Method for processing pre-contrast and contrast images representative of a body part before and after injection of at least a first dose of contrast agent, respectively - Patents.com

Abstract

The invention relates to a method for medical imaging comprising performing, by a data processor (11b) of a second server (1b), the steps of: (a) obtaining at least one candidate pre-contrast and contrast images representative of a body part before and after injection of a first dose of contrast agent, respectively, where the first dose is higher than a predetermined low dose and the predetermined reduced dose is lower than a predetermined standard dose, a Convolutional Neural Network, CNN, trained to reconstruct a standard dose contrast image representative of the body part after injection of the standard dose of contrast agent from at least pre-contrast and low dose contrast input images representative of the body part before and after injection of the low dose of contrast agent, respectively; and (b) simulating an injection of a second dose of contrast agent higher than the standard dose, comprising applying the CNN to the at least one candidate pre-contrast and contrast image to generate a synthetic contrast image.

Description

[Technical field]

[0001] The field of the invention is that of machine learning / deep learning.

[0002] More particularly, the present invention relates to a method for processing at least pre-contrast and contrast images representing a body part before and after, respectively, injection of a first standard dose of contrast agent, and in particular to a method for generating a synthetic contrast image simulating said body part after injection of a second dose of contrast agent, which is higher than the standard dose. [Background technology]

[0003] Contrast agents are substances used in medical imaging to enhance the contrast of structures or fluids within the body.

[0004] Contrast agents typically absorb or modify extraneous radiation emitted by medical imaging devices. In X-rays, contrast agents increase the radiodensity of target tissues or structures. In MRI, contrast agents alter the relaxation times of nuclei in body tissues, changing the contrast of the image.

[0005] Contrast agents are commonly used to enhance the visibility of lesions, particularly in neuroimaging for early diagnosis and treatment planning of brain tumors such as gliomas, brain metastases, and meningiomas.

[0006] Dynamic susceptibility contrast (DSC) and dynamic contrast enhancement (DCE), which exploit the T2 and T1 effects, respectively, are the two most common techniques for MRI. In both cases, a gadolinium-based contrast agent (GBCA) is injected intravenously into the patient and rapid repetitive imaging is performed to obtain a temporal sequence of images.

[0007] Injection of GBCAs increases the sensitivity of MRI, for example by allowing identification of smaller malignant nodules and earlier initiation of treatment, improving patient survival and quality of life. One solution to further increase the sensitivity of MRI is to inject more GBCA.

[0008] However, based on precautionary considerations, recent clinical guidelines recommend the use of the minimum dose that provides sufficient contrast enhancement, and therefore the use of GBCAs should be minimized.

[0009] Nonetheless, three complementary research avenues for improving MRI contrast and MRI lesion detection performance can be identified. - First, new contrast agents with improved chemical properties may be able to improve image contrast without increasing the injected gadolinium dose, but such agents are still under investigation and may be even more expensive than GBCAs. - Second, novel MRI sequences, with or without contrast agents, may replace or complement the usual contrast-enhanced T1 sequences. In particular, turbo spin-echo T1 sequences with variable flip angles (TSE) have been demonstrated to be more sensitive than gradient echo (GRE) sequences and are currently recommended for brain tumor imaging, although some qualitative limitations have been recognized in addition to longer minimum scan times that may promote motion artifacts. - Third, promising deep learning approaches have been proposed to automatically detect and delineate tumors, with the potential to achieve uniform accuracy levels with virtually instantaneous reading times. However, these deep learning algorithms are still in the research stage, and their performance is insufficient for immediate large-scale deployment. Summary of the Invention [Problem to be solved by the invention]

[0010] Thus, there remains a need for new methods to further improve the sensitivity of contrast-enhanced medical imaging. [Means for solving the problem]

[0011] For these purposes, the invention provides a method of medical imaging, comprising the steps of: (a) obtaining (i) at least one candidate pre-contrast image and candidate contrast image representing the body part before and after injection of a first dose of contrast agent, respectively, where the first dose of contrast agent is higher than a low dose of contrast agent, and the low dose is lower than a standard dose of contrast agent; and (ii) a convolutional neural network, CNN, trained to reconstruct a standard dose contrast image representing the body part after injection of the standard dose of contrast agent from at least one pre-contrast input image and a low dose contrast input image representing the body part before and after injection of the low dose of contrast agent, respectively; (b) applying the CNN to at least one candidate pre-contrast image and the candidate contrast-enhanced image to generate a composite contrast-enhanced image. The present invention provides a method for producing a medicament for a cellular environment, the method comprising:

[0012] Preferred, but non-limiting, features of the present invention are as follows:

[0013] The low dose is 10 to 50% of the standard dose.

[0014] The low dose is 1 / 5 to 1 / 3 of the standard dose, preferably about 25%.

[0015] The first dose is at least 50% of the standard dose.

[0016] The first dose is at least 80% of the standard dose, preferably about 100% of the standard dose.

[0017] The second dose is at least 240% of the standard dose, preferably 320% to 400% of the standard dose.

[0018] At least one candidate pre-contrast image and candidate contrast-enhanced image are acquired by a medical imaging device, in particular an MRI scanner, connected to the second server.

[0019] The at least one pre-contrast candidate image includes a T1 weighted pre-contrast image, and the candidate contrast image is a T1 weighted image.

[0020] The at least one pre-contrast image includes three candidate pre-contrast images including a T1 weighted pre-contrast image, a T2-flair weighted pre-contrast image, and an ADC pre-contrast map.

[0021] The CNN is trained to reconstruct a T1-weighted low-dose contrast image representative of the body-part after injection of the standard dose of contrast agent from input T1-weighted and T2-flair-weighted pre-contrast images, ADC pre-contrast input maps, and a T1-weighted contrast input image representative of the body-part before and after injection of the low dose of contrast agent, respectively.

[0022] Step (b) involves applying the CNN to the three candidate pre-contrast images and one candidate contrast image to generate a composite T1-weighted contrast image.

[0023] The CNN includes an encoder branch followed by a decoder branch with a skip connection between the encoder branch and the decoder branch.

[0024] The method includes a preliminary step of training, by a data processor of a first server, the CNN from a base of a sequence of at least one training pre-contrast image, a first training contrast image and a second training image representing a body part before, after injection of the low dose of contrast agent and after injection of the standard dose of contrast agent, respectively.

[0025] According to a second and third aspect, the present invention provides a computer program product comprising code instructions for performing the method of medical imaging according to the first aspect and a computer readable medium having stored thereon a computer program product comprising code instructions for performing the method of medical imaging according to the first aspect.

[0026] The above and other objects, features and advantages of the present invention will become apparent from the following detailed description of illustrative embodiments thereof which is to be read in connection with the accompanying drawings. [Brief description of the drawings]

[0027] [Figure 1] 1 shows an example of an architecture in which the method according to the invention is carried out; [Diagram 2] 1 shows an example of a CNN training pipeline and an operation pipeline in a method according to the present invention. [Diagram 3] 1 shows a preferred architecture of a CNN used in the method according to the invention. [Figure 4] 1 illustrates an embodiment of a method according to the present invention. [Diagram 5] T1c and T1c+ contrast images acquired using the method according to the invention are compared. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0028] architecture The invention proposes a method for medical imaging, in particular a method for processing, by means of a Convolutional Neural Network, CNN, at least one candidate pre-contrast image and one candidate contrast-enhanced image representative of a body part before and during or after injection of a first dose of contrast agent, respectively.

[0029] A pre-contrast or "plain" image means an image representative of a given (monitored) body-part of a human or animal before injection of a contrast agent. A contrast-enhanced image means an image representative of said body-part during or after injection of a contrast agent.

[0030] In other words, if there is a temporal sequence of images, the first image is a pre-contrast image and each subsequent image is a contrast image. Note that the contrast images can be images of a given phase (e.g., arterial, portal, delayed, etc.) or full dynamic contrast enhancement (DCE) images.

[0031] In the following description, the preferred example of brain imaging is considered, i.e. the body part is the brain.

[0032] The images (pre-contrast or contrast-enhanced) are either directly acquired by a scanner-type medical imaging device or are derived from directly acquired images.

[0033] The imaging by injection of contrast agent may be as follows. - CT (Computed Tomography) → Medical imaging device is a rotating X-ray scanner capable of tomographic reconstruction. - MRI (Magnetic Resonance Imaging) → The medical imaging device is an MRI scanner. - Mammography → Medical imaging device is an X-ray mammography machine. - others.

[0034] Acquisition of said images may require the injection of a contrast agent such as Gadolinium (GBCA) for MRI or a suitable X-ray contrast agent.

[0035] Note that an "image" can be a 2D object (having two spatial dimensions), but also a 3D object (having three spatial dimensions), i.e. a volume made up of a stack of 2D images as "slices" along the third spatial dimension (in other words, there are 2+1 spatial dimensions).

[0036] Additionally, there may be multiple pre-contrast and / or candidate contrast images.

[0037] Preferred MRI embodiments include at least T1 weighted pre-contrast and T1 weighted contrast images, in particular GRE (gradient echo) T1 weighted pre-contrast / contrast images, but may further include: - T2-flair weighted pre-contrast imaging (especially TSE (turbo spin echo)), and / or - ADC (apparent diffusion coefficient) pre-contrast map (automatically converted from EPI (echo planar) DWI (diffusion weighted)).

[0038] Note that TSE (Turbo Spin Echo) T1 weighted contrast images are also available but are not used in this application.

[0039] The above-mentioned method is implemented by a first and / or second server 1a, 1b in an architecture as shown in Fig. 1. The first server 1a is a training server (implementing the CNN training method) and the second server 1b is a processing server (implementing the processing method). It is quite possible that these two servers can be integrated.

[0040] Each of these servers 1a, 1b is typically a remote computer equipment connected to an extended network 2 such as the Internet for data exchange. Each server includes a data processing means 11a, 11b of processor type (the data processor 11a of the first server 1a has powerful computing power, especially since the learning is long and complex compared to the usual use of trained models) and optionally a storage means 12a, 12b such as a computer memory, for example a hard disk. The second server 1b can be connected to one or more medical imaging devices 10 as client equipment providing images to be processed and receiving parameters.

[0041] It is noted that the imaging device 10 includes an injector for performing the injection of contrast agent, said injector being assumed to apply the injection parameters.

[0042] The memory 12a of the first server 1a stores a set of images called a training database, i.e. training images (different from the so-called input images that are required to be processed correctly). Each image of the database can be a pre-contrast or contrast-enhanced image, the contrast-enhanced images being labelled with respect to the dose of contrast agent injected. It is noted that images corresponding to the same injection (i.e. forming a temporal sequence) are associated in a sequence.

[0043] Dose minimization CNN In the following description, reference is made to a "low dose" and a "standard dose" of injected contrast agent. Both doses are predetermined. For convenience, an enhanced image representing a body part after injection of a standard dose of contrast agent will be referred to as a "standard dose contrast image" and an enhanced image representing a body part after injection of a low dose of contrast agent will be referred to as a "low dose contrast image."

[0044] The standard dose, or "full dose", is the recommended dose typically used for medical testing. Although side effects are possible, such doses are not particularly dangerous to the patient and allow a level of image quality sufficient for analytical / diagnostic purposes. For GBCAs, the standard dose is 0.1 mmol / kg.

[0045] A low dose, i.e. a "reduced dose", is a dose that is lower than the standard dose and has less impact on the patient's health. The low dose can be any percentage of the standard dose, preferably 1 / 10 to 1 / 2 (10 to 50%) of the standard dose, 1 / 5 to 1 / 3 (20 to 33%) of the standard dose, preferably about 1 / 4 (25%). In the case of GBCA, the low dose is, for example, 0.025 mmol / kg (25% of 0.1 mmol / kg).

[0046] Clearly, low-dose contrast-enhanced images have lower contrast than standard-dose contrast-enhanced images. Moreover, simply amplifying the contrast enhancement of x% low-dose CE-MRI by a factor of 100 / x leads to a degradation of image quality, spreading noise and obscuring structures.

[0047] The CNN is trained to reconstruct a standard-dose contrast image representative of the body-part after injection of the standard dose of contrast agent from at least one pre-contrast input image and a low-dose contrast input image representative of the body-part before and after injection of the low-dose of contrast agent, respectively.

[0048] Such CNNs are known to those skilled in the art (see, for example, WO 2019074938 or Ammari S, Bone A, Balleyguier C, et al. Can Deep Learning Replace Gadolinium in Neuro-Oncology? Invest Radiol 2021; Publish Ah(00):1-9 doi:10.1097 / rli.00000000000008) and allow the dose of contrast agent to be reduced. That is, by applying the CNN to candidate pre-contrast and low-dose contrast images representing a body part before and after injection of only a low dose of contrast agent, a synthetic contrast image can be generated that simulates said body part after injection of a standard dose.

[0049] In other words, this CNN is capable of predicting standard-dose contrast images from pre-contrast and low-dose contrast images, and can therefore be called a "dose-minimizing CNN."

[0050] In fact, one can inject only a low dose of contrast agent and still obtain the quality of a standard dose contrast image.

[0051] A possible embodiment of this CNN and training method is described below.

[0052] Preferably, the CNN uses at least a T1-weighted pre-contrast / contrast image (and optionally also a T2-flair weighted pre-contrast image and / or an ADC pre-contrast map) as input and another T1-weighted contrast image as output.

[0053] The method proposes the judicious use of this dose-minimizing CNN to virtually increase the dose by applying a dose-reducing CNN to a candidate contrast-enhanced image representing a body part after injection of a "first dose" of contrast agent higher than the low dose, and possibly up to (but not exceeding) the standard dose, thereby generating a synthetic contrast-enhanced image believed to represent said body part after injection of a "second dose" or "super dose" of contrast agent higher than the standard dose.

[0054] The result is a very high quality image, which appears as if a high dose of contrast agent (which can be dangerous to the patient) had been injected, when in reality only a standard dose of contrast agent had been injected.

[0055] In other words, this CNN is not used for the task it was actually trained for, which is quite peculiar. With reference to the two example pipelines in Figure 2, the standard dose contrast image is directly used as the first dose contrast image.

[0056] In this figure, T1-weighted pre-contrast images are designated as "T1", T2-flair weighted pre-contrast images are designated as "T2", ADC pre-contrast maps are designated as "ADC", low-dose T1-weighted images are designated as "T1c-", standard-dose T1-weighted images are designated as "T1c", and "ultra-dose" T1-weighted images are designated as "T1c+".

[0057] Under the hypothesis that the CNN had learned to primarily amplify the contrast difference between pre-contrast and post-contrast inputs, it was predicted that by replacing T1c- images with T1c images during inference, it would be possible to synthesize T1c+ images with almost four times the dose.

[0058] For example, low dose (D c- ) is the standard dose (D c ), then D c =x / 100×D c It becomes.

[0059] Because the same CNN is used, the first and second doses (D1, D2) can be written in the same formula: D1=x / 100×D2.

[0060] Therefore, D c- <D1≦<D c x / 100×D c <x / 100×D2≦D c D c <D2≦100 / x×D cand D. c+ =100 / x×D c It becomes.

[0061] As a result, in an embodiment where the low dose is 25% of the standard dose, if the standard dose is used as the first dose, the second dose can reach 400% of the standard dose (see FIG. 2). In other words, if the standard dose is 0.1 mmol.kg, the second dose can reach 0.4 mmol / kg (a virtual replacement infusion of 0.3 mmol / kg), which is why it is called a "super dose."

[0062] It should be noted that said first dose is advantageously at least 80% of the standard dose, preferably about 100% of the standard dose, which means that the second dose is typically at least 160% (80% / 50%) of the standard dose, preferably at least 240% (80% / 33%), preferably 320%-400% (80% / 25%-100% / 25%) of the standard dose.

[0063] 3D UNet An exemplary architecture of a U-Net type dose-minimizing CNN is shown in Figure 3. Note that for 3D pre-contrast / contrast images, the CNN itself is a 3D network, as in the example of Figure 2, processing 4D feature maps.

[0064] U-Net is an encoder-decoder type neural network. It includes an encoder branch (or "contraction path") that maps inputs (at least one pre-contrast image and low-dose / first-dose images) to a high-level representation, and then a decoder branch (or "expansion path") that generates output images (standard-dose / second-dose images) from the high-level representation.

[0065] U-Net further includes skip (or "lateral") connections between the encoder and decoder branches.

[0066] The encoder branch acts as a backbone and can be considered as a conventional feature extraction network that can be of many types, in particular a conventional CNN, preferably a fully convolutional neural network (direct successive blocks of convolutional layers and nonlinear layers such as ReLU (Regulated Linear Units), which in particular alternate residual convolutional blocks and strided convolutional blocks for downsampling). From the input image, the encoder branch extracts multiple initial feature maps that represent the input image at various scales. More precisely, the backbone consists of multiple successive convolutional blocks, where a first block generates a first initial feature map from the input, a second block generates a second initial feature map from the first initial feature map, and so on.

[0067] In convolutional neural networks, successive maps are traditionally understood to have smaller scales (i.e., lower resolution, "smaller" feature maps and therefore less detailed), but with increasing semantic depth because higher levels of image structure are being captured. Specifically, the initial feature maps have an increasing number of channels as the spatial size decreases.

[0068] In practice, a pooling layer is placed between the two blocks to halve the scale (typically a 2x2x2 convolution with stride 2 for downsampling in the case of 3D images), and for each block, the number of filters in the convolutional layer used (typically 3x3x3 convolutions) is increased (preferably doubled).

[0069] For a 5-level standard U-Net, for example, the number of consecutive channels are 32, 64, 128, 256 and 512, and the consecutive map space sizes (for an input image of 160x192x160) are 160x192x160, 80x96x80, 40x48x40, 20x24x20, 10x12x10. It can be seen that the input already has four channels, i.e., three pre-contrast images (T1, T2-flair, ADC) and one contrast image (T1c- in training, T1c in behavior), and the output has a single channel (T1c in training, T1c+ in behavior).

[0070] The feature maps obtained by the encoder branch are called initial maps because they are reprocessed by the decoder branch. Indeed, as explained, the "low-level" maps have a higher spatial resolution but a lower semantic depth. The decoder branch aims to increase the semantic depth by incorporating information from the "high-level" maps.

[0071] The decoder branch of a CNN therefore has a symmetric architecture to the encoder branch: from an initial feature map, it generates multiple enriched feature maps that again represent the input image at different scales, but incorporating information from the initial feature maps at smaller or equal scales while reducing the number of channels.

[0072] In other words, the decoder branch also consists of multiple successive convolutional blocks, but in the reverse order, for example, the second block generates the second enriched feature map from the third enriched feature map and the second initial feature map, and then the first block generates the first enriched feature map (from which the output image can be directly generated) from the second enriched feature map and the first initial feature map. The decoder branch is also preferably a fully convolutional neural network (direct successive blocks of convolutional layers and nonlinear layers such as ReLU, which in particular alternate between residual convolutional blocks and strided convolutional blocks for upsampling).

[0073] More specifically, each i-th enriched map has the scale (i.e., approximately the same spatial size) of the corresponding i-th initial map, but incorporates the information of all j-th maps, for each j≧i. In practice, each i-th enriched map Di is generated according to each corresponding i-th initial map Ei and / or the next (i+1) enriched map, thus obtaining the "contraction and expansion" nature of the branches (i.e., the "U" shape). That is, the initial maps are obtained in ascending order, and the enriched maps are obtained in descending order.

[0074] In effect, the highest semantic level is obtained at the "smallest scale" map, from which each map is enriched again on its way down with information from already enriched maps. Skip connections between the encoder and decoder branches provide the decoder branches with different initial maps.

[0075] Typically, generation of an enriched map based on a corresponding initial map and a smaller-scale enriched map involves rescaling the enriched map, usually doubling the scale (if the scale has been halved in the encoder branch), i.e. upsampling the enriched feature map with a 2x2x2 convolution ("upconvolution") to halve the number of feature channels, then concatenating it with the corresponding initial map Ei (with unnecessary parts trimmed so that both maps are substantially the same scale) to double the number of channels again, and for each block, the number of filters in the used convolutional layer (usually 3x3x3 convolution) is reduced again (preferably further halved).

[0076] The downsampling and upsampling convolution blocks rely on 2x2x2 kernels, all other kernels are 3x3x3 except for the final 1x1x1 convolution. All activation functions are 0.2 except for the final sigmoid (LeakyReLU).

[0077] It should be noted that the present invention is not limited to a particular U-Net architecture: any CNN trained to reconstruct a training contrast-enhanced image representative of a body part after injection of the standard dose of contrast agent from at least one training pre-contrast image and a training contrast-enhanced image representative of the body part before and after injection of the low dose of contrast agent, respectively, may be suitable.

[0078] Training is described below.

[0079] Medical Imaging Methods As shown by FIG. 4, the method according to the present invention starts with step (a) of acquiring the pre-contrast and contrast-enhanced images to be processed (referred to as candidate pre-contrast and contrast-enhanced images), in particular T1-weighted pre-contrast / contrast images, from a medical imaging device 10 preferably connected to a second server 1b.

[0080] The candidate pre-contrast image and the candidate contrast image represent the body part before and after injection, respectively, of a first dose of contrast agent, the first dose being higher than a predetermined low dose, which is lower than a predetermined standard dose.

[0081] In a preferred MRI embodiment, step (a) involves acquiring three candidate pre-contrast images: a candidate T1 weighted pre-contrast image, a candidate T2-flair weighted pre-contrast image, and a candidate ADC pre-contrast map.

[0082] It should be noted that this step (a) may be performed by the data processor 11b of the second server 1b and / or by the medical imaging device 10.

[0083] In a main step (b), performed by the data processor 11b of the second server 1b, a CNN is applied to the candidate pre-contrast and contrast images to generate a synthetic contrast image which is believed to simulate said body-part after injection of a second dose of contrast agent, which is higher than the standard dose.

[0084] In a preferred MRI embodiment, the CNN is applied to three candidate pre-contrast images and one candidate contrast image.

[0085] training method The method advantageously comprises a preliminary step (a0) of training a CNN executed by the data processor 11a of the first server 1a.

[0086] Training means determining optimal values ​​for the parameters and weights of the CNN. Note that, alternatively, a CNN can be directly employed as "off the shelf" with pre-set values ​​for parameters and weights.

[0087] The training method may be performed according to conventional techniques and may use any suitable training protocol known to those skilled in the art.

[0088] It should be understood that the CNN is trained to perform its original "dose minimization" task, i.e., to reconstruct a training contrast-enhanced image representative of a body-part after injection of the standard dose of contrast agent from at least one training pre-contrast image and one training contrast-enhanced image representative of said body-part before and after injection of the low dose of contrast agent, respectively.

[0089] Indeed, it is not possible to directly train a CNN from at least one training pre-contrast image and training contrast image representing a body part before and after injection, respectively, of a first dose of contrast agent higher than a low dose, to generate a synthetic contrast image simulating said body part after injection of a second dose of contrast agent higher than a standard dose, such training would require training images after injection of the second dose, which would result in a total injected dose higher than that recommended by current clinical guidelines.

[0090] In contrast, to learn a "dose minimization" task, the training base may include multiple sequences of at least one training pre-contrast image (although as described there may optionally be three pre-contrast images) representing a body part before, after and after injection of the low dose of contrast agent, respectively, an "initial" training contrast image (a low dose contrast image) and a "final" training contrast image (a standard dose contrast image).

[0091] The CNN is trained to predict the final training contrast image (as ground truth) from training pre-contrast and early contrast images of the same sequence.

[0092] More specifically, to generate training examples, a normal protocol for acquiring contrast-enhanced images representative of the body part after injection of the standard dose of contrast agent is performed, but the administration of the standard dose is simply divided into two successive injections: (1) a low dose, and (2) the difference between the low dose and the standard dose, e.g., 0.025 mmol / kg and 0.1-0.025=0.075 mmol / kg of GBCA as described. In between, an additional contrast-enhanced image is acquired as an "initial" contrast image, which in fact represents the body part after injection of the low dose of contrast agent. It should be noted that two-injection MRI protocols have recently been included in consensus guidelines for glioma imaging, with the first injection serving as a preload bolus for a perfusion sequence that may be performed later. It is therefore harmless.

[0093] In summary, for patients: - Pre-contrast images are obtained; - A low dose is injected and a first contrast image is acquired; - The difference between the standard and low doses is injected (so the patient receives the standard dose in total) and a second contrast image is obtained; - A sequence of pre-contrast images, where two consecutive contrast images are labeled as training examples.

[0094] test Figure 5 shows T2-Flair, T1c, T1c+ and tse-T1c images for four exemplary MRI studies from either brain metastasis or glioma test samples. It can be seen that the T1c+ images (synthetic hyper-dose contrast images) show superior quality and are easier to interpret.

[0095] Table 1 reports the mean image quality (IQ) scores for T1c, T1c+ and tse-T1c images from the 79 examinations of the test sample. The ratings are expressed on a 4-point Likert scale ranging from 1 (poor) to 4 (excellent). Standard deviations are given in parentheses. Differences between readers and between post-contrast MRI images were compared using two-tailed t-tests and p-values ​​are reported. When a significance threshold of 5% was met, the best indicators are highlighted using bold type.

[0096] The table shows that when no quality differences were found between the T1c and tse-T1c images, the combined T1c+ images were preferred by the two readers (neuroradiologists with 10 and 11 years of experience, respectively) for their overall image quality. On average between readers, the T1c and tse-T1c were rated 2.7 / 4 (average to good) and the T1c+ was rated 3.4 / 4 (good to excellent).

[0097] [Table 1]

[0098] Table 2 reports the mean contrast-to-noise ratio (CNR), lesion-to-brain ratio (LBR) and contrast enhancement ratio (CEP) performance metrics for T1c, T1c+ and tse-T1c images from 52 examinations of test samples with at least one reference lesion. Standard deviations are given in brackets. Differences between readers and between post-contrast MRI images were compared using two-tailed t-tests and p-values ​​are reported. When a significance threshold of 5% was met, the best metrics are highlighted using bold type.

[0099] The combined T1c+ images had a mean CNR of 44.5, LBR of 1.66, and CEP of 112.4%, outperforming both T1c and tse-T1c images with respect to all metrics considered.

[0100] [Table 2]

[0101] Finally, the mean lesion detection performance reached when reading T1c images was compared with that when reading the corresponding post-processed T1c+ images together. In both reading scenarios, pre-contrast T2-Flair images were also available to the readers. For reading with access to T2-Flair, T1c and tse-T1c images, 187 reference lesions with a median long axis length of 9.2 mm (interquartile range 10.7 mm) were defined. Depending on the minimum lesion size to be included (10 mm, 5 mm or 0 mm, in which case all lesions were considered), three branched assessment configurations were considered:

[0102] Access to T1c+ images increased lesion detection sensitivity (SE) for both readers in all evaluation configurations. Averaged across readers, the overall SE when T1c+ images were available was 75% (59% for T1c images only, P < .001). * ), and 85% for lesions larger than 5 mm (70% for T1c, P < .001 * ), and 96% for lesions larger than 10 mm (88% for T1c, P = .008). * ) No differences were found in the FDR (false discovery rate), less than 0.19 / examination, across readers, reading scenarios, and evaluation configurations. No differences were found between readers for either SE or FDR.

[0103] PPV was greater than 90% for all configurations. On average across readers, the quantitative difference in PPV was at most 3% between the two reading scenarios. F1 was greater than 90% across readers and readings for lesions larger than 10 mm, greater than 80% for lesions larger than 5 mm, and greater than 70% when all lesions were included. On average across readers, F1 was greater when T1c+ was available to the reader, by a margin of 4%–11%, depending on the range of lesion sizes considered.

[0104] Computer Program Products In a second and third aspect, the present invention provides a computer program product comprising code instructions for executing (in particular on a data processor 11a, 11b of the first and / or second server 1a, 1b) the method according to the first aspect of the invention and storage means (the memory of the first or second server 1a, 1b) readable by a computing device containing this computer program product.

Claims

1. In a method of medical imaging, a data processor (11b) of a second server (1b) (a) obtaining (i) at least one candidate pre-contrast image and candidate contrast image representing a body-part before and after injection of a first dose of contrast agent, respectively, wherein the first dose of contrast agent is higher than a low dose of contrast agent, and the low dose is lower than a standard dose of contrast agent; and (ii) a convolutional neural network, CNN, which is trained to reconstruct a standard dose contrast image representing the body-part after injection of the standard dose of contrast agent from at least one pre-contrast input image and a low dose contrast input image representing the body-part before and after injection of the low dose of contrast agent, respectively; (b) generating a synthetic contrast-enhanced image by applying the CNN to the at least one candidate pre-contrast image and the candidate contrast-enhanced image; The method according to claim 1, further comprising:

2. The method of claim 1 , wherein the low dose of contrast agent is between 10% and 50% of the standard dose of contrast agent.

3. 3. The method of claim 1 or 2, wherein the first dose of contrast agent is at least 50% of the standard dose of contrast agent.

4. 4. The method of claim 2 or 3 in combination, wherein the low dose of contrast agent is ⅕ to ⅓ of the standard dose of contrast agent and the first dose of contrast agent is at least 80% of the standard dose of contrast agent.

5. The method according to any one of claims 1 to 4, wherein the at least one candidate pre-contrast image and the candidate contrast-enhanced image are acquired by a medical imaging device (10), in particular an MRI scanner, connected to the second server (1b).

6. The method of any one of claims 1 to 5, wherein the at least one candidate pre-contrast image comprises a T1 weighted pre-contrast image, and the candidate contrast image is a T1 weighted image.

7. - said at least one candidate pre-contrast image comprises said T1 weighted pre-contrast image, a T2-flair weighted pre-contrast image and an ADC pre-contrast map; said at least one pre-contrast input image comprises a T1 weighted pre-contrast input image, a T2-flair weighted pre-contrast input image and a training ADC pre-contrast input map; said low-dose contrast input image is a T1-weighted contrast image; said standard dose contrast image is a T1 weighted contrast image, The method according to claim 6, wherein said synthetic contrast-enhanced image is a synthetic T1-weighted contrast-enhanced image.

8. The method according to any one of claims 1 to 7, wherein the CNN comprises an encoder branch followed by a decoder branch with a skip connection between the encoder branch and the decoder branch.

9. The method according to any one of claims 1 to 8, comprising a preliminary step of training, by a data processor (11a) of a first server (1a), said CNN from the base of a sequence of at least one training pre-contrast image, a first training contrast image and a second training image representing a body part before, after injection of said low dose of contrast agent and after injection of said standard dose of contrast agent, respectively.

10. A computer program product comprising code instructions for carrying out the method of medical imaging according to any one of claims 1 to 9, when said computer program is run on a computer.

11. A computer readable medium having stored thereon a computer program product comprising code instructions for carrying out the method of medical imaging according to any one of claims 1 to 9.

Images