An improved method and protocol for utilisation of contrast agent in imaging applications

Abstract

There is provided a method for contrast agent protocol selection for a subject under assessment, comprising the steps of: a) obtaining, from a scan acquisition process performed by a medical imaging

Description

The present invention relates to an improved method of, and apparatus for, utilisation of contrast agent in imaging applications. More particularly, the present invention relates to an improved method of, and apparatus for, recommending selective utilisation of contrast agent in imaging applications. In embodiments, the present invention relates to scan protocol selection and acquisition in response to one or more indicators of pathology and one or more indicators of contrast enhancement. At least one indicator may comprise a pathology indicating metric or quantitative parameter for contrast enhancement. A known non-invasive technique for imaging brain and other body regions is computerised tomography (CT) scans which combine a series of X-ray images taken from different angles around the body and use computational methods to create cross-sectional images (slices) of the bones, blood vessels and soft tissues inside the body. Another technique is positron emission tomography (PET) which can be used to produce detailed 3-dimensional images of the inside of a body. A PET scan utilises radiotracers, which are molecules comprising a small amount of radioactive material that can be detected on the PET scan. They are designed to accumulate in cancerous tumours or regions of inflammation. They can also be made to bind to specific proteins in the body. A further known technique for patient imaging is structural magnetic resonance imaging (MRI). MRI is a non-invasive technique for examining the physical structure of the body (for example, calculation of the volumes of tissue). This is of high value for monitoring tissue damage and, particularly, neurodegenerative diseases. As is well known, MRI is based on the magnetisation properties of atomic nuclei. A large and uniform external magnetic field aligns proteins within the water nuclei of the tissue under examination. The alignment is then perturbed by an external radio frequency (RF) signal. The nuclei return to a resting alignment through numerous different relaxation processes, during which RF signals are emitted. By varying the emission and detection sequences of RF pulses, different characteristics of the tissue under examination can be measured. Repetition Time (TR) is the amount of time between successive pulse sequences applied to the same slice. Time to Echo (TE) is the time between the delivery of the RF pulse and the receipt of the echo signal. Numerous MRI techniques are available. T1 -weighted and T2-weighted scans are common. T1 (longitudinal relaxation time) is a time constant determining the rate at which excited protons realign with the applied external magnetic field. T2 (transverse relaxation time) is a time constant determining the rate at which excited protons lose phase coherence with the nuclei having a spin perpendicular to the applied external magnetic field. T1 -weighted images are characterised by short TE and TR times. Conversely, T2-weighted images are produced by using longer TE and TR times. An increasingly used sequence is the Fluid Attenuated Inversion Recovery (FLAIR). FLAIR is similar to a T2-weighted image except that the TE and TR times are very long. Using this approach, abnormalities remain bright but cerebrospinal fluid is attenuated and so is dark in the obtained image. Such a system is also referred to as a T2 FLAIR scan. Diffusion-weighted imaging (DWI) is a form of MRI based upon measuring a random Brownian motion of water molecules within a voxel (volume pixel) of measured tissue. In general, highly cellular tissues or those with cellular swelling exhibit lower diffusion coefficients. Diffusion is particularly useful in tumour characterisation and cerebral ischaemia. A further technique is a gradient echo (GRE) MRI scan. This is produced by a single RF pulse in combination with a gradient reversal. Following an RF pulse, a first negative part of a lobe of the gradient causes a phase dispersion of the precessing spins. Upon reversal of this gradient is reversed, the spins refocus and form a gradient (recalled) echo. A GRE scan typically involves short TR and short TE values and so provides for rapid signal acquisition. Consequently, GRE sequences can provide rapid imaging and MR angiographic techniques. Susceptibility weighted imaging (SWI) is a 3D high-spatial resolution velocity-corrected gradient echo MRI sequence which utilises compounds having paramagnetic, diamagnetic, and ferromagnetic properties. Such paramagnetic compounds include deoxyhaemoglobin, ferritin and haemosiderin. Diamagnetic compounds include bone minerals and dystrophic calcifications. A variant of GRE is a susceptibility weighted angiography (SWAN). SWAN allows for high-resolution visualisation of both cerebral veins and arteries in one sequence without application of contrast agent and with significantly shortened scan time than other arrangements. Finally, turbo spin echo (TSE) (also known as Fast Spin Echo - FSE) is an adaptation of the normal spin-echo (SE) acquisition technique designed to reduce imaging time. In a standard SE sequence, a single echo is measured during each repetition time (TR). IN TSE, multiple echoes are recorded after each TR. This is achieved by transmitting a series of 180-degree inversion pulses at predetermined intervals and recording the corresponding echo signal according to a different phase encoding gradient. In this way, multiple phase-encoded steps can be encoded after a single 90-degree pulse. The use of contrast agents has become widespread in imaging applications such as CT, PET and MRI. It is of particular interest and proves highly effective in brain imaging and in detecting and characterizing brain lesions. Contrast agents may take a number of forms. Known contrast agents may include magnesium-based contrast agents, iron oxide-based contrast agents, superparamagnetic iron oxide nanoparticle contrast agents such as ferumoxytol, manganese-based contrast agents and protein-based contrast agents. One suitable and popular class of contrast agent for use in MRI applications is gadolinium-based contrast agents (GBCAs). Two techniques that utilise GBCAs are contrast-enhanced MRI (CE-MRI) and Dynamic Contrast-Enhanced MRI (DEC-MRI). In CE-MRI, Gadolinium contrast is used to improve the clarity of the imaging available. CE-MRI can improve the visibility of inflammation, blood vessels, and tumours, for example. DCE-MRI, multiple MR images are acquired sequentially following administration of contrast. This allows the contrast of contrast (“wash in” and “wash out”) to be monitored, enabling improved detection of, for example, vascular lesions and tumours. GBCAs are in fact used in up to 35% of brain MR imaging examinations to outline vessels or, most commonly, to exploit the consequences of blood-brain barrier (BBB) disruption to detect certain brain pathologies. For example, the leakage of GBCAs in the extracellular space is visible as T1 post-contrast enhancement thanks to the T1 shortening effects of extravascular GBCAs. Based on the presence, location and pattern of contrast enhancement, it is possible to detect and characterize many brain pathologies for diagnostic and prognostic purposes, including primary and secondary brain tumours, inflammatory and infectious processes as well as some vascular pathologies such as infarct and haemorrhage. However, there are known safety concerns regarding the use of GBCAs. It has been observed that the intravenous injection of GBCAs is associated with nephrogenic systemic fibrosis (NSF) in patients with advanced renal disease. In addition, varying degrees of deposition and retention of gadolinium in the brain and other organs, irrespective of renal function or integrity of the BBB, has culminated with the suspension of some GBCAs in Europe in 2018 and in the update of recommendations by the Food and Drug Agency (FDA) and the European Medical Agency (EMA) restricting the use of GBCAs to situations of clinical necessity. On top of the minimisation of the side effects, the reduction of GBCA administration could improve patient burden, as well as positively impact the costs and logistics within radiology clinics and reduce the environmental footprint. This also has applicability for other forms of imaging modality and contrast agent form. Therefore, there exists a technical problem that existing methods for improving contrast in medical scans require the use of potentially problematic contrast agents. Two known solutions are simply a blanket reduction of contrast agent dosage, and the generation of synthetic contrast images using computational techniques. Neither approach is ideal in a clinical setting, and both ignore the reality that in specific clinical situations contrast agent use is a preferable method. The present invention, in embodiments, addresses these issues. Therefore, there exists a need in the art for an improved process for governing contrast agent usage. The present invention, in embodiments, addresses the above issues. The present invention as described and claimed below may be used with, but is not limited to, all the above described techniques. According to a first aspect of the present invention there is provided a method for contrast agent protocol selection for a subject under assessment, the method being executed by at least one hardware processor and comprising the steps of: a) obtaining, from a scan acquisition process performed by a medical imaging system, medical image data representative of the anatomical structure and appearance of at least a part of the subject; b) utilising the medical image data in a first machine learning model trained to classify anatomical features in the medical image data to generate an output identifying any quantitative pathological indicators representative of the presence of pathological features in the medical image data; c) utilising the medical image data in a second machine learning model trained to predict an effect of contrast agent on the medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features derived from the medical image data; d) comparing the outputs from the first and second machine learning models to determine any correlation between any identified quantitative pathological indicators and any identified quantitative contrast indicators and, if any correlation is identified, the method further comprises: e) generating a recommendation for a scan acquisition process using contrast agent protocols. In one embodiment, the output from the first machine learning model comprises a first feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative pathological indicators have an associated spatial location on the first feature map. In one embodiment, the output from the second machine learning model comprises a second feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative contrast indicators have an associated spatial location on the second feature map. In one embodiment, the second machine learning model is configured to generate synthetic contrast agent enhanced medical image data from the medical image data. In one embodiment, the second feature map comprises a contrast difference map representative of the difference between the synthetic contrast agent enhanced medical image data and the medical image data. In one embodiment, step d) comprises determining any correlation comprises determining whether the spatial location of any identified quantitative pathological indicators overlaps with the spatial location of any identified quantitative contrast indicators. In one embodiment, if no quantitative pathological indicators are identified in step b), then the method further comprises: f) utilising a third computational model to reduce the synthetic contrast in regions of the synthetic contrast agent enhanced medical image data related to anatomically normal features to generate calibrated synthetic contrast agent enhanced medical image data. In one embodiment, subsequent to step f), the method further comprises: g) utilising a fourth computational model to segment the calibrated synthetic contrast agent enhanced medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features in the medical image data; and if any quantitative contrast indicators are identified, the method comprises: h) generating a recommendation for a scan acquisition process using contrast agent protocols. In one embodiment, subsequent to a recommendation for a further scan acquisition process using contrast agent protocols in step e), the method further comprises: i) determining the ratio of the size of the pathological features associated with the identified quantitative pathological indicators to the size of the contrast-enhanced pathological features associated with the identified quantitative contrast indicators; j) determining, based on the ratio, one or more quantitative pathological indicators representative of specific pathological features; and k) recommending a contrast dosing regime in dependence upon the determined quantitative pathological indicators. In one embodiment, if at step b) any quantitative pathological indicators representative of the presence of acute pathological features are identified the method further comprises: I) generating a notification indicative of the probable need for urgent treatment of the subject. In one embodiment, step a) further comprises: m) performing a scan acquisition process on a subject in the medical imaging system to obtain the medical image data. In one embodiment, the scan acquisition process comprises a plurality of different scan types to generate the medical image data. In one embodiment, the medical image data utilised in step b) and / or step c) comprises medical image data generated from one or more of the different scan types. In one embodiment, the medical image data utilised in step b) comprises medical image data from a first set of scan types and the medical image data utilised in step c) comprises medical image data from a second set of scan types, wherein the first and second sets are different. In one embodiment, the subject remains in the medical imaging system for steps b) to e). In one embodiment, step e) further comprises a recommendation for a scan acquisition process using contrast agent protocols to be performed whilst the subject remains in the medical imaging system. In one embodiment, the medical imaging system comprises an MRI scanner and the plurality of scan types are selected from the group of: T1 -weighted; T2-weighted; T2-FLAIR; SWI / T2* and DWI b1000. In one embodiment, the medical imaging system comprises a CT scanner and the imaging scan data comprises computed tomography (CT) data or the medical imaging system comprises a PET scanner and the imaging scan data comprises PET data. In one embodiment, the recommended contrast agent protocol comprises the use of a contrast agent selected from the group of: gadolinium-based contrast agents; magnesiumbased contrast agents, iron oxide-based contrast agents, superparamagnetic iron oxide nanoparticle contrast agents; manganese-based contrast agents; and protein-based contrast agents. According to a second aspect of the present invention, there is provided a method for generating a synthetic contrast-enhanced medical image for a subject under assessment, the method being executed by at least one hardware processor and comprising the steps of: a) obtaining, from a scan acquisition process performed by a medical imaging system, medical image data representative of the anatomical structure and appearance of at least a part of the subject; b) utilising the medical image data in a first machine learning model trained to classify anatomical features in the medical image data to generate an output identifying any quantitative pathological indicators representative of the presence of pathological features in the medical image data; c) utilising the medical image data in a second machine learning model trained to generate synthetic contrast agent enhanced medical image data from the medical image data to identify any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features derived from the medical image data; d) comparing the outputs from the first and second machine learning models to determine any correlation between any identified quantitative pathological indicators and any identified quantitative contrast indicators and, if any correlation is identified, the method further comprises: e) generating a synthetic contrast-enhanced medical image data for subsequent analysis. In one embodiment, if no quantitative pathological indicators are identified in step b), then the method further comprises: f) utilising a third computational model to reduce the synthetic contrast in regions of the synthetic contrast agent enhanced medical image data related to anatomically normal features to generate calibrated synthetic contrast agent enhanced medical image data. In one embodiment, subsequent to step f), the method further comprises: g) utilising a fourth computational model to segment the calibrated synthetic contrast agent enhanced medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features in the medical image data; and if any quantitative contrast indicators are identified, the method comprises: h) generating a synthetic contrast-enhanced medical image data for subsequent analysis. In one embodiment, the output from the first machine learning model comprises a first feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative pathological indicators have an associated spatial location on the first feature map. In one embodiment, the output from the second machine learning model comprises a second feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative contrast indicators have an associated spatial location on the second feature map. In one embodiment, the medical imaging system comprises an MRI scanner and the plurality of scan types are selected from the group of: T1 -weighted; T2-weighted; T2-FLAIR; SWI / T2* and DWI b1000. In one embodiment, the medical imaging system comprises a CT scanner and the imaging scan data comprises computed tomography (CT) data or the medical imaging system comprises a PET scanner and the imaging scan data comprises PET data. In one embodiment, the recommended contrast agent protocol comprises the use of a contrast agent selected from the group of: gadolinium-based contrast agents; magnesiumbased contrast agents, iron oxide-based contrast agents, superparamagnetic iron oxide nanoparticle contrast agents; manganese-based contrast agents; and protein-based contrast agents. According to a third aspect of the present invention, there is provided a computer readable medium comprising instructions which, when executed by a computer, cause the computer to perform a method according to any one of the preceding claims. According to a fourth aspect of the present invention, there is provided a computer system comprising: a processing device, a storage device and a computer readable medium of the third aspect. Embodiments of the present invention will now be described in detail with reference to the accompanying drawings, in which: Figure 1 illustrates an example computing system 10 forming part of the present invention; Figure 2 illustrates a model 100 according to an embodiment; Figure 3 shows a schematic arrangement of a U-net neural network; Figure 4 shows a schematic view of the operation of the fourth model of an embodiment; Figure 5 shows a flow chart of a method of the present invention; Figure 6 shows a flow chart of a method of a further embodiment of the present invention; Figures 7a and 7b show box plots of enhancing volume to total lesion volume ratio box for low grade and high grade gliomas using true post contrast T1 measurements and using the method of the present invention respectively; Figure 8 shows a scatter plot between the true enhancing volume and pseudo enhancing volume for HGG only; Figure 9 shows a flow chart of a method of a further embodiment of the present invention; Figure 10 shows a flow chart of a method of a further embodiment of the present invention; and Figure 11 shows a workflow diagram of the present invention. The research performed by the inventors indicates that selective and informed use of GBCAs can be the preferred approach in a clinical setting. In embodiments, the present invention enables an accurate and appropriate selection of situations where GBCA use is preferable and situations where it can be avoided to reduce patient risk. As described below, this may take the form of a generated indication as to whether GBCA use is beneficial in a particular clinical setting or, alternatively, where it is unlikely to be of benefit or where simulated GBCA data is likely to be sufficiently accurate for diagnostic and analytical purposes to enable the use of GBCA to be avoided in the clinical setting. MEDICAL IMAGING AND ANALYSIS SYSTEM CONFIGURATION Figure 1 illustrates a computing system 10 forming part of the present invention. The present invention relates to a medical analysis computing system 10. The computing system 10 may take any suitable form and may comprise, for example, a cloud-based computing system, a remote computer connected over a network, a local workstation or a dedicated medical imaging console. Figure 1 shows a computing system 10 according to an embodiment. The computing system 10 comprises one or more physical processors 12, a computer-readable physical memory 14 and a non-transitory storage device 16 such as a hard disk drive or solid-state drive. The one or more physical processors 12 may comprise any suitable processor types; for example, central processing units (CPUs), graphical processing units (GPUs) or any other suitable processor such as, for example, Field Programmable Gate Arrays (FPGAs) and stream processors. The storage device 16 may take any suitable form and may include storage device(s) local to the computing system 10 and / or storage devices external to the computing system 10. For example, the storage device 16 may comprise cloud or other virtual networked storage. The computing system 10 further comprises an interface 18 through which image data is received. The interface 18 may take any suitable form and may be in the form of a communications interface and / or a network connection to a suitable source of input medical image data. A computing application 20 is run on the computing system and is executable to cause the computing system 10 to perform the method of the present invention. The computing application 20 comprises an analysis module 22 and a training database 24 as will be described below. The computing application 20 comprises one or more machine learning algorithms to analyse and process medical image data. In embodiments, in use, the computing application 20 comprises two aspects - a training stage and an operational stage. These will be described below. The computing application 20 is operable to communicate through the interface 18 with the medical imaging device 50. In embodiments, the medical imaging device may take the form of an MRI scanner 50. The interface 18 and medical imaging device 50 are in data communication with a storage device 60 in the form of a picture archiving and communication system (PACS). The PACS 60 is an industry standard device and format for medical imaging. However, this is not intended to be limiting and other configurations may be used. The PACS 60 is operable to handle image and other data in the Digital Imaging and Communications in Medicine (DICOM) format. However, other proprietary or non-proprietary formats may be used. A reading station 62 is in communication with the PACS 60 and can be used by a medical practitioner to read, interpret and analyse image data. It is noted that the computing system 100, in this embodiment, communicates through a Radiologist Information System or Hospital Information System (HIS), shown generally at 124. However, this is not material to the present invention and other systems may be used as appropriate. ANALYSIS MODULE ARCHITECTURE The computing application 20 will now be described with reference to Figures 2 and 3. In embodiments, the analysis module 22 comprises a computational model 100. The computational model 100 may itself comprise a plurality of computational models. Figure 2 shows a general schematic overview of the computational model 100 and the inputs, outputs and components thereof. Figure 3 shows a schematic diagram of the basic workflows of the model 100 and machine learning algorithms. In embodiments, the analysis module 22 comprise a computational model 100 which is operable to generate one or more quantitative parameters to enable real-time decisions on contrast administration for a patient to be made. The computational model 100 comprises a plurality of machine learning modules. Each module comprises one or more trained machine learning algorithms. Figure 2 shows a schematic of the configuration of the model 100. The computational model 100 comprises four models 102, 104, 106, 108. Each model 102, 104, 106, 108 may each comprise one or more models. For example, any of the models 102, 104, 106, 108 may themselves comprise ensemble models. FIRST MODEL 102 The first model 102 comprises one or more algorithms operable to utilize one or more medical images 110 as inputs to determine an output 112 relating to the structure and appearance of an input anatomical structure which is the subject of the input medical images 110. In embodiments, the medical images 110 comprise medical images acquired by the medical imaging device 50. In embodiments, the medical images 110 may comprise MRI images where the medical imaging device 50 is an MRI scanner 50. In specific embodiments, the medical images 110 may comprise medical image data from multiple MRI scan types. In specific embodiments, the medical images may comprise MRI data from: input T1 -weighted; T2-weighted; and T2-FLAIR. Optionally, MRI data from SWI / T2* and DWI b1000 sequences may also be utilised. In embodiments, the input anatomical structure comprises the brain of a patient under analysis, the first model comprises a mass-occupying lesions detection model and the output 112 may comprise one or more quantitative parameters representative of the presence or absence of mass-occupying lesions within the brain. In embodiments, the quantitative parameter is representative of the presence or absence of mass-occupying lesions within the brain which may require contrast injections. In embodiments, the output 112 may comprise a lesions map from which potential lesion structures can be identified to determine the presence or absence of mass-occupying lesions within the brain which may require contrast. In embodiments, the first model 102 may comprise one or more machine learning models which utilize the one or more medical images 110 as inputs. In embodiments, the first model 102 may comprise one or more neural networks. Any suitable neural network may be used, for example, deep neural networks (DNNs) or other techniques. In embodiments, the neural networks may comprise convolutional neural networks (CNN). A CNN comprises an input layer, an output layer a sequence of encoding layers therebetween. The encoding layers of a typical CNN may comprise repeated applications of convolution layers, non-linear activation functions and pooling layers for down sampling, followed by one or more dense (fully connected) layers. In embodiments, the first model 102 may comprise one or more U-Net deep learning algorithms. A U-net is a specific form of CNN comprising an encoding part and a decoding part. The encoding part comprises a contracting path and the decoding part comprises an expansive path. A specific embodiment of the first model 102 will now be described. In specific embodiments, the first model 102 utilises an aggregated voxel-wise classification from a U-Net deep learning algorithm for medical image segmentation. An example of such a system is disclosed in Ronneberger 0., Fischer P., Brox T. (2015) “U-Net: Convolutional Networks for Biomedical Image Segmentation”. However, this is not intended to be limiting and other voxel-level classifiers could be used in its place. The building block of deep neural networks is an artificial neuron or node. Each input has an associated weight. The sum of all weighted inputs is then passed through a nonlinear activation function f, to transform the pre-activation level of the neuron to an output. The output then serves as input to a node in the next layer. Several activation functions are available, which differ with respect to how they map a preactivation level to an output value. The most commonly activation functions used are the rectifier function (where neurons that use it are called rectified linear unit (ReLU)), the hyperbolic tangent function, the sigmoid function and the softmax function. The latter is commonly used in the output layer as it can compute the probability of multiclass labels. For each mode j in the first hidden layer, a nonlinear function is applied to the weighted sum of the inputs. The result of this transformation serves as input for the second hidden layer. The information is propagated through the network up to the output layer, where the softmax function yields the probability of a given observation belonging to each class. Typically, convolutional networks are utilised for classification tasks, where the output to an image is a single class label. However, in many visual tasks, especially in biomedical image processing, the desired output can include localisation, i.e., a class label assigned to each voxel. In an exemplary embodiment, the network architecture of a U-net algorithm is shown in Figure 3 and is taken from reference Ronneberger 0., Fischer P., Brox T. (2015,) “U-Net: Convolutional Networks for Biomedical Image Segmentation”. The algorithm corresponds to a convolutional neural network (CNN) and comprises an encoding part and a decoding part. The encoding side comprises a contracting path (shown on the left side) and the decoding part comprises an expansive path (shown on the right side). The contracting path follows a typical architecture of a convolutional network and consists of the repeated application of two 3x3 convolutions, each followed by a rectified linear unit (ReLU) and a 2x2 max pooling operation with stride 2 for down-sampling. At each down-sampling step the number of feature channels is doubled. Every step in the expansive path consists of an up-sampling of the feature map followed by a 2x2 convolution (“up-convolution”) that halves the number of feature channels, a concatenation with the correspondingly cropped feature map from the contracting path, and two 3x3 convolutions, each followed by a ReLU. In embodiments, if a CNN using a U-net algorithm is utilised in the first model 102, this may comprise different forms. For example, a conventional 3D U-net may be used with different parameters from the examples described above. Any suitable number and size of convolutions may be utilised, having any suitable kernel size and stride size. For example, a stride of 1 may be used. The ReLUs may comprise, in non-limiting examples: softplus approximations; noisy ReLUs (comprising Gaussian noise); leaky ReLUs (which allow a small, positive gradient when the ReLU is not active); parametric ReLUs (where the leakage coefficient within the leaky ReLU is a learned parameter of the network); and / or exponential Linear Units (ELUs). Finally, the pooling layer may utilise max pooling with any suitable filter and stride size. Alternatives to a conventional U-net algorithm may be used. For example, a variational autoencoder (VAE) may additionally be utilised. As a further alternative, a cascade U-net VAE may be used in which two sets of U-net VAE models are interlinked such that the output of the first U-net is concatenated to the input image and fed into the second U-net. Any number of U-nets may be used in the first model 102. Three specific examples of the first model 102 are provided below. In the first example, a 3D U-Net comprises an encoder-decoder structure with skip connections between matching scales of encoder and decoder. Two convolution layers (12 filters each) with a dropout (0.25 rate) in between the input and the U-Net are used. Each scale in an encoder represents a max-pooling layer of size 2, and in the decoder represents an up-sampling layer. Trilinear interpolation is used for up-sampling. In between two scales, two convolution layers followed by an instance normalisation layer is used. Each convolution layer is equipped with a leaky ReLU (negative slope 1e-2). A patch size of 192x192x192 is used. 12 filters are used in the first layer followed by 24, 48, and 56 filters in the encoder. Similar numbers of filters are used in the decoder but in reverse order. All convolutions are performed with a kernel size of 3 and stride 1. In the end, a sigmoid activation function is used to convert the logits into probabilities of belonging to foreground and background. In a second example, the first model 102 comprises a U-net VAE and utilises a VAE to regularize the latent space. The VAE branch starts from the last layer of the encoder. The first layer of the encoder generates mean and standard deviation of a normal distribution from which a 32-dimensional vector is sampled. The vector is then used to reconstruct the input. The sampled vector layer is then connected to a layer of size matching the last layer of the encoder. Following which four convolution layers with kernel size 1 and four up-sampling (factor 2) layers are used to reconstruct the input. In a third example, three sets of U-Net VAE models with the same architecture as the second example are used. The output of the first U-Net is concatenated to the input image and fed into the second network. In the first U-Net, the input images are scaled by a factor of 0.5 by a scaling layer at the input. The output layer is then scaled by 2 to match the size of the input. The networks are trained sequentially. Note that while the second network is trained, the first network is fixed. While training the second model, only the U-Net part of the first model is used and the VAE branch is discarded since it does not have any influence on the segmentation during inference. In summary, the first model 102 receives medical image data 110 (which may be MRI images 110 from the Medical imaging device 50) and generates an output 112 which comprises one or more quantitative parameters comprising indicators of pathology. In embodiments, the indicators of pathology may comprise the presence or absence of massoccupying lesions within the imaged brain. In embodiments, the indicators of pathology may comprise the presence or absence of infarcts, intracranial tumours and intracranial haemorrhages. If an MRI scanner 50 is used, this may be based on a FLAIR sequence, T2w sequence and, optionally, MRI data from SWI / T2* and / or DWI b1000. In embodiments, the output 112 may comprise a lesions map from which potential lesion structures can be identified to determine the presence or absence of mass-occupying lesions within the brain which may require contrast. SECOND MODEL 104 The second model 104 comprises one or more algorithms operable to utilize one or more medical images 110 as inputs to determine an output 114 relating to the structure and appearance of an input anatomical structure which is the subject of the input medical images 110. In embodiments, the medical images 110 comprise medical images acquired by the medical imaging device 50. In embodiments, the medical images 110 may comprise MRI images where the medical imaging device 50 is an MRI scanner 50. In specific embodiments, the medical images 110 may comprise medical image data from multiple MRI scan types. In specific embodiments, the medical images may comprise MRI data from: input T1 -weighted; T2-weighted; and T2-FLAIR. In embodiments, the input anatomical structure comprises the brain of a patient under analysis and the second model 104 is operable to generate an output 114 in the form of a contrast difference map. A contrast difference map is a two- or three-dimensional estimated representation of the difference between pre- and post-contrast imaging. In other words, the second model 104 receives medical images 110 and is operable to generate a simulated post-contrast image therefrom. In embodiments, the input data comprises a concatenation of the received medical images 110, namely a concatenation of medical image data 110 relating to T2 FLAIR, T2-weighted, and pre-contrast T1 -weighted sequences. The simulated post-contrast image comprises a synthesised pseudo-contrast image. In embodiments, this may comprise a synthesised contrast-enhanced T1 image sequence (i.e. a T1 c image sequence). From the synthesised post-contrast image a contrast difference map can be generated which is representative of the appearance and structure of anatomical features which would be represented differently under contrast enhanced imaging. The output 114 in the form of the contrast difference map may comprise one or more quantitative parameters representative of the presence or absence of enhanced lesions within the brain. In embodiments, the quantitative parameter is representative of the presence or absence of mass-occupying lesions within the brain which may be enhanced through the use of contrast injections. In embodiments, the second model 104 may comprise one or more machine learning models which utilize the one or more medical images 110 as inputs. In embodiments, the second model 104 may comprise one or more neural networks. Any suitable neural network may be used, for example, deep neural networks (DNNs) or other techniques. In embodiments, the neural networks may comprise convolutional neural networks (CNN). A CNN comprises an input layer, an output layer a sequence of encoding layers therebetween. The encoding layers of a typical CNN may comprise repeated applications of convolution layers, non-linear activation functions and pooling layers for down sampling, followed by one or more dense (fully connected) layers. In embodiments, the second model 104 may comprise one or more U-Net deep learning algorithms having an encoding part comprising a contracting path and a decoding part comprising an expansive path. In embodiments, one or more U-net algorithms may be used in the second model 104 such as that shown in Figure 3 and taken from reference Ronneberger 0., Fischer P., Brox T. (2015) “U-Net: Convolutional Networks for Biomedical Image Segmentation”. In specific embodiments, the second model 104 comprises an ensemble of three U-Net models. In specific embodiments with an identical architecture, each adapted to a unique MRI sequence. The U-Net architecture may, in embodiments, be fixed for each model in the ensemble of the second model 104. In embodiments, each U-net of the ensemble follows the general structure shown in Figure 3. However, in embodiments, the structure and form of each U-net may be adapted as required. In embodiments, instance normalization layers may be utilised due to small batch size. In embodiments, the same padding may be utilised to preserve image dimensions within a convolution block. In embodiments, the convolution may be transposed to incorporate learnable up sampling. In embodiments, a depth of five may be used to increase the number of low-resolution feature channels. In embodiments, deep supervision may be utilised to regularize feature maps at multiple resolutions. In embodiments, a self-attention layer may be implemented immediately following the input layer to assign importance weights to the three MRI sequences used as input data 110. Finally, in embodiments, a sigmoid activation function at the end of the network may be applied rather than a softmax activation. One or more of the above modifications may be made within the scope of the disclosed embodiments as required. In summary, the ensemble second model 104 receives as an input a concatenation of medical image data 110 relating to different scan types (which, in embodiments where an MRI scanner 50 is used, may comprise T2 FLAIR, T2-weighted, and pre-contrast T1 -weighted sequences) and computes an output 114 in the form of one or more contrast-difference maps in the coordinate spaces of each of the three acquisition sequences. THIRD MODEL 106 The third model 106 comprises a normal contrast calibrator. The normal contrast calibrator comprises a registration-based algorithm to reduce synthesized contrast as generated by the second model 104 in anatomically normal areas. The third model is operable to access a normal contrast image library 106L. Given a target pseudo-contrast prediction from and a fixed library of three outputs 114 from the second model 104 on randomly selected, anatomically normal scans, the target image is non-linearly registered to input medical image data 110 using a suitable machine learning model. A suitable model is disclosed in Hoffman, M et al “SynthMorph: Learning Contrast-Invariant Registration Without Acquired Images”, IEEE Transactions on Medical Imaging, volume 41, issue 3, March 2022. This reference discloses a convolutional U-Net to predict a stationary velocity field (SVF) v0 from an input {m,f}. The exemplary encoder comprises 4 blocks consisting of a stride-2 convolution and a LeakyReLU layer (parameter 0.2), that each halve the resolution relative to the inputs. The decoder features 3 blocks that each include a stride-1 convolution, an upsampling layer, and a skip connection to the corresponding encoder block. The SVF v0 is obtained after 3 further convolutions at half resolution, and the warp ¢0 after integration and upsampling. The operation of the third model 106 is shown in Figure 4. To reduce the prediction of contrast uptake signal in anatomically normal areas, the target synthesized contrast uptake image (i.e. the output 114 from the second model 104) is non-linearly registered by the third model 106 to a fixed template library of three randomly selected templates 114T. The templates 106T comprise randomly selected second model 104 outputs from anatomically normal scans. Contrast uptake from any of the registered library images is then removed from the target images 114 received as inputs by the third model 106. In other words, a predicted contrast enhancement signal in any of the three registered templates is interpreted by the third model 106 as comprising likely normal contrast uptake and removed from the target contrast image. The output 118 from the third model 106 is then one or more calibrated pseudo contrast images 120 (see Figure 4). In other words, the third model 106 is operable to use normal contrast calibration to reduce overprediction of contrast in anatomically normal regions using a suitable machine learning model to perform non-linear registration of the synthesised contrast images 114 generated by the second model 104 to a library of normal and / or non-pathological contrast uptake templates 114T in order to generate an output 118 comprising calibrated synthetic contrast images 120. FOURTH MODEL 108 The fourth model 108 comprises an enhancing lesion segmentation algorithm. The fourth model 108 comprises a classifier in the form of a segmentation model operable to segment the calibrated synthetic contrast images 120 forming part of the output 118 to generate one or more segmentation predictions 122. In embodiments, the fourth model 108 is operable to segment a whole tumour (plus oedema), a tumour core, and enhance tumour core areas based on T1w, T1c (ground truth or synthesized as per experiment), T2w and FLAIR sequences. The fourth model 108 may take any suitable form operable to classify image features and appearance in order to generate appropriate segmentation predictions 122. Any suitable segmentation algorithm or protocol may be used. In embodiments, the fourth model 108 may perform segmentation operations using image processing techniques and mathematical modelling such as thresholding, edge detection, region-based segmentation and / or clustering methods. Thresholding methods utilize a threshold value to transform a greyscale image to a binary one. Known techniques for determining threshold values include histogram thresholding, maximum entropy, maximum variance and k-means clustering. For edge detection, gradientbased or histogram methods are commonly used. Edges have a rapid change in intensity and these are often linked to form closed object boundaries. Segmentation may also be applied on edge-detected images themselves. Clustering algorithms comprise unsupervised algorithms which cluster pixels having common attributes into groups allocated to a particular segment. K-means clustering is one variant which considers all the pixels and clusters the pixels into “k” classes. Alternatively, the fourth model 108 may utilize probabilistic techniques such as boundary fitting or other specific model-based techniques. Alternatively, the fourth model 108 may utilise machine learning algorithms. In specific embodiments, the fourth model may utilise a U-net configuration. In one example, the open-source picture-nnunei package pre-op multi-class glioma segmentation may be used to segment a whole tumour (plus oedema), a tumour core, and enhance tumour core areas based on T1w, T1c (ground truth or synthesized as per experiment), T2w and FLAIR sequences. The fourth model 108 may take as an input the medical image data 110 and the calibrated synthetic contrast images 120 forming part of the output 118 and generate one or more segmentation predictions 122 used to determine contrast injection protocols. MODEL TRAINING The models 102, 104, 106, 108 are trained prior to operational use. Embodiments of methodologies which may be used are described below. FIRST MODEL TRAINING The first model 102 may be trained in a training phase in which a subset of the available data known as the training set is used to optimise the parameters of the algorithm of the first model 102 to perform the necessary operations to determine the one or more quantitative parameters comprising indicators of pathology comprising the output 112 from the first model 102. The model may be trained from scratch without any pre-training. Following the training phase, there is implemented a so-called testing phase in which the remainder subset which is known as the test set is used to assess whether the trained model can blind-predict the class of new observations. When the amount of available data is limited, it is also possible to run the training and testing phases several times on different training and test splits of the original data and then estimate the average performance of the model - an approach known as cross-validation. The algorithm of the first model 102 is trained using a training set containing a set of medical scan tuples with associated ground truth annotations, to classify each voxel in the medical scan into one of several candidate classes. These may be from any suitable image acquisition technique or method. A plurality of input images is utilised which form a training set containing a set of medical scan tuples with associated ground truth annotations to classify each voxel in the medical scan into one of several candidate classes. The data may be pre-processed. For example, the input images may be resampled to achieve an isotropic voxel size. To address problems such as infarct segmentation, a diffusion sequence may be resampled such that the images are, firstly, isotropic and then the fluid attenuated inversion recovery (FLAIR) and / or susceptibility weighted imaging (SWI) sequences may be reformatted to the resampled diffusion space. An apparent diffusion coefficient can then be computed. In a similar fashion, for the problem of haemorrhage segmentation, an SWI space may be used. After resampling, the images may be z-normalized and fed into the neural network for training and inference. The data may be cropped to the region of non-zero values after the resampling using, for example, Otsu median filtering. The above examples are intended to be exemplary and non-limiting. The skilled person would be readily aware of other input images and / or pre-processing methods which could be used with the present invention. In addition, the U-Net training of the first model 102 is improved by artificial extension using data augmentation. Such data augmentation may be desirable to enable the network to be trained sufficiently where limited input images are available. A number of transformations may be used to augment the dataset of training images. For example, input images may be extended with transformations based on a random velocity field. Additionally or alternatively, a random selection of transformation (which may include no augmentation or transformation) may be used for each training step. The input images so transformed may be weighted with a cost function having a predetermined factor to account for discrepancies and artifacts. In an embodiment, the predetermined factor is 0.75. and artifacts. Alternative transformations may be used and these may include one or more of: rotations, elastic transformation, gaussian noise, gaussian blur, contrast, gamma correction, and mirroring. For models with a VAE branch only intensity transformation may only be used due to the image reconstruction regularization. Finally, post-processing of images may be used to improve image classification performance. In embodiments, a morphological filtering may be used to remove random predictions in the images. For architecture selection, blobs that are smaller than a particular size may be removed. In embodiments, blobs smaller than 5 mm3 may be removed. Other sizes may be used. For example, classification performance may be optimised by removal of blobs that are smaller than 125 mm3. The first model 102 is trained using the training set of images a database (not shown). In general the first model 102 functions by taking the values of specific features (independent variables or predictors, in regression) in an example (the set of independent variable values) and predicts the class that that example belongs to (the dependent variable). In the present medical imaging context, the features comprise voxels and the class may be indicative of a particular probable pathology. The first model 102 is required to learn a number of parameters from the training data. The classifier is essentially a model of the relationship between the features and the class label in the training set. More formally, given an example x, the classifier is a function f that predicts the label y = f(x). Typically, neural networks can learn through a gradient descent-based algorithm. The gradient descent algorithm aims to find the values of the network weights that best minimise the error (difference) between the estimated and true outputs. It should be noted that these methods are not actual classifiers themselves; instead, they are networks that are pre-trained to learn useful patterns in the data and then fed to a real classifier at the final layer. In an exemplary embodiment, the training set of input images and corresponding segmentation maps are used to train the algorithmic network with a stochastic gradient descent method. The following is a general example of the training functionality as it may be carried out in an embodiment. An energy function is computed by a pixel-wise soft-max over a final feature map combined with the cross-entropy loss function. The soft-max is defined as set out in equation 1): 1) / EE (E^E p-xp^ where ak(x) denotes the activation in feature channel k at the pixel position xeQ with '; . k is the number of classes and pk(x) is the approximated maximum-function, i.e. pk(x)»1 for values of k that has the maximum activation ak(x) and pk(x)«0 for all other k. The cross entropy then penalises at each position the deviation of p'(x)(x) from 1 as per 2) below: where I: {1 is the true label of each pixel and w : HR is a weight map that is introduced to give some pixels more importance in the training. A weight map is pre-computed for each ground truth segmentation to compensate for the different frequency of pixels from a certain class in the training data set, and to force the network to learn the small separation borders that are introduce between adjacent test cells. The weight map is then computed as per equation 3): where wc: HR is the weight map to balance the class frequencies, di : Q-> IR denotes the distance to the border of the nearest cell, and d2 : Q-> IR the distance to the border of the second nearest cell. In deep networks with many convolutional layers and different paths through the network, it is important to initialise the weights correctly, otherwise parts of the network might give excessive activations, while other parts never contribute. Ideally, the initial weights should be adapted such that each feature map in the network has approximately unit variance. For a network with alternating convolution and ReLU layers, this can be achieved by drawing the initial weights from a Gaussian distribution with a standard deviation 0N2 / N, where N denotes the number of incoming nodes of one neuron. For example, for a 3x3 convolution and 64 feature channels in the previous layer N= 9-64 = 576. However, alternative approaches may be used. In an embodiment, the U-net or similar neural network is trained using dice as the loss function. In an example, a kernel weight regularisation of 1e-6 may be used on all layers. A dice loss function functionally similar to reference [1 d] may be used and is formulated as set out in equation 4): , . where u is the predicted segmentation, v is the ground truth, and K is the number of classes. Both predicted segmentation and ground truth are one-hot encoded. Sigmoid is used as the activation function. In an embodiment, optimisation may be achieved through use of an adaptive learning rate approach such as adaptive moment estimation or Adam. The Adam process utilises estimations of first and second moments of a gradient to adapt the learning rate for each weight of the neural network. In an embodiment, the Adam methodology is utilised with a learning rate of 1 e - 1 as the optimiser. In embodiments, the model may be trained for 100 epochs, where one epoch comprises 750 steps. In non-limiting embodiments, a learning rate scheduler may be used to decrease the learning rate by 5% if there is no increase in the validation dice for the previous 10 epochs. The epoch with the best overall validation dice may then be used to infer finally. A validation dice score then consists of two scores: a) an overall dice for the foreground, and b) dice only on images with foreground. A mean of these two dice scores constitute the overall dice. Whilst other methods may be used and such a configuration is non-limiting, such a combination has the advantage of the presence of a large set of non-foreground images with diffused pathology which may lead to a high number of false positives. Once trained, the first model 102 can then be used to determine whether the features used contain information about the class of the example. This relationship is tested by using the learned classifier on a different set of the test data. In embodiments, identification of infarcts in brain MRI were carried out based on tuples of T2-weighted fluid attenuation inversion recovery (FLAIR), diffusion weighted imaging (DWI), SWI, susceptibility weighted angiography (SWAN) and / or gradient echo (GRE) MRI sequences. In embodiments, the first model 102 was trained using manually annotated findings in the medical scans, annotated by in-house radiographers. The medical scans were annotated by consulting already existing clinical reports with radiological findings and annotating the voxels corresponding to these findings in the scans. Once trained, the first model 102 is ready to be implemented in an operational workflow. SECOND MODEL 104 THAINING The training and evaluation of the second model 104 will now be described. In embodiments, the second model 104 is trained using a training set of images. In embodiments, the training images comprise medical images acquired using one or more MRI scan protocols. In embodiments, the training set of images may be optionally pre-processed. In embodiments, different versions of the second model 104 may be trained on different datasets; for example, to cover different MRI acquisition techniques or sequences. In embodiments, the second model 104 is trained through a gradient descent-based algorithm which aims to find the values of the network weights that best minimise the error (difference) between the estimated and true outputs. The following is a general example of the training functionality as it may be carried out in an embodiment. Training data for second model 104 In embodiments, for the development of the second model 104, a volume of MRI medical image data is obtained for training. In specific embodiments, a total of 2,598 unique adult brain MRI studies with T2-FLAIR, T2-weighted, pre- and post-contrast T1 -weighted sequences and paired radiology reports retrospectively collected from 8 different institutions across United States, Brazil, Israel, India, Ukraine and Denmark is used for training. Additionally, in embodiments, training data may be enriched by utilisation of other data sources. In specific embodiments, the training data for the second model 104 is enriched with 7 open-source datasets: Brain Tumour Segmentation; (BraTS) glioma; meningioma and metastasis challenge data; the University of Pennsylvania glioblastoma cohort; the University of California San Francisco preoperative diffuse glioma; MRI dataset; amide proton transfer technique diffuse glioma data; and supratentorial meningioma data. In embodiments, the variability in MRI resolutions and acquisition planes across different acquisition sequences (e.g. T2-FLAIR, T2-weighted and pre-contrast T1 -weighted sequences) was addressed by training different machine learning models on different datasets. In embodiments, three different datasets are used to train three different machine learning models, where each model conformed to one of the three sequence spaces. Data pre-processing In embodiments, the data may be pre-processed to facilitate training. In embodiments, the pre-processing steps may be identical across the three training datasets. In embodiments, the pre-processing steps may comprise the following steps: 1. MRI sequences may be skull-stripped using a SynthStrip algorithm; 2. Primary sequences may be resampled to isotropic space; 3. Non-primary sequences may be resampled to isotropic primary sequence space; 4. Non-primary sequences may be registered to primary sequence using a “SynthMorph” affine registration algorithm; 5. Sequences may be concatenated and cropped to brain-only bounding box; 6. Sequence intensity values may be scaled using min-max normalization to [0,1] range; 7. Contrast difference may be computed as a non-zero intensity difference between post and pre-contrast T1 -weighted sequences and may be scaled using min-max normalization to [0,1 ] range. Note that one or more of the seven steps above may be omitted as required, or additional steps may be introduced. Training data was split to training (85%) and validation (15%) with respect to data origin, MRI scanner and sequence parameters, as well as diagnosed pathologies. Second model 104 training workflow In embodiments, a single 40GB NVIDIA A100 Tensor Core GPU is used for model training. In embodiments, all training schemes may use a batch size of 3, where: 1. The first element in the batch may be either a normal study, or a study which did not contain any contrast-enhancing pathology (such as a chronic infarction or an arachnoid cyst); 2. The second element in the batch may be a study which contained a glioma (such as an astrocytoma or an oligodendroglioma); 3. The third element in the batch may be a study which contained a non-glioma contrastenhancing pathology (such as a meningioma or a schwannoma). For model training, in embodiments, the output contrast-difference maps may be passed through one or more loss functions. In embodiments, the output contrast-difference maps may be passed through a linear combination of a multi-scale structural similarity index measure loss, a global mean-absolute error loss, as well as a local mean-absolute error loss. The local mean-absolute error loss was computed on regions of non-zero ground-truth contrast-difference connected components of at least 100 mm3 in volume, excluding CSF and ventricles (segmentation of CSF and ventricles may be obtained using the SynthSeg algorithm). Additionally, in embodiments, L2 regularization loss with 10-5 regularization factor on the convolution kernel parameters may be added. In embodiments, gradient-based optimisation may be used. In specific embodiments, gradient-based optimisation is performed using the Adam optimizer with Pi = 0. 9 p2 = 0. 99 and a learning rate of 10-5. In embodiments, a comprehensive suite of spatial and intensity data augmentation operations may be applied probabilistically during training. In embodiments, this included, but was not limited to, axis reflection, resizing, rotations, elastic transformations, Gaussian blurring and Gaussian noising. In embodiments, each of the three U-Net models for the second model 104 were first pretrained for 750,000 steps (250 steps per epoch for 3000 epochs) on both the internal and the external training data, with all loss coefficients equal to 1, and were then fine-tuned for 250,000 steps on internal data only, with local loss coefficient increased to 10. However, any suitable variations to this may be utilised during the training process. THIRD MODEL TRAINING The third model 106 may comprise a machine learning aspect. This, in embodiments, may require training. As noted above, a suitable model configuration for the third model 106 and for training the third model 106 may be as disclosed in Hoffman, M et al “SynthMorph: Learning Contrast-Invariant Registration Without Acquired Images”, IEEE Transactions on Medical Imaging, volume 41, issue 3, March 2022. In embodiments, the third model 106 is operable to enable registration of real images both within and across contrasts. The third model 106 may be trained on only synthetic data that far exceeds the realistic range of medical images. In an exemplary training process, training images may be synthesised from label maps, whereas registration requires no label maps at test time. First, a generative model is utilised to generate random label maps of variable geometric shapes. Secondly, conditioned on these maps, or optionally given other maps of interest, images are synthesised with arbitrary contrasts, deformations, and artifacts. This strategy enables the use of a contrast-agnostic loss that measures label overlap, instead of an image-based loss. This leads to two network variants that yield substantial generalizability, both capable of registering any contrast combination tested without retraining. In specific embodiments, Let m and f be a moving and a fixed 3D image, respectively. An unsupervised learning-based registration is used as a framework on which to focus on deformable (non-linear) registration. These frameworks may use a CNN h0 with parameters 0 that outputs the deformation 0 e = he (m, f) for image pair {m, f}. In embodiments, at each training iteration, the network he is given a pair of images {m, f}, and parameters are updated by optimizing a loss function L(0 ; m, f, q> 0 ) similar to classical cost functions, using stochastic gradient descent. Typically, the loss contains an image dissimilarity term LdiS (m ^0 ,f) that penalizes differences in appearance between the warped image and the fixed image, and a regularization term Lreg (q>) that encourages smooth deformations: 5) C(fh m, f, ¢0) == Cju (m 0 f) + (^¼). where q> 0 = h0 (m, f ) is the network output, and A controls the weighting of the terms. However, networks trained this way only predict reasonable deformations for images with contrasts and shapes similar to the data observed during training. To address this, two paired 3D label maps {sm , s f} can be generated using a function g s (z) = {sm , s f}, given random seed z. Alternatively or additionally, if anatomical labels are available, these can be used nstead of synthesizing segmentation maps. Another function g I (sm , s f, ~z) = {m, f} can be generated to synthesize two 3D intensity volumes {m, f} based on the maps {sm , s f} and seed ~z. This generative process has two advantages. First, training a registration network h0 (m, f) using the generated images exposes the network to arbitrary contrasts and shapes at each iteration, removing the dependency on a specific MRI contrast. Secondly, because label maps are synthesized first, a similarity loss that measures label overlap independent of image contrast, thereby obviating the need for a cost function that depends on the contrasts being registered at that iteration. It is noted that the above training method is exemplary and other training methods may be used. When the model is trained, the third model 106 is ready for use. FOURTH MODEL TRAINING In embodiments, the fourth model 108 may take the form of a machine learning algorithm. In this case, the fourth model 108 may, optionally, require training. The fourth model 108 may be trained on multiple class segmentations for each tumour. Test data may be used to assess the performance of the model. If missing sequences are present in the training data, sparsified training may be applied. To avoid gaps in the training data, empty (zero-filled) scans could be utilised in place of missing sequences. It is noted that the above training method is exemplary and other training methods may be used. When the model is trained, the fourth model 108 is ready for use. METHOD An exemplary method workflow according to an embodiment will now be described with reference to Figure 5. Step 200: Prepare scan session At step 200, the medical imaging device 50 is prepared for the scan. This may comprise determining the scan sequence and protocol that is needed for a specific portion of region of a subject’s body. This may be derived automatically, or a medical practitioner may specify a particular scan sequence. In an embodiment, the computing application 20 may, based on the initial input for the scan sequence, derive a suitable sequence. This may also be informed by other data, such as empirical data. For example, semantic or natural language analysis of patient records, empirical knowledge of the most common pathologies for a given body part, or other data could be used to inform an initial “best guess”. This acts as a placeholder for subsequent analysis or further scan sequences. The following exemplary embodiments refer to scans of the brain. However, this is not intended to be limiting and in principle any suitable body part or region of the body could be scanned using the described method. In embodiments, the initial scan sequence does not comprise any scans utilising contrast enhancement. The initial scan sequence and subsequent processing steps will determine whether any further scan sequences involving contrast enhancement are required. Step 202: Execute scan sequence At step 202, the initial scan sequence is executed. In this embodiment, the initial scan sequence comprises a plurality of different image acquisition sequences which form part of the first or Level 1 scan sequence. It is noted that the following order of sequences shown and described in this embodiment may be varied as appropriate and different orders, or different scan techniques, may be used as appropriate. In this embodiment, if the medical imaging device 50 is an MRI scanner, the initial scan sequence comprises the following techniques: T1 T2 DWI 3D SWI T1 TSE These techniques are selected so as to provide a coarse indicator of the following potential pathologies: • Tumour (general) • Glioma (most prevalent type of tumour) • Granuloma • Abscess • Haemorrhage • Microhaemorrhage • Infarct • Ischemic changes • Demyelination • Neoplasm • Hydrocephalus • Infections The order of the acquired sequences can be determined based on importance. For example, if SWI is acquired in 3D, SWI can be used to assess both calcifications and haemorrhages / haemosiderin (iron). No other sequences are therefore needed for a coarse indicator of the potential pathologies listed above. However, the order and timing of the sequences is not material and any suitable ordering of sequences may be selected as appropriate. A generalised example of a typical acquisition process will now be described. This is intended to be purely exemplary and non-limiting. Initially a localiser is utilised for 20 to 30 seconds. Then, the T1 scan acquisition is carried out, with this process taking approximately 3 to 5 minutes. The data so acquired is sent via a DICOM router from the scanner 50 to the computing application 20 via the interface 18. The next scan acquisition is a DWI scan process which last for approximately 2 minutes. Concurrently with this scan, image processing and analysis may be carried out using the model 100 during this stage. This will be described in the next step although temporally steps 202 and later steps may occur simultaneously or at least in part concurrently. Usefully, the scan interpretation data from the T2 FLAIR process may be obtained prior to the end of the DWI sequence. Once this scan is complete, the data is sent via the DICOM router to computing application 20 and model 100 for processing and identification of potential pathologies and contrast enhancement decision making. The next stage in the Level 1 scan acquisition according to this embodiment is the determination of a SWI scan. This typically takes around 3-4 minutes. If required, a further scan in the form of a T1 TSE (not shown in the timeline figure) may also be carried out. It is noted that for each scan acquisition image scan data is obtained. This may comprise any suitable form; for example, the scan data may comprise image data of a single image, multiple images, or a combination of image data and metadata. The image scan data may comprise two-dimensional images (either obtained directly as two-dimensional images or as two-dimensional images obtained from three-dimensional data) or three-dimensional images. Alternatively, the image scan data may comprise data and metadata which is not directly image related but is utilised in forming an image of a body part of the subject. In summary, at step 202 medical image data is obtained from a scan acquisition process performed by a medical imaging system (in embodiments, the scanner 50), where the medical image data is representative of the anatomical structure and appearance of at least a part of the subject under investigation. It is noted here that steps 200 and 202 may be performed at a different location or time from the later steps of the method. There is no explicit requirement for the initial data acquisition to be concurrent with the later steps, although in embodiments this may have advantages and be desirable. In other words, all that is required for subsequent analysis steps is that suitable medical image data obtained from a scan acquisition process is provided and obtained in this step. The method proceeds to step 204. Step 204: Process data in first model 102 At step 204, the first model 102 receives medical image data 110 from the images 110 from the Medical imaging device 50 generated in step 202 and generates an output 112 which comprises one or more quantitative parameters comprising indicators of pathology. In embodiments, the indicators of pathology may comprise the presence or absence of massoccupying lesions within the imaged brain. In embodiments, the indicators of pathology may comprise the presence or absence of infarcts, intracranial tumours and intracranial haemorrhages based on (for an MRI scanner 50) a T2w and a FLAIR sequence and, optionally, MRI data from SWI / T2* and DWI b1000. In embodiments, the output 112 may comprise a lesions map from which potential lesion structures can be identified to determine the presence or absence of mass-occupying lesions within the brain which may require contrast. Step 206: Process data in second model 104 At step 206, the second model 104 receives as an input a concatenation of medical image data 110 relating to multiple scan types (if an MRI scanner is used, these may comprise T2 FLAIR, T2-weighted, and pre-contrast T1 -weighted sequences) obtained in step 202. Based on this data, the second model 104 computes an output 114 in the form of one or more contrast-difference maps in the coordinate spaces of each of the three acquisition sequences. In embodiments, the three contrast difference maps may be resampled in this step to a single coordinate space. In embodiments, the three contrast difference maps are resampled to the T1 -weighted coordinate space. This provides a single output for further use. Note that step 206 may occur concurrently with step 204, and optionally may occur during at least a part of the scan acquisition in step 202. Step 208: Indicator of pathology found by first model 102? At step 208 it is determined whether the output 112 from the model 102 comprises one or more quantitative parameters comprising indicators of pathology. In embodiments, the indicators of pathology may be indicative of the presence or absence of identified pathological features. For example, the pathological features may comprise mass-occupying lesions within the imaged brain. If the generated indicators of pathology have identified any mass-occupying lesions, then the method proceeds to step 212. If not, the method proceeds to step 214. Alternatively, if the identified pathological features correspond to one or more acute findings (for example, infarcts, haemorrhages or other acute conditions or findings), the method may proceed directly to step 222. It is noted that the acute finding aspect may be optional in certain embodiments. In summary, steps 204 and 206 describe that the medical image data is utilised in a first machine learning model trained to classify anatomical features in the medical image data to generate an output identifying any quantitative pathological indicators representative of the presence of pathological features in the medical image data. Step 210: Indicator of contrast enhancement found by second model 104? At step 210 it is determined whether the output 114 from the second model 104 in the form of one or more contrast-difference maps in the coordinate spaces of each of the three acquisition sequences (or a single combined contrast map in a single coordinate space as described above) have identified one or more indicators of contrast enhancement. In embodiments, the one or more indicators of contrast enhancement may comprise the detection or determination of one or more enhanced lesions. In other words, steps 206 and 210 comprise utilising the medical image data in a second machine learning model trained to predict an effect of contrast agent on the medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features derived from the medical image data. In embodiments, if one or more enhanced lesions are identified, then a determination may be made that the patient analysis may benefit from contrast enhancement .In embodiments, this determination may be supported by additional analysis as described below. If the generated indicators of contrast enhancement have identified that contrast enhancement may be beneficial, then the method proceeds to step 212. If not, the method proceeds to step 220 and no contrast enhancement is recommended. Step 212: Overlap? At step 212, the outputs from the first and second machine learning models 102, 104 are compared to determine any correlation between any identified quantitative pathological indicators and any identified quantitative contrast indicators. If any correlation is identified, the method proceeds to step 218 to recommend contrast. In embodiments, determining any correlation comprises determining whether there is overlap between the results found from steps 208 and 210. In other words, in embodiments, in step 212 it is determined if there is any contrast enhancement detected by the second model 104 in step 210 in a segmentation area where indicators of pathology are identified by the first model 102 in step 208. If correlation in the form of overlap is detected, the method proceeds to step 218 to recommend contrast. If not, the method proceeds to step 220 to recommend no contrast. Step 214: Further processing of contrast enhanced data If no indicators of pathology (for example, mass-occupying lesions) are identified in step 208 by the first model 102 based on the data generated in step 204, then a further process is triggered whereby the third and fourth models 106, 108 are engaged to process the data generated by the second model 104 in step 206 irrespective of the decision in step 208. As part of step 214, the third model 106 (i.e. the normal contrast calibrator) is operable to reduce synthesized contrast in the contrast difference maps generated by the second model 104 in anatomically normal areas. During this operation, the third model is operable to access the normal contrast image library 106L. In embodiments, the third model 106 reduces the prediction of the contrast uptake signal in anatomically normal areas. In operation, the target synthesized contrast uptake image (i.e. the output 114 from the second model 104) is non-linearly registered by the third model 106 to the fixed template library of three randomly selected templates 114T. The templates 106T comprise randomly selected second model 104 outputs from anatomically normal scans. In accordance with the method, the contrast uptake from any of the registered library images is then removed from the target images 114 received as inputs by the third model 106. In other words, a predicted contrast enhancement signal in any of the three registered templates is interpreted by the third model 106 as comprising likely normal contrast uptake and removed from the target contrast image. The third model 106 then outputs the output 118 in the form of one or more calibrated pseudo contrast images 120. Consequently, in this step, the third model 106 is operable to use normal contrast calibration to reduce overprediction of contrast in anatomically normal regions using a suitable machine learning model to perform non-linear registration of the synthesised contrast images 114 generated by the second model 104 to a library of normal and / or non-pathological contrast uptake templates 114T in order to generate an output 118 comprising calibrated synthetic contrast images 120. Once the calibrated synthetic contrast images 120 are generated as part of the output 118, this data is passed to the fourth model 108. The fourth model 108 comprises a classifier in the form of a segmentation model operable to segment the calibrated synthetic contrast images 120 forming part of the output 118 from the third model 106 to generate one or more segmentation predictions 122. In embodiments, the fourth model 108 is operable to segment a whole tumour (plus oedema), a tumour core, and enhance tumour core areas based on T1w, T1c (ground truth or synthesized as per experiment), T2w and FLAIR sequences. In embodiments, the fourth model 108 may perform segmentation operations using image processing techniques and mathematical modelling such as thresholding, edge detection, region-based segmentation and / or clustering methods. Alternatively, the fourth model 108 may utilise machine learning algorithms. In specific embodiments, the fourth model may utilise a U-net configuration. As part of step 214, the fourth model 108 takes as inputs the medical image data 110 generated in step 202 and the calibrated synthetic contrast images 120 forming part of the output 118 generated by the third model 106 and generates one or more segmentation predictions 122 used to determine contrast injection protocols. The method proceeds to step 216 Step 216: Contrast enhancement detected? At step 216, the data generated in step 214 is analysed. If an indicator of contrast enhancement (such as, for example, a lesion or tumour) is identified following the enhancement protocols of step 214, contrast is to be recommended and the method proceeds to step 218. If not, the method proceeds to step 220 for a recommendation of no contrast. Step 218: Recommend contrast At step 218, it is recommended that further scans are carried out with contrast. This comprises generating a recommendation for a scan acquisition process using contrast agent protocols. In embodiments, this may comprise gadolinium contrast agent protocols. Step 220: No contrast needed At step 220, it is recommended that no contrast is required for any further scans that may be necessary. Step 222: Further action following acute findings If the indicators of pathology derived in step 208 indicate probable or possible acute findings, clinical action may need to be planned or, in specific cases, may be needed urgently. In embodiments, a range of different actions may be taken as part of step 222 at this stage. For example, if a positive indication of a probable pathology is identified at step 208, then a notification may be provided during the medical scan session. In other words, a notification relating to, corresponding to, or indicative of the probable pathology may be indicated whilst the subject is still in the medical imaging device 50 during the medical scan session. The notification may be provided in any suitable way. For example, it may be provided as an audible signal, or it may be provided on a display or on multiple displays. The notification may be provided to a medical professional or to an operator of the medical imaging device 50. In embodiments, the notification may be indicative of the probable need for urgent treatment of the subject. The notification may therefore provide an indication or information to a medical professional or operator which the medical professional or operator can then use to make a suitable diagnosis regarding whether urgent treatment of the subject is needed whilst the subject is in medical imaging device 50. In embodiments, the one or more quantitative indicators may provide an indication of one or more of: Haemorrhage; Microhaemorrhage; or infarct, in which case a suitable notification may indicate to a medical professional the probable presence of stroke in the subject and the medical professional can then make a diagnosis of the condition and the treatment required. In embodiments, the notification may be indicative of the probable need for: thrombolytic treatment; blood thinning medication; or treatment avoiding blood thinning medication. By providing an indication of the probable need for such treatments, the present invention provides a notification which alerts and facilitates a medical practitioner to make a fast diagnosis and corresponding appropriate treatment of a subject still in the medical imaging device 50. Alternatively, if the probable acute condition does not need urgent treatment, the process may simply notify the medical professionals of the identification of a provable acute condition to aid treatment planning. In embodiments, the method may also recommend subsequent scans or scan sequences which could be used to provide further information on diagnosis. An exemplary method workflow according to a further embodiment will now be described with reference to Figure 6. In Figure 6, steps 300 to 316, 320 and 322 correspond to steps 200 to 216, 220 and 222 respectively of the embodiment of Figure 5 and, for clarity and brevity, will not be described further here. The method of the embodiment of Figure 6 differs from that of the method of the embodiment of Figure 5 in that, following a decision that contrast is appropriate, a further sequence of steps are utilised to determine the appropriate level of contrast to be used following a finding that contrast may be beneficial. Step 318A: Enhancing volume fraction Steps 318A to 318C replace step 218 in the first embodiment. Step 318A is reached via either route A or route B. Route A comprises detection of an overlap between the finding of indicators of pathology (non-acute) in step 308 and the finding of indicators of contrast enhancement in step 310. Route B comprises a finding of no indicators of pathology in step 308 but an indicator of contrast enhancement (such as, for example, a lesion or tumour) is identified in step 316 following the enhancement protocols of step 314 (corresponding to that described in step 214 of the previous embodiment). In embodiments, the inventors have recognised that, from a workflow perspective, the decision to use low / normal dosage of contrast can be of clinical benefit. In doing so they have identified that the methodology of the present invention can be useful in determining the need of low-dose contrast (for low-grade gliomas (LGGs)) and normal-dose contrast (for high-grade gliomas(HGGs)). The inventors have identified that an enhancing-to-lesion ratio can be a discriminatory feature in recognising HGGs vs LGGs. Figures 7a and 7b show examples of this. Figure 7a shows a box plot of enhancing volume to total lesion volume ratio box plots for low grade and high grade gliomas using true post contrast T1 measurements. The true enhancing volume was computed using the PIC embodiment of the fourth model 108 on true contrast images. Figure 7b shows a box plot of enhancing volume to total lesion volume ratio box plots for low grade and high grade gliomas using the present invention. In this scenario, the first model 102 selected the tumours and so the use of the fourth model 108 was not necessary (i.e. these scenarios were derived via Route A described above). The mean dice for cases with glioblastoma tumours was 0.27± 0.19 and the median dice was 0.25. For brain metastasis, the mean dice was 0.41 ± 0.22. All cases where the method of the present invention did not predict any enhancement were excluded. Since many detection algorithms tend to be biased by lesion volume size, the discriminatory ability of just tumour volume was also studied. Figure 8 shows a scatter plot between the true enhancing volume (i.e. Figure 7a) and pseudo enhancing volume (i.e. Figure 7b) for HGG only. The Pearson correlation coefficient is r=0.47 with a p-value <0.00001. LGGs are not considered due to the lack of presence of enhancing lesions in majority of the cases. As expected, the volume was moderately discriminatory with a large number of false positives (specificity of 28% with original contrast images) indicating that the enhancing-to-lesion ratio may contain more information for the classification. In summary, step 318A involves computing the enhancing lesion volume ratio (i.e. ratio of enhancing volume to total lesion volume) to enable a probabilistic determination. Once this has been determined the probability of an HGG (i.e. class 1) for a given sample can, in embodiments, be determined from equation 6) below: P(lugh grade glioma i X) 6) • age 4- ■ sex 4- Aslvr ■ ELVR 4- &) Where age is the age of the individual in years, sex is a binary variable where 1 represents female and 0 represents make and ELVR is the enhancing volume ratio. The recited p values represent coefficients for each parameter. In embodiments, page is 0.075, Psex is =0.164, Pelvr is 2.837. b is the bias intercept term which, in embodiments, takes a -1 / -1 - 1 value of =3.700. Finally, ' represents the sigmoid function that maps the linear combination to a probability. As a result, the logistic regression model computes the probability of a tumour being a HGG in accordance with equation 7): Pfhigh grade glioma j X) - ...... 7) Once the enhancing volume to total lesion volume ratio is computed in step 318A, the method proceeds to step 318B. Step 318B: Determine indicator of pathology At step 318B, based on the results of step 318A, an indicator of pathology is determined. In embodiments, the indicator of pathology is determined as a probable LGG or HGG based on the analysis of enhancing volume to total lesion volume ratio from step 318A. An indicator of pathology indicative of an LGG would indicate a low dose contrast regime is most appropriate. An indicator of pathology indicative of an HGG would indicate a normal dose contrast regime is most appropriate. Step 318C: Recommend contrast regime At step 318C the contrast regime determined by the indicator of pathology in step 318B can be recommended and subsequently implemented by a medical professional. An exemplary method workflow according to a further embodiment will now be described with reference to Figure 9. Figure 9 is a method relating to a follow up regime where a priori knowledge of a tumour in the subject under examination is available. As a result, the first model 102 is not required in this embodiment. Step 402: Execute scan sequence At step 402, the initial scan sequence is executed. In this embodiment, the initial scan sequence comprises a plurality of different image acquisition sequences which form part of the initial scan sequence. It is noted that the following order of sequences shown and described in this embodiment may be varied as appropriate and different orders, or different scan techniques, may be used as appropriate. In this embodiment, the initial scan sequence comprises the following techniques (for an MRI scanner): T1w T2w T2FLAIR It is noted that for each scan acquisition image scan data is obtained. This may comprise any suitable form; for example, the scan data may comprise image data of a single image, multiple images, or a combination of image data and metadata. The image scan data may comprise two-dimensional images (either obtained directly as two-dimensional images or as two-dimensional images obtained from three-dimensional data) or three-dimensional images. Alternatively, the image scan data may comprise data and metadata which is not directly image related but is utilised in forming an image of a body part of the subject. Step 402: Obtain data from HER As an alternative or addition to step 400, data previously acquired may be provided from the patient’s electronic health record (EHR). This may include previous scan data as well as other empirical data such as patient sex and age. Step 404: Process data in second model 104 At step 404, the second model 104 receives as an input a concatenation of medical image data 110 relating to the sequences obtained in step 400 or from the EHR in step 404. Based on this data, the second model 104 computes an output 114 in the form of one or more contrast-difference maps in the coordinate spaces of each of the three acquisition sequences. In embodiments, the three contrast difference maps may be resampled in this step to a single coordinate space. In embodiments where MRI is used, the three contrast difference maps are resampled to the T1 -weighted coordinate space. Step 406: Indicator of contrast enhancement found by second model 104? At step 406 it is determined whether the output 114 from the second model 104 in the form of one or more contrast-difference maps in the coordinate spaces of each of the three acquisition sequences (or a single combined contrast map in a single coordinate space as described above) have identified one or more indicators of contrast enhancement. In embodiments, the one or more indicators of contrast enhancement may comprise the detection or determination of one or more enhanced lesions. In embodiments, if one or more enhanced lesions are identified, then a determination may be made that the patient analysis may benefit from contrast enhancement. If the generated indicators of contrast enhancement have identified that contrast enhancement may be beneficial, then the method proceeds to step 410. If not, the method proceeds to step 408 and no contrast enhancement is recommended. Step 408: No contrast enhancement At step 408, it is determined that no contrast enhancement is beneficial in this case and no contrast enhancement is recommended. Step 410: Enhancing volume fraction At step 410 an enhancing-to-lesion ratio is utilised as a discriminatory feature in recognising HGGs vs LGGs. Once the enhancing volume to total lesion volume ratio is computed in step 410, the method proceeds to step 412. Step 412: Determine indicator of pathology At step 412, based on the results of step 410, an indicator of pathology is determined. In embodiments, the indicator of pathology is determined as a probable LGG or HGG based on the analysis of enhancing volume to total lesion volume ratio from step 412. An indicator of pathology indicative of an LGG would indicate a low dose contrast regime is most appropriate. An indicator of pathology indicative of an HGG would indicate a normal dose contrast regime is most appropriate. Step 414: Recommend contrast regime At step 414, the contrast regime determined by the indicator of pathology in step 318B can be recommended (e.g. dose range or dose value) and subsequently implemented by a medical professional. An exemplary method workflow according to a further embodiment will now be described with reference to Figure 10. In Figure 10, steps 500 to 516 correspond to steps 200 to 216respectively of the embodiment of Figure 5 and, for clarity and brevity, will not be described further here. Step 518 is unique to this embodiment will be described below. Step 518: Generate synthetic contrast image The above embodiments are operable to determine whether contrast should be used or not, and to determine the level to use. , the inventors have further recognised that the analysis and decision processes leading to this stage may result in synthetic pseudo-contrast medical images of a high quality and which have been validated through identification of potential pathologies utilising the first model 104 (i.e. Route A) or via a sequential processing of synthetic contrast data to produce a detailed pseudo-contrast image (Route B). Therefore, from either from Route A or Route B, high quality and validated synthetic contrast image data is obtained. Thus, in this embodiment, step 518 comprises generating a synthetic contrast-enhanced image for further analysis. Further analysis may comprise analysis and / or diagnosis by a medical professional. OVERALL WORKFLOW The above embodiments of the model structure and workflow methods described above are summarised in the workflow diagram shown in Figure 11. This diagram sets out the processing workflow for different stages of analysis and individuals, including the radiologist, technician, MRI scan by the medical scanner 50, processing by the model 100, and subsequent analysis and diagnosis by a physician. EXPERIMENTAL RESULTS The performance of an embodiment of the present invention was tested to determine the effectiveness of the methodology and systems in detecting lesions with uptake of GBCAs. A system in accordance with an embodiment was tested across all lesions in both the Indian and Danish cohorts against the GT obtained via the annotated data. A summary of the 5 evaluation metrics per cohort as well as the Pooled results are reported in Table 1. Results were comparable in the two cohorts. India fn= 350) Denmark (n = 469) Amsterdam fn = 432) Aos. yrs 36.8 ± 20.6 61.8 + 17.8 54,1 + 14.9 Sex, females 145 (43.3%) 255 (54.4%) 177(41.0%) Contrast uptake 100 (28.6%) 156 (33,3%) 355 (82.2 %) Norma! 119 182 0 Brain Tumor 77 154 432 Cyst 8 17 : X'? X Meningioma to 26 : X'?x Craniopharyng-oma 9 0 Pituitary adenoma 13 0 Schwannoma 4 ixA Astrocytoma 1 0 39 Glioblastoma muitiforme 12 6 230 Low-grade glioma 9 8 75 Metastases 9 84 67 Other 78 12 21 infectious disorders 52 4 Granuloma 36 0 x3 x Cerebral abscess 5 0 X X Meningitis o 0 X X. Encephalitis 5 4 : x\x\ Meningo-encephalitis 1 0 : x / x Other 2 inflammatory and autoimmune disorders 4 0 x <.ix Multiple sclerosis 2 0 x < Other 2 0 o Vascular pathology 45 87 : xx-'-x Acute infarct 15 13 : xx^x Hemorrhage 13 13 fxV'x Vascular malformations 1 Other 16 Neurodegeneration IS X <o X <o Toxic and aeguired metabolic conditions 4 : x\+x Other pathologies 31 42 Table 1. Demographics and pathology distributions per cohort. Continuous variables are expressed as mean ± standard deviation. Categorical variables are expressed as n, percentage (%). Age and sex were not available (NA) for n=14 and n=15 respectively for India dataset and n=21 and n= 8 respectively for Denmark dataset. When pooling the data, the proportion of patients who had a true contrast uptake and were correctly recommended by the algorithm for GBCAs administration was 74% (sensitivity) 10 while the proportion of patients who did not have contrast uptake and were correctly not recommended for GBCAs administration ranged 82% (specificity). The proportion of patients recommended for GBCAs administration who actually had contrast uptake was 63% (positive predictive value, PPV), while the proportion of patients not recommended for GBCAs administration who indeed did not show any contrast uptake was 89% (negative predictive value, NPV). The proportion of patients who did not have contrast uptake but were incorrectly recommended for GBCAs administration by the algorithm ranged 18% (false positive rate, FPR); vice versa the proportion of patients who showed contrast uptake but were incorrectly not recommended for GBCAs administration by the algorithm ranged was 26% (false negative rate, FNR). Finally, the proportion of patients recommended for GBCAs administration who did not actually have any contrast uptake was 37% (false discovery rate, FDR). India Denmark pooled sensitivity 0.74 (0.65, 0.62) 0.73 (0.66, 0.80) 0.74 (0.65, 0.83) specificity 0.82 (0.77, 0.8?) 0.76 (0.72, 0.81) 0.82 (0.78, 0.87) PPV 0.63(0.54,0.71) 0.61 (0.54, 0 68) 0.63 (0.54,0.71) NPV 0.89 (0.84, 0.92) 0.85 (0.80, 0.89) 0.89 (0.85, 0.93) FPR 0.18(0.13, 0.23) 0.24 (0.19, 0.28) 0.18 (0.13, 0.22) FDR 0.37 (0.29, 0.48) 0.39 (0.32, 0.45) 0.37 (0.29, 0.46) FNR 0.26 (0.18, 0.35) 0.27 (0.20, 0.34) 0.26 (0.17, 0.35) Table 2. Summary of evaluation metrics assessing Apollo SmartGAD algorithm performance in predicting contrast uptake (yes / no). For each of the evaluation metrics, 95¾ confidence intervals are indicated in brackets. Acronyms: positive predictive value (PPV), negative predictive value (NPV), false positive rate (FPR), false discovery rate (FDR), false negative rate (FNR). An overview of the incorrectly classified cases, i.e., false negatives (FN) and false positives (FP), is depicted in Table 3. Most of the false negatives were naturally under the brain tumour category. Out of the 190 tumours that showed real contrast uptake, 41 were wrongly called as negatives. Among the 41 cases, the most missed were metastasis and meningiomas. The maximum lesion size among the missed lesions was 2.8cm. No reports of oedema were found in the radiological reports for these lesions. The first model 102 alone picked 4 / 11 meningiomas as tumours but due to the design of the model 100 they were deemed to not be recommended for contrast. Among the 20 FN metastases, 18 were from the Danish cohort and all were reported as small lesions. Among the 20 metastases missed, the largest lesion missed was 2.1 cm in diameter. The first model 102 predicted 17 cases as tumours but since the model 100 did not enhance due to its size, contrast recommendations were not made for these cases. Additionally, 11 vascular pathologies were marked to have contrast uptake. However, they were due to the ischemia-induced blood-brain barrier breakdown and warrant a different kind of contrast procedure. Of these 11 cases, 3 cases had DWI sequences and the first model 102 predicted infarctions on all three. 12 infections were missed, of which 7 were granulomas (a condition specific to the Indian subcontinent). Among the false positives (n=116), 37 cases were that of normal cases (32% of FPs). All cases from the Danish cohort had compatible SWI and DWI sequences missing which could be the reason for a large number of false positives. Among the 13 false positives from the Indian cohort, four cases were considered to be borderline upon visual inspection. There was one paediatric patient, one patient with traumatic brain injury with large regions of vasogenic oedema (report incorrect), one patient with a large pituitary macroadenoma (report incorrect leading to wrong ground truth) and one patient with a doubtful pituitary microadenoma (report incorrect leading to wrong ground truth) - all determined from visual inspection by an expert. In addition, 17 cases of false positives had a brain tumour. Among them, 3 were low-grade gliomas (LGGs), 7 were cysts, 3 metastasis, and 2 central nervous system (CNS) tumours. Vascular lesions accounted for 45 false positives (39%), a substantial amount of these may be addressed by using the predictions of the first model 102 of Infarct / Haemorrhages which will be addressed in future work. Among the vascular lesions, 6 were acute infarcts, 12 haemorrhages, and 20 were a mix of different vascular malformations. Finally, Infectious disorders accounted for a small number of false positives. In embodiments of the present invention, the determination of the necessity of administration of gadolinium-based contrast agents (GBCAs) during initial brain MR examinations could significantly reduce the number of contrast scans performed globally. As a result, the present invention has the advantages that patient discomfort and potential adverse effects can be minimised, while also decreasing the environmental footprint of GBCAs. False Negatives Fai.se Positives n Total Ratio (%) n Ratio (%) | Brain tumor 41 190 21.58 1? 14.86 Cyst 4 R 80.00 6.03 | Pituitary adenoma 4 •J ft 40.00 0 .•V S Meningioma 11 36 30.58 0 s 0 o G 2 2.59 | Ependymoma 0 1 6.86 j iGiiobiastoma muitiforme rj 0 1 0.88 Primary CMS lymphoma x 6 16.87 7 1.72 Metastasis so 22.22 2,59 i Otter ■: 2 50.00 Vascular pathology 11 23 47.83 45 38.70 oacute intarc- 6 13 43.15 8 5 17 Hemorrhage •s 8 50.00 12 10 34 Other 2 100.00 20 17.24 Vascular i Malformations 0 2 0 0 0 Infectious disorders 12 48 30.00 4 3.45 Granuloma 7 26 26.92 Cerebral abscess X 25.00 0 1 Meningitis 3 33.33 0 0 Meningo-encsoh aiitis * 1 100.00 4 3.45 iOther 3 33.33 ft inflammatory and autoimmune disorders 1 100.60 .. 0.86 Other 3 8 37.50 11 9.48 Normal p. 8 0 37 31 0 Table 3. Definition of false negatives (FN) and false positives (FP) for each method. 5 To achieve this, embodiments of the present invention are operable to assess the need for GBCA administration in real-time during first-line MR examinations. The model 100 of embodiments is operable to generate contrast recommendations based on pseudo-contrast images derived from pre-contrast brain MRI scans, without prior knowledge of patient diagnoses. The model 100 was validated using retrospective multicentric cohorts from Denmark and India. Additionally, the utility of the model 100 was examined in follow-up scenarios for known brain tumours using a brain tumour dataset from the Netherlands, specifically assessing the decision for low or normal dosage of the model 100. The findings from the inventors indicate that real-time decision-making regarding GBCA administration based on pseudo-contrast imaging is feasible and has the potential to offer significant clinical and economic benefits. By reducing unnecessary GBCA use, this approach enhances patient safety, streamlines radiology workflows, and contributes to more efficient and sustainable medical imaging practices. An aspect of the claimed invention is its high negative predictive value (NPV), which was 89%. The NPV indicates the proportion of negative results (i.e., no GBCA administration recommended) that are correctly identified as true negatives (i.e., cases where GBCAs are genuinely unnecessary). By reliably determining when GBCAs are unnecessary, the present invention can significantly improve the efficiency and safety of brain MR imaging, as it could ensure that patients who do not require GBCAs are correctly identified thereby minimizing the potential risks associated with GBCAs and contributes to the efficient use of medical resources, reducing costs and improving the logistical efficiency of radiology clinics. In aspects, the embodiments described herein relate to a method of extracting information from a digital image. However, the embodiments described herein are equally applicable as an instruction set for a computer for carrying out said method or as a suitably programmed computer. While the invention has been described with reference to the preferred embodiments depicted in the figures, it will be appreciated that various modifications are possible within the scope of the invention as defined in the following claims. Variations are possible. For example, whilst the above embodiments have been shown and described with reference to MRI scans and scan data, the present invention is applicable to other non-invasive imaging techniques for imaging at least a part of the body of a subject. For example, another non-invasive technique for imaging brain and other body regions is computerised tomography (CT) scans which combine a series of X-ray images taken from different angles around the body and use computational methods to create cross-sectional images (slices) of the bones, blood vessels and soft tissues inside the body. Another technique to which the present invention is applicable is positron emission tomography (PET) which can be used to produce detailed 3-dimensional images of the inside of a body. A PET scan utilises radiotracers, which are molecules comprising a small amount of radioactive material that can be detected on the PET scan. They are designed to accumulate in cancerous tumours or regions of inflammation. They can also be made to bind to specific proteins in the body. Whilst examples have referenced the use of gadolinium-based contrast agents, any suitable contrast agent may be used. Contrast agents suitable for use with the present invention may include, in non-limiting examples: magnesium-based contrast agents, iron oxide-based contrast agents, superparamagnetic iron oxide nanoparticle contrast agents such as ferumoxytol, manganese-based contrast agents and protein-based contrast agents. In this specification, unless expressly otherwise indicated, the word "or" is used in the sense of an operator that returns a true value when either or both of the stated conditions are met, as opposed to the operator "exclusive or" which requires only that one of the conditions is met. The word "comprising" is used in the sense of "including" rather than to mean "consisting of". Where applicable, various embodiments provided by the present disclosure may be implemented using hardware, software, or combinations of hardware and software. Also, where applicable, the various hardware components and / or software components set forth herein may be combined into composite components comprising software, hardware, and / or both without departing from the spirit of the present disclosure. Where applicable, the various hardware components and / or software components set forth herein may be separated into sub-components comprising software, hardware, or both without departing from the scope of the present disclosure. In addition, where applicable, it is contemplated that software components may be implemented as hardware components and vice-versa. Software, in accordance with the present disclosure, such as program code and / or data, may be stored on one or more computer readable mediums. It is also contemplated that software identified herein may be implemented using one or more general purpose or specific purpose computers and / or computer systems, networked and / or otherwise. Where applicable, the ordering of various steps described herein may be changed, combined into composite steps, and / or separated into sub-steps to provide features described herein. While various operations have been described herein in terms of “modules”, “units” or “components,” these terms should not be limited to single units or functions. In addition, functionality attributed to some of the modules or components described herein may be combined and attributed to fewer modules or components. The methods described herein are, in use, executed on a suitable computer system or device running one or more computer programs formed in software and / or hardware and operable to execute the above method. A suitable computer system will generally comprise hardware and an operating system. The term ‘computer program’ is taken to mean any of (but not necessarily limited to) an application program, middleware, an operating system, firmware or device drivers or any other medium supporting executable program code. The methods described herein may be embodied in one or more pieces of software and / or hardware. The software is preferably held or otherwise encoded upon a memory device such as, but not limited to, any one or more of, a hard disk drive, RAM, ROM, solid state memory or other suitable memory device or component configured to software. The methods may be realised by executing / running the software. Additionally or alternatively, the methods may be hardware encoded. The method encoded in software or hardware is preferably executed using one or more processors. The memory and / or hardware and / or processors are preferably comprised as, at least part of one or more servers and / or other suitable computing systems. The term ‘hardware’ may be taken to mean any one or more of the collection of physical elements that constitutes a computer system / device such as, but not limited to, a processor, memory device, communication ports, input / output devices. The term ’firmware’ may be taken to mean any persistent memory and the program code / data stored within it, such as but not limited to, an embedded system. The term ‘operating system’ may taken to mean the one or more pieces, often a collection, of software that manages computer hardware and provides common services for computer programs.

Claims

1. A method for contrast agent protocol selection for a subject under assessment, the method being executed by at least one hardware processor and comprising the steps of:a) obtaining, from a scan acquisition process performed by a medical imaging system, medical image data representative of the anatomical structure and appearance of at least a part of the subject;b) utilising the medical image data in a first machine learning model trained to classify anatomical features in the medical image data to generate an output identifying any quantitative pathological indicators representative of the presence of pathological features in the medical image data;c) utilising the medical image data in a second machine learning model trained to predict an effect of contrast agent on the medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features derived from the medical image data;d) comparing the outputs from the first and second machine learning models to determine any correlation between any identified quantitative pathological indicators and any identified quantitative contrast indicators and, if any correlation is identified, the method further comprises:e) generating a recommendation for a scan acquisition process using contrast agent protocols.

2. A method according to claim 1, wherein the output from the first machine learning model comprises a first feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative pathological indicators have an associated spatial location on the first feature map.

3. A method according to claim 1 or 2, wherein the output from the second machine learning model comprises a second feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative contrast indicators have an associated spatial location on the second feature map.

4. A method according to claim 3, wherein the second machine learning model is configured to generate synthetic contrast agent enhanced medical image data from the medical image data.

5. A method according to claim 4, wherein the second feature map comprises a contrast difference map representative of the difference between the synthetic contrast agent enhanced medical image data and the medical image data.

6. A method according to claim 3, 4 or 5 when dependent on claim 2, wherein determining any correlation in step d) comprises determining whether the spatial location of any identified quantitative pathological indicators overlaps with the spatial location of any identified quantitative contrast indicators.

7. A method according to any one of claims 4 or 5, wherein if no quantitative pathological indicators are identified in step b), then the method further comprises:f) utilising a third computational model to reduce the synthetic contrast in regions of the synthetic contrast agent enhanced medical image data related to anatomically normal features to generate calibrated synthetic contrast agent enhanced medical image data.

8. A method according to claim 7, wherein, subsequent to step f), the method further comprises:g) utilising a fourth computational model to segment the calibrated synthetic contrast agent enhanced medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features in the medical image data; and if any quantitative contrast indicators are identified, the method comprises:h) generating a recommendation for a scan acquisition process using contrast agent protocols.

9. A method according to any one of the preceding claims, wherein, subsequent to a recommendation for a further scan acquisition process using contrast agent protocols in step e), the method further comprises:i) determining the ratio of the size of the pathological features associated with the identified quantitative pathological indicators to the size of the contrast-enhanced pathological features associated with the identified quantitative contrast indicators;j) determining, based on the ratio, one or more quantitative pathological indicators representative of specific pathological features; andk) recommending a contrast dosing regime in dependence upon the determined quantitative pathological indicators.

10. A method according to any one of the preceding claims, wherein if at step b) any quantitative pathological indicators representative of the presence of acute pathological features are identified the method further comprises:I) generating a notification indicative of the probable need for urgent treatment of the subject.

11. A method according to any one of the preceding claims, wherein step a) further comprises:m) performing a scan acquisition process on a subject in the medical imaging system to obtain the medical image data.

12. A method according to claim 11, wherein the scan acquisition process comprises a plurality of different scan types to generate the medical image data.

13. A method according to claim 12, wherein the medical image data utilised in step b) and / or step c) comprises medical image data generated from one or more of the different scan types.

14. A method according to claim 12, wherein the medical image data utilised in step b) comprises medical image data from a first set of scan types and the medical image data utilised in step c) comprises medical image data from a second set of scan types, wherein the first and second sets are different.

15. A method according to any one of claims 11 to 14, wherein the subject remains in the medical imaging system for steps b) to e).

16. A method according to claim 15, wherein step e) further comprises a recommendation for a scan acquisition process using contrast agent protocols to be performed whilst the subject remains in the medical imaging system.

17. A method according to any one of the preceding claims, wherein the medical imaging system comprises an MRI scanner and the plurality of scan types are selected from the group of: T1 -weighted; T2-weighted; T2-FLAIR; SWI / T2* and DWI b1000.

18. A method according to any one of the preceding claims, wherein the medical imaging system comprises a CT scanner and the imaging scan data comprises computed tomography (CT) data or the medical imaging system comprises a PET scanner and the imaging scan data comprises PET data.

19. A method according to any one of the preceding claims, wherein the recommended contrast agent protocol comprises the use of a contrast agent selected from the group of: gadolinium-based contrast agents; magnesium-based contrast agents, iron oxide-based contrast agents, superparamagnetic iron oxide nanoparticle contrast agents; manganese-based contrast agents; and protein-based contrast agents.

20. A method for generating a synthetic contrast-enhanced medical image for a subject under assessment, the method being executed by at least one hardware processor and comprising the steps of:a) obtaining, from a scan acquisition process performed by a medical imaging system, medical image data representative of the anatomical structure and appearance of at least a part of the subject;b) utilising the medical image data in a first machine learning model trained to classify anatomical features in the medical image data to generate an output identifying any quantitative pathological indicators representative of the presence of pathological features in the medical image data;c) utilising the medical image data in a second machine learning model trained to generate synthetic contrast agent enhanced medical image data from the medical image data to identify any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features derived from the medical image data;d) comparing the outputs from the first and second machine learning models to determine any correlation between any identified quantitative pathological indicators and any identified quantitative contrast indicators and, if any correlation is identified, the method further comprises:e) generating a synthetic contrast-enhanced medical image data for subsequent analysis.

21. A method according to claim 20, wherein if no quantitative pathological indicators are identified in step b), then the method further comprises:f) utilising a third computational model to reduce the synthetic contrast in regions of the synthetic contrast agent enhanced medical image data related to anatomically normal features to generate calibrated synthetic contrast agent enhanced medical image data.

22. A method according to claim 21, wherein, subsequent to step f), the method further comprises:g) utilising a fourth computational model to segment the calibrated synthetic contrast agent enhanced medical image data to generate an output identifying any quantitative contrast indicators representative of the presence of contrast-enhanced pathological features in the medical image data; and if any quantitative contrast indicators are identified, the method comprises:h) generating a synthetic contrast-enhanced medical image data for subsequent analysis.

23. A method according to any one of claims 20 to 22, wherein the output from the first machine learning model comprises a first feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative pathological indicators have an associated spatial location on the first feature map.5 24. A method according to any one of claims 20 to 23, wherein the output from the secondmachine learning model comprises a second feature map representative of the spatial distribution of anatomical features in the medical image data and wherein any identified quantitative contrast indicators have an associated spatial location on the second feature map.io25. A computer readable medium comprising instructions which, when executed by a computer, cause the computer to perform a method according to any one of the preceding claims.15 26. A computer system comprising: a processing device, a storage device and a computerreadable medium as claimed in claim 25.

Images