Diffusion-weighted magnetic resonance imaging

Abstract

A computer system is provided for analysing diffusion-weighted magnetic resonance images of an object. The computer system is configured to receive diffusion-weighted images of the object acquired by a magnetic resonance imaging scanner, and is programmed with a neural network which filters the acquired images to produce one or more predicted images from the acquired images. The diffusion-weighted images acquired by the magnetic resonance imaging scanner depict contiguously successive slices through the object, each acquired image being an image depicting a given slice at a given b-value. The neural network has an input layer that is configured to receive a set of diffusion-weighted images from the acquired images respectively depicting at least three successive slices, and is trained such that it generates therefrom, at an output layer, one or more predicted images depicting a middle one of the successive slices with an improved CNR relative to the diffusion-weighted image received at the input layer depicting the middle slice.

Description

[0001] 008614430

[0002] 1

[0003] DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGING

[0004] Field of the Invention

[0005] The present disclosure relates to diffusion-weighted magnetic resonance imaging.

[0006] Background

[0007] Diffusion-weighted MR-imaging (DWI) is a non-invasive tool used for staging and response evaluation in oncologic practice. Whole-body DWI (WBDWI) is at the core of emerging response criteria in advanced prostate and breast cancers [1-3] and has also been incorporated into the National Institute for Health and Care Excellence and International Myeloma Working group guidelines for assessing myeloma-related bone disease [4,5]. This technique sensitizes MR-imaging contrast to the diffusion rate of water within tissues through application of magnetic gradients, which can be manipulated to adjust the magnitude of diffusion-weighting within images. The contrast generated between diseased and healthy tissues using this technique provides radiologists with a sensitive tool for reviewing the extent of bony disease within the skeleton. The degree of diffusion weighting within an image can be manipulated at will through modification of the so-called ‘b-value’, which encapsulates the timing and strengths of the diffusionweighting gradients into a single variable (typically in the range 0-5000 s / mm2on clinical systems). By acquiring images for the same anatomical location at two or more b-values, WBDWI offers voxel-wise quantification of the ‘apparent diffusion coefficient’ (ADC) of water, providing a potential marker for tumour response assessment [6].

[0008] WBDWI is typically achieved using a series of sequential imaging stations from the head to the mid-thigh, with each station comprising 30-50 equally-spaced axial sections or slices, with images acquired using two to three diffusion weightings and typically many signal averages [7]. Therefore, WBDWI accounts for more than 50% of the acquisition time of conventional whole-body MRI studies with a 1-hour duration. In the context of the ever-increasing capacity pressures on MRI departments, reducing acquisition times would facilitate the wider adoption of clinical WBDWI, reduce costs, and improve patient acceptance. DWI is also embedded into consensus MRI protocols across almost all tumour types including primary prostate and breast cancers, metastatic liver disease, gynaecological and gastrointestinal cancers [8-12] where time savings would also be beneficial.

[0009] One approach for DWI acceleration is based on supervised learning where the 2D “noisy” image of a slice acquired using a smaller number (NeX) of signal excitations is given as input and the “clinical” image reconstructed by averaging a larger number (NeX ® 9-16) of signal excitations is used as the ground truth. A convolutional neural network (CNN), usually resembling the U-net architecture, is used to denoise the image and is trained by a loss function to minimize the difference between the output and ground truth, thereby improving its CNR (contrast to noise ratio). Zormpass-Petridis et al. used a deep learning approach to improve the image quality of subsampled images (number of acquisitions = 1) to reduce whole-body diffusion-weighted MRI (WBDWI) acquisition times [13, 14]. In particular this approach 008614430

[0010] 2

[0011] (henceforth called “DNIF") demonstrated that by using only 1 diffusion encoding direction and 1 signal average using b-values such as 50, 600 or 900 s / mm2it is possible to achieve up to 50% reduction in whole-body MRI acquisition times by reducing whole-body DWI to less than five minutes. Wessling et al. utilized a variational network to accelerate a breast diffusion sequence by 40%

[0015] by reducing averages from 4 to 2 for low b-value and from 16 to 8 for high b-value images. Kaye et al. accelerated prostate DWI using a CNN with a mean-squared-error loss function to denoise sub-sampled data of two signal averages

[0016] . Afat et al. used a variational network to accelerate live DWI by 40% by reducing the signal averages from 12 to 6 for the high b-value images and from 2 to 1 for the low b-value images

[0017] . Selfsupervised approaches, such as Noise2Noise

[0018] which model the noise and do not require labelled data could be incorporated. Kawamura et al. used a supervised learning approach with a CNN to predict the residual noise instead of the denoised image directly

[0019] . Maosong et al. used a residual encoderdecoder Wasserstein generative adversarial network along with perceptual similarity loss function to denoise 3D MRI, but not DWI

[0020] . Instead of reducing NeX, other approaches include shot-to-shot phase reconstruction

[0021] and k-space undersampling

[0022] .

[0012] However, further reductions in acquisition times and improvements in image quality are still sought.

[0013] The present invention has been devised in light of the above considerations.

[0014] Summary of the Invention

[0015] A first aspect of the present invention provides a computer system for analysing diffusion-weighted magnetic resonance images of an object, the computer system being configured to receive diffusion- weighted images of the object acquired by a magnetic resonance imaging scanner, and being programmed with a neural network which filters the acquired images to produce one or more predicted images from the acquired images; wherein the diffusion-weighted images acquired by the magnetic resonance imaging scanner depict contiguously successive slices through the object, each acquired image being an image depicting a given slice at a given b-value; and wherein the neural network has an input layer that is configured to receive a set of diffusion- weighted images from the acquired images respectively depicting at least three successive slices, and is trained such that it generates therefrom, at an output layer, one or more predicted images depicting a middle one of the successive slices with an improved CNR relative to the diffusion-weighted image received at the input layer depicting the middle slice.

[0016] Advantageously, the slices to either side of the middle slice introduce spatial context that can significantly improve the predictive performance of the neural network. In addition, in images of the object formed by taking sections through a stack of the predicted images depicting the successive slices, this enhanced spatial context can help smooth the transitions between neighbouring images of the stack. 008614430

[0017] 3

[0018] The invention includes the combination of the aspects and preferred features described except where such a combination is clearly impermissible or expressly avoided.

[0019] Image voxels forming each acquired image may conveniently be transformed using a log-transform. Advantageously, performing this transformation before acquired images are input into the neural network can reduce the influence of any high-valued pixels, thereby reducing the influence of noisy outliers in the filtering performed by the neural network.

[0020] Preferably, the given b-value of the acquired images received by the input layer has the same diffusion direction for the successive slices. In this way, the acquired images depicting the successive slices at the given b-value are consistent in their b-value diffusion direction.

[0021] The one or more predicted images depicting the middle slice may be trace-weighted images. Thus advantageously the neural network can receive at its input a diffusion-weighted image of the middle slice acquired from a single excitation, but is trained to predict a trace-weighted image for that slice even though conventional derivation of a trace-weighted image would require the synthesis of images from excitations at different b-value diffusion directions.

[0022] Advantageously, multiple b-values can be processed simultaneously by the neural network. More particularly, the input layer may receive plural (for example three) acquired images at respective given b- values depicting each slice of the at least three successive slices. Moreover, the predicted images depicting the middle slice can include corresponding diffusion-weighted images at the respective given b- values. For example, when the input layer receives acquired images depicting three successive slices at three different b-values for each slice, there are thus 3 (for the three b-values) x 3 (for the three slices) = 9 acquired images. In addition, the neural network can then generate 3 (for the three b-values) predicted images at the output layer. Preferably, the one or more predicted images include an apparent diffusion coefficient map of the middle slice. Conveniently, when the predicted images depicting the middle slice include corresponding diffusion-weighted images at the respective given b-values, these predicted images can then be used to calculate the apparent diffusion coefficient map. However, this does not exclude that the apparent diffusion coefficient map can be generated directly by the neural network as a predicted image without first predicting diffusion-weighted images at the respective given b-values.

[0023] The neural network may be a spatially variant filter. Thus the degree of smoothing performed by the network within a certain region of the acquired images can be dependent on the position of that region within the entire imaging field. Particularly in the context of an object that is a human or animal subject, this allows the network to learn anatomical position in order to tune the degree of smoothing it performs at a particular anatomical location.

[0024] The neural network may be a convolutional neural network.

[0025] The computer system may be further programmed such that the neural network receives further sets of diffusion-weighted images from the acquired images depicting further sets of three successive slices and 008614430

[0026] 4 generates corresponding predicted images therefrom, the computer system thereby producing a 3D representation of the object from the predicted images. Particularly in the context of an object that is a human or animal subject, a radiologist can then examine the 3D representation for indications of diseased tissue.

[0027] A second aspect of the present invention provides an imaging system for performing diffusion-weighted magnetic resonance imaging, system including: a magnetic resonance imaging scanner for acquiring diffusion-weighted images of an object; and the computer system according to the first aspect operatively connected to the scanner to receive the acquired images.

[0028] A third aspect of the present invention provides a computer-implemented method of analysing diffusion- weighted magnetic resonance images of an object acquired by a magnetic resonance imaging scanner, the method including: receiving the acquired diffusion-weighted images of the object; and using a neural network which filters the acquired images to produce one or more predicted images from the acquired images; wherein the diffusion-weighted images acquired by the magnetic resonance imaging scanner depict contiguously successive slices through the object, each acquired image being an image depicting a given slice at a given b-value; and wherein the neural network has an input layer that is configured to receive a set of diffusion- weighted images from the acquired images respectively depicting at least three successive slices, and is trained such that it generates therefrom, at an output layer, one or more predicted images depicting a middle one of the successive slices with an improved CNR relative to the diffusion-weighted image received at the input layer depicting the middle slice.

[0029] Thus the method of the third aspect may be performed using the computer system of the first aspect. Moreover, optional features of the computer system of the first aspect discussed above pertain individually or in any combination to the method of the third aspect.

[0030] The method may further include: acquiring the diffusion-weighted images using the magnetic resonance imaging scanner.

[0031] The method may further include: transforming image voxels forming each acquired image using a logtransform.

[0032] The method may further include: displaying the one or more predicted images.

[0033] The object may be a human or animal subject. In this case, the method may include a further step of: analysing the one or more predicted images for assessment of disease extent (e.g. cancer extent) in the human or animal subject. 008614430

[0034] 5

[0035] A fourth aspect of the present invention provides a computer program comprising code which, when the code is executed on a computer, causes the computer to perform the method of the third aspect.

[0036] A fifth aspect of the present invention provides a non-transitory computer readable storage medium storing the computer program of the fourth aspect.

[0037] A sixth aspect of the present invention provides a method of training a neural network programmed to produce one or more predicted images by filtering diffusion-weighted images depicting contiguously successive slices through an object acquired by a magnetic resonance imaging scanner, the neural network having an input layer that is configured to receive acquired images respectively depicting at least three successive slices, and having an output layer that generates therefrom one or more predicted images depicting a middle one of the successive slices, the method including: providing training data including diffusion-weighted images at a given b-value of plural objects, the provided images of each object having been acquired by a respective magnetic resonance imaging scanner at contiguously successive slices through that object, and the provided images of each slice depicting that slice at different b-value diffusion directions; for each slice of at least a subset of the slices, combining the provided images of the slice at the different b-value diffusion directions to form a respective ground truth, trace-weighted, diffusion-weighted image depicting that slice at the given b-value; pairing each ground truth image with a respective set of images selected from the provided images at the given b-value, the set of selected images sharing the same b-value diffusion direction and depicting at least three successive slices through the respective object, a middle one of the successive slices being the same slice as that depicted by the ground truth image; and inputting at the input layer each set of selected images, and training the neural network to minimise a loss function that measures similarity between a predicted image thereby generated at the output layer and the ground truth image paired with the inputted set of selected images.

[0038] Thus the method of this other aspect may be used to train the network of the first or third aspect.

[0039] The image voxels forming the provided images of each slice in the training data may be transformed using a log-transform.

[0040] Preferably, the provided images depicting each slice at different b-value diffusion directions are single excitation images. In this way, the network can learn to predict high CNR, ground truth, trace-weighted, diffusion-weighted images from low CNR input images.

[0041] The subset of slices may be every Nthslice through each object, where N (an integer) is three or more. In particular, when each set of selected images sharing the same b-value diffusion direction depicts M (an integer) successive slices through the respective object, conveniently N = M. In this way, the images of each set are not shared with the images of neighbouring sets, i.e. the sets of selected images do not overlap. 008614430

[0042] 6

[0043] In the method of training a neural network:

[0044] — the input layer of the neural network may be configured to receive plural (for example three) acquired images at respective given b-values depicting each slice of the at least three successive slices, and the output layer may generate therefrom plural predicted images depicting the middle slice respectively corresponding to the given b-values,

[0045] — for each of the objects, the provided training data may include diffusion weighted images at each of the given b-values, the provided images of each slice depicting that slice at different b-value diffusion directions for each given b-value,

[0046] — for each slice of the subset of slices, a respective ground truth, trace-weighted, diffusion- weighted image depicting that slice may be formed for each of the given b-values, and

[0047] — the set of images selected from the provided images and paired with each ground truth image may be at the respective given b-value of the ground truth image.

[0048] The inputting at the input layer can then performed by: for each slice of the subset of slices, inputting at the input layer the sets of selected images at the given b-values having a corresponding middle slice, and training the neural network to minimise a loss function that measures similarity between the predicted images corresponding to the given b-values thereby generated at the output layer and the respective ground truth images. In this case:

[0049] — the neural network may be further configured to calculate a predicted apparent diffusion coefficient map from the predicted images generated at the output layer,

[0050] — the method may further include calculating a respective ground truth apparent diffusion coefficient map of each slice of the subset of slices from the ground truth images formed of that slice, and

[0051] — the loss function may also measure similarity between the predicted and corresponding ground truth apparent diffusion coefficient maps.

[0052] The training can thus force the network to make accurate predictions of the apparent diffusion coefficient map, as well as b-value images.

[0053] The provided training data may include diffusion-weighted images of human or animal subjects.

[0054] The term "computer readable storage medium" may represent one or more devices for storing data, including read only memory (ROM), random access memory (RAM), magnetic RAM, core memory, magnetic disk storage mediums, optical storage mediums, flash memory devices and / or other machine readable mediums for storing information. The term "computer-readable storage medium" includes, but is not limited to portable or fixed storage devices, optical storage devices, wireless channels and various other mediums capable of storing, containing or carrying instruction(s) and / or data.

[0055] Furthermore, embodiments may be implemented by hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware or microcode, the program code or code segments to perform the necessary tasks may be stored in a computer readable medium. One or more processors may perform the necessary tasks. A code segment may represent a procedure, a function, a subprogram, a program, a routine, a 008614430

[0056] 1 subroutine, a module, a software package, a class, or any combination of instructions, data structures, or program statements. A code segment may be coupled to another code segment or a hardware circuit by passing and / or receiving information, data, arguments, parameters, or memory contents. Information, arguments, parameters, data, etc. may be passed, forwarded, or transmitted via any suitable means including memory sharing, message passing, token passing, network transmission, etc.

[0057] Summary of the Figures

[0058] Embodiments and experiments illustrating the principles of the invention will now be discussed with reference to the accompanying figures in which:

[0059] Figure 1 shows schematically the core architecture of an Enhanced-quickDWI deep-learning neural network model with linear output activation.

[0060] Figure 2 shows a workflow illustrating quantitative improvement in image quality using the Enhanced- quickDWI approach relative to the original noisy input data and an earlier DNIF approach. Delineations for different body regions are obtained, and the pixel-wise root-mean-square error (RMSE) of ADC values and / or b-value images within delineated body regions are computed, comparing each of the following image pairs: (i) noisy input data versus clinical images, (ii) Enhanced-quickDWI data versus clinical images, and (iii) DN I F-processed data versus clinical images.

[0061] Figure 3 shows training and validation curves for the whole-body Enhanced-quickDWI model, and the abdomen-pelvis Enhanced-quickDWI model.

[0062] Figure 4 shows results from an example whole-body validation patient for heavily sub-sampled noisy input data, ground-truth data, Enhanced-quickDWI-generated and D / V / F-generated images. Left top: high b-value images (b = 900 s / mm2) presented alongside calculated ADC maps. Left bottom: sagittal plane reconstructions from the high b-value images and the ADC maps. Right: Maximum Intensity Projection reconstructions from the high b-value images.

[0063] Figure 5 shows quantitative comparisons of value prediction for 16 validation patients. Box plots indicate the range of RMSE calculated for each body region for respectively: ground truth values vs noisy input, ground truth values vs D / V / F-generated and ground truth values vs Enhanced-quickDWI generated.

[0064] Figure 6 shows results from a qualitative reader study assessing overall image quality.

[0065] Detailed Description of the Invention

[0066] Aspects and embodiments of the present invention will now be discussed with reference to the accompanying figures. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned in this text are incorporated herein by reference. 008614430

[0067] 8

[0068] In the following we describe a study using a context-aware, deep learning neural network which is arranged to denoise heavily sub-sampled diffusion-weighted imaging data. The network produces images with equivalent clinical image quality, achieved by retaining the network’s 2D nature but introducing spatial context through slices neighbouring to the slice of interest (an approach which is henceforth called “Enhanced-quickDWr). An advantageous optional feature of the neural network is that it can also be made multi-b-value, processing simultaneously all respective b-values. A further advantageous optional feature is a special loss function that can be used during training of the neural network, the loss function being based on calculation of ADC values.

[0069] Due to the quantitative nature of DWI and the potential of ADC to assess intra-lesion changes to monitor treatment response, the study includes a validation of the Enhanced-quickDWI network-generated images which goes beyond the assessment of contrast. In particular, the validation includes thorough quantitative and qualitative analysis. The study also compares the results obtained by performing Enhanced-quickDWI against the DNIF approach of Zormpass-Petridis et al. [13, 14] discussed earlier. In particular the study:

[0070] (i) blindly compares radiological image quality of Enhanced-quickDWI with conventional DWI in a large patient cohort, and

[0071] (ii) quantitatively compares Enhanced-quickDWI results against DNIF results.

[0072] Materials and Methods

[0073] Image Acquisition

[0074] Patients were split into training (N = 84), validation (N = 16) and test (N = 84) groups. Validation data were used for quantitative evaluation of model losses / metrics during training, whilst test data were used for a qualitative radiological evaluation to identify the model which achieved the best quantitative performance. All images were acquired on one of three 1.5T scanners (two Aera™ and one Sola™, Siemens Healthcare GmbH). Multi-directional diffusion-weighted (MDDW) protocols were employed for acquiring whole-body and abdomen-pelvis DWI examinations, the parameters for which are defined in Table 1 . 2D distortion correction was applied to all source DWI images.

[0075] 008614430

[0076] 9

[0077] Table 1 : Parameters of MDDW protocols for acquiring whole-body and abdomen-pelvis DWI examinations

[0078] Deep Learning Model

[0079] The core architecture of the Enhanced-quickDWI neural network consists of an encoder-decoder path using residual blocks and skip connections, as shown in Figure 1 . Each convolution layer consists of a 3x3 filter with ReLU activation and the weights incident to each hidden unit are constrained to have a norm value of less than or equal to 3. The number of filters for each block were 64, 128, 256, 512 and 1024 respectively.

[0080] The input to the Enhanced-quickDWI model provides images of three contiguously successive slices, namely the middle slice of interest and the neighbouring slices to either side of that slice. The neighbouring slices are used by the model as enhanced spatial context for predicting the 2D middle slice rather than for predicting a 3D volume corresponding to all three slices.

[0081] In addition, the Enhanced-quickDWI model in this study receives, for each slice, images at three different b-values. Thus the Enhanced-quickDWI model has nine input channels: (3 successive slices where the slice of interest is in the middle) x (3 corresponding b-values for every slice). Moreover, as the output of core architecture is three predicted images at the respective b-values for the slice of interest, it has three 008614430

[0082] 10 output channels. For comparison, the DNIF model has one input channel for an image depicting one slice at one b-value, and one corresponding output channel.

[0083] Image voxels of the input image depicting each slice are transformed using a log-transform (any zero values being set to 1). This helps to reduce the influence of noisy outliers in the input image (i.e. it reduces the influence of a few high-valued pixels). Image standardization is then performed by the Enhanced-quickDWI model according to the following equations: where Xi is the log-transformed, as-acquired magnitude of the ithvoxel out of a total of N voxels in the image, and Zi is the standardised log-transformed magnitude of the ithvoxel.

[0084] The Enhanced-quickDWI model in this study also has an extra, non-trainable layer which calculates the ADC map for the slice of interest from the three predicted b-value images from the output channels. Advantageously, this ADC map can also be included in the overall loss function for the model. In particular, the loss function used in the study is the sum of mean-absolute-error (MAE) between log- transformed predicted images and ground truth images (discussed below) for the three b-values and the corresponding ADC, with the b-values and ADC being treated equally in the loss function. Thus the model is regularized by forcing the network to make accurate predictions of the ADC map during training, as well as b-value images.

[0085] The network training parameters and data preprocessing information are presented in Table 2. The network was trained separately on the whole-body (b = 50, 600 and 900 s / mm2) data and subsequently on the abdomen-pelvis (b = 100, 600 and 1050 s / mm2) data. The abdomen-pelvis model was initialized using pretrained weights from the trained whole-body model. Also included in Table 2 for comparison are the corresponding data and information for the DNIF model.

[0086] 008614430

[0087] 11

[0088] Table 2: Details of differences between training parameters for DNIF (whole-body and abdomen-pelvis), Enhanced-quickDWI (whole-body) and Enhanced-quickDWI (abdomen-pelvis) models

[0089] To produce the training and validation data, every third slice from the successive slices of the MDDW data was identified as a slice of interest. For each of the three b-values of that slice, an image depicting the slice from a single diffusion direction and a single signal excitation was randomly selected from the MDDW data. Three-slice slabs were then created placing the slice of interest in the middle. For each of the neighbouring slices to either side of the middle slice and for each b-value, a single signal excitation image was selected from the MDDW data depicting that slice and having the same diffusion direction as the middle slice image at the corresponding b-value. Thus each slab provided three sets of images for the three b-values, each set being formed of three signal excitation images depicting the middle and the two neighbouring slices for a single diffusion direction of that b-value.

[0090] Each set of three images for a given b-value was then paired with a respective denoised, ground truth image for that b-value formed by combining all the MDDW images of the middle slice at the b-value (i.e. all excitations and all diffusion directions) in a respective trace-weighted diffusion-weighted image. The 008614430

[0091] 12 training and validation data thus had pairings of nine single excitation images (i.e. three slices x three b- values) with three ground truth images (i.e. three b-values) for each slice of interest to match the nine input channels and three output channels of the Enhanced-quickDWI neural network. The initial randomized selection from the MDDW data ensured that the network received random direction information.

[0092] This procedure resulted in a total of 16480 input / output data pairs for the whole-body training images, 5469 data pairs for the abdomen-pelvis training images, 2960 data pairs for the whole-body validation images, and 1122 data pairs for the abdomen-pelvis validation images. In addition, the three ground truth images (for the three b-values) of each middle slice of both the training and validation data were used to calculate a respective ground truth ADC map for that slice.

[0093] The Enhanced-quickDWI model was compared against the DNIF model [13, 14], where the input is a 1- channel standardized (but not log transformed) image and the output is also a 1 -channel image (see Table 2). However, the core architecture of the DNIF neural network (i.e. the encoder-decoder path) is the same as that of the Enhanced-quickDWI network. The loss function for training the DNIF model was the mean-absolute error (MAE) between a predicted and a ground truth image. To produce the training and validation data for the DNIF model, for each slice from the MDDW data and each b-value of that slice, an image depicting the slice from a single diffusion direction and a single signal excitation was randomly selected from the MDDW data. This was then paired with a denoised, ground truth image for that b-value formed by combining all the MDDW images of the slice at the b-value (i.e. all excitations and all diffusion directions) in a trace-weighted diffusion-weighted image. This procedure resulted in a total of 75198 image pairs for training, and 7068 available for validation.

[0094] Experimental Design

[0095] To quantitatively evaluate the precision of predicted DWI and ADC values, an externally validated deep learning methodology

[0023] was utilized that automatically delineates skeleton and soft-tissue organ regions within both whole-body and abdomen-pelvis validation patients. This allowed the evaluation to be focused on body regions of interest, and also allowed comparisons between body regions. For each model, the corresponding voxel values for each pair of corresponding predicted and ground truth images (i.e. b-value image pair or ADC map pair) in the validation data were compared. More specifically, the voxel-wise root-mean-square-error (RMSE) in the predicted values relative to the ground truth values was calculated within each delineated region, with the hypothesis that a better model will reduce the RMSE across all b-values and for ADC maps. As a reference baseline for these predicted / ground truth RMSEs, the RMSEs in the original noisy data (i.e. randomly selected single diffusion direction and a single signal excitation b-value images, and corresponding ADC maps calculated from three such images at the different b-values) relative to the ground truth images was also calculated within each delineated region.

[0096] Figure 2 shows a workflow, including a list of delineated body regions, illustrating this quantitative evaluation. 008614430

[0097] 13

[0098] Turning to qualitative evaluation, to assess the clinical utility of our generated images, predictions of the Enhanced-quickDWI model were evaluated by three expert radiologists using a combination of the validation and test datasets, comprising a total of 100 patients. For each patient study, the corresponding pairs of diffusion-weighted images and ADC maps for ground truth target datasets and Enhanced- quickDWI model predictions were uploaded to a secure cloud-based GDPR-compliant radiological image viewing platform (Collective Minds Radiology, www.cmrad.com) for review by independent, external radiologists using a fully anonymized and randomized process. At least two weeks elapsed between consecutive reads of the ground truth and Enhanced-quickDWI processed images (presented in random order) for the same patient study to reduce the risk that a radiologist would recognize a given study from an earlier image. The images were also re-anonymised (double-blinded), so that reads of one image type could not inform the reading of the other. Readers scored overall image quality (OIQ) on a 5-point Likert scale for diffusion-weighted images and ADC maps separately.

[0099] To perform statistical evaluation of the qualitative analysis, a non-inferiority experimental design was used which aimed to provide evidence that rejects the null hypothesis that the average Likert score for image quality from Enhanced-quickDWI processed images is significantly poorer than for clinical, trace-weighted scans (independently for b-value images and ADC maps). A non-inferiority margin of 0.5 was chosen by an internal expert panel of radiologists to define a non-inferiority threshold.

[0100] Results and Discussion

[0101] Model Training

[0102] The Enhanced-quickDWI model achieved adequate convergence after 200 epochs for the whole-body model, and 40 epochs for the tuned abdomen-pelvis model, as shown in Figure 3. However, the wholebody model exhibited several successive steps, indicating that it requires careful training.

[0103] Upon visual inspection, the Enhanced-quickDWI model successfully de-noised the heavily sub-sampled noisy input data and improved their image quality, i.e. increased the CNR. A representative patient example is presented in Figure 4. The Enhanced-quickDWI-generated images appear sharper and with more accurate contrast over the D / V / F-generated images. Additionally, in sagittal plane and whole-body projection reconstructions there is a noticeably smoother transition between neighboring slices (indicated by block arrows at left-bottom and right in Figure 4) which is likely attributed to the inclusion of spatial context in the Enhanced-quickDWI model compared to the DNIF model trained on single slices.

[0104] Quantitative Model Evaluation

[0105] Enhanced-quickDWI was superior to DNIF in more than 90% of cases based on the Enhanced-quickDWI RMSE value for whole-body or abdomen-pelvis images being reduced compared to the DNIF RMSE values. Moreover, both models show decreased RMSE values compared to the noisy input MDDW data on all occasions. These results are observed across all b-values and ADC maps independently, indicating consistent improvement in CNR. An overview of the quantitative evaluation of the precision of predicted values for each of the delineated body regions is demonstrated in Figure 5. In most cases a negative 008614430

[0106] 14 trend is observed indicating that image quality is better for a given region when using the whole-body (top) and abdomen-pelvis (bottom) Enhanced-quickDWI model predictions compared to the DNIF model predictions. Additionally, both models are more accurate compared with the noisy input MDDW data. Based on these results the Enhanced-quickDWI model was selected over DNIF for the independent qualitative reader study.

[0107] Although the same core architecture is shared by the Enhanced-quickDWI and DNIF models, this marked improvement in performance can be explained by the inclusion of neighboring slices as spatial context in the Enhanced-quickDWI model, making image transitions smoother between slices. The inclusion of the ADC calculation inside the loss function also acts as an additional regularizer to the values produced by the network, whereby the loss function is more relevant to the DWI context. Another significant feature is the simultaneous processing by the network of all b-values.

[0108] Qualitative Model Evaluation

[0109] The Enhanced-quickDWI model demonstrated non-inferiority to the ground truth images in all cases for both whole-body and abdomen-pelvis versions for the 100 patients of the validation and test set. This held true for all b-value images and calculated ADC maps. An overview of the reader overall image quality assessment is presented in Figure 6, which also demonstrates that sufficient data were available in this study to justify the non-inferiority claim (95% confidence interval bars shown to be within the noninferiority margin). In particular, results for the whole-body and abdomen-pelvis Enhanced-quickDWI model predictions indicate that non-inferiority may be inferred for ADC maps (dashed bars), and for the for b-value images (solid bars) as all cases lie to the left of the rightmost dashed line (non-inferiority limit).

[0110] Conclusion

[0111] The Enhanced-quickDWI model is a context-aware, deep learning neural network for the de-noising of single excitation (NeX=1) DWI data that results in images of equivalent clinical quality. Advantageously, it can also be made to accept multiple b-value images. Our results indicate that the images produced by the model could be clinically used as a substitute to high-quality clinical images, thereby significantly reducing acquisition times. We performed both quantitative and qualitative analysis in 100 patients providing the rigorous validation needed for the clinical translation of such quantitative imaging applications.

[0112] ***

[0113] The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for obtaining the disclosed results, as appropriate, may, separately, or in any combination of such features, be utilised for realising the invention in diverse forms thereof.

[0114] While the invention has been described in conjunction with the exemplary embodiments described above, many equivalent modifications and variations will be apparent to those skilled in the art when given this disclosure. Accordingly, the exemplary embodiments of the invention set forth above are considered to 008614430

[0115] 15 be illustrative and not limiting. Various changes to the described embodiments may be made without departing from the spirit and scope of the invention.

[0116] For the avoidance of any doubt, any theoretical explanations provided herein are provided for the purposes of improving the understanding of a reader. The inventors do not wish to be bound by any of these theoretical explanations.

[0117] Any section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.

[0118] Throughout this specification, including the claims which follow, unless the context requires otherwise, the word “comprise” and “include”, and variations such as “comprises”, “comprising”, and “including” will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

[0119] It must be noted that, as used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” one particular value, and / or to “about” another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by the use of the antecedent “about,” it will be understood that the particular value forms another embodiment. The term “about” in relation to a numerical value is optional and means for example + / - 10%.

[0120] References

[0121] A number of publications are cited above in order to more fully describe and disclose the invention and the state of the art to which the invention pertains. Full citations for these references are provided below. The entirety of each of these references is incorporated herein.

[0122] 1 . Padhani AR et al. METastasis reporting and data system for prostate cancer: practical guidelines for acquisition, interpretation, and reporting of whole-body magnetic resonance imaging-based evaluations of multiorgan involvement in advanced prostate cancer. Eur Urol 2017;71 (1 ):81 — 92

[0123] 2. Eiber M, Holzapfel K, Ganter C, et al. Whole-body MRI including diffusion weighted imaging (DWI) for patients with recurring prostate cancer: technical feasibility and assessment of lesion conspicuity in DWI. J Magn Reson Imaging 2011 ;33(5):1160-1170

[0124] 3. Padhani AR, Koh DM, Collins DJ. Whole-body diffusion-weighted MR imaging in cancer: current status and research directions. Radiology 2011 ;261 (3):700-718

[0125] 4. Chantry A, Kazmi M, Barrington S, et al. Guidelines for the use of imaging in the management of patients with myeloma. Br J Haematol 2017;178(3):380-393

[0126] 5. Myeloma diagnosis and management: NICE guideline [NG35] and appendices. National Institute for Health and Care Excellence Web site, https: / / www.nice.org.uk / guidance / ng35. Published February 2016. Last updated October 2018. Accessed October 2018 008614430

[0127] 16

[0128] 6. Evans R, Taylor S, Janes S, et al. Patient experience and perceived acceptability of whole-body magnetic resonance imaging for staging colorectal and lung cancer compared with current staging scans: a qualitative study. BMJ Open 2017;7(9):e016391

[0129] 7. Evans RE, Taylor SA, Beare S, et al. Perceived patient burden and acceptability of whole body MRI for staging lung and colorectal cancer; comparison with standard staging investigations. Br J Radiol 2018;91 (1086):20170731

[0130] 8. Nougaret, Stephanie, et al. "Endometrial cancer MRI staging: updated guidelines of the European Society of Urogenital Radiology." European radiology 29 (2019): 792-805

[0131] 9. Chernyak, Victoria, et al. "Liver Imaging Reporting and Data System (LI-RADS) version 2018: imaging of hepatocellular carcinoma in at-risk patients." Radiology 289.3 (2018): 816-830

[0132] 10. Baltzer, Pascal, et al. "Diffusion-weighted imaging of the breast — a consensus and mission statement from the EUSOBI International Breast Diffusion-Weighted Imaging working group." European radiology 30 (2020): 1436-1450

[0133] 11. Turkbey, Baris, et al. "Prostate imaging reporting and data system version 2.1 : 2019 update of prostate imaging reporting and data system version 2." European urology 76.3 (2019): 340-351

[0134] 12. Beets-Tan, Regina GH, et al. "Magnetic resonance imaging for clinical management of rectal cancer: updated recommendations from the 2016 European Society of Gastrointestinal and Abdominal Radiology (ESGAR) consensus meeting." European radiology 28 (2018): 1465-1475

[0135] 13. Zormpas-Petridis, Konstantinos, et al. "Accelerating Whole-Body Diffusion-weighted MRI with Deep Learning-based Denoising Image Filters." Radiology: Artificial Intelligence 3.5 (2021): e200279

[0136] 14. Published patent application WO2021 / 052838

[0137] 15. Wessling, Daniel, et al. "Novel deep-learning-based diffusion weighted imaging sequence in 1.5 T breast MRI." European Journal of Radiology (2023): 110948

[0138] 16. Kaye, Elena A., et al. "Accelerating prostate diffusion-weighted MRI using a guided de-noising convolutional neural network: retrospective feasibility study." Radiology: Artificial Intelligence 2.5 (2020): e200007

[0139] 17. Afat, Saif, et al. "Acquisition time reduction of diffusion-weighted liver imaging using deep learning image reconstruction." Diagnostic and Interventional Imaging 104.4 (2023): 178-184

[0140] 18. Lehtinen, Jaakko, et al. "Noise2Noise: Learning image restoration without clean data." arXiv preprint arXiv:1803.04189 (2018)

[0141] 19. Kawamura, Motohide, et al. "Accelerated Acquisition of High-resolution Diffusion-weighted Imaging of the Brain with a Multi-shot Echo-planar Sequence: Deep-learning-based De-noising." Magnetic Resonance in Medical Sciences 20.1 (2021): 99-105

[0142] 20. Ran, Maosong, et al. "Denoising of 3D magnetic resonance images using a residual encoderdecoder Wasserstein generative adversarial network." Medical image analysis 55 (2019): 165-180

[0143] 21. Hu, Yuxin, et al. "RUN-UP: Accelerated multishot diffusion-weighted MRI reconstruction using an unrolled network with U-Net as priors." Magnetic resonance in medicine 85.2 (2021): 709-720

[0144] 22. Aamir, Fariha, et al. "Accelerated diffusion-weighted MR image reconstruction using deep neural networks." Journal of Digital Imaging 36.1 (2023): 276-288 008614430

[0145] 17

[0146] 23. Candito, Antonio, et al. "Deep learning assisted atlas-based delineation of the skeleton from Whole-Body Diffusion Weighted MRI in patients with malignant bone disease." Biomedical Signal Processing and Control 92 (2024): 106099

Claims

00861443018Claims:1 . A computer system for analysing diffusion-weighted magnetic resonance images of an object, the computer system being configured to receive diffusion-weighted images of the object acquired by a magnetic resonance imaging scanner, and being programmed with a neural network which filters the acquired images to produce one or more predicted images from the acquired images; wherein the diffusion-weighted images acquired by the magnetic resonance imaging scanner depict contiguously successive slices through the object, each acquired image being an image depicting a given slice at a given b-value; and wherein the neural network has an input layer that is configured to receive a set of diffusion- weighted images from the acquired images respectively depicting at least three successive slices, and is trained such that it generates therefrom, at an output layer, one or more predicted images depicting a middle one of the successive slices with an improved CNR relative to the diffusion-weighted image received at the input layer depicting the middle slice.

2. The computer system according to claim 1 , wherein image voxels forming each acquired image are transformed using a log-transform.

3. The computer system according to claim 1 or 2, wherein the given b-value of the acquired images received by the input layer has the same diffusion direction for the successive slices.

4. The computer system according to any one of the previous claims, wherein the one or more predicted images depicting the middle slice are trace-weighted images.

5. The computer system according to any one of the previous claims, wherein the input layer receives plural acquired images at respective given b-values depicting each slice of the at least three successive slices.

6. The computer system according to claim 5, wherein the one or more predicted images depicting the middle slice include corresponding diffusion-weighted images at the respective given b-values.

7. The computer system according to claim 5 or 6, wherein the one or more predicted images include an apparent diffusion coefficient map of the middle slice.

8. The computer system according to any one of the previous claims, wherein the neural network is a spatially variant filter.

9. The computer system according to any one of the previous claims, wherein the neural network is a convolutional neural network.

10. The computer system according to any one of the previous claims, wherein the computer system is further programmed such that the neural network receives further sets of diffusion-weighted images00861443019 from the acquired images depicting further sets of three successive slices and generates corresponding predicted images therefrom, the computer system thereby producing a 3D representation of the object from the predicted images.

11. An imaging system for performing diffusion-weighted magnetic resonance imaging, the system including: a magnetic resonance imaging scanner for acquiring the diffusion-weighted images of an object; and the computer system according to any one of the previous claims operatively connected to the scanner to receive the acquired images.

12. A computer-implemented method of analysing diffusion-weighted magnetic resonance images of an object acquired by a magnetic resonance imaging scanner, the method including: receiving the acquired diffusion-weighted images of the object; and using a neural network which filters the acquired images to produce one or more predicted images from the acquired images; wherein the diffusion-weighted images acquired by the magnetic resonance imaging scanner depict contiguously successive slices through the object, each acquired image being an image depicting a given slice at a given b-value; and wherein the neural network has an input layer that is configured to receive a set of diffusion- weighted images from the acquired images respectively depicting at least three successive slices, and is trained such that it generates therefrom, at an output layer, one or more predicted images depicting a middle one of the successive slices with an improved CNR relative to the diffusion-weighted image received at the input layer depicting the middle slice.

13. The method according to claim 12, wherein the object is a human or animal subject.

14. The method according to claim 13, further including: analysing the one or more predicted images for assessment of disease extent in the human or animal subject.

15. A computer program comprising code which, when the code is executed on a computer, causes the computer to perform the method of claim 12.

16. A non-transitory computer readable storage medium storing the computer program of claim 15.

17. The method according to any one of claims 12 to 14, further including: acquiring the diffusion-weighted images using the magnetic resonance imaging scanner.

18. The method according to any one of claims 12 to 14 or 17, further including: displaying the one or more predicted images.0086144302019. A method of training a neural network programmed to produce one or more predicted images by filtering diffusion-weighted images depicting contiguously successive slices through an object acquired by a magnetic resonance imaging scanner, the neural network having an input layer that is configured to receive acquired images respectively depicting at least three successive slices, and having an output layer that generates therefrom one or more predicted images depicting a middle one of the successive slices, the method including: providing training data including diffusion-weighted images at a given b-value of plural objects, the provided images of each object having been acquired by a respective magnetic resonance imaging scanner at contiguously successive slices through that object, and the provided images of each slice depicting that slice at different b-value diffusion directions; for each slice of at least a subset of the slices, combining the provided images of the slice at the different b-value diffusion directions to form a respective ground truth, trace-weighted, diffusion-weighted image depicting that slice at the given b-value; pairing each ground truth image with a respective set of images selected from the provided images at the given b-value, the set of selected images sharing the same b-value diffusion direction and depicting at least three successive slices through the respective object, a middle one of the successive slices being the same slice as that depicted by the ground truth image; and inputting at the input layer each set of selected images, and training the neural network to minimise a loss function that measures similarity between a predicted image thereby generated at the output layer and the ground truth image paired with the inputted set of selected images.

20. The method according to claim 19, wherein the provided images depicting each slice at different b-value diffusion directions are single excitation images.

21. The method according to claim 19 or 20, wherein the subset of slices is every Nthslice through each object, where N is three or more.

22. The method according to any one of claims 19 to 21 , wherein: the input layer of the neural network is configured to receive plural (for example three) acquired images at respective given b-values depicting each slice of the at least three successive slices, and the output layer generates therefrom plural predicted images depicting the middle slice respectively corresponding to the given b-values, for each of the objects, the provided training data includes diffusion weighted images at each of the given b-values, the provided images of each slice depicting that slice at different b-value diffusion directions for each given b-value, for each slice of the subset of slices, a respective ground truth, trace-weighted, diffusion-weighted image depicting that slice is formed for each of the given b-values, the set of images selected from the provided images and paired with each ground truth image are at the respective given b-value of the ground truth image, and for each slice of the subset of slices, inputting at the input layer the sets of selected images at the00861443021 given b-values having a corresponding middle slice, and training the neural network to minimise a loss function that measures similarity between the predicted images corresponding to the given b-values thereby generated at the output layer and the respective ground truth images.

23. The method according to claim 22, wherein: the neural network is further configured to calculate a predicted apparent diffusion coefficient map from the predicted images generated at the output layer, the method further includes calculating a respective ground truth apparent diffusion coefficient map of each slice of the subset of slices from the ground truth images formed of that slice, and the loss function also measures similarity between the predicted and corresponding ground truth apparent diffusion coefficient maps.

24. The method according to any one of claims 19 to 23, wherein the provided training data include diffusion-weighted images of human or animal subjects.

Images