A method of automated slice planning of a thorax

Abstract

A method of automated slice planning of a thorax is disclosed. The method comprises receiving, retrieving, or generating a three-dimensional model of a thorax, generating a three-dimensional probability map for each of at least three points, each point associated with a known anatomical structure in the thorax by applying a trained model to the three-dimensional model of the thorax, identifying the coordinate in each three-dimensional probability map with the highest probability for the respective points, and using the identified coordinates to define a scanning plane in the thorax.

Description

[0001] A method of automated slice planning of a thorax

[0002] Field

[0003] The present invention relates to a method of automated slice planning of a thorax. Specifically, the present invention relates to applying a model to a three-dimensional model of a thorax to generate the automates slice plans. The present invention further relates to training the model which is applied to a three-dimensional model of the thorax. The present invention further relates to a method of assessing the quality of a scan of a thorax.

[0004] Background

[0005] Cardiac magnetic resonance imaging (CMR) scanning allows doctors to create detailed images of the heart. The acquisition of high-quality cardiac magnetic resonance (CMR) images is contingent on the precise positioning of scan planes and the minimization of movement artefacts due to arrhythmia or breathing. Unidentified or incorrectly managed misplanning and movement artefacts degrade image quality, decrease diagnostic confidence, and may necessitate patient recall in severe cases.

[0006] During a scan, radiographers take pictures of different parts of the heart, each of which must be precisely positioned. The need for expert radiographers to perform each scan can make CMR's delivery difficult and can result in a patient experiencing a delay of more than half a year for a scan. Currently, radiographers must manually inspect each acquired image to confirm diagnostic quality and decide whether reacquisition is warranted, for example if the image is not clear enough due to patient movement, patient breathing, or other misalignment reasons.

[0007] Summary

[0008] According to a first aspect of the invention, there is provided a method of automated slice planning of a thorax. The method comprises receiving, retrieving, or generating a three-dimensional model of a thorax. The method comprises generating a three- dimensional probability map for each of at least three points, each point associated with a known anatomical structure in the thorax by applying a trained model to the three-dimensional model of the thorax. The method comprises identifying the coordinate in each three-dimensional probability map with the highest probability for the respective points, and using the identified coordinates to define a scanning plane in the thorax.

[0009] Using such a method, the necessary slices (planes) of a thorax required to perform a cardiac magnetic resonance imaging (CMR) scan can be planned from a three- dimensional model of the thorax (for example, interpolated from a transaxial stack of images), reducing the need for iterative planning or additional localiser scans. Thus, this method may allow for users (e.g., healthcare professionals) to create detailed images of the heart more quickly and efficiently.

[0010] The three-dimensional model of a thorax may comprise a scale, for example a voxel size, or be a three-dimensional cardiovascular magnetic resonance model.

[0011] The plane may be defined as intersecting the at least three identified points, or may be defined as orthogonal to the plane intersecting the at least three identified points, or defined as a plane intersecting one point with a normal vector passing between two identified points.

[0012] The trained model may be an artificial neural network (ANN), for example, a convolutional neural network, a transformer network, or other suitable ANN.

[0013] The thorax may be a human thorax.

[0014] The three-dimensional model of the thorax may comprise or may be generated from a transaxial stack of images of the thorax.

[0015] The number of images in in the transaxial stack of images may be between 10 and 100, for example, there may be 20 or 30 images in the transaxially stack. In some circumstances where a very detailed three-dimensional model is required or desired, there may be more than 100 images in the transaxial stack, for example, between 100 and 1000 images.

[0016] The three-dimensional model may further comprise or may be derived from a voxel size for the transaxial stack or a distance between at least two of the transaxial stack images.

[0017] The three-dimensional model of the thorax may be interpolated from the stack of two- dimensional images of the thorax, and the and the known scale.

[0018] The three-dimensional model may be an isotropic 3D model.

[0019] The three-dimensional model of a thorax may be a three-dimensional magnetic resonance model or volume.

[0020] The method may comprise defining a region of interest in the scanning plane using the identified coordinates.

[0021] For example, the identified coordinates may be used to define the limits of a square, a rectangle, or other suitable shape, on the defined scanning plane. The region of interest may sometimes be referred to as a shimming area.

[0022] By identifying a region of interest, it is possible to ensure the magnetic resonance imaging (MRI) machine shims and minimises magnetic field inhomogeneities over the heart, increasing or optimising cardiac image quality.

[0023] The method may comprise defining a volume of interest based on the region of interest.

[0024] The volume of interest may be defined by increasing the thickness of the region of interest orthogonal to the plane of the region of interest.

[0025] The method may comprise receiving or setting a phase encoding direction of a magnetic resonance imaging scanner used to perform scans defined by the planes, and may comprise defining a field of view by expanding the region of interest in the phase encoding direction to encompass the whole of the body containing the thorax. Setting an appropriate phase encoding direction for the scanner can aid the reduction of wrapping or aliasing effects in the image, increasing image quality.

[0026] The location of the body containing the thorax in the scanner can be determined from the three-dimensional model or transaxial stack of images of the thorax. Further points associated with the edges of the whole body can be identified using the trained model, which may allow for the accurate determination of the edge(s) of the whole body.

[0027] The method may comprise converting the identified probability map coordinates into real-world coordinates using a or the voxel size, and, optionally, an offset representing the location of the three-dimensional model in real-world coordinates.

[0028] At least one of the known anatomical structures associated with the points may be from the list of: the left ventricular anterior wall; the left ventricular inferior wall; the left ventricular anteroseptal wall; the left ventricular inferolateral wall; the left ventricular inferoseptal wall; the left ventricular anterolateral wall; the left ventricular apex; the right ventricular lateral wall; the right ventricular septal wall; the right ventricular apex; an aortic valve coaptation point; and the centre of the left ventricular cavity.

[0029] According to a second aspect of the invention, there is provided a method of refining the location coordinates of at least one point associated with a known anatomical structure in the thorax. The method comprises using the method of the first aspect to define a scanning plane in the thorax, receiving, retrieving, or generating an image from this scanning plane, generating a two-dimensional probability map for each of one or more points in the image, each point associated with a known anatomical structure in the thorax by applying a second trained model to the image, and identifying the two-dimensional coordinate in each two-dimensional probability map with the highest probability for the respective point.

[0030] The image receiving, retrieving, or generating from this scanning plane may be a two- chamber image.

[0031] The method may comprise updating the three-dimensional coordinate of the one of more points using the identified two-dimensional coordinate and the scanning plane. The method may comprising using the updated three-dimensional coordinate to define a scanning plane in the thorax.

[0032] According to a third aspect of the invention, there is provided a method of training a model comprising receiving, retrieving or generating a plurality of images of at least part of a thorax along with the associated scan plane, for each image, identifying whether a point associated with a known anatomical structure in the thorax is present, identifying a three-dimensional coordinate for each point by mapping the identified points associated with a known anatomical structure in the thorax into three- dimensional space using the voxel size, and using the three-dimensional coordinates to train the model.

[0033] In other words, whether a point associated with a known anatomical structure in the thorax is present in a series of two-dimensional images may be identified using a suitable method (e.g., using a separate model, which may also be an artificial neural network, e.g., a convolutional neural network or a transformer network, or using intersecting images with known identified points), these points may then be mapped three-dimensional space using the voxel size of the transaxial stack the images are from, or the scan plane the image was acquired from. A model is then trained using the three-dimensional locations of the identified points. This means that the model can be trained without any manual labelling, reducing the time it takes to train a network, and can also increase the quality of the trained network by using more consistent identification of points.

[0034] The images may be transaxial stack images, and have a known voxel size associated with each image.

[0035] The plurality of images may be a plurality of transaxial stacks of images and an associated voxel size for each transaxial stack.

[0036] Identifying whether a point associated with a known anatomical structure in the thorax is present may be performed using a second trained model.

[0037] The second trained model may also be a network-based model, for example, an artificial neural network, a convolutional neural network, or a transformer network.

[0038] Identifying whether a point associated with a known anatomical structure in the thorax is present may comprise comparing three different images from the same subject which have a common anatomical structure, and identifying the coordinates in each of the three images where the images intersect.

[0039] The method may comprise, for at least one image, manually identifying a point associated with a known anatomical structure in the thorax, mapping the manually identified point associated with a known anatomical structure in the thorax into three- dimensional space using the scan plane, and using the manually-identified three- dimensional point to train the model.

[0040] The method may comprise outputting the weights of the trained model.

[0041] Each image or transaxial stack of images may be of a different thorax. The thoraxes may be human thoraxes.

[0042] The trained model may be the model of the third aspect.

[0043] According to a fourth aspect of the invention, there is provided a method of assessing the quality of a scan of a thorax comprising: receiving or retrieving a series of cardiac magnetic resonance imaging images or a cardiac magnetic resonance imaging video, and scoring the series of images or video using a trained model for at least one domain.

[0044] The scoring may be unbounded and continuous.

[0045] The domain may be a positional domain, i.e., the position of the scan used to take the series of images or video, or it may be a movement domain, i.e., whether a movement artefact is present in the series of images, for example, from patient breathing and / or suffering arrhythmia. Other quality domains may be assessed, for example the sharpness or blur of the image, the contrast of the image. Other quality domains may include freedom from other artefacts, including radiofrequency artefact, "wrap-around" (aliasing) artefact, susceptibility artefact, crosstalk artefact, moire fringes etc. Quality domains may include, but not be limited to those listed at: https: / / radiopaedia.org / articles / mri-artifacts-l?lang = us.

[0046] The method may comprise normalising the scoring of the series of images or video between zero and one. The normalising of the scoring of the series of images or video may be performed using a sigmoid activation function.

[0047] The method may comprise applying a threshold to the score given to the series of images or video to classify the series of images or video into either a first category or a second category.

[0048] For example, if the score is above a threshold, the video or series of images is accepted as high enough quality, and if the score is below that threshold, the video or series of images is rejected as being too low a quality. There may be more than one threshold applied, for example, each series of images or video may be placed into one of three categories: reject outright, send to healthcare professional for review, and accept. Such a "traffic light" system can speed up the acquisition of high-quality images while reducing the need for repeated scans where image quality is sufficient.

[0049] The first category may be for rejection, and the second category may be for acceptance. Accepted videos or series of images may then be analysed by a healthcare professional.

[0050] The method may comprise, in response to the scoring indicating that the series of images or video is too low a quality, outputting an indication of the domain of the trained model.

[0051] The method may comprise, wherein the domain indicated is for movement artefacts, outputting electrocardiography data corresponding to the series of images or video.

[0052] According to a fifth aspect of the invention, there is provided a method of training a model for at least one domain. The method comprises receiving or retrieving a plurality of series of cardiac magnetic resonance imaging images or a plurality of cardiac magnetic resonance imaging videos, each series of images or video associated with a manually assigned quality score for the domain, and training the model for the domain using the series of images or video and the associated quality score.

[0053] According to a sixth aspect of the invention, there is provided a method of the fourth aspect wherein the trained model is the model of the fifth aspect.

[0054] The method of any of the first to sixth aspects, where the model is a two-dimensional or a three-dimensional model. For example, the model may be a two-dimensional or a three-dimensional network model, for example, an artificial neural network, a convolutional neural network, a transformer network etc.

[0055] According to a seventh aspect of the invention, there is provided a computing device performing the method steps of any of the first to sixth aspects.

[0056] According to an eighth aspect of the invention, there is provided a computer readable medium having program instructions for performing the method of any of the first to sixth aspects.

[0057] According to a ninth aspect of the invention, there is provided a computer system comprising: memory, and at least one processing unit; wherein the processor is configured to perform the method of any of the first to sixth aspect.

[0058] Brief description of the drawings

[0059] Certain embodiments of the present invention will now be described, by way of example, with reference to the accompanying drawings, in which:

[0060] Figure 1 is a system block diagram of a computer system and a magnetic resonance imaging system;

[0061] Figure 2 is a process flow diagram across the MRI scanner, the slice planning Al device, and the Al quality control and quality assessment device;

[0062] Figure 3 is a process flow diagram of planning scan planes;

[0063] Figure 4 is a coronal or frontal plane MRI image of a human thorax;

[0064] Figure 5 is a first horizontal MRI image or "slice" of a human thorax;

[0065] Figure 6 is a second horizontal MRI image or "slice" of a human thorax;

[0066] Figure 7 is a third horizontal MRI image or "slice" of a human thorax;

[0067] Figure 8 is a stack of the first, second and third horizontal slices form Figures 5-7;

[0068] Figure 9 is a series of horizontal MRI image slices with key points identified;

[0069] Figure 10 is an interpolated image of a 2-chamber slice using the identified key points;

[0070] Figure 11 is an MRI image of a 2-chamber slice using the identified key points;

[0071] Figure 12(a) is an MRI image with an example of wrap artefact in a chest wall;

[0072] Figure 12(b) is an MRI image without wrap artefact in a chest wall;

[0073] Figure 13 is an MRI image with four identified key points;

[0074] Figure 14 is an MRI image showing a field of view and a region of interest;

[0075] Figure 15 is a process flow diagram of updating scan planes;

[0076] Figure 16 is a process flow diagram of training a model to identify key points from 2D images;

[0077] Figure 17 is a process flow diagram of assessing quality of an MRI image;

[0078] Figure 18 is a process flow diagram of training a model for assessing quality of an MRI image for a domain;

[0079] Figure 19 are four scatter plots of the correlation between Al-judged and expert- judged image quality; and

[0080] Figure 20 are four plots showing the true positive rate against the false positive rate for Al judging clinically inadequate images.

[0081] Detailed description of certain embodiments

[0082] Current slice planning methods and workflow

[0083] Cardiac magnetic resonance imaging (MRI or CMR) requires radiographers to take a series of pictures and videos of the heart in different orientations. Each specific orientation is termed a view. Common views include "2 chamber", "3 chamber", "4 chamber" and "short axis".

[0084] The views must be positioned precisely to ensure the correct cardiac structures are shown in a way that allows important clinical measurements and judgments to be made.

[0085] Because every person's heart is of different morphology and orientations, the exact positioning of images must be personalised to each patient. The process of obtaining the positioning of images is called "planning".

[0086] In current practice, planning requires several bespoke images to be taken, termed localizers. These are a series of images of fixed orientation (in the axial, coronal or sagittal planes), which are used to locate the heart within the patient's chest. Using these images, further localizers are positioned and obtained, through an iterative process, until the first set of clinically useful views can be captured. Later views in the scan are planned using information from earlier views (e.g., the left ventricular outflow tract (LVOT) coronal view is planned from the 3-chamber view).

[0087] This process means the acquisition and interpretation of localizers by radiographers, and the resultant slice planning, enforces a delay before clinically useful images are obtained.

[0088] Slice planning

[0089] Alongside the localizers, there is another sequence performed in every patient which does not require any prior knowledge of the patient's cardiac anatomy. It is acquired at the start of the scan and is called the 'transaxial stack', or 'anatomy stack'.

[0090] The transaxial stack may be a series of axial (cross-sectional) slices which begin at the neck and continue down to the upper abdomen. It therefore captures the entire heart. This is not a localizer - it is a clinically useful sequence which doctors use to examine things like the aorta and the lungs. However, it is unique in that it does not require localizers to be taken beforehand, because it images the entire chest, and is positioned in the axial plane, rather than being personalised to the patient's anatomy. Therefore, the transaxial stack may be acquired for each patient before planning begins. By taking all the slices of the transaxial stack and interpolating them in three- dimensions (3D), an isotropic model of a patient's thorax (e.g., a human thorax) can be generated. This model can then be used by an model (for example, an artificial intelligence (Al) model) to identify the location and orientation of key cardiac structures. From these structures, all the cardiac views needed in a cardiac MRI scan can be calculated immediately.

[0091] Because the transaxial stack is often already acquired for each patient, there is no additional time lost by using this for slice planning. Furthermore, the need for localizers (and their processing by radiographers) may be reduced or even completely abolished. This can reduce the time needed for a cardiac MRI scan and may also improve image quality.

[0092] Referring to Figure 1, a computer system 1 comprising a processor 2, one or more interfaces 3, an image processor 4 (e.g., a graphics processing unit), a monitor 5, memory 6, and storage 7, for example for data files and images. A bus system 8 or other suitable interface connects the computer system 1 to a magnetic resonance imaging scanner system (MRI system scanner) 10, and allows data transfer between the computer system 1 and the MRI scanner system 10.

[0093] The MRI scanner system 10 may be an MRI scanner system suitable for obtaining MRI images of the thorax. The MRI scanner system comprises an MRI scanner 12, which comprises a magnet 13, a shim coil 14, an RF coil 15, and gradient coils 16. The MRI scanner system 10 further comprises gradient amplifiers 19, which may include Gx, GY, GZ amplifiers, 20, 21, 22 and a gradient interface 23 operatively connected to the gradient coils 16. The MRI scanner system 10 also comprises am RF amplifier operatively connected to the RF coil 15. The MRI scanner system 10 further includes a system control module or device 27 for controlling the amplifiers and the MRI scanner, which in turn comprises a processor 28, an interface 29, a pulse generator 30, a transceiver 31 and memory 32. The system control 27 also includes storage 33, for example, for storing instructions, setting, data, or images. There may be other parts to the MRI scanner system 10 to allow MRI images to be acquired from a subject, for example a human.

[0094] The MRI control system l can send instructions to the scanner 12 and amplifiers 19, 25 to obtain scan images of the subject, for example, coronal or frontal images, or horizontal or transaxial images of a subject's thorax. These scan images may then be sent to the computer system 1 via the bus system 8 for further processing using the processors 2, 4.

[0095] Referring to Figure 2, the process of acquiring and processing scan images is illustrated. The MRI scanner 10 obtains planning images 36 of a subject. These planning images may be, for example a series of horizontal or transaxial images of a subject's thorax which include images of a subject's heart. The planning images 36 may optionally include a coronal or frontal scan of the subject's thorax which may be sent to the computer system 1 to be assessed by a radiographer or other healthcare provider, or alternatively be assessed using a model running on the system 1, to ensure that the subject is positioned correctly in the MRI scanner to enable the correct transaxial images 36 to be acquired. The planning images 36 may be a series of transaxial images separated by a known distance, for example 6 mm, 10 mm, or 12 mm, have a scale, or may be converted into a three-dimensional model of a thorax. Thus, the planning images 36 are a three-dimensional representation of at least part of the thorax of the subject.

[0096] The planning images 36 are then sent to the computer system 1 using the interface 8. As will be explained in more detail later, a model is applied to the planning images 36 which identifies key points 37 in three-dimensional space, each point 37 associated with a known anatomical structure in the thorax. The model may be a trained model, for example, an artificial neural network-based model (ANN). The model may be, for example, a convolutional neural network, and / or a three-dimensional neural network, a transformer network, or other suitable ANN. The model may identify any suitable number of key points 37 in the three-dimensional representation of the thorax. The key points 37 do not have to be located on one of the transaxial images 36, they may be in the space between the transaxial slices 36.

[0097] Using at least three of the identified key points 37, a slice through the thorax can be defined, also referred to as slice planning. The slice planning will typically define all the slices which are clinically useful for the health care provider, for example a radiographer or cardiologist.

[0098] Once the slices are planned, the slice parameters 38 are defined and the necessary details that the MRI scanner 10 requires are sent to the scanner 10 via the interface 8. The scanner 10 then plans the cardiac views using the slice parameters 38, and acquires the model-defined images 39, for example, a 2-chamber image, a 4-chamber image etc. Optionally, the model-defined images 39 are then sent to either the same computer system li or a second computer system 12, for quality control analysis and quality assessment. The quality control and quality assessment may use a second trained model to determine whether the acquired images 39 are in the correct plane, are sufficiently clear for clinical analysis and assessment. A quality control readout 40 of the assessment is then generated which can be viewed by a user, e.g., the healthcare provider, and / or the readout 40 or assessment can be sent to the MRI scanner 10 which can either take more scans, or is the quality is sufficiently low, can start the process again by acquiring a new transaxial stack of images for the computer system li to analyse.

[0099] Referring to Figure 3, the method of automated slice planning of a thorax begins with receiving, retrieving, or generating a three-dimensional model of a thorax (step SI). As mentioned previously, the three-dimensional model may be a series of transaxial images 36 with a scale (e.g., a voxel size) or a known distance between them or may be a three-dimensional cardiovascular magnetic resonance model of the thorax including image information.

[0100] A trained model is then applied to the stack of images or the model to generate a three-dimensional probability map for each of at least three points (step S2). Each of these points is associated with a known anatomical structure in the thorax. The known anatomical structures of the thorax may be, for example, the left ventricular anterior wall, the left ventricular inferior wall, the left ventricular anteroseptal wall, the left ventricular inferolateral wall, the left ventricular inferoseptal wall, the left ventricular anterolateral wall, the left ventricular apex, the right ventricular lateral wall, the right ventricular septal wall, the right ventricular apex, an aortic valve coaptation point, and / or the centre of the left ventricular cavity.

[0101] The three-dimensional coordinates associated with the highest probability in each three-dimensional probability map for each of the respective points is then identified (step S3). Optionally, a field of view which encompasses the whole of the body containing the thorax is defined (step S4) which can help with scan planning. In some cases or implementations, the identified three-dimensional coordinates in image or model space are converted into real-world three-dimensional coordinates (step S5), which again can help with scan planning in certain circumstances. The identified three- dimensional coordinates are then used to define a scanning plane of the thorax (step S6). Using such a method, the necessary slices (planes) of a thorax required to perform a cardiac magnetic resonance imaging (CMR) scan can be planned from a three- dimensional model of the thorax which is already acquired at the start of a cardiac MRI scanning session, which may reduce the need for iterative planning or additional localiser scans. Thus, this method may allow for users (e.g., healthcare professionals) to create detailed images of the heart more quickly and efficiently.

[0102] The scanning plane may be defined as intersecting the at least three identified points or may be defined as orthogonal to the plane intersecting the at least three identified points or defined as a plane intersecting one point with a normal vector passing between two identified points.

[0103] Referring to Figure 4, an example of a coronal or frontal image of a human thorax is shown with the heart in the centre.

[0104] Referring to Figures 5 to 7, three transaxial images 36 of the same thorax are shown at different locations, but the same horizonal orientation. Referring also to Figure 8, these images can form a stack of images with a known distance between them.

[0105] As described earlier, the three-dimensional model of the thorax may comprise or may be generated from a transaxial stack of images of the thorax. The number of images in in the transaxial stack of images may be between 10 and 100, for example, there may be 20 or 30 images in the transaxially stack. In some circumstances where a very detailed three-dimensional model is required or desired, there may be more than 100 images in the transaxial stack, for example, between 100 and 1000 images. The three-dimensional model may further comprise or may be derived from a voxel size for the transaxial stack or a distance between at least two of the transaxial stack images. The three-dimensional model of the thorax may be interpolated from the stack of two-dimensional images of the thorax and the known scale. The three- dimensional model may be an isotropic 3D model.

[0106] Referring to Figure 9, twelve tansaxial images from a stack of images 36 are shown, and in each image, the two-dimensional location of a three-dimensional key point is indicated. This is Figure shows a step that is no performed under normal operation, but is optional, and can be of use if the returned scans in the defined planes are not as expected. Al inference for slice olanninq

[0107] The basis of the the model may be a three-dimensional convolutional neural network (3D CNN). A 3D CNN may be a "fully convolutional" neural network, similar to a UNet, but using 3D rather than traditional 2D convolutions.

[0108] Network input

[0109] As described above, the CNN requires a three-dimensional model of the thorax to perform slice planning. This can either be a 'true' three-dimensional CMR acquisition, or, as is often the case, an interpolated 3D model that has been constructed from a 'stack' of two-dimensional images. The transaxial anatomy stack serves this purpose. The transaxial anatomy stack may be interpolated into an isotropic 3D model, and then fed into the network.

[0110] Network output

[0111] As described above, the network is a three-dimensional analogue to well established two-dimensional convolutional neural networks used for image segmentation (e.g., UNet). However, in the present method, the CNN is not performing a segmentation task - instead, it is producing a probability map for various key points 37 that can be used for slice planning.

[0112] This is because many scan planes that we want to plan slices for can be defined as the plane that cuts through three known anatomical structures. For example, a 4-chamber image can be defined as the plane which cuts through the basal inferoseptum, the basal anteroseptum, and the cardiac apex. A 2-chamber image can conversely be defined as the plane which cuts through the basal anterior wall, basal inferior wall and the cardiac apex.

[0113] Other scan planes can be defined from a combination of other scan planes and / or key points 37. For example, the LVOT coronal scan plane is orthogonal to the 3-chamber scan plane, whilst capturing the midpoint of the left ventricle.

[0114] For every key point 37 the model is trained to identify, it produces a tensor (matrix) of numbers of the same size as the input three-dimensional model. The value of every voxel represents the likelihood of the key point 37 lying at that location.

[0115] Therefore, to retrieve the exact spatial coordinates, the coordinate of the voxel with the highest value (probability) needs to be identified for each of the output tensors (where each output represents each key point 37). Referring to Figure 10, optionally, the model is able to "slice" through the interpolated stack (or model) to "predict" what the videos will look like when the scanner 12 acquires the model-planned images 39. These tend to be low resolution images but allow a user to eye-ball the images, and in turn can allow them to determine whether the images the scanner 12 will acquire will be sensible. Referring to Figure 11, an actual scan of the interpolated stack is shown which shows a much higher resolution image which may be suitable for clinical assessment.

[0116] Thus, using this method, there is no need for traditional slice planning which requires dedicated localisers and iterative slice planning. For example, to plan an LVOT coronal view, a user currently has to go through multiple steps of planning, with localisers followed by a short axis localizers, followed by a 3-chamber view, followed by the desired LVOT coronal. The present method instead is able to provide "one shot" planning from a single sequence.

[0117] Further, views can be acquired in any order. Following on from the above, because slice planning is not iterative (where one view is defined from the last), images can be acquired in any order, and undesired views can be omitted.

[0118] Scan times can also be reduced. Because the present method plans images uses a sequence of images that would be acquired anyway (the transaxial stack), and localizers and iterative planning is no longer needed, and users can now choose to omit certain views without this impacting planning, scans can be performed more quickly.

[0119] Planning using the present method may also be of a higher quality and resultant images are more accurate. Because the heart is a 3D model, traditional iterative planning performed in 2D can mean key points are not captured exactly. This results in suboptimal images. The most common example of this is probably "foreshortening", where the cardiac apex is not captured on long axis images, and the heart appears 'shorter' than it truly is. This results in inaccurate measurements being made. Because the present method plans the scan in 3D, it is able to capture the true location of the apex.

[0120] Automated shimming

[0121] MRI scanners 12 rely on magnetic fields to take pictures. The assumption is that these magnetic fields are perfectly uniform, or "homogeneous". However, in practice the magnetic field inevitably has irregularities, termed magnetic field inhomogeneities.

[0122] These inhomogeneities can introduce artefacts into an image, degrading image quality.

[0123] To combat this, MRI scanner systems 10 can undertake a process termed "shimming", where the MRI machine 10 attempts to minimise a particular area's inhomogeneities. In cardiac MRI, ideally, the MRI machine 10 shims over where the heart is located. The location of the heart within the patient's chest in three-dimensional space can be identified using the model, for example, by identifying coordinates of known landmarks in the chest. With this information, it can be possible to ensure the MRI machine 12 shims and minimises magnetic field inhomogeneities over the heart and increase or optimise cardiac image quality.

[0124] As mentioned earlier, a region of interest may be defined in the scanning plane using the identified coordinates. For example, the identified coordinates may be used to define the limits of a square, a rectangle, or other suitable shape, on the defined scanning plane of the thorax. The region of interest may sometimes be referred to as a shimming area. By identifying a region of interest, it is possible to ensure the magnetic resonance imaging (MRI) machine shims and minimises magnetic field inhomogeneities over the heart, increasing or optimising cardiac image quality.

[0125] A volume of interest based on the region of interest may be defined by increasing the thickness of the region of interest orthogonal to the plane of the region of interest.

[0126] A phase encoding direction may be set (or a setting received) of the magnetic resonance imaging scanner 12 used to perform scans defined by the planes and may include defining a field of view by expanding the region of interest in the phase encoding direction to encompass the whole of the body containing the thorax. The setting of an appropriate phase encoding direction for the scanner 12 can aid the reduction of wrapping or aliasing effects in the image, increasing image quality.

[0127] The location of the body containing the thorax in the scanner 12 can be determined from the three-dimensional model or transaxial stack of images of the thorax. Further points associated with the edges of the whole body can be identified using the trained model, which may allow for the accurate determination of the edge(s) of the whole body. This can aid locating the heart within the thorax in image space, but also in the real world, planning particular scan planes, assessing whether unusual or rarely required scans would be useful, and / or scanning the whole area volume of the thorax at the greatest resolution possible.

[0128] The identified probability map coordinates may be converted into real-world or scanner 12 coordinates using a known voxel size, and, optionally, an offset representing the location of the three-dimensional model in real-world coordinates. The key point coordinates 37 are relative to the transaxial stack and are not 'scanner' coordinates. For example, if the model indicates the apex is at the coordinate (200,100,150) in three-dimensional space, this is referring to the coordinates in the transaxial stack, not where the subject's apex lies within the MRI scanner. Therefore, the transaxial coordinates need to be converted to scanner coordinates that the MRI scanner 12 can understand. This can be done, for example, by adding an offset (representing the location of the transaxial stack in 3D space) and multiplying by the size of each voxel (or other suitable scale). For example, if each voxel in the interpolated transaxial stack is 2 mm3, and the location of the apex in the transaxial stack is at voxel (200, 100, 150), then the cardiac apex is 400mm left, (200 x 2), 200mm (100 x 2) posteriorly and 300m cranial (150 x 2) to the starting coordinates of the transaxial stack.

[0129] Now that the coordinates are in scanner 12 coordinates, the scan planes can be defined in scanner coordinates, which is typically defined as the 'normal' vector (a, b, c) and a constant (d), resulting in the equation ax + by + cz = d. Using the three key points 37 for a particular scan plane of interest, two different pairs of these points can be subtracted to find two vectors that must lie on that plane. The cross-product of these vectors provide the normal vector to plane of interest (a, b, c), and d can be found by substituting any key point 37 into the equation. These planes are the slice parameters 38 which can be sent to the MRI scanner 12 via the interface 8.

[0130] Prevention of wrapping artefact

[0131] Referring to Figure 12(a) and Figure 12(b), choosing the right angle and location for a slice is not the only aspect of slice planning. The "field of view" must also be identified. The field of view is simply the spatial region that will be captured in the image.

[0132] One of the common artefacts that can occur if the field of view is not appropriately selected is known as "wrapping" or "aliasing". This phenomenon occurs when a part of a subject (e.g., a patient) that lies outside of the field of view 'wraps around' to the opposite side. This can be seen in the MRI scanner image of a left chest wall in Figure 12(a). The field of view of this MRI scanner image is too small, and so the 'right' of the image, which has been cut off, has 'wrapped' around, and appeared on the 'left' of the image as a wrapping artifact. The outline of this wrapping artefact is indicated by a dashed line in Figure 12(a). The MRI scanner image of Figure 12(b) shows the same left chest wall but with a wider field of view. The 'right' of the image, which was cut off and appeared as the wrapping artefact in Figure 12(a), is present in Figure 12(b) and indicated by the dashed line. The use of the wider field of view removes the wrapping artefact from the image of Figure 12(b).

[0133] In other examples, the 'bottom' of an image may be cut off due to the chosen field of view being too small. In which case, the 'bottom' of the image may appear at the 'top' of the image as a wrapping artefact.

[0134] Therefore, it can be advantageous to choose a field of view big enough that it prevents wrapping artefacts that might obscure important cardiac structures. However, using an unnecessarily large field of view is not always the optimal solution, because the heart in the resultant image will be very small and detail will be lost. It can therefore be important to identify the smallest field of view possible that includes the entire heart and minimises wrap.

[0135] The present model can identify the smallest field of view that includes the entire heart, free of wrapping artefact, for every image it plans. The model identifies the borders of the thoracic cavity using the transaxial stack and calculates if and where wrap may occur within an image.

[0136] The problem of warping can be slightly more complicated than this because wrap only happens in one dimension of an image, which is based on scanner settings. This is because every two-dimensional image is designated to have a phase-encoding direction and a frequency-encoding direction. For example, the horizontal rows of an image might be phase encoded, and the vertical columns frequency encoded, or vice versa. This is important because wrapping artefact only occurs in the direction of phase encoding.

[0137] The present model can optimise the scanner settings, so the optimal phase encoding dimension is selected, by calculating the relative benefits of eliminating wrap in each dimension, and the one with the largest benefit being designated as the frequency encoding dimension.

[0138] Processing of network outputs into shimming area Referring to Figure 13, a shimming area on an MRI machine 10 (e.g., a Siemens MRI machine) is rectangular (if 2D) and cuboidal (if 3D). Ideally, the cuboidal area should be the smallest area that encompasses the region of interest. For each view, a region of interest can be defined using the identified key points 37 as borders. Figure 13 outlines this process. The ideal shimming area of this 4-chamber view is bounded by the anterolateral wall (point 1), the RV lateral wall (point 2), the roof of the atria (point 3), and the cardiac apex (point 4). By identifying these four points using the slice planning model, it is possible to calculate a rectangular shimming area in the outlined box.

[0139] To calculate a shimming volume, the thickness of the slice we are imaging is added to the 2D box. For example, if the slice is 8mm thick, a shimming volume which extends 4mm above and 4mm below the centre of the slice would be defined.

[0140] Processing of network outputs into field of view

[0141] Referring to Figure 14, when calculating the field of view, the region of interest desired is calculated first in the same way as above for shimming. This step provides the region of interest encapsulating the heart within the image. This is calculatable from key points 37 specific to each view (see "Processing of network outputs into shimming area" above for more details).

[0142] However, if the shimming area was simply chosen as the field of view, then wrapping artefact would be present (see "Prevention of wrapping artefact" above). A wrapping artefact only occurs in one dimension, however - which is the direction of phase encoding (set on the scanner).

[0143] To prevent wrapping artefact, the field of view is extended so that it encapsulates the patient's entire body in the phase encoding direction. Referring still to Figure 14, an example of this is shown: the field of view box has been extended so the patient's entire body is captured in the direction of phase encoding (the horizontal axis in the image). However, there is no need to extend the field of view vertically, because this is the phase encoding direction.

[0144] Therefore, the ideal field of view can be calculated form the region of interest (calculatable from the key points 37 from the planning model 37), the extent of the patient's body in the scanner (which we can work out from the transaxial stack), and the direction of phase encoding. These three sets of data serve as the inputs into the algorithm, and the coordinates of the field of view as the outputs. Using the above methods, images can also be automatically optimised by automated shimming and field of view optimisation. Traditionally radiographers have had to make educated guesses based on what field is view is needed to prevent wrap before the image is acquired. However, because the present methods use a 3D model of the thorax to plan images, it can also find the borders of the thoracic cavity.

[0145] Refinement of landmark locations

[0146] Referring to Figure 15, it is possible to refine the location coordinates of at least one of the key points 37 associated with a known anatomical structure in the thorax using a second trained model. Using a scanning plane (step S21), an image of a thorax can be acquired using that scan plane using the method described above (step S22). A second two-dimensional probability map for each of one or more points in that image associated with a known anatomical structure in the thorax is then generated by applying the second trained model to the image (step S23). The second trained model can identify the two-dimensional coordinate in each two-dimensional probability map with the highest probability for the respective key point (step S24). It is then possible to update the three-dimensional coordinate of the one of more originally identified points using the newly identified two-dimensional coordinate and the scanning plane (step S25). Then, the scanning plane can be updated using the updated three- dimensional coordinates.

[0147] Training

[0148] Automated labelling strategies using ground truth knowledge from previously acquired scan planes and transaxial stacks.

[0149] Referring to Figure 16, a model can be trained for use in the methods described above.

[0150] To train a neural network to find these points as described above, a high volume of labelled data ( / .e., example volumes and the ground truth of where the key points lie) is desirable.

[0151] However, manually labelling structures in three-dimensions is very difficult and timeconsuming as it adds an extra dimension beyond labelling an image, the images are low resolution and blurry because they are made up of quite thick slices, and there are multiple key points that must be labelled for each scan. The present model however uses a novel labelling strategy. The model was trained using historic cases where both a three-dimensional volume could be created (the input data), and the desired scan planes where acquired. By definition, the desired anatomic points that define these planes must lie within these images (e.g., if a 4-chamber view is defined by the plane that cuts through the basal inferoseptum, basal anterior wall and cardiac apex, these three points must be visible on the 4-chamber picture acquired by radiographers). Once each scan plane has been acquired, the key points can be found using three different techniques:

[0152] 1. Manually labelling points on the two-dimensional images (still means a human need to label, but this is much easier than labelling in three-dimensions);

[0153] 2. Using an existing model (e.g., and artificial intelligence model) that can find the points in the two-dimensional images; and / or

[0154] 3. Defining a point by the position three different views intersect. For example, the aortic valve is visible on three different views - the 3-chamber, the LVOT coronal view, and the aortic valve view. If the coordinates of where these three views intercept are found, that is, by definition, the tips of the aortic valve.

[0155] The present method(s) can use a combination of these three techniques. This means that the model can be trained without any manual labelling.

[0156] Once the locations of the anatomical points in the two-dimensional images are known, it is possible to 'back project' the two-dimensional coordinates into three-dimensional coordinates using the header information in the scans (e.g., if the apex is known at a certain coordinate in a two-dimensional image, and the position and orientation of the image in some absolute term is known, and the position and orientation of the three- dimensional model is the same absolute terms is also known, then it is possible to work out the three-dimensional coordinate of the apex, also). These 3D coordinates can then be used to train the model (e.g., a three-dimensional convolutional neural network).

[0157] Thus, a method of training a model which can be used to identify the three- dimensional locations of key points associated with known anatomical structures includes receiving a plurality of images of a thorax together with the associated scan plane (step S31). Then for each image, identifying whether a point associated with a known anatomical structure in the thorax is present (step S32), and if present, identifying a three-dimensional coordinate for each point by mapping the identified points associated with a known anatomical structure in the thorax into three- dimensional space using the voxel size (step S33), and using the three-dimensional coordinates to train the model (step S34).

[0158] In other words, whether a point associated with a known anatomical structure in the thorax is present in a series of two-dimensional images may be identified using a suitable method (e.g., using a separate model, which may also be an artificial neural network, e.g., a convolutional neural network or a transformer network, or using intersecting images with known identified points), these points may then be mapped to real-world three-dimensional space using the voxel size of the transaxial stack the images are from, or the scan plane the image was acquired from. A model is then trained using the three-dimensional locations of the identified points. This means that the model can be trained without any manual labelling, reducing the time it takes to train a network, and can also increase the quality of the trained network by using more consistent identification of points.

[0159] A combination of manually and automated training can be used to train the model. For example, a manually identifying a point associated with a known anatomical structure in the thorax can be mapped into three-dimensional space using the scan plane and using the manually identified three-dimensional point in real-world three-dimensional (scanner) space to train the model.

[0160] This model may be refined, or a second model may be generated using scan planes with the locations of known anatomical structures obtained using the methods described above, and for these known locations along with the scan plane to be used to either retrain the existing model, or to train a new model.

[0161] Usinq continuous quality scores to qauqe imaqe quality for separate quality domains. Previous work has aimed to classify images by whether artefact is present as a binary classification task. However, artefacts are almost always continuous in nature, and thresholding these into good / bad makes labelling difficult and the task for the neural network harder. For example, one doctor may find an image adequate, whilst another may want it repeated. Furthermore, the quality required for an image may be specific to the patient and the type of artefact. An image which does not visualise the aorta properly may be acceptable for a patient who is having their second scan within a few months to see if their heart function is improving, but it would not be for someone with suspected aortic disease.

[0162] The present invention approaches the image quality task in two novel ways: • Instead of classification, image quality is judged by providing continuous quality scores. If binary classification is desired, this quality could be simply thresholded. However, more complex classifications are possible, e.g. a traffic light system, where quality scores above e.g. 0.8 are deemed adequate and not needing human review, scores below 0.3 are deemed inadequate and require repeat imaging, and images between 0.3 and 0.8 require manual radiographer review.

[0163] • Image quality is assessed separately across different domains. For example, a 4-chamber image may be poor quality because of being incorrectly positioned (misplanned), having lots of movement artefact due to the patient breathing or suffering arrhythmia, or a combination of the two. If the present model categorises the image quality sufficiently low that an image must be repeated, it must know why it was low quality. For example, if it was misplanned (because the patient's position has changed), then the planning processed should be repeated. However, if there is extensive movement artefact, a free breathing sequence may be chosen.

[0164] Referring to Figure 19, to assess the quality of a scan image of a thorax, a scan image of a thorax is generated, received, or retrieved (step S41), then using a trained model, scoring the series of images or video for at least one domain (step S42). The scoring may be unbounded and continuous.

[0165] The domain may be a positional domain, i.e., the position of the scan used to take the series of images or video, or it may be a movement domain, i.e., whether a movement artefact is present in the series of images, for example, from patient breathing and / or suffering arrhythmia. Other quality domains may be assessed, for example the sharpness or blur of the image, the contrast of the image. Other quality domains may include freedom from other artefacts, including radiofrequency artefact, "wrap-around" (aliasing) artefact, susceptibility artefact, crosstalk artefact, moire fringes etc. Quality domains may include, but not be limited to those listed at: https: / / radiopaedia.org / articles / mri-artifacts-l?lang = us. The scoring of the image or video may be normalised, for example, scored between zero and one, for example, by using a sigmoid activation function.

[0166] A threshold may be applied to the score given to the image, series of images, or video to classify the image, series of images, or video into either a first category or a second category. For example, if the score is above a threshold, the video or series of images is accepted as high enough quality, and if the score is below that threshold, the video, or series of images is rejected as being too low a quality. There may be more than one threshold applied, for example, each series of images or video may be placed into one of three categories: reject outright, send to healthcare professional for review, and accept. Such a "traffic light" system can speed up the acquisition of high-quality images while reducing the need for repeated scans where image quality is sufficient.

[0167] The first category may be for rejection, and the second category may be for acceptance. Accepted videos or series of images may then be analysed by a healthcare professional.

[0168] In some examples, the image quality task may involve indicating the cause of the low quality of an image, series of images, or video. In such examples, the image, series of images, or video is scored by a given model such that it is considered too low a quality. Then, the domain for which the model is trained is indicated. In this way, it can be determined whether the low quality is due to, for example, the position of the scan (positional domain) or from patient breathing and / or suffering arrhythmia (movement domain).

[0169] Movement artefacts due to patient breathing and movement artefacts due to patient suffering arrhythmia are visually very similar. In some examples, an identification of the movement domain is accompanied by electrocardiography data (e.g. an electrocardiogram) of the patient corresponding to the low-quality image, series of images, or video. The electrocardiography data can indicate arrhythmia in the patient and, in this way, can be used to distinguish between movement artefacts due to breathing and movement artefacts due to arrhythmia.

[0170] Training quality control model

[0171] Referring to Figure 20, training a model for quality control and quality assessment for at least one domain is performed by receiving or retrieving a plurality of series of cardiac magnetic resonance imaging images or a plurality of cardiac magnetic resonance imaging videos, each series of images or video associated with a manually assigned quality score for the domain (step S51). The domains may be any of the domains described and references above. A model is then trained for the domain using the series of images or video and the associated quality score (step 52). The quality control model (e.g., a three-dimensional convolutional neural network) was designed and trained to detect and quantify misplanning and movement artefacts in long-axis cines. A unified network incorporating four distinct output heads was utilised for assessment of image quality across four tasks: (1) 4-chamber misplanning, (2) 2-chamber misplanning, (3) 3-chamber misplanning, (4) arrhythmia or breathing artefact identification. Back-propagation was selectively performed across these heads based on the labels present for each cine.

[0172] Cases were randomised in a 3: 1 : 1 ratio to the training, validation and testing datasets.

[0173] The quality of each video was graded on a continuous scale by a level three CMR expert, focusing separately on planning and movement artefacts. Each image in the testing set was also quadruple reported by four level three CMR experts, providing a consensus decision on their clinical adequacy.

[0174] The model's (e.g., a 3D CNN's) assessment of image quality was evaluated using Spearman's rho. The ability to identify images requiring reacquisition was assessed using the area under receiver operating characteristic (AUROC).

[0175] Quality control results

[0176] A total of 2037 cines across 1791 studies were included (Table 1). 1260, 398, and 379 cines were assigned to the training, validation, and testing sets, respectively.

[0177] Referring to Figure 19, the model (e.g., Al algorithm) exhibited strong correlation with human expert judgement of image quality, with Spearman's rho of 0.82, 0.77, 0.81 and 0.83 for 4 / 2 / 3-chamber planning, and freedom of arrhythmia or breathing artefacts respectively.

[0178] Referring to Figure 20, the model (Al model) also showed high efficacy in flagging clinically inadequate cines (AUROC 0.91, 0.88 and 0.94 for misplanning of 4 / 2 / 3- chamber cines, respectively; 0.92 for excessive breathing or arrhythmia artefact; Fig 2).

[0179] Conclusions

[0180] Al can independently assess distinct key domains of long-axis CMR cine quality: planning accuracy and freedom from arrhythmia or breathing artefacts. These ratings can be used to identify cases where acquisition is warranted and could guide specific corrective actions to optimise image quality, such as replanning, prospective gating, or real-time imaging.

[0181] Table 1

[0182] Distribution of LAX cines and their corresponding studies across different labels (2- chamber misplanning, 3-chamber misplanning, 4-chamber misplanning and arrhythmia or breathing) for training, validation and testing datasets.

[0183] Using this method, radiographers do not need to review every image during workflow, saving a lot of time. Because these methods produce continuous quality scores, certain images can be prioritised for review during the scan, whereas images the model judges to be perfect can be deprioritised.

[0184] The present method is not only able to identify images that need to be repeated due to low quality, but it can identify why the quality is low (the exact cause of the artefact). This allows the automatic replanning of images where that is the appropriate step, or switch to a different sequence type that is more robust to the artefact present. Thus, saving time for the user and the subject.

[0185] Modifications

[0186] It will be appreciated that various modifications may be made to the embodiments hereinbefore described. Such modifications may involve equivalent and other features which are already known in the design and use of methods for identifying features in images and which may be used instead of or in addition to features already described herein. Features of one embodiment may be replaced or supplemented by features of another embodiment.

[0187] Although claims have been formulated in this application to particular combinations of features, it should be understood that the scope of the disclosure of the present invention also includes any novel features or any novel combination of features disclosed herein either explicitly or implicitly or any generalization thereof, whether or not it relates to the same invention as presently claimed in any claim and whether or not it mitigates any or all of the same technical problems as does the present invention. The applicants hereby give notice that new claims may be formulated to such features and / or combinations of such features during the prosecution of the present application or of any further application derived therefrom.

Claims

Claims1. A method of automated slice planning of a thorax, the method comprising: receiving, retrieving, or generating a three-dimensional model of at least part of a thorax; generating a three-dimensional probability map for each of at least three points, each point associated with a known anatomical structure in the thorax by applying a trained model to the three-dimensional model of the thorax; identifying the coordinate in each three-dimensional probability map with the highest probability for the respective points; and using the identified coordinates to define a scanning plane in the thorax.

2. The method of claim 1, wherein the three-dimensional model of the thorax comprises or is generated from a transaxial stack of images of the thorax.

3. The method of claim 2, wherein the three-dimensional model further comprises or is derived from a voxel size for the transaxial stack or a distance between at least two of the transaxial stack images.

4. The method of claim 1 wherein the three-dimensional model of a thorax is a three-dimensional magnetic resonance model or volume.

5. The method of any of claims 1 to 4 comprising: defining a region of interest in the scanning plane using the identified coordinates.

6. The method of claim 5 comprising defining a volume of interest based on the region of interest.

7. The method of claims 5 or 6, comprising: receiving or setting a phase encoding direction of a magnetic resonance imaging scanner used to perform scans defined by the planes; and defining a field of view by expanding the region of interest in the phase encoding direction to encompass the whole of the body containing the thorax.

8. The method of any of claims 1 to 7, comprising:converting the identified probability map coordinates into real-world coordinates using a or the voxel size, and, optionally, an offset representing the location of the three-dimensional model in real-world coordinates.

9. The method of any of claims 1 to 8 wherein at least one of the known anatomical structures associated with the points is from the list of: the left ventricular anterior wall; the left ventricular inferior wall; the left ventricular anteroseptal wall; the left ventricular inferolateral wall; the left ventricular inferoseptal wall; the left ventricular anterolateral wall; the left ventricular apex; the right ventricular lateral wall; the right ventricular septal wall; the right ventricular apex; an aortic valve coaptation point; and the centre of the left ventricular cavity.

10. A method of refining the location coordinates of at least one point associated with a known anatomical structure in the thorax, the method comprising: retrieving or generating a scanning plane defined using the method of any of claims 1 to 9; receiving, retrieving, or generating an image from the scanning plane; generating a two-dimensional probability map for each of one or more points in the image, each point associated with a known anatomical structure in the thorax by applying a second trained model to the image; and identifying the two-dimensional coordinate in each two-dimensional probability map with the highest probability for the respective point.

11. The method of claim 10, comprising: updating the three-dimensional coordinate of the one of more points using the identified two-dimensional coordinate and the scanning plane.

12. The method of claim 11, comprising: using the updated three-dimensional coordinate to define a scanning plane in the thorax.

13. A method of training a model comprising: receiving, retrieving, or generating a plurality of images of at least part of a thorax along with the associated scan plane of each image; for each image, identifying whether a point associated with a known anatomical structure in the thorax is present; identifying a three-dimensional coordinate for each point by mapping the identified points associated with a known anatomical structure in the thorax into three-dimensional space using the voxel size; and using the three-dimensional coordinates to train the model.

14. The method of claim 13, wherein the images are transaxial stack images, and have a known voxel size associated with each image.

15. The method of claim 13 or 14, wherein identifying whether a point associated with a known anatomical structure in the thorax is present is performed using a second trained model.

16. The method of any of claims 13 to 15, wherein identifying whether a point associated with a known anatomical structure in the thorax is present comprises comparing three different images from the same subject which have a common anatomical structure, and identifying the coordinates in each of the three images where the images intersect.

17. The method of any of claims 13 to 16 comprising: for at least one image, manually identifying a point associated with a known anatomical structure in the thorax; mapping the manually identified point associated with a known anatomical structure in the thorax into three-dimensional space using the scan plane; and using the manually-identified three-dimensional point to train the model.

18. The method of any of claims 13 to 17 comprising: outputting the weights of the trained model.

19. The method of any of claims 1 to 9, wherein the trained model is the model of any of claims 13 to 18.

20. A method of assessing the quality of a scan of a thorax comprising:receiving or retrieving a series of cardiac magnetic resonance imaging images or a cardiac magnetic resonance imaging video; and scoring the series of images or video using a trained model for at least one domain.

21. The method of claim 20, comprising: normalising the scoring of the series of images or video between zero and one.

22. The method of claim 20 or 21, comprising: applying a threshold to the score given to the series of images or video to classify the series of images or video into either a first category or a second category.

23. The method of any one of claims 20 to 22, comprising: in response to the scoring indicating that the series of images or video is too low a quality, outputting an indication of the domain of the trained model.

24. The method of claim 23, comprising: wherein the domain indicated is for movement artefacts, outputting electrocardiography data corresponding to the series of images or video.

25. A method of training a model for at least one domain comprising: receiving or retrieving a plurality of series of cardiac magnetic resonance imaging images or a plurality of cardiac magnetic resonance imaging videos, each series of images or video associated with a manually assigned quality score for the domain; and training the model for the domain using the series of images or video and the associated quality score.

26. The method of claim 20, wherein the trained model is the model of claim 25.

27. The method of any of claims 1 to 25 wherein the model is a two-dimensional or a three-dimensional model.

28. A computing device performing the method steps of any of claims 1 to 27.

29. A computer readable medium having program instructions for performing the method of any of claims 1 to 27.

30. A computer system comprising: memory; and at least one processing unit; wherein the processor is configured to perform the method of any one of claims 1 to 27.

Images