Medical Image Enhancement

Abstract

A method and apparatus for enhancing magnetic resonance images to generate contrast-enhanced images without the need to administer contrast to a patient, the image processing apparatus utilizing a trained machine learning algorithm, preferably a generative adversarial network, as an image processor to generate images from contrast-free magnetic resonance images, the generated images having similar appearance, better image quality, and better pathological sensitivity than actually acquired contrast-enhanced images, and being able to distinguish more pathological conditions than actually acquired contrast-enhanced images.

Description

[Technical field]

[0001] The present invention relates to methods and apparatus for generating improved computed contrast-enhanced medical images from dedicated quantitative imaging data sets without reliance on injectable contrast agents. [Background technology]

[0002] Medical imaging is a ubiquitous technique for assisting clinicians by producing images of a patient's anatomy and physiological processes within the patient's body. There are a wide variety of medical imaging techniques or modalities, such as x-ray or computed tomography (CT) imaging, nuclear medical imaging such as positron-emission tomography (PET) or single-photon emission computerized tomography (SPECT), and magnetic resonance (MR) imaging (MRI).

[0003] The present invention relates to MRI. In general, an MR image is a two-dimensional representation of contrast resulting from differences in the magnetic properties of tissues in a patient. *Many different types of MRI sequences or "modalities" are well known, such as T2-weighted MRI and T1-weighted MRI. These sequences are routinely preprogrammed into the MRI scanner and use different combinations of static and oscillating magnetic fields and radio frequency (RF) pulses, the effect of which is that the contrast in the resulting images depends more strongly on the choice of the desired modality (e.g., T1 or T2) and parameters. In addition to the MR imaging modalities mentioned above, other known modalities are fat-water imaging, double inversion recovery T2-weighted imaging, STIR (short-tau inversion recovery) T2-weighted imaging, tagged CMR (cardiovascular magnetic resonance) imaging, strain-encoded imaging, diffusion-weighted imaging, diffusion tensor imaging, and arterial spin labelling imaging. Additionally, in cardiac MRI, so-called "cine images" may be acquired by imaging the patient over several heartbeats to produce a stack of images that may be formed into frames of video images showing the heartbeat, blood flow, and other types of motion.

[0004] An important development in medical imaging has been the introduction of contrast-enhanced (CE) imaging, in which images are acquired following administration of a contrast agent to the patient. In such images, the contrast in the image reflects the location of the contrast agent and can enhance the distinction between tissues. Furthermore, the contrast in the image reflects the wash-in and wash-out of the contrast agent and can therefore give additional pharmacokinetic information regarding the pathophysiology of the patient's tissues. Contrast agents can be used with X-ray (CT) and MR imaging, and nuclear medical imaging essentially involves the administration of a radiotracer, which for these purposes can be considered as a contrast agent that produces a contrast-enhanced image.

[0005] In cardiac MR imaging, paramagnetic contrast agents, such as gadolinium-based contrast agents (GBCAs), are commonly used. The most common applications of GBCAs include first pass perfusion, early gadolinium enhancement (EGE), late gadolinium enhancement (LGE) imaging, and angiographic studies. In particular, LGE imaging has become the gold standard in clinical cardiology to assess myocardial viability and differentiate ischaemic from nonischemic heart disease. This may affect clinical decision-making on whether a patient should undergo an invasive procedure (such as open-heart surgery to open up blockages in the coronary arteries or invasive percutaneous coronary intervention (PCI), or simpler noninvasive treatments). Clinicians have become very adept at interpreting CE MR images and therefore readily desire to refer to such CE medical images to aid them in diagnosing various medical conditions. Combining CE results with non-contrast imaging allows the calculation of other quantities of interest such as extracellular volume (ECV), salvageable risk area in acute myocardial infarction, and viable myocardium.

[0006] However, CE imaging typically requires intravenous cannulation of the subject and administration of contrast, which is not only invasive but also lengthens the scanning procedure, typically to over 45 minutes for LGE imaging. Long scanning procedures not only make it difficult for the subject to remain motionless, a requirement for good quality images, but also reduce patient tolerance inside the scanner. Furthermore, in some applications, such as cardiac MRI, the subject's heart rate and ability to hold their breath can significantly limit the clinically acceptable duration of the sequence.

[0007] Contrast agents may also be contraindicated in some classes of patients (e.g., patients with severe renal impairment who are unable to effectively excrete the contrast agent and may develop severe irreversible complication nephrogenic systemic fibrosis). Contrast agents such as GBCAs may be less desirable in others (e.g., pregnant or lactating women, and children). For example, evidence of GBCA accumulation in the brains of patients who have undergone multiple MRI examinations has recently been published without any clear prediction regarding the clinical significance of these brain accumulations. Many patients require serial examinations to monitor disease and response to treatment, and thus the potential risk of complications may increase substantially. GBCAs are given as an example, with similar limitations for other contrast agents, such as iron oxide, manganese in MRI, and iodine contrast agents in CT.

[0008] Despite these challenges, CE medical imaging is routinely used in clinical practice. For example, LGE has traditionally been considered to contain crucial and irreplaceable information for differential diagnosis, and thus GBCAs are routinely administered despite the additional cost, (albeit small) risk, and burden to the patient. The ubiquity of LGE allows many clinicians to recognize LGE images and to map observed patterns to specific symptoms.

[0009] Traditionally, interest in MRI images was limited to the variation of contrast across the image, rather than the absolute values ​​of the magnetic parameters (T1 or T2); more recently, based on the acquisition and analysis of variably weighted MR data sets, the following parameters have been determined: T1, T1-ρ, T2, T2 *MRI sequences that allow quantitative mapping of certain fundamental magnetic properties of materials, such as T1 mapping, T2 mapping, T3 mapping, T4 mapping, T5 mapping, T6 mapping, T7 mapping, T8 mapping, T9 mapping, T1 ... Mapping techniques have shown that sensitivity to earlier disease than LGE patterns can be observed (Ferreira et al., "Native T1-mapping detects the location, extent and patterns of acute myocarditis without the need for gadolinium contrast agents," J Cardiovasc Magn Reson 16, 36 (2014); Dass et al., "Myocardial Tissue Characterization Using Magnetic Resonance Noncontrast T1 Mapping in Hypertrophic and Dilated Cardiomyopathy," Cardiovascular Imaging. 2012;5:726-733). [Prior art documents] [Patent documents]

[0010] [Patent Document 1] US Patent Application Publication No. 2013 / 0322713 [Non-patent literature]

[0011] [Non-Patent Document 1] Ferreira et al., “Native T1-mapping detects the location, extent and patterns of acute myocarditis without the need for gadolinium contrast agents,” J Cardiovasc Magn Reson 16, 36 (2014) [Non-Patent Document 2] Dass et al., “Myocardial Tissue Characterization Using Magnetic Resonance Noncontrast T1 Mapping in Hypertrophic and Dilated Cardiomyopathy”, Cardiovascular Imaging.2012;5:726-733 [Non-Patent Document 3] Abdula et al., "Synthetic late gadolinium enhancement cardiac magnetic resonance for diagnosing myocardial scar," Scandinavian Cardiovascular Journal, 52:3, pp. 127-132. [Non-Patent Document 4] Phillip Isola et al., “Image-to-Image translation with Conditional Adversarial Networks,” Berkeley AI Research (BAIR) Laboratory, UC Berkeley [Non-Patent Document 5] Ting-Chun Wang et al., “High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs”, NVIDIA Corporation, CVPR 2018 Summary of the Invention [Means for solving the problem]

[0012] According to a first aspect of the present invention there is provided a method of generating a computed contrast-enhanced medical image comprising receiving an image dataset of a subject, the image dataset comprising native quantitative mapping images acquired by performing a contrast-free magnetic resonance imaging procedure, inputting the image dataset into an image processor and processing the image dataset using the image processor to generate a computed contrast-enhanced medical image, the image processor comprising a machine learning processor trained on a training dataset comprising a set of images, each set of images comprising a contrast-free native quantitative mapping image having the same quantitative mapping as the quantitative mapping images of the image dataset and a corresponding acquired contrast-enhanced medical image.

[0013] Thus, with the present invention, the need to administer contrast to the subject is eliminated by training an image processor to predict what the calculated contrast-enhanced medical image will be from the contrast-free quantitative mapping image. This was previously thought to be impossible because the pharmacokinetic aspects of contrast-enhanced imaging are different from those of native imaging. Now, the inventors have discovered that contrast-free quantitative mapping gives the same pathological sensitivity as CE medical images or better pathological sensitivity than CE medical images. As a result, by training a machine learning processor on a set of images, one image of the set being a quantitative mapping image acquired from the subject without contrast and another image of the set being a corresponding medical image acquired after administration of contrast, the trained machine learning processor can generate a calculated CE medical image from the contrast-free quantitative mapping image. The calculated CE medical image has better image quality than the acquired CE medical image, and in some cases, as described below, has been assessed as having a better sensitivity to disease, e.g., edema, than the acquired CE medical image.

[0014] The computed CE medical images provide a standardized presentation of quantitative mapping images and a way to achieve different disease sensitivities by specifying the training dataset and MR modality.

[0015] Computed CE medical images are free of common CE artifacts (e.g., LGE artifacts) and are more sensitive to pathological changes (including edema and global expansion of intracellular matrix in diffuse pathology) than CE (known to be related to focal fibrosis and blood-brain barrier breakdown in the brain). Thus, computed CE medical images that show a similar appearance to CE (e.g., LGE) have better image quality than real CE (e.g., LGE) and can detect more extensive pathology than real CE (e.g., LGE). This is believed to result from using machine learning to present them with a similar appearance to CE images while at the same time leveraging information in native quantitative mapping. This is achieved without the need to administer contrast agents to the subject and the associated redundant scanning procedures.

[0016] The machine learning processor is trained to process the contrast-free quantitative mapping images and generate a calculated CE medical image that matches to a predetermined degree with a corresponding acquired contrast-enhanced medical image as a target or "ground truth", e.g., by minimizing the difference (or reducing the difference below a predetermined threshold), with the goal of making the calculated CE medical image indistinguishable from the corresponding actually acquired CE medical image by a simultaneously trained machine learning classifier. Because the acquired CE medical image is used as a target or ground truth image in the machine learning, the calculated CE medical image is effectively the same type of image as the acquired CE medical image.

[0017] As is well known, machine learning processors use data processing algorithms to generate target output data from input data, and the algorithms are improved in an iterative training process in which input data from a training dataset (for which the target data is known) is processed by the machine learning processor, the output is compared to the target data, and the algorithm is adjusted (e.g., by backpropagation in deep learning), and this process is repeated until the output is sufficiently close to the target, whereupon the machine learning processor is considered to have been trained and can be used to process new input data.

[0018] The contrast-free quantitative mapping image may be, for example, a T1 mapping image, including T1-ρ mapping or stress T1 mapping. The contrast-free quantitative mapping image may alternatively be a quantitative mapping image other than T1 mapping, for example, T2 mapping or T2 * It may have a mapping.

[0019] The image processor may also utilize information in one or more additional contrast-free MR images. Thus, the method may include acquiring magnetic resonance images of the subject from a number of different contrast-free sequences, inputting the images into an image processor, and processing the images to generate enhanced (computed CE) images to standardize the appearance of the mapping technique.

[0020] The corresponding acquired CE medical image (and therefore effectively the calculated CE medical image) may be a CE magnetic resonance image, such as an LGE image.

[0021] In one example, the corresponding acquired CE medical image may be a quantitative mapping image having the same or different quantitative mapping as the quantitative mapping image of the image dataset, which is itself useful for standardizing the appearance of T1 mapping for immediate clinical diagnosis.

[0022] It is foreseeable to use synthetic LGE images (Abdula et al., "Synthetic late gadolinium enhancement cardiac magnetic resonance for diagnosing myocardial scar", Scandinavian Cardiovascular Journal, 52:3, pp. 127-132) and segmented lesion masks (Ferreira et al., "Native T1-mapping detects the location, extent and patterns of acute myocarditis without the need for gadolinium contrast agents", J Cardiovasc Magn Reson 16, 36 (2014)), or a combination of such derived images with contrast-enhanced images, in training.

[0023] However, in alternative embodiments, the corresponding acquired CE medical image (and therefore effectively the computed CE medical image as well) may be a computed contrast-enhanced image of a non-MR modality, e.g. a contrast-enhanced CT image, such as a late iodinated-contrast enhancement CT image, a contrast-enhanced PET image, a contrast-enhanced SPECT image, or a contrast-enhanced ultrasound image. In general, it may be any type of imaging that employs the administration of a contrast agent or the introduction of physiological stress to generate additional enhancement, particularly in any form already accepted as standard in medical applications.

[0024] In one type of embodiment, the image dataset of the subject may only include native quantitative mapping images. This may be useful when the method is applied to standardize quantitative mapping. In another type of embodiment, the image dataset of the subject may further include other images of the subject and / or further data that is not image data. Thus, the image processor may include one or more convolution streams for taking one or more image modalities as input, and / or streams for taking MR metadata or non-imaging diagnostic information as input. Next, some non-limiting examples that may be used in any combination are described.

[0025] The image dataset of the subject may further include at least one further magnetic resonance image acquired by a contrast-free magnetic resonance modality other than the quantitative mapping of the quantitative mapping image of the image dataset, in which case the set of images of the training dataset further includes at least one further magnetic resonance image acquired by the other contrast-free magnetic resonance modality.

[0026] In some instances, the at least one additional magnetic resonance image may include at least one of a raw magnetic resonance image, an image that is a fusion of the raw magnetic resonance image, or an image that is a derivative of the raw magnetic resonance image.

[0027] In some instances, the at least one additional magnetic resonance image is a T1 mapping image, a T1 weighted image (including a T1 inversion recovery weighted image), a T2 weighted image, a T2 * Weighted image, T2 mapping image, T2 * The images may include at least one of a mapping image, or a cine CMR image.

[0028] In some instances, the at least one additional magnetic resonance image includes at least one of a STIR image, a tagged CMR image, a strain-encoded image, a diffusion-weighted image, a diffusion tensor image, an arterial spin labeling image, a PD-weighted image, or a fat-water separation image.

[0029] The image dataset of the subject may further include at least one non-MR image without contrast, in which case the set of images of the training dataset further includes at least one non-MR image of the same type as the at least one non-MR image of the image dataset.

[0030] In some instances, the at least one non-magnetic resonance image may include at least one of an echocardiogram, a nuclear perfusion image, a CT image, an electrophysiological cardiac map image, or a chest x-ray.

[0031] At least one non-magnetic resonance image can be in a native baseline, stress or pharmacologically induced state.

[0032] The method may further comprise the step of inputting into the image processor at least one acquired CE medical image already available from a previous visit or study of the same subject.

[0033] The CE medical images may be substituted by derived segmented masks based on the CE medical images or other clinically actionable human segmentations of the images. Such processed masks may be derived from such images for clinically meaningful parameters (such as indicative of levels of edema, fibrosis, infarction, microangiopathy, or similar disease pathology).

[0034] The image dataset may include further data that is not image data. In general, the further data may be data related to the subject and / or data related to the captured image.

[0035] In some instances, the additional data may include at least one of imaging metadata, such as image acquisition parameters, or at least one non-imaging diagnostic test result.

[0036] The imaging metadata may include at least one of the following: MR device manufacturer model name or designation, magnetic field strength, software version, inversion time, and the like.

[0037] The non-imaging diagnostic test results may include at least one of MR spectroscopy results, blood test results, an electrocardiogram, clinical characteristics of the subject (e.g., medical condition, medications, symptoms, subject risk factors, subject history, physical exam findings), or reasons for subject referral.

[0038] The image processor includes a trained machine learning algorithm, process, or processor, which may be, for example, one of a trained Generative Adversarial Network (GAN), a trained Conditional GAN, a trained Fully Convolutional Neural Network, a trained variational autoencoder, a trained U-Net, a trained V-Net. The image processor may have multiple input streams to match the number of input modalities, and / or a non-convolutional stream for taking non-image information as input.

[0039] The computed CE medical images provide a standardized presentation of quantitative mapping images and a way to achieve different disease sensitivities by specifying the training dataset and the CMR modality.

[0040] The technique is data-driven and inherits information from contrast-free mapping, which means that the machine learning processor can be trained on a training dataset of multiple corresponding acquired CE medical images with different characteristics (e.g., with edema and without myocardial edema, myocarditis, acute myocardial infarction, and / or chronic myocardial infarction). The image dataset can then be processed with the image processor to generate multiple computed contrast-enhanced medical images such that the computed contrast-enhanced medical images have different sensitivities (e.g., to detect only chronic MI scar, or to detect chronic MI scar and edema).

[0041] Optionally, such multiple computed contrast-enhanced medical images can be combined to generate a combined contrast-enhanced medical image for further enhancement to generate a new comprehensive computed CE medical image that can distinguish disease states (e.g., chronic MI scar from edema) better than a single computed CE medical image or a single legacy CE medical image.

[0042] As an alternative to combining only the calculated CE medical images, it is also possible to include but not be limited to combining both the calculated CE medical images and the acquired CE medical images to generate a combined calculated CE image.

[0043] According to a first aspect of the invention, the machine learning processor has been trained as described above. Optionally, the step of training the machine learning processor may be part of the method.

[0044] According to a first aspect of the invention, an image dataset of a subject is received and input to an image processor. Optionally, the step of acquiring the image dataset of the subject may be part of the method, for example performing a contrast-free magnetic resonance imaging procedure on the subject to acquire an image dataset including native quantitative mapping images, and also performing other imaging procedures to acquire other images, when other images are used.

[0045] According to a second aspect of the present invention there is provided an image processor adapted to generate a calculated contrast-enhanced medical image, the image processor comprising an input for receiving an image dataset of a subject including quantitative mapping images acquired by performing a contrast-free magnetic resonance imaging procedure, and a data processor for processing the input image dataset to generate a calculated contrast-enhanced medical image, the data processor comprising a machine learning processor trained on a training dataset including a set of images, each set of images including a contrast-free native quantitative mapping image having the same quantitative mapping as a quantitative mapping image of the image dataset and a corresponding acquired contrast-enhanced medical image.

[0046] The image processor of the second aspect of the invention performs an equivalent method to the first aspect of the invention, and therefore the various features discussed above are equally applicable to the second aspect of the invention.

[0047] According to a third aspect of the present invention there is provided a method of training said data processor comprising: a) receiving a training data set comprising corresponding sets of images including a contrast agent-free native quantitative mapping image of the subject, the corresponding acquired contrast-enhanced images of the subject and, optionally, at least one further image of the subject and / or further data which are not images; b) inputting into a data processor each set of contrast agent-free native quantitative mapping images and, if present, the at least one further image of the subject and / or further data of each set, and processing same with the data processor using a generative image processing function to generate a calculated contrast-enhanced medical image and comparing the calculated contrast-enhanced medical image with the corresponding acquired contrast-enhanced image from the set; c) modifying the processing performed by the data processor so as to reduce a difference between the calculated contrast-enhanced medical image and the corresponding acquired contrast-enhanced image from the set; and d) repeating steps b) and c) until the difference between the calculated contrast-enhanced medical image and the corresponding acquired contrast-enhanced image is below a predefined threshold.

[0048] The method of the third aspect of the invention may be used to train the image processor in the first and second aspects of the invention and therefore the various features discussed above may be equally applied to the third aspect of the invention.

[0049] Preferably, the method further comprises operating an adversarial discriminator to distinguish between the computed contrast-enhanced medical images generated by the generative image processing function and the corresponding acquired contrast-enhanced images by classifying each image as either a computed contrast-enhanced image generated by the generative image processing function or a contrast-enhanced image, the processing performed by the generative image processor being modified in step c) to reduce a cost function that includes both the difference between the computed contrast-enhanced medical image and the corresponding acquired contrast-enhanced image and the classification confidence. The adversarial discriminator is simultaneously trained to increase its classification accuracy by reducing a separate loss function of the classification, for example a cross-entropy loss.

[0050] According to a further aspect of the present invention there is provided a computer program executable by a computing device which, when executed, is configured to cause the computing device to perform a method according to the first aspect of the present invention, a computer readable storage medium storing such a computer program, and a computing device configured to perform a similar method.

[0051] The invention may also provide a magnetic resonance scanner system or scanner operating system adapted to perform or control a contrast-free MR imaging procedure and to process the obtained MR images to provide an in-line enhanced medical image on the scanner according to the method of the invention. Thus, the invention includes a scanner capable of providing a calculated CE medical image as one of the available output options for the operator.

[0052] The invention will now be further described, by way of example only, with reference to the accompanying drawings in which: [Brief description of the drawings]

[0053] [Figure 1(A)] FIG. 1 illustrates a schematic diagram of an apparatus for processing images of a single MR native (i.e., contrast-free) T1 map according to an embodiment of the present invention; [Figure 1(B)] FIG. 5 shows a schematic diagram of an apparatus for processing MR native TI map images complemented by an additional MR sequence according to a second embodiment of the invention. [Figure 2(A)] FIG. 2 illustrates a schematic diagram of a process for training an apparatus for processing MR native T1 map images according to an embodiment of the present invention. [Figure 2(B)] FIG. 2(A) shows the training data set fed into the training process shown in FIG. [Figure 3(A)] FIG. 4 illustrates diagrammatically the steps of training a generator and a discriminator in an apparatus for processing MR native T1 map images according to another embodiment of the present invention. [Figure 3(B)] FIG. 4 illustrates diagrammatically the steps of training a generator and a discriminator in an apparatus for processing MR native T1 map images according to another embodiment of the present invention. [Figure 4(A)] FIG. 4 illustrates diagrammatically exemplary steps of training a generator and a discriminator in an apparatus for processing MR native T1 map images according to another embodiment of the present invention. [Figure 4(B)] FIG. 4 illustrates diagrammatically exemplary steps of training a generator and a discriminator in an apparatus for processing MR native T1 map images according to another embodiment of the present invention. [Diagram 5] FIG. 5 shows a schematic illustration of an example of a generator architecture for use in the apparatus of the embodiment of FIG. 4. [Figure 6] FIG. 13 illustrates an alternative generator architecture that utilizes four streams of input information. [Figure 7] FIG. 5 illustrates a schematic of a discriminator architecture used in the embodiment of FIG. 4; [Figure 8]FIG. 1 illustrates training multiple image processor units on different datasets to gain sensitivity to different pathologies. [Figure 9] FIG. 1 illustrates combining multiple c-CE images, each generated by a differently trained image processor unit, to generate a combined c-CE image that is capable of distinguishing between different types of pathologies. [Figure 10(A)] FIG. 1 shows a schematic diagram of an MRI procedure according to the prior art. [Figure 10(B)] FIG. 1 illustrates a schematic diagram of an MRI procedure, according to one embodiment of the present invention. [Figure 11(A)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 11(B)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 12] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 13] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 14] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 15] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 16(A)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 16(B)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 17]FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 18(A)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 18(B)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 19(A)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 19(B)] FIG. 1 illustrates a comparison of a native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 20] 1 is a graph showing the match between the calculated LGE and LGE in detecting the presence of LGE enhancement. [Figure 21] 13 is a graph of image quality for c-CE images and acquired CE images. [Figure 22] FIG. 1 shows a comparison of a weighted native T1 scan, a calculated LGE scan, and an LGE scan obtained according to one embodiment of the present invention. [Figure 23] FIG. 1 shows combined computed CE images that can distinguish between normal tissue, myocardial edema, and myocardial infarction scar. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0054] In this text, reference is made to a magnetic resonance imaging or scanning "procedure". A procedure is performed according to a protocol and may include one or several scan sequences, each sequence producing differently weighted raw images and their fusion and derivatives such as quantitative maps, and may include non-MR imaging steps such as blood tests and contrast administration. The protocol may also set out aspects relevant to the subject, such as rest times and breath-holding requirements.

[0055] FIG. 1 is a schematic diagram of an apparatus implementing a method for processing contrast-free magnetic resonance T1 mapping images to generate computed contrast-enhanced medical images according to a first embodiment of the present invention.

[0056] The device 1 comprises an input 3 for receiving an image dataset 100, which in this example comprises only T1 mapping images 101 (also called "T1 map images" or "T1 maps") previously acquired by performing a contrast-free magnetic resonance imaging procedure, e.g. MOLLI or ShMOLLI T1 mapping images in grey scale or colour maps, which may be any number of raw input T1 weighted images and phase maps with any associated metadata (e.g. inversion time) acquired using an experiment for T1 mapping. The T1 data may be replaced by T2 data or any other MR image dataset sufficient to predict the target type of the CE images.

[0057] In these examples, the T1 mapping image 101 (also the T1 mapping image 21 used in training and described below) is an example of an image with quantitative mapping. Any T1 mapping may be used, including T1-ρ mapping or stress T1 mapping. More generally, the T1 mapping image 101 (also the T1 mapping image 21 used in training and described below) may be, for example, a T2 mapping image or a T2 *The mapping images may be replaced by images with any other quantitative mapping. Such quantitative mapping is provided by the acquisition and analysis of variably weighted MR data sets. The resulting images are classified into T1, T1-ρ, T2, T2 * etc., giving a quantitative mapping of certain fundamental magnetic properties of the materials in the image.

[0058] The MR image data set 100 is passed to an image processor 5, which is a data processor that uses a trained machine learning (or artificial intelligence) process. The image processor 5 processes the input image data set 100 to generate a computed contrast-enhanced image 6 that has a similar appearance and pathological sensitivity to an acquired contrast-enhanced image that would have been acquired if the subject had been scanned after administration of a contrast agent. This image is referred to herein as a computed CE medical image (c-CE image) 6. The c-CE image may be displayed on a display 7 or otherwise output.

[0059] In the example of FIG. 1(A), the image processor 5 is trained on a training dataset that includes a set of images, each set of images including a contrast-free T1 mapping image having the same quantitative mapping as the T1 mapping image 101 of the image dataset 100 and a corresponding acquired contrast-enhanced medical image.

[0060] The contrast-free T1 mapping images of the training dataset are acquired by performing the same contrast-free magnetic resonance imaging procedure as the T1 mapping images 101 of the image dataset 100, for example a MOLLI or ShMOLLI acquisition.

[0061] The corresponding acquired contrast-enhanced medical images of the training dataset are the target or “ground truth” images for machine learning, so the computed c-CE images6 are effectively of the same type as the acquired CE medical images.

[0062] FIG. 1(B) shows a second embodiment of the present invention, in which the apparatus 1 is the same as the apparatus 1 of FIG. 1(A), except that the image dataset 100 is extended to include further data in addition to the T1 mapping image 101.

[0063] The further data of the image dataset 100 may include at least one further magnetic resonance image 102. The further magnetic resonance image 102 is acquired by a contrast-free magnetic resonance modality other than the quantitative mapping acquisition of the T1 mapping image 101. In some instances, the further magnetic resonance image 102 is acquired by a contrast-free magnetic resonance modality other than the quantitative mapping acquisition of the T1 mapping image 101. * weighted image, cine MR image, STIR, tagged CMR, strain image, diffusion image, PD weighted image, DTI, ASL, etc. Other examples are given above. The further magnetic resonance images 102 may be acquired by a single contrast-free magnetic resonance modality or by multiple different contrast-free magnetic resonance modalities.

[0064] An image dataset 100 is received at an input 3 and passed to an image processor 5 which operates as described above to generate a c-CE medical image 6 which is displayed on a display 7 or otherwise output. In this case, the image processor 5 has been trained on the training dataset as described above, except that the set of images of the training dataset further comprises at least one further magnetic resonance image acquired with another contrast-free magnetic resonance modality which corresponds to the contrast-free magnetic resonance modality of the further magnetic resonance image 102 of the image dataset 100.

[0065] Further data of the image dataset 100 may include at least one non-magnetic resonance image 103. The non-magnetic resonance image 103 is acquired by a non-contrast non-magnetic resonance modality. Some examples are given above. The non-magnetic resonance image 103 may be acquired by a single modality or by multiple different modalities.

[0066] An image dataset 100 is received at the input 3 and passed to the image processor 5 which operates as described above to generate a c-CE medical image to be displayed on the display 7 or otherwise output. In this case, the image processor 5 has been trained on the training dataset as described above, except that the set of images of the training dataset further comprises at least one non-magnetic resonance image acquired by a non-magnetic resonance modality without contrast, corresponding to the modality of the non-magnetic resonance image 103 of the image dataset 100. The further data of the image dataset 100 may comprise further data 104 which is not image data. The further data 104 may be data relating to the subject and / or data relating to the input image. The additional data 104 may be imaging metadata, such as image acquisition parameters, or at least one non-imaging diagnostic test result for the subject, such as blood test results, electrocardiogram, clinical characteristics (e.g., medical condition, medications, symptoms, risk factors, history, physical examination findings), the reason for the subject's referral for imaging, and other imaging diagnostic tests (echocardiogram, nuclear perfusion imaging, CT scan, electrophysiological cardiac mapping, chest x-ray, etc.). Other examples are provided above.

[0067] Although three different types of further data are shown, namely further magnetic resonance images 102, non-magnetic resonance images 103 and further data 104. However, it is not mandatory to use all these types of further data and more generally any one or any combination of the different types of further data may be used, in which case the training data set will contain the corresponding combination of data.

[0068] An image dataset 100 is received at input 3 and passed to image processor 5 which operates as described above to generate a c-CE medical image that is displayed on display 7 or otherwise output. In this case, image processor 5 has been trained on a training dataset as described above, except that the training dataset further comprises further data corresponding to that of further data 104 of image dataset 100.

[0069] 1(B) may be used together in any combination, i.e. at least one further magnetic resonance image 102, at least one non-magnetic resonance image 103, and further data 104 may be used in any combination, and the training data set includes a corresponding combination of data.

[0070] The image processor 5 is thus a trained machine learning process that has been trained on a training dataset whose contents correspond to the stream of data of the image dataset 100 input to the apparatus 1 together with corresponding acquired contrast-enhanced medical images acquired through the use of a contrast agent.

[0071] The acquired CE medical images may be MR images. In this case, the acquired CE medical images may be MR images that are not quantitative mapping images. Alternatively, the acquired CE medical images may be MR images that are quantitative mapping images, optionally with the same quantitative mapping as the T1 map image 101 of the image dataset 100 or with a different quantitative mapping. In this case, the images are acquired by performing a contrast-enhanced magnetic resonance imaging procedure.

[0072] Alternatively, the acquired contrast-enhanced medical images of the training data set may be contrast-enhanced medical images of a non-magnetic resonance modality, for example contrast-enhanced CT, PET or SPECT images, or ultrasound images acquired based on the administration of a contrast agent (injected, orally, inhaled, etc.) or using the introduction of physiological stresses to generate additional enhancement depending on pathophysiological tissue properties, in which case they are acquired by performing a relevant contrast enhancement procedure of the non-magnetic resonance modality.

[0073] Examples of suitable machine learning algorithms that may be employed in the image processor 5 are fully convolutional neural networks, variational autoencoders, U-Net, dense U-Net, V-Net, conditional generative adversarial networks (GANs), cascade refinement networks, including variants thereof such as cycle GANs.

[0074] The image processor 5 may be trained prior to execution of the method performed by the device 1 shown in FIG. 1(A) or FIG. 1(B) to provide a trained machine learning processor.

[0075] Figure 2(A) shows how the image processor 5 is trained. In particular, Figure 2(A) shows the functional blocks that perform the training method, as follows.

[0076] 2B shows a training data set 20 used in training. The training data set 20 includes a set of images 25, each set of images 25 including the following data:

[0077] As described above, each set of images in the training dataset 20 includes a contrast-free T1 mapping image 21 having the same quantitative mapping as the T1 mapping image 101 in the image dataset 100, and an acquired contrast-enhanced medical image 26. The contrast-free mapping image 21 and the contrast-enhanced medical image 26 in each set correspond to each other in that they are acquired from the same subject or patient, e.g., an image acquired before a contrast agent is administered to the patient and an image acquired after a contrast agent is administered to the patient.

[0078] The acquired CE medical images and contrast-free T1 mapping images of each set 25 of images of the training dataset 20 may be acquired in the same MR scanning procedure, with images acquired at the same position relative to the subject, such that the images contain matching pathological structures. However, the images may alternatively be acquired from different procedures at different times. The images may be acquired, for example, using different scanners (if one procedure requires, for example, a different protocol or scanner configuration than another procedure). As described above, the acquired CE medical images may be T1 mapping images, but the acquired CE medical images may alternatively be MR images that are of a different modality than the T1 mapping (including, but not limited to, quantitative mapping images) or may be contrast-enhanced medical images generated by a non-magnetic resonance modality. The images may be from different procedures of the same or different imaging techniques, different procedures in the same or different patient positions, e.g., combining short-axis CMR images with long-axis CMR images to predict short-axis LGE using given image position and orientation information, and / or different procedures at the same or different times, e.g., procedures from previous visits of the same patient.

[0079] Typically, the set of images 25 of the training dataset 20 (or, in the context of using a trained processor, the input data) is data registered with the same patient.

[0080] Multiple sets of images 25 are used. In principle, any number of sets of images 25 may be used, but training is improved by increasing the number of sets of images 25 and the variation among the sets of images 25. Preferably, the training set includes 25 pairs of images from many different patients with different clinical conditions, a suitable training set may be obtained, for example, from the Hypertrophic Cardiomyopathy Registry, which has over 4000 suitable pairs of quality-controlled pre-contrast CMR images and corresponding LGE (late gadolinium enhancement) images.

[0081] When using the apparatus of FIG. 1(A), the set of images 25 of the training data set 20 includes only the T1 mapping image 21 and the acquired contrast-enhanced medical image 26, and does not include the further data described below. In this case, the "set" 25 of images is a pair of images. However, the training data set 20 may include further data, as follows:

[0082] When using the apparatus of FIG. 1(B), further data is used as follows.

[0083] If the image dataset 100 includes at least one further MR image 102, the set of images 25 of the training dataset 20 further includes at least one further magnetic resonance image 22 acquired by another contrast-free magnetic resonance modality that corresponds to the contrast-free magnetic resonance modality of the further magnetic resonance image 102 of the image dataset 100.

[0084] If the image dataset 100 includes at least one non-MR image 103, the set of images 25 of the training dataset 20 further includes at least one non-magnetic resonance image 23 acquired by a non-magnetic resonance modality without contrast corresponding to the modality of the non-magnetic resonance image 103 of the image dataset 100.

[0085] If the image dataset 100 comprises further data 104 , the set of images 25 of the training dataset 20 further comprises further data 24 that corresponds to the further data 104 of the image dataset 100 .

[0086] Training is performed using techniques known in the art of machine learning, as follows: A training data set 20 is received at an input 30 to a training processor 35, which includes a machine learning algorithm 32 that processes the contrast-free T1 mapping images 21 of each set of images 25 (as well as at least one further magnetic resonance image 22, at least one non-magnetic resonance image 23 and / or further data 24, if used) to generate a calculated CE image 31 of the same patient.

[0087] The computed CE images 31 are compared with the corresponding actual CE images 26 in the training dataset 20 and a cost function calculator 33 calculates a measure of their difference. The machine learning algorithm 32 is then iteratively modified, for example by backpropagation, to reduce the difference between the computed CE images 31 and the actual CE images 26. The data in the training dataset 20 may be augmented by rotation, translation, reflection, scaling, distortion, adding noise, etc. to improve the robustness of the learning process.

[0088] If it is determined that the calculated CE image 31 is sufficiently close to the actual CE image 26, the machine learning algorithm 32 is considered to be trained, and the machine learning algorithm 32 can then be used to process new contrast-free MR images in the device 1 to generate a calculated CE medical image 6.

[0089] 3(A) and 3(B) illustrate the training process for a particular example of machine learning algorithm 32, namely, a conditional generative adversarial network (conditional GAN). Conditional GANs are described, for example, in Phillip Isola et al., “Image-to-Image translation with Conditional Adversarial Networks,” Berkeley AI Research (BAIR) Laboratory, UC Berkeley, and Ting-Chun Wang et al., “High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs,” NVIDIA Corporation, CVPR 2018 (the teachings of which may be applied herein).

[0090] In this case, the machine learning algorithm 32 comprises the following functional blocks for implementing a training method: In particular, the machine learning algorithm 32 comprises a generator 51 and a discriminator 52. Fig. 3(A) shows the training step of the generator 51. The generator 51 is trained to process the contrast-free T1 mapping images 21 of each set of images 25 (also at least one further magnetic resonance image 22, at least one non-magnetic resonance image 23 and / or further data 24, if used) to generate a calculated CE image 31 that a) has a minimum pixel-wise difference with the acquired CE image 26 and b) is for the discriminator 52 indistinguishable from the acquired CE image 26.

[0091] Fig. 3(B) shows the training step of the discriminator 52 using backpropagation of the classification loss value 55. The discriminator 52 is trained to distinguish between acquired CE medical images 26 from the training dataset 20 and computed CE medical images 31 generated by the generator 51. The discriminator 52 is not provided with any information indicating which computed CE image 31 corresponds to which acquired CE image 26, it is only trained to distinguish between the two types of images and give an indication of how confident the discriminator 52 is about that distinction.

[0092] The difference block 53 derives the difference between the calculated CE image 31 and the acquired CE image 26, which is fed to a generator loss function block 54 together with a classification loss value 55 of the discriminator 52. The generator loss function block 54 trains the generator 51 by using backpropagation of the generation losses to iteratively modify the processing of the generator 51 to minimize the difference between the calculated CE medical image 31 and the corresponding acquired CE medical image 26 and increase the classification loss value 55 of the discriminator 52.

[0093] Once the calculated CE medical images 31 are sufficiently close to the actual CE images in the training data set, the generator 51 can be used to process new contrast-free MR images in one embodiment of the apparatus shown in Figure 1. The discriminator 52 does not need to be used in the image processor 5 of Figure 1.

[0094] 4(A) shows an example of the conditional GAN ​​method and training of a machine learning image processor using a T1 mapping image 21 and an acquired contrast-enhanced medical image 26, in this example an LGE MR image, as ground truth for training. The T1 mapping image 21 can be in grayscale or color map. A T1 color map is a T1 mapping image in which T1 values ​​are mapped to colors in the image according to a predefined mapping, such as that described in U.S. Patent Application Publication No. 2013 / 0322713.

[0095] As explained above with reference to FIG. 3, the generator 51 generates calculated CE images 31, in this case calculated LGE images, which are compared to actual acquired CE images 26, in this case LGE images, from the training dataset 20 such that the generator 51 may be trained to minimize the L1 distance loss between corresponding pairs of calculated CE medical images 31 and acquired CE medical images 26, and the calculated CE images 31 are also provided to the discriminator 52 such that the generator 51 may be trained to maximize the classification loss value 55 determined by the discriminator 52.

[0096] FIG. 4(B) shows an example of jointly training a machine learning discriminator 52 to discriminate between the computed CE image 31 and the acquired CE image 26 by minimizing a classification loss value 55.

[0097] FIG. 5 shows an example of a generator architecture used for generator 51, which is a standard U-Net structure, where the input to generator 51 from the training set 20 is the T1 mapping images 21, which are processed by convolutional neural network layers and concatenated as input for the U-Net.

[0098] Figure 6 shows a modified architecture for a generator 51, where the input to the generator 51 from the training dataset 20 is a T1 mapping image 21, at least one further magnetic resonance image 22, at least one non-magnetic resonance image 23 and further data 24.

[0099] Figure 7 shows the discriminator architecture used for the discriminator 52 in the embodiment of Figure 4. The computed CE medical images 31 and the acquired CE medical images 26 are processed by convolutional and fully connected layers to generate labels that distinguish between real or computed CE images with a confidence score.

[0100] 8 illustrates training the image processor 5 with different training datasets to obtain sensitivity to different disease conditions. In this example, the image processor 5 includes multiple units 5-1, 5-2, etc. trained on respective training datasets 20-1, 20-2, etc. having different characteristics, such as different sensitivity to different disease conditions (e.g., with edema and without myocardial edema, myocarditis, acute myocardial infarction, and / or chronic myocardial infarction). The multiple units 5-1, 5-2, etc. then process the image dataset 100 to generate c-CE images 6-1, 6-2, etc. having similarly different sensitivities to different diseases.

[0101] Such c-CE medical images 6-1, 6-2, etc. may be combined to generate a combined contrast-enhanced medical image for further enhancement. Figure 9 shows a specific example of generating a combined c-CE image 6-C, which may be implemented in the image processor 5. As shown, a T1 mapping image 101 is processed by two units 5-1, 5-2 of the image processor 5 to generate two respective c-CE images 6-1, 6-2. Figure 9 shows an example in which a first c-CE image 6-1 is sensitive to edema and a second c-CE image 6-2 is not sensitive to edema.

[0102] The c-CE images 6-1, 6-2 are fed to an oedema region of interest (ROI) generator 61, which generates an oedema region of interest (ROI) 62 as a myocardial region having a higher signal intensity in the first c-CE image 6-1 (which is sensitive to edema) than in the second c-CE image 6-2 (which is not sensitive to edema). The edema ROI 61 is color coded. The edema ROI 61 and one or both of the c-CE images 6-1, 6-2 (the second c-CE image 6-2 is shown as an example in FIG. 9) are fed to an image combination unit 63, which combines them to generate a combined c-CE image 6-C. As shown in FIG. 9, the combined c-CE image 6-C detects and distinguishes between normal myocardium (black), edema (pink), and myocardial infarction scar (white) in a single computed LGE image (c-LGE image), which conventional LGE images are unable to achieve.

[0103] Thus, converting quantitative T1 mapping into an LGE-like image provides a standardized presentation and further allows for direct combination of multiple modalities with trained c-LGE images 6-1, 6-2, etc., to derive a comprehensive combined c-LGE image 6-C. The combined cLGE 6-C, which does not require contrast administration, can distinguish more pathologies than conventional CE images that require contrast administration. As an example, FIG. 23 shows two examples of a T1 mapping image 101, a c-LGE image 6-1 that is sensitive to edema, a c-LGE image 6-2 that is not sensitive to edema, and a combined c-LGE image 6-C that detects and distinguishes between normal myocardium (black), edema regions of interest (pink), and myocardial infarction scar (white), as pointed out by the arrows in FIG. 23.

[0104] 10(A) and 10(B) respectively show a general example of a (prior art) cardiovascular magnetic resonance (CMR) tissue characterization protocol imaging procedure and a non-invasive contrast-free procedure according to one embodiment of the present invention.

[0105] As shown in Figure 10(A), a prior art imaging procedure may involve a long series of different imaging sequences to generate various modality MR images. The procedure shown is: (1) Pilot and Planning (3 min), (2) Cine (HLA, VLA, LVOT long axis) (5 min), (3) Native T1 map (5 min), (4) Native T2 map (5 min), (5) T2 * (5) post-GBCA-delayed gadolinium enhancement (LGE) at 10 minutes after administration of the GBCA (5 minutes), and (6) post-contrast T1 map at 15 minutes after administration (5 minutes). This type of procedure may last, for example, around 45 minutes, which may be too long for some subjects or may be inappropriate for some patient groups (e.g., acute or unstable patients).

[0106] 10(B) shows a modified protocol according to one embodiment of the present invention. Essentially, the procedure can be stopped prior to administration of contrast and the image processor of the present invention utilizes native T1 mapping to generate a calculated CE image, in this case a calculated LGE image, and thus an LGE image that does not need to be physically acquired. This can reduce the length of the procedure to approximately 25 minutes.

[0107] One embodiment of the present invention was tested by applying the embodiment to a CE-free MR T1 mapping image from step 4 of a procedure such as that shown in Figure 10(B) and comparing the resulting calculated CE image (in this case the calculated LGE) from step 7 of such procedure with the real CE image (in this case the LGE). This embodiment of the present invention was used as the input native T1 mapping image and the generator shown in Figure 5 as the image processor was trained using the conditional GAN ​​method shown in Figure 4.

[0108] For comparison purposes, Figures 11-19 show a native T1 mapping image, an actual LGE image, and an acquired computed LGE image. In each of Figures 11-19, the image on the left is a T1 image that is color-mapped in the original image (although the colors are not visible in the black and white version of the drawings), the image in the middle is a computed LGE image generated by an embodiment of the invention, and the image on the right is an actual LGE image from a subject after administration of contrast agent.

[0109] 11(A) and 11(B) show that the computed LGE images are highly similar to LGE images for various pathologies in individuals with no significant LGE indicative of viable (living) myocardium.

[0110] FIG. 12 shows comparative images for an individual with diffuse fibrosis in hypertrophic cardiomyopathy demonstrated by fuzzy LGE (arrowed).

[0111] FIG. 13 shows focal LGE at the inferior insertion point between the right and left ventricles (marked with an arrow).

[0112] FIG. 14 shows large areas of patchy LGE in the basal anterior wall (arrowed), and FIG. 15 shows diffuse fibrosis in the basal inferolateral wall (arrowed).

[0113] In contrast-enhanced imaging, LGE images are acquired using inversion recovery (IR) or phase-sensitive inversion recovery (PSIR) techniques to reveal abnormal myocardial enhancement. An appropriate inversion time (TI) must be selected to “null” normal myocardial tissue and represent it as black in the LGE image. If the scanning operator makes an error by selecting the wrong TI, the LGE images will have poor quality or reduced diagnostic value. The optimal TI changes as the GBCAs wash out of the tissue, which also means that the reproducibility of the LGE images will decrease. Figures 19(A) and 19(B) show an example where real LGE images were acquired with suboptimal TI and incorrect nulling of the myocardial tissue. The calculated CE, in comparison, does not have a mechanism to be sensitive to the operator selecting the correct TI, and the images all appear to have properly nulled (black) myocardium.

[0114] Because the calculated CE procedure is much shorter than the actual CE scanning procedure, the technique is potentially more robust to artifacts caused by, for example, patient fatigue or the patient's inability to tolerate remaining in the scanner.

[0115] Figure 17 shows a comparison image, where the LGE image was acquired at the end of a 45-minute scanning procedure. Many patients will be fatigued by the end of the scan and will no longer be able to hold their breath, which will steadily result in breathing artifacts (marked with arrows) on the LGE image. The native T1 mapping images, on which the computed CE images are based, have short breath-hold times (<10 seconds per slice) and are performed at the beginning of the scanning procedure. The entire left ventricle of the heart is covered in 8-10 slices (i.e., <5 minutes, including the exposure time and the time between slices to allow the patient to breathe), which may reduce the chance of motion artifacts or patient intolerance.

[0116] Another example of a calculated CE image that demonstrates better image quality and less noise than the actual CE image is shown in Figures 18(A) and 18(B).

[0117] Fig. 21 shows the average quality scores of the computed LGE (virtual LGE) images and the real LGE images as assessed by human observers, which shows that the computed LGE has significantly better image quality than the traditionally acquired LGE.

[0118] Contrast-free MRI modalities such as T1 mapping convey rich information and have better sensitivity for some pathologies, such as diffuse change and edema, than LGE images.

[0119] Figures 16(A) and 16(B) show an example where the computed LGE image shows pathology in the hypertrophic (thickened) septum, but the actual LGE image (marked with an arrow) shows none.

[0120] Figure 20 shows that the calculated LGE showed generally good agreement with the acquired LGE in detecting the presence of LGE enhancement and calculating the percentage of LGE enhancement, subject to potentially diagnostically significant differences.

[0121] Figure 22 shows an example where the calculated LGE image detects edema. In comparison, the real LGE shows no or detects a smaller extent of edema. This indicates that the calculated LGE shows better sensitivity in detecting edema (highlighted by the arrow).

[0122] Figure 23 shows that a computed LGE image generated by combining multiple computed LGE images using the method in Figure 9 detects and differentiates between multiple conditions, e.g., edema (highlighted by arrows), chronic myocardial infarction scar, demonstrating that the computed LGE image contains more information than traditionally acquired LGE images and is capable of detecting many pathologies.

[0123] The above examples demonstrate the effectiveness of the method when applied to the example of T1 mapping applied to a native T1 mapping image 101. It is understood that the benefits arise from the use of quantitative mapping and therefore may be used for other variants of T1 mapping including, for example, T1-ρ mapping or stress T1 mapping, and indeed other non-contrast quantitative mappings other than T1 mapping, such as T2 mapping or T2 * A similar effect is expected for mapping.

[0124] The device 1 shown in Figures 1(A) and 1(B), which may also perform the combination described with reference to Figures 8 and 9, may be a computer device implementing the steps of the method. Similarly, the functional blocks used for training and shown in Figures 2 to 4(A) may also be a computer device implementing the steps of the method. To achieve this, a computer program executable by a computer device may be provided. The computer program is configured such that, when executed, it causes the computer device to perform the relevant steps of the method.

[0125] The computing device, when used, can be any type of computer system, but is generally a computer system of conventional configuration. The computer program can be written in any suitable programming language. The computer program can be stored on any type of computer readable storage medium, for example, a storage medium that can be inserted into a drive of a computing system and can store information magnetically, optically or magneto-optically, a fixed storage medium of a computer system, such as a hard drive, or a computer memory.

[0126] As part of the method, the MR image data set 100 may be acquired by performing a contrast-free imaging magnetic resonance procedure to provide a T1 mapping image, or, if other images are used, by performing suitable contrast-free imaging procedures to provide those other images.

Claims

1. 1. A method for generating a computed contrast enhanced medical image, the method comprising: receiving an image dataset of the subject, the image dataset including native quantitative mapping images acquired by performing a contrast-free magnetic resonance imaging procedure; inputting the image data set into an image processor and processing the image data set with the image processor to generate a computed contrast-enhanced medical image; Including, The method, wherein the image processor comprises a machine learning processor trained on a training dataset including a set of images, each set of images including a contrast-free quantitative mapping image having the same quantitative mapping as the quantitative mapping image of the image dataset and a corresponding acquired contrast-enhanced medical image.

2. The native quantitative mapping image is a T1 mapping image, a T2 mapping image, or a T2 ＊ The method of claim 1 which is a mapping image.

3. the corresponding acquired contrast-enhanced medical image is a contrast-enhanced magnetic resonance image; the corresponding acquired contrast-enhanced medical image is a contrast-enhanced quantitative mapping image having the same quantitative mapping as the quantitative mapping image of the image dataset, or the corresponding acquired contrast-enhanced medical image is a contrast-enhanced image of a non-magnetic resonance modality; 3. The method of claim 1 or claim 2, wherein the contrast-enhanced image of a non-magnetic resonance modality is one of a contrast-enhanced computed tomography image, a contrast-enhanced PET image, a contrast-enhanced SPECT image, or an ultrasound image.

4. The method of claim 1 , wherein the image dataset of the subject includes only the native quantitative mapping images.

5. the image dataset of the subject further comprises at least one further magnetic resonance image acquired by a contrast-free magnetic resonance modality other than the quantitative mapping of the quantitative mapping images of the image dataset, and the set of images of the training dataset further comprises at least one further magnetic resonance image acquired by the other contrast-free magnetic resonance modality; The at least one further magnetic resonance image at least one of a raw magnetic resonance image, an image that is a fusion of the raw magnetic resonance images, or an image that is a derivative of the raw magnetic resonance images; T1 mapping image, T1 weighted image, T2 weighted image, T2 ＊ Weighted image, T2 mapping image, T2 ＊ at least one of a mapping image, or a cine CMR image; at least one of a STIR image, a tagged CMR image, a strain-encoded image, a diffusion-weighted image, a diffusion tensor image, an arterial spin labeling image, a PD-weighted image, or a fat-water separation image; The method according to any one of claims 1 to 3, comprising at least one of the following steps:

6. the image dataset of the subject further comprises at least one non-magnetic resonance image, and the set of images of the training dataset further comprises at least one non-magnetic resonance image of the same type as the at least one non-magnetic resonance image of the image dataset; 6. The method of claim 1, 2, 3 or 5, wherein the at least one non-magnetic resonance image comprises at least one of an echocardiogram, a nuclear perfusion image, a CT image, an electrophysiological heart map image, or a chest x-ray.

7. the image dataset includes further data which is not image data, and the training dataset includes further training data associated with each set of images of the same type as the further data which is not image data; The method of claim 1 , 2, 3, 5 or 6, wherein the further data includes at least one of imaging metadata, image acquisition parameters, or non-imaging diagnostic test results.

8. 8. The method of claim 1, further comprising the step of inputting into the image processor at least one CE image already available from a previous visit or study of the same subject.

9. The method according to any one of claims 1 to 8, wherein the image is a cardiac image.

10. 10. The method of claim 1, wherein the image processor is one of a trained generative adversarial network: a trained variational autoencoder, a trained fully convolutional neural network, a trained U-Net, a trained V-Net, or a trained conditional generative adversarial network.

11. The image data set is processed using the image processor to generate a plurality of computed contrast-enhanced medical images having different disease sensitivities; The method of claim 1 , further comprising combining the plurality of computed contrast-enhanced medical images to generate a combined computed contrast-enhanced medical image.

12. the machine learning processor is trained by processing the contrast-free quantitative mapping images of the training data set to generate computed contrast-enhanced medical images; and / or 12. The method of claim 1, wherein the trained machine learning algorithm is trained to minimize differences between the calculated contrast-enhanced medical image generated by the image processor and the corresponding acquired contrast-enhanced image.

13. The method of claim 1 , further comprising the step of training the machine learning processor.

14. 14. The method of claim 1, further comprising performing a contrast-free magnetic resonance imaging procedure on a subject to obtain the image dataset including the native quantitative mapping image.

15. A computer program executable by a computing device, the computer program being configured, when executed, to cause the computing device to carry out a method according to any one of claims 1 to 13.

Images