments provide a computer program configured for causing a computing system to perform the above method when the computer program is executed on the computing system. Embodiments provide a computer program product comprising one or more non-transitory computer-readable storage media having stored thereon program instructions that can be read by a computing system to cause the computing system to perform the same method. However, the (computer) program can be executed on any computing system (see below). The program can be implemented as a standalone module, as a plug-in for an existing software program (e.g., a configuration application or an imaging application), or even directly in an existing software program.

[0222] In general, similar considerations apply if the program is structured in a different way or additional modules or functions are provided, provided that it still falls within the scope of the claims. In particular, the program can take any suitable form that is appropriate for use by the computing system, whereby it is configured to perform the desired operations; the program can be in the form of external or resident software, firmware, or microcode (either in object code or in source code), e.g., to be compiled or interpreted. Moreover, it is possible to provide the program on any computer-readable storage medium. The storage medium is any tangible medium that can retain and store instructions for use by the computing system (inherently different from a transitory signal). For example, the storage medium can be of electronic, magnetic, optical, electromagnetic, infrared, or semiconductor type; examples of such storage media are a fixed disk (where the program can be preloaded), a removable disk, a memory key (e.g., USB), 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 forward it for storage into one or more storage devices of the computing system. In any case, the solution according to embodiments of the present disclosure is even suitable for being implemented with a hardware structure (e.g., by means of electronic circuits integrated on one or more chips of a semiconductor material) or with a combination of suitably programmed or otherwise configured software and hardware.

[0223] Embodiments provide a computing system comprising components configured to perform the steps of the above-described method. Embodiments provide a computing system comprising circuitry for performing each step of the same method (i.e., any suitably configured hardware, e.g., by software). However, the computing system can be of any type (e.g., a configuration computer for training a machine learning model implemented by a server, a virtual machine, a cloud service, etc., a control computer for each scanner implemented by a PC, and a control unit for each scanner using the machine learning model being trained, etc.).

[0224] In general, similar considerations apply if the computing system has a different structure, comprises equivalent components, or has other operational characteristics, provided that it remains within the scope of the claims. In any case, each of its components can be separated into more elements, or two or more components can be combined into a single element; and each component can be replicated to support the parallel execution of the corresponding operations. Moreover, unless otherwise specified, any interaction between different components need not be continuous and can be direct or indirect through one or more intermediaries.

[0225] Embodiments provide a medical method applied to a body part of a patient. However, the medical method can be applied to any body part of any patient (see above).

[0226] In embodiments, the medical method comprises acquiring an operational baseline image representative of the body part. However, the operational baseline image can be acquired in any way (e.g., before administering the contrast agent, administering the contrast agent at an operational baseline dose lower than the operational administration dose, etc.).

[0227] In embodiments, the medical method comprises administering a contrast agent to the patient at an operational administration dose. However, the contrast agent can be administered in any way (e.g., with a syringe, an infusion pump, in advance, shortly before acquiring the operational administration images, continuously during the acquisition of the images, etc.).

[0228] In embodiments, the medical method comprises acquiring one or more operational administration images representative of the body part in response to said administering the contrast agent to the patient (the corresponding operational simulation images are simulated from the operational baseline image and the operational administration images, and the representation of the body part based on the operational simulation images is output according to the above-described method). However, the operational administration images can be of any number and acquired in any way (e.g., with any delay after administering the contrast agent, during the administration of the contrast agent, continuously, at specific times, etc.).

[0229] In embodiments, the medical method includes performing a medical procedure related to the body part in accordance with the representation of the body part. However, the medical procedure can be of any type (e.g., a diagnostic procedure, a therapeutic procedure, a surgical procedure, etc.).

Claims

1. A method (400) to train a machine learning model (305) for use in medical imaging applications, wherein the method (400) comprises, under control of a computing system (130): providing (403-410), to the computing system (130), for each of a plurality of imaging procedures relating to a corresponding body part of a subject, a plurality of incomplete sample sets, each of the incomplete sample sets comprising at least one sample target image representative of the corresponding body part of the subject having been administered a contrast agent at a sample target dose, and at least one sample baseline image representative of the corresponding body part of the subject having no contrast agent or having been administered a contrast agent at a sample baseline dose lower than the sample target dose, simulating (411-445; 450-456), by the computing system (130), a plurality of sample source images, at least one of the sample source images being simulated from each of at least one of the incomplete sample sets of each of the imaging procedures to represent a corresponding body part mimicking administration of a contrast agent to the corresponding subject at a sample source dose lower than the sample target dose, wherein a ratio between the sample source dose and the sample target dose is equal to a downscaling factor, generating (446; 457), by the computing system (130), for each of the imaging procedures, one or more complete sample sets, each of the complete sample sets comprising the sample baseline image and the sample target image of one of the incomplete sample sets of the imaging procedure and the sample source image simulated from another of the incomplete sample sets of the imaging procedure, training (458-473), by the computing system (130), the machine learning model (305) to optimize an ability of the machine learning model (305) to generate the sample target image of each of the complete sample sets from the sample baseline image and the sample source image of the complete sample set, and deploying (475), by the computing system (130), the machine learning model (305) trained for use in medical imaging applications to mimic an increase in a contrast agent dose administered to a corresponding patient according to a zooming factor corresponding to an inverse of the downscaling factor.

2. The method (400) of claim 1, wherein the sample source dose is intermediate between the sample baseline dose and the sample target dose.

3. The method (400) of claim 1 or 2, wherein the target dose is a full dose of contrast agent that is standard in clinical practice.

4. The method (400) of any one of claims 1 to 3, wherein at least a portion of the subjects are animals, and wherein the patient is a human.

5. The method (400) of any one of claims 1 to 4, wherein the method (400) comprises: receiving (403), by the computing system (130), a plurality of sample images for each of the imaging procedures, the plurality of sample images for each of the imaging procedures including one or more of the sample baseline images and one or more of the sample target images, and creating (404-410), by the computing system (130), an incomplete sample set for each imaging procedure based on the sample images for that imaging procedure.

6. The method (400) of claim 5, wherein the method (400) comprises: generating (407-408), by the computing system (130), one or more new sample baseline images and / or one or more new sample target images for each of one or more of the imaging procedures from the respective received corresponding sample baseline images and / or sample target images.

7. The method (400) of claim 6, wherein the method (400) comprises: generating (407-408), by the computing system (130), the new sample baseline images and / or the new sample target images for each imaging procedure such that the number of sample baseline images is equal to the number of sample target images for that imaging procedure.

8. The method (400) of any one of claims 5-7, wherein the method (400) comprises: discarding (405-406), by the computing system (130), one or more of the sample baseline images and / or one or more of the sample target images for each of one or more of the imaging procedures.

9. The method (400) of claim 8, wherein the method (400) comprises: discarding (405-406), by the computing system (130), sample baseline images and / or sample target images for each imaging procedure such that the number of sample baseline images is equal to the number of sample target images for that imaging procedure.

10. The method (400) of any one of claims 1-9, wherein the method (400) comprises: creating (410), by the computing system (130), an incomplete sample set for each of one or more of the imaging procedures by: combining a single one of the sample baseline images for that imaging procedure with each of a plurality of sample target images, combining the same number of sample baseline images with sample target images one-to-one for that imaging procedure, or combining sample baseline images with sample target images for that imaging procedure in all possible ways.

11. The method (400) of any one of claims 1-10, wherein the method (400) comprises: generating (446; 457), by the computing system (130), a complete sample set for each of one or more of the imaging procedures by: combining the incomplete sample set one-to-one with sample source images simulated from another half of the incomplete sample set of the imaging procedure, combining the incomplete sample set in all possible ways with sample source images simulated from other incomplete sample sets of the incomplete sample set of the imaging procedure, or combining the incomplete sample set one-to-one with sample source images simulated from other incomplete sample sets of the incomplete sample set of the imaging procedure.

12. The method (400) of any one of claims 1 to 11, wherein each of the sample baseline images, each of the sample target images, and each of the sample source images comprises a plurality of sample baseline values, a plurality of sample target values, and a plurality of sample source values, respectively, the method (400) comprising: computing (418; 436), by the computing system (130), each of the sample source values of each of the sample source images by applying a simulation formula that depends on the reduction factor, the simulation formula being derived from a signal law that expresses an amplitude of a response signal of the body part as a function of a local concentration of the contrast agent. 426； 13. The method (400) of claim 12, wherein the simulation formula is derived from a signal law that is linearized with respect to the local concentration of the contrast agent.

14. The method (400) of claim 13, wherein the simulation formula is derived from a signal law that assumes a local concentration of the contrast agent to be proportional to a dose.

15. The method (400) of any one of claims 12 to 14, wherein the sample baseline values, the sample target values, and the sample source values represent response signals of corresponding locations of the body part, the method (400) comprising: modulating (417; 425), by the computing system (130), a reduction factor to be used for computing each of the sample source values of each of the sample source images as a function of an indication of a local concentration of the contrast agent at the corresponding location derived from the corresponding sample target value.

16. The method (700) of any one of claims 1 to 15, wherein the method (400) comprises: injecting (420-421; 438-442), by the computing system (130), artificial noise into each of the sample source images, the artificial noise having a statistical distribution that depends on the reduction factor.

17. The method (400) of any one of claims 1 to 16, wherein the method (400) comprises: 428-429； training (450), by the computing system (130), a further machine-learned model (370) to optimize an ability of the further machine-learned model (370) to generate the sample source images of each of the at least a portion of the complete sample sets from corresponding sample baseline images and sample target images, and ​ ​ generating (451-456), by the computing system (130), a refined version of each of the sample source images by applying a sample baseline image and a sample target image of a corresponding incomplete sample set to the further machine learning model (370) being trained.

18. The method (400) of any one of claims 1 to 17, wherein the method (400) comprises: repeating (474), by the computing system (130), the simulating (411-445; 450- 456) of sample source images, the generating (446; 457) of complete sample sets, and the training (458-473) of machine learning models (305) for a plurality of values of a downscaling factor, and deploying (475), by the computing system (130), the machine learning models (305) trained with the values of the downscaling factor in corresponding configurations for selecting one or more corresponding values of an upscaiiing factor in each of the medical imaging applications.

19. The method (400) of any one of claims 1 to 18, wherein the machine learning models (305) are neural networks.

20. A method of using a machine learning model (305) trained with the method (400) of any one of claims 1 to 18 in a medical imaging application for imaging a body part of a patient, wherein the method comprises, under the control of a computing system (115): receiving, by the computing system (115), one or more operational administration images representative of a body part of a patient to which contrast agent has been administered at an operational administration dose, and at least one operational baseline image representative of a body part of a patient to which no contrast agent or contrast agent has been administered at an operational baseline dose lower than the operational administration dose, simulating, by the computing system (115), from the operational baseline image and the operational administration images, a corresponding operational simulation image representative of a body part of a patient mimicking administration of contrast agent at an operational simulation dose higher than the operational administration dose, with a ratio between the operational simulation dose and the operational administration dose equal to an upscaiiing factor corresponding to an inverse of a downscaling factor of the machine learning model, and outputting, by the computing system (115), a representation of the body part based on the operational simulation image.

21. A computer program (300) configured for causing a computing system (130; 115) to perform the method (500) of any one of claims 1 to 20 when the computer program (500) is executed on the computing system (130; 115).

22. A computer program product comprising one or more non-transitory computer-readable storage media having program instructions stored on the readable storage media, the program instructions being loadable into an internal memory of a computing system, to cause the computing system to perform the method of any one of claims 1 to 20 when the program instructions are executed on the computing system.

23. A computing system (130; 115) comprising means (500) configured to perform the steps of the method according to any one of claims 1 to 20.

24. A computing system comprising circuitry for performing each step of the method according to any one of claims 1 to 20.

25. A medical method applied to a body part of a patient, wherein the medical method comprises: acquiring at least one operational baseline image representative of the body part, administering a contrast agent to the patient at an operational administration dose, acquiring one or more operational administration images representative of the body part in response to the administration of the contrast agent to the patient, a corresponding operational simulation image being simulated from the operational baseline image and the operational administration image, and a representation of the body part based on the operational simulation image being output according to the method of claim 20, and performing a medical procedure related to the body part according to the representation of the body part.

Images