Method for scatter correction of medical images

Abstract

A method for correcting scatter in a reconstructed medical image comprises inputting the reconstructed volume 302 into a trained machine learning model 304 to output a correction volume 306. The corre

Description

Field of the Invention The present disclosure relates to a method of training a machine learning (ML) model to predicta correction volume for correcting scatter in reconstructed medical images. The present disclosure also relates to a method of correcting scatter in reconstructed medical images. The present disclosure also relates to a training node, a processing node, and to a computer program product configured, when run on a computer, to carry out methods of training a ML model and of correcting scatter in reconstructed medical images. Background of the Invention Radiotherapy (RT) is one of the cornerstones of cancer treatment, using ionizing radiation to eradicate tumor cells. A total radiation dose for a patient is typically divided into 3-30 daily fractions to optimize its effect. As the surrounding normal tissue is also sensitive to radiation, highly accurate delivery is a key part of effective RT. Image guided RT (IGRT) is a technique to capture the anatomy of the patient at the time of dose fraction delivery, using in-room imaging in order to align the treatment beam with the tumor location. A Computed Tomography (CT) reconstruction is a three-dimensional image showing the internal structure of a patient or an object. It is prepared by taking x-ray projections through the patient or object, which are then used in a computational process that calculates a structure that would have led to the collection of x-ray projections that was used. Cone Beam Computed Tomography (CBCT) is the most widely used imaging modality for IGRT. In CBCT, a cone of radiation is directed towards the patient or object and detected after attenuation by a two-dimensional flat-panel detector to yield a number of two-dimensional projection images. The radiation source and the detector are then rotated around the patient or object to give the necessary collection of images from multiple directions. These can then be used to reconstruct a three-dimensional volume image. Regardless of the type of CT scanning, the mathematical algorithms used to create the images from the projections assume that the photons which arrive at a specific location in the projection image have been attenuated along a straight-line path from a point-like source of radiation. In practice, and especially for fan-beam and cone-beam CT, this is not the case due to scattering. When an x-ray photon interacts with matter, it can be attenuated (i.e. absorbed), or it can be scattered. In the latter case, the photon is re-emitted in a random direction, and therefore may be detected giving rise to an inaccurate measurement of the attenuation elsewhere in the projection image. Fortunately, x-ray scattering is a well-characterised phenomenon and is therefore relatively predictable given knowledge of the nature of the beam and the matter that it will be interactingwith. Many approaches to scatter correction already exist using hardware (e.g. antiscatter grids) or software solutions. Software corrections can broadly be classified as methods that work in projection space, in image space, or in both (e.g. iteratively). These approaches can be further classified as using either classical (e.g. analytic or kernel based), neural network or simulation (e.g. Monte Carlo simulation) approaches. For example, a Monte Carlo-type simulation can be run to compute the outcome of a large number of random interactions between x-ray photons and the target to produce projection images of just the scattered radiation. These images can then be subtracted from the actual projection images captured by the apparatus, producing a set of clean (substantially scatter-free) projection images. These can be used to reconstruct a substantially scatter-free CBCT image. Summary of the Invention Neural network-based approaches (in both projection and image space) have shown considerable benefit in their combination of speed and performance, making them a highly promising candidate for industrial use. However, the ‘black box’ nature of neural networks, and the difficulty of making guarantees about their behaviour (for example the difficulty of ruling out the possibility of so-called ‘hallucinations’) presents a challenge that must be overcome for safe adoption in medical settings. In particular, applying neural networks directly to reconstructed images (and directly predicting ‘corrected’ images) is risky, as there is a possibility of altering the images in unpredictable ways. It is an aim of the present disclosure to at least partially address one or more of the challenges mentioned above. The invention is defined in the independent claims, to which reference should now be made. Further features are set out in the dependent claims. According to a first aspect of the present disclosure, there is provided a computer-implemented method of training a machine learning model to predicta correction volume for correcting scatter in reconstructed medical images. The method comprises obtaining an input training dataset comprising a plurality of pairs of reconstructed volumes, each pair comprising: (i) a first reconstructed volume comprising scatter artifacts, and (ii) a second reconstructed volume comprising fewer scatter artifacts than the first reconstructed volume, wherein the second reconstructed volume corresponds to the first reconstructed volume. The method further comprises training the machine learning model based on the input training dataset. The trained machine learning model is configured to accept a reconstructed volume and output a correction volume which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume. Values of the correction volume are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts. In some embodiments, the first reconstructed volume is obtained by adding scatter artifacts to the second reconstructed volume. In some embodiments, adding scatter artifacts comprises using Monte-Carlo techniques. In some embodiments, adding scatter artifacts comprises randomly adding smooth dark and / or light patches to the second reconstructed volume. In some embodiments, the plurality of pairs of reconstructed volumes comprise simulated reconstructed volumes. In some embodiments, values of the correction volume are constrained to be smoothly varying across the correction volume. In some embodiments, constraining the values of the correction volume to be smoothly varying across the correction volume comprises enforcing a maximum spatial gradient across the correction volume. In some embodiments, constrainingthe values of the correction volume to be smoothly varying across the correction volume comprises predicting parameters of gaussians in 3D space and calculating a weighted sum of the gaussians to determine the correction volume. In some embodiments, values of the correction volume are constrained such that they do not produce negative values when applied to the reconstructed volume. In some embodiments, values of the correction volume are constrained such that they do not change values of the reconstructed volume by more than a predetermined amount, and optionally wherein the predetermined amount is 1%, 3%, 5%, 10%, 15% or20%. In some embodiments, values of the correction volume are constrained such that a number of regions within the correction volume having increasing values is at or below a predetermined number, and optionally wherein the predetermined number is 1,2, 3, 4 or 5. In some embodiments, values of the correction volume are constrained such that a number of regions within the correction volume having decreasing values is at or below a predetermined number, and optionally wherein the predetermined number is 1,2, 3, 4 or 5. In some embodiments, the reconstructed volume comprises a CBCT reconstructed volume. According to a second aspect of the present disclosure, there is provided a computer-implemented method of correcting scatter in reconstructed medical images. The method comprises obtaining a reconstructed volume. The method further comprises inputting the reconstructed volume into a trained machine learning model which is trained to output a correction volume, which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume. Values of the correction volume are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts. The method further comprises applying the correction volume to the reconstructed volume to obtain a final corrected volume. In some embodiments, the trained ML model has been trained in accordance with the first aspect of the present disclosure. In some embodiments, applying the correction volume comprises multiplying the reconstructed volume by the correction volume and / or adding the correction volume to the reconstructed volume. In some embodiments, the method further comprises displaying the final corrected volume to a user. In some embodiments, the reconstructed volume has already undergone a first scatter correction and the inputting and applying steps comprise a second scatter correction of the reconstructed volume. In some embodiments, the reconstructed volume has not been previously corrected for scatter. In some embodiments, values of the correction volume are constrained to be smoothly varying across the correction volume. In some embodiments, constraining the values of the correction volume to be smoothly varying across the correction volume comprises enforcing a maximum spatial gradient across the correction volume. In some embodiments, constrainingthe values of the correction volume to be smoothly varying across the correction volume comprises predicting parameters of gaussians in 3D space and calculating a weighted sum of the gaussians to determine the correction volume. In some embodiments, values of the correction volume are constrained such that they do not produce negative values when applied to the reconstructed volume. In some embodiments, values of the correction volume are constrained such that they do not change values of the reconstructed volume by more than a predetermined amount, and optionally wherein the predetermined amount is 1%, 3%, 5%, 10%, 15% or20%. In some embodiments, values of the correction volume are constrained such that a number of regions within the correction volume having increasing values is at or below a predetermined number, and optionally wherein the predetermined number is 1,2, 3, 4 or 5. In some embodiments, values of the correction volume are constrained such that a number of regions within the correction volume having decreasing values is at or below a predetermined number, and optionally wherein the predetermined number is 1,2, 3, 4 or 5. In some embodiments, the reconstructed volume comprises a CBCT reconstructed volume. According to a third aspect of the present disclosure, there is provided a computer program product comprising a computer readable medium, the computer readable medium having computer readable code embodied therein, the computer readable code being configured such that, on execution by a suitable computer or processor, the computer or processor is caused to perform a method in accordance with any of the aspects / embodiments described above. According to a fourth aspect of the present disclosure, there is provided a training node for training a ML model to predict a correction volume for correcting scatter in reconstructed medical images, the training node comprising processing circuitry configured to cause the training node to carry out a method in accordance with the first aspect described above. According to a fifth aspect of the present disclosure, there is provided a processing node for correcting scatter in reconstructed medical images, the processing node comprising processing circuitry configured to cause the processing node to carry out a method in accordance with the second aspect described above. According to a sixth aspect of the present disclosure, there is provided a radiotherapy treatment apparatus comprising at least one of: a training node in accordance with the fourth aspect described above and a processing node in accordance with the fifth aspect described above. Other features of the disclosure are described below. The invention may be implemented in digital electronic circuitry, or in computer hardware, firmware, software, or in combinations thereof. The invention may be implemented as a computer program or a computer program product, i.e. a computer program tangibly embodied in a non-transitory information carrier, e.g. in a machine-readable storage device or in a propagated signal, for execution by, or to control the operation of, one or more hardware modules. A computer program may be in the form of a stand-alone program, a computer program portion, or more than one computer program, and may be written in any form of programming language, including compiled or interpreted languages, and it may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a data processing environment. The invention is described in terms of particular embodiments. Other embodiments are within the scope of the following claims. For example, the steps of the invention may be performed in a different order and still achieve desirable results. Elements of the invention have been described using the terms “processor” etc. The skilled person will appreciate that such functional terms and their equivalents may refer to parts of the system that are spatially separate but combine to serve the function defined. Equally, the same physical parts of the system may provide two or more of the functions defined. For example, separately defined means may be implemented using the same memory and / or processor as appropriate. Brief Description of the Drawings Embodiments of the invention will now be further described by way of example only and with reference to the accompanying drawings, wherein like reference numerals refer to like parts, and wherein: Figure 1 is a flow chart illustrating process steps in a method of training a ML model to predict a correction volume for correcting scatter in reconstructed medical images; Figure 2 is a flow chart illustrating process steps in a method of correcting scatter in reconstructed medical images; Figure 3 is a schematic overview of the method of correcting scatter in reconstructed medical images; Figure 4 shows example results from experiments; Figure 5 illustrates an example two-dimensional slice of a correction volume; Figure 6A is a block diagram illustrating functional modules in an example training node; Figure 6B is a block diagram illustrating functional modules in an example processing node; Figure 7 is a diagram of the geometric arrangement for cone beam projection scanning and reconstruction; Figure 8 is a diagram of a cone beam CT scanner suitable for obtaining a reconstructed volume for use in a method for correcting scatter of the reconstructed volume; and Figure 9 is a radiotherapy system, suitable for carrying out embodiments of the invention. Detailed Description The present disclosure proposes methods to correct for the effects of scatter in reconstructed medical images using a trained neural network, where the scatter correction is constrained. This is achieved by inputting a reconstructed image volume to be corrected into a trained neural network. The trained neural network then outputs a correction volume (also referred to herein as a “multiplier volume”), a volume which, when multiplied by (or added to, or otherwise applied to) the reconstructed image volume, gives a corrected reconstructed image volume. The correction volume is constrained so that it can only make changes to the reconstructed image volume that correspond to plausible scatter artifacts. This method works directly in image space (rather than projection space). Working in image space is advantageous because it is possible to better constrain predicted values directly knowingthe voxel value instead of the prior-to-reconstruction pixel values. In image space it is also possible to leverage three-dimensional topological info (e.g. proximity of structures where it is known scatter will be decreasing / increasing). In image space it is also possible to focus on specific three-dimensional regions known to be affected by scatter related artifacts (e.g. cupping artifacts at the center of field of view). Importantly, the scatter correction is constrained to allow for guarantees about its behaviour, which is crucial when applying neural networks to medical image data. The approach is also designed to work well as a ‘secondary correction’. For example, it can be applied to a reconstructed volume on which projection space scatter correction was already used to correct / reduce any remaining scatter artifacts. Embodiments of the present disclosure use neural networks to predict scatter correction of CBCT images but constrain the prediction to avoid hallucinations by the neural networks. This is especially important in the medical field. The present disclosure combines the following ideas (which are discussed in more detail below): 1. Predict a correction / multiplier volume instead of directly predicting the corrected image. 2. Constrain the correction / multiplier volume in a way that aligns with the kinds of artifacts resulting from scatter. 3. If applying this approach as a secondary scatter correction (or in situations where scatter is minimal) then tighter constraints may be used on the correction / multiplier volume, resulting in tighter error bounds. Various aspects and details of these principal concepts will be described below with reference to Figures 1 to 9. The proposed approach is to train a neural network which takes as input the reconstructed image volume to be corrected (V) and predicts a correction / multiplier volume (M). M is then applied to (e.g. multiplied (elementwise) with) V to produce the final corrected volume. Thus, in the case of applying by multiplication, the final corrected volume is V*M. Additionally, the correction / multiplier volume is constrained so that the network cannot hallucinate detailed structures and can only make predictions that differ from the input in ways that could feasibly result from scatter. Figure 1 is a flow chart illustrating process steps in a computer implemented method 100 of training a ML model to predict a correction volume for correcting scatter in reconstructed medical images. The method may be performed by a training node, which may comprise a physical or virtual node, and may be implemented in a computer system, treatment apparatus, such as a radiotherapy treatment apparatus, computing device, or server apparatus, and / or may be implemented in a virtualized environment, for example in a cloud, edge cloud, or fog deployment. Examples of a virtual node may include a piece of software or computer program, a code fragment operable to implement a computer program, a virtualised function, or any other logical entity. The training node may encompass multiple logical entities, as discussed in greater detail below. Referring to Figure 1, the method 100 comprises, in step 102, obtaining an input training dataset comprising a plurality of pairs of reconstructed volumes. Each pair comprises: (i) a first reconstructed volume comprising scatter artifacts, and (ii) a second reconstructed volume comprising fewer scatter artifacts than the first reconstructed volume (advantageously, the second reconstructed volume may have no scatter artifacts, e.g. the second reconstructed volume is “scatter-free”). The second reconstructed volume corresponds (e.g. spatially corresponds) to the first reconstructed volume (i.e. the second reconstructed volume is a version of the first reconstructed volume, but with fewer scatter artifacts, e.g. the second reconstructed volume may be a scatter-free version of the first reconstructed volume). In some examples, the reconstructed volumes are CBCT reconstructed volumes. In some examples, the training dataset comprises simulated reconstructed volumes. In some examples, the first reconstructed volume is obtained by adding scatter artifacts to the second reconstructed volume. Adding scatter artifacts may be achieved using Monte-Carlo techniques and / or by randomly adding smooth dark and / or light patches to the second reconstructed volume. Example approaches of how simulated data may be used to train the network include the following. For example, if a is the ground truth scatter-corrected volume (referred to as the “second reconstructed volume”) and b is that same volume with scatter artifacts (referred to as the “first reconstructed volume”), e.g., from Monte Carlo scatter simulation, then one approach is to learn to map from a*z+b*(1-z) to a, for z randomly sampled in [0,1] (and randomlyre-sampled each batch). This then simulates scatter being randomly somewhere between totally uncorrected and fully corrected. The distribution over [0,1] from which z is sampled can be adjusted based on, for example, whether the network is intended for use as a secondary correction. Another approach is to randomly add smooth dark / light patches to the second reconstructed volume a to produce a corrupted volume p, and then learn to map for p back to a. When adding smooth dark / light patches, a user may decide how smooth these random perturbations should be. A reasonable estimate of realistic smoothness is possible either qualitatively or quantitively. In practice, preliminary investigation shows that either of (or a combination of both of) these two approaches can produce good results. In step 104, the method comprises trainingthe ML model based on the input training dataset. The ML model may be any neural network(e.g. conventional neural networks, transformers, etc.). The trained ML model is configured to accept a reconstructed volume and output a correction volume which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume. Values of the correction volume are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts. A correction volume is a 3D dataset used to correct for the impact of scattered radiation in medical imaging. The values of the correction volume are values applied to the reconstructed volume which relate to the predicted amount of scatter in each voxel of the volume. Embodiments of the present disclosure allow the amount the predicted corrected volume can differ from the initial volume V to be directly controlled by controlling the multiplier / correction volume M. For example, if M is forced to be 1 everywhere then the final prediction is exactly the same as the input (i.e., no correction was made). Similarly, if all elements of M are constrained to be in [0.9,1.1] then the values of the prediction can’t deviate from the input volume by more than 10% (and thus, for example, no lung tissue could be converted to fat). These examples show that restricting the possible values M can take allows the space of possible final predictions to be bound. Thus, for safe scatter correction, the aim is to constrain M so that it can only make changes to the volume that correspond to plausible scatter artifacts. Scatter artifacts (especially after an initial correction step) generally present as smooth and localized dark or bright regions in the reconstructed volume. Therefore, M may be restricted to be smooth. This can be done in a number of ways, such as enforcing a maximum spatial gradient (e.g., by predicting and then up-sampling a low resolution multiplier volume with restrictions on the maximum and minimum values) or by directly predicting the parameters of a small number of gaussians in 3D space, the weighted sum of which (added to a volume of ones) yields the multiplier volume. Many other strategies are also possible. Restricting the spatial gradient removes the ability to hallucinate any high-frequency structure, such as additional anatomical structures. The less scatter there is to correct, the more strongly it is possible to restrict the multiplier volume, and thus the more the possibility of error is restricted. For example, if only small residual scatter errors are expected (e.g., after a previous preliminary scatter correction has already been performed) then it is possible to be stricter with the maximum allowed spatial gradient in M. As discussed above, the way the values are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts may be chosen based on prior knowledge of how scatter artifacts present. The constraints may be applied via choices in the architecture of the network itself. As an example, in the training phase, the first reconstructed volume may be prepared by adding scatter to the second reconstructed volume by means of Monte-Carlo or similar. Once the first reconstructed volume is generated, then one could evaluate a specific metric such as maximum spatial gradient (it is known that scatter is a smooth function), maximum values (e.g. maximum values of the correction volume, or maximum values of the final corrected volume), structure specific scatter gradient (e.g. bone / tissue or air / body interfaces). These could be used as global or region-specific constraints during the inference phase. These constraints may be learnt by the ML model heuristically (based on the inputs of the first and second reconstructed volumes), or may be input into the ML model during the training phase. For example, it may be desirable to calculate scatter in a thorax scan. It is known that properties of scatter change at the interfaces between soft tissue and the lungs. Since it is known from the original reconstructed volume (the volume to be corrected) where soft tissue and the lungs are, it is possible to constrain the scatter to smoothly transition from high value in soft tissue to low values in the lungs. In some examples, the values of the correction volume may be constrained to be smoothly varying across the correction volume. In some examples, constraining the values of the correction volume to be smoothly varying may comprise enforcing a maximum spatial gradient across the correction volume. In some examples, constraining the values of the correction volume to be smoothly varying across the correction volume comprises predicting parameters of gaussians in 3D space and calculating a weighted sum of the gaussians to determine the correction volume. In some examples, the values of the correction volume may be constrained such that the correction volume will not produce negative values when applied to the reconstructed volume. In some examples, the values of the correction volume maybe constrained such that the reconstructed volume’s values will not change by more than a certain percentage (e.g. in the case of applying the correction volume by multiplication, the correction volume values may be restricted to be within the range 0.9 to 1.1 such that the reconstructed volume’s values will not change by more than 10%). In other examples the percentage could be 1% (for example, if the scatter correction is a secondary scatter correction, the constraints can be tighter), 3%, 5%, 15%, 20%, between 1 and 10%, between 1 and 20%, between 10 and 20%, etc. The choice of the tightness of the constraints can be made based on the specifics of the case, e.g. whether it is a secondary correction. In some examples, the values of the correction volume may be constrained so that the number of areas / regions within the correction volume having reducing / decreasing values is at or below a certain value (e.g. the certain value may be 1,2, 3,4 or 5 and may be chosen based on knowledge of the scatter artifacts present). In some examples, the values of the correction volume may be constrained so that the number of areas / regions within the correction volume having increasing values is at or below a certain value (e.g. the certain value maybe 1,2, 3,4 or 5 and maybe chosen based on knowledge of the scatter artifacts present). For example, Figure 5 shows a two-dimensional slice of a simplified example 3D correction volume which has one area having increasing values 502, and one area having decreasing values 504. For example, the correction volume could be constrained to have a maximum of one area having increasing values and a maximum of two areas having decreasing values. This example is purely for illustrative purposes. Any of these examples of constraints may be used in combination (e.g. M may be constrained to be smoothly varying and have values between 0.9 and 1.1). Further, in some examples there may be different ML models trained for different regions of the volume. For example, a pelvic field of view (FOV) scan may be divided into regions. At least one region may encompass the centre of the FOV where it is known that scatter is greater and the scatter-related cupping artifact produces darker areas (i.e. that area suffers from the greatest scatter effect). Then, several ML models may be trained so the regions have different scatter corrections (higher or lower). If the result of the trained ML models on a pelvic scan is output to a user, the user may see an underestimation of scatter in the centre of the FOV as it may still be darker after correction, then a new ML may be trained and used with “higher” correction in that region. In other words, multiple ML models may be trained for different regions, and if one region has sub-optimal scatter correction, the ML model responsible forthat region may be re-trained. Using the training dataset to train the ML model may comprise usingthe training dataset to update values of trainable parameters of the ML model. This may comprise updating values of the trainable parameters so as to minimise, fora given second reconstructed volume, a loss function based on a difference between the second reconstructed volume and the first reconstructed volume. The ML model may be operable to receive the following as inputs: a reconstructed volume (i.e. a 3D medical image such as a CBCT scan). The ML model may process the input in accordance with values of its trainable parameters, and output a correction volume which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume, wherein values of the correction volume are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts. The method 100 is for training a ML model. For the purposes of the present disclosure, the term “ML model” encompasses within its scope the following concepts: machine learning algorithms, comprising processes or instructions through which data may be used in a training process to generate a model artefact for performing a given task, or for representing a real-world process or system; and the model artefact that is created by such a training process, and which comprises the computational architecture that performs the task. Figure 2 is a flow chart illustrating process steps in a computer implemented method 200 of correcting scatter in reconstructed medical images. The method may be performed by a processing node, which may comprise a physical or virtual node, and may be implemented in a computer system, treatment apparatus, such as a radiotherapy treatment apparatus, computing device, or server apparatus, and / or may be implemented in a virtualized environment, for example in a cloud, edge cloud, or fog deployment. Examples of a virtual node may include a piece of software or computer program, a code fragment operable to implement a computer program, a virtualised function, or any other logical entity. The processing node may encompass multiple logical entities, as discussed in greater detail below. Referring to Figure 2, the method 200 comprises, in step 202, obtaining a reconstructed volume. In some examples, the reconstructed volumes are CBCT reconstructed volumes. The reconstructed volume may have already undergone a first scatter correction (e.g. in projection space). In this case, steps 204 and 206 described below may comprise a second scatter correction of the reconstructed volume (i.e. this scatter correction of method 200 may be a secondary correction). Using this approach as a secondary correction is advantageous as the correction volume may be more tightly constrained. Alternatively, the reconstructed volume may not have been previously corrected for scatter (i.e. this scatter correction of method 200 is the reconstructed volume’s first scatter correction). In step 204, the method comprises inputting the reconstructed volume into a trained ML model which is trained to output a correction volume, which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume, wherein values of the correction volume are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts. The trained ML model may have been trained in accordance with method 100 described above. As described in relation to method 100, the way the values are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts maybe chosen based on prior knowledge of how scatter artifacts present. The examples discussed in relation to method 100 also apply to method 200. For example, the values may be constrained to be smoothly varying, to not produce negative values, to limit change to 10% or less, and / or to constrain the number of increasing / decreasing areas. In step 206, the method comprises applying the correction volume to the reconstructed volume to obtain a final corrected volume. Applying the correction volume may comprise multiplying the reconstructed volume by the correction volume and / or adding the correction volume to the reconstructed volume. The method may further comprise displaying the final corrected volume to a user. At this point, the user may decide whether or not they are satisfied with the scatter correction. If the user is not satisfied, they may choose to put the output volume back into the ML model to further scatter correct the volume. In examples where different ML models are used for different regions of the scan (as described above), a user may segment or otherwise identify areas or regions where the scatter correction is sub-optimal. In such a case, the user may initiate scatter correction again using a new ML model where all the other regions are kept the same except for those segmented or identified regions where the scatter correction was sub-optimal. The new ML model may have higher or lower scatter correction in the segmented or identified regions. Figure 3 shows a schematic overview of the proposed approach. (A) shows the details of the proposed pipeline, where a network 304 predicts a multiplier / correction volume 306 that is then multiplied (elementwise) with the input reconstructed volume 302 to produce the final scatter-corrected output 308. (B) shows a use-case for which this approach is well suited: performing a final correction on an already corrected reconstruction 310, in which some scatter artifacts remain. In this case, the proposed approach 312 is applied to the scatter corrected volume 310 which produces a final corrected volume 314. As a neural network-based approach, embodiments of the present disclosure are fast and perform well. This has been demonstrated in preliminary experiments, the results of which are shown in Figure 4. Figure 4 shows two rows of 2D slices of reconstructed volumes. The top row 402 are the input reconstructed volumes (i.e. the obtained reconstructed volumes of method 200). The bottom row 404 are the output reconstructed volumes (i.e. the final corrected volumes of method 200). As can be seen in the Figure, the inputs 402 have scatter artifacts such as dark patches and light patches. In the outputs 404, these scatter artifacts are no longer present as they have been corrected by the correction volume. The constraints used to constrain the correction volume in the example shown in Figure 4 were max spatial gradients and non-negative values (i.e. the correction volume could not have any negative values). However, the key point is that these fast and accurate results are achieved whilst safely restricting the possible behaviour (i.e. preventing hallucinations by constraining the correction volume). Example methods according to the present disclosure allow fast scatter correction of medical images using a neural network but constrain the correction to avoid hallucinations by the neural network. The methods described above offer speed advantages associated with use of machine learning while mitigating the risks of using Al in medical fields (i.e. by preventing hallucinations) to produce high quality, accurate images. This combination of speed and quality can support both the planning and delivery of radiotherapy treatment, for example in the form of online Adaptive Radiotherapy (ART). The safety mechanism provided by methods of the present disclosure can support the provision of online ART, in which CBCT is used to capture patient imaging at the start of each visit of the treatment fraction. This up-to-date imaging data can enable clinicians to track changes in patient anatomy, includingforexample, tumourshrinkage overthe course of the radiotherapy treatment, allowing for online target localisation and plan adaptation without the constraints of diagnostic CT imaging. The safety mechanism offered by methods according to the present disclosure may result in many additional medical treatment benefits (including improved accuracy of radiotherapy treatment, reduced exposure to unintended radiation, reduced treatment duration, etc.). The methods presented herein may be applicable to a variety of medical treatment and diagnostic settings or radiotherapy treatment equipment and devices. As discussed above, the method 100 may be performed by a training node, and the present disclosure provides a training node that is adapted to perform any or all of the steps of the above discussed methods. The training node may comprise a physical or virtual node, and may be implemented in a computer system, treatment apparatus, such as a radiotherapy treatment apparatus, computing device, or server apparatus, and / or may be implemented in a virtualized environment, for example in a cloud, edge cloud, or fog deployment. Examples of a virtual node may include a piece of software or computer program, a code fragment operable to implement a computer program, a virtualised function, or any other logical entity. The training node may encompass multiple logical entities, as discussed in greater detail below. Figure 6A is a block diagram illustrating an example training node 600 which may implement the method 100, as illustrated in Figure 1, according to examples of the present disclosure, for example on receipt of suitable instructions from a computer program 608. Referring to Figure 6A the training node 600 comprises a processor or processing circuitry 602, and may comprise a memory 606 and interfaces 604. The processing circuitry 602 is operable to perform some or all of the steps of the method 100 as discussed above with reference to Figure 1. The memory 606 may contain instructions executable by the processing circuitry 602 such that the training node 600 is operable to perform some or all of the steps of the method 100, as illustrated in Figure 1. The instructions may also include instructions for executing one or more telecommunications and / or data communications protocols. The instructions may be stored in the form of the computer program 608. In some examples, the processor or processing circuitry 602 may include one or more microprocessors or microcontrollers, as well as other digital hardware, which may include digital signal processors (DSPs), special-purpose digital logic, etc. The processor or processing circuitry 602 may be implemented by any type of integrated circuit, such as an Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA) etc. The memory 606 may include one or several types of memory suitable for the processor, such as readonly memory (ROM), random-access memory, cache memory, flash memory devices, optical storage devices, solid state disk, hard disk drive, etc. The interfaces 604 may be operable to communicate with other nodes or systems, including for example, a processing node 650. In some examples as discussed above, the training node may be incorporated into treatment apparatus, and examples of the present disclosure also provide a treatment apparatus comprising either or both of a training node as discussed above and / or a processing node as discussed below, and / or a planning node operable to implement a method for developing and / or adaptinga radiotherapy treatment plan. As discussed above, the method 200 may be performed by a processing node, and the present disclosure provides a processing node 650 that is adapted to perform any or all of the steps of the above discussed methods. The processing node may comprise a physical or virtual node, and may be implemented in a computer system, treatment apparatus, such as a radiotherapy treatment apparatus, computing device, or server apparatus, and / or may be implemented in a virtualized environment, for example in a cloud, edge cloud, or fog deployment. Examples of a virtual node may include a piece of software or computer program, a code fragment operable to implement a computer program, a virtualised function, orany other logical entity. The processing node may encompass multiple logical entities, as discussed in greater detail below. Figure 6B is a block diagram illustrating an example processing node 650 which may implement the method 200, as illustrated in Figure 2, according to examples of the present disclosure, for example on receipt of suitable instructions from a computer program 658. Referring to Figure 6B the processing node 650 comprises a processor or processing circuitry 652, and may comprise a memory 656 and interfaces 654. The processing circuitry 652 is operable to perform some or all of the steps of the method 200 as discussed above with reference to Figure 2. The memory 656 may contain instructions executable by the processing circuitry 652 such that the processing node 650 is operable to perform some or all of the steps of the method 200, as illustrated in Figure 2. The instructions may also include instructions for executing one or more telecommunications and / or data communications protocols. The instructions may be stored in the form of the computer program 658. In some examples, the processor or processing circuitry 652 may include one or more microprocessors or microcontrollers, as well as other digital hardware, which may include digital signal processors (DSPs), special-purpose digital logic, etc. The processor or processing circuitry 652 may be implemented by any type of integrated circuit, such as an Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA) etc. The memory 656 may include one or several types of memory suitable for the processor, such as readonly memory (ROM), random-access memory, cache memory, flash memory devices, optical storage devices, solid state disk, hard disk drive, etc. In some examples as discussed above, the processing node may be incorporated into a treatment apparatus, and examples of the present disclosure also provide a treatment apparatus comprising either or both of a processing node and / or a training node as discussed above, and / or a planning node operable to implement a method for developing and / or adapting a radiotherapy treatment plan. Figures 1 and 2 discussed above provide an overview of methods which may be performed according to different examples of the present disclosure. These methods may be performed by a training node and a processing node respectively, as illustrated in Figures 6A and 6B. Figure 7 illustrates the geometry of a typical cone beam projection acquisition setup 700, as may be used to obtain projection data which is reconstructed to obtain reconstructed volumes for use in embodiments. For simplicity, only the x-y plane is illustrated; the skilled reader will appreciate that the beam of rays extends also into the z-axis, orthogonal to the x-y plane. In this example, the acquisition setup captures projection data of object 702. X-ray source 704 generates and emits X-rays 706 towards object 702. In CBCT, X-rays may be considered as a beam of rays, emitted from a point source. Detector (or detectors) 708 capture projections, which are sets of line integrals along paths that radiate from the source 704. Multiple projections of the image may be acquired from different angles by rotating the source 704 and detector 708 around the centre of the image 710. In the present example, the source 704 and detector 708 may be rotated arcuately along orbital path 712. In this way, the x-y plane may be rotated counterclockwise around the point of origin (or centre of the image 710) in a manner that keeps the mutual positional relationship between source 704 and detector 708 when passing through the orbital path 712. Other configurations of imaging systems are, of course, feasible; for instance, the source may be configured to rotate and a complete ring of detectors may be configured to capture projections, or there may be multiple sources arranged circumferentially a round a complete ring of detectors. Further, the imaging system may rotate the source along a helical path, so as to capture projection data along the axis of the object 702 of interest (along the z-axis in the present example). The attenuation of the intensity of the rays that pass through the object 702 may be measured by processing signals received from the detector 708. By making projective measurements at a series of different projection angles through the object 702, a sinogram may be constructed from the projection data, mapping the spatial dimension of the detector array to the projection angle dimension. The intensity attenuation resulting from a particular volume within the object will trace out a sine wave for the spatial dimension along the detector perpendicular to the rotation axis of the system. Volumes of the object farther from the centre of rotation correspond to sine waves with greater amplitudes than those corresponding to volumes nearer the centre of rotation. The phase of each sine wave in the sinogram corresponds to the relative angular positions with respect to the rotation axis. By performing an image reconstruction technique (such as an inverse Radon transform) on the projection data in the sinogram, one may reconstruct an image, where the reconstructed image corresponds to a cross-sectional slice of the object 702. The detector 708 may comprise underlying detector elements (such as pixels or bins of pixels), rather than a single active region. The measured ray intensities may then be taken as a function of underlying intensity signals acquired at detector elements of the detector. This is advantageous in that one reduces the impact of random noise, which may plague an intensity (particularly a low intensity) measured at a single detector element. For example, the measured ray intensities may be an average (mean, mode or median) of underlying single detector element intensities. Figure 8 illustrates the geometry of a cone beam projection acquisition setup, as incorporated into a medical CT scanner 800. Radiation sources 802 emit beams of X-ray radiation, which may pass through the patient supported on a couch within the CT scanner (not depicted). Detectors 804 capture attenuated X-rays. The radiation sources 802 and detectors 804 may be configured to rotate within a gantry of the CT scanner, so as to acquire projective measurements at a series of different projection angles. As shown, the angular separation between measurement orientations need not be equal. Referring back to the methods of the disclosure, the obtained reconstructed volume may be acquired using the set up described above in relation to Figures 7 and 8. Scatter artifacts in CBCT occur due to the scattering behaviour of x-ray photons at diagnostic energies. Figure 9 is a block diagram of an implementation of a radiotherapy system 900, suitable for executing methods according to embodiments. The example radiotherapy system 900 comprises a computing system 910 within which a set of instructions, for causing the computing system 910 to perform the method (or steps thereof) discussed herein, may be executed. For example, the computing system 910 may comprise one or both of the training node and the processing node described above in relation to Figures 6A and 6B. The computing system 910 may implement an image reconstruction system. The image reconstruction system may use any conventional image reconstruction technique suitable for use with projection data. Examples of known suitable reconstruction techniques include a known Feldkamp, Davis, and Kress (FDK) reconstruction method, an iterative reconstruction method, an Al reconstruction method, a deep-learning based reconstruction method and a Polyquant reconstruction method. The computing system 910 may also be referred to as a computer. In particular, the methods described herein may be implemented by a processor or controller circuitry 911 of the computing system 910. The computing system 910 shall be taken to include any number or collection of machines, e.g., computing device(s), that individually or jointly execute a set (or multiple sets) of instructions to perform anyone or more of the methods discussed herein. That is, hardware and / or software may be provided in a single computing device, or distributed across a plurality of computing devices in the computing system. In some implementations, one or more elements of the computing system may be connected (e.g., networked) to other machines, for example in a Local Area Network (LAN), an intranet, an extranet, or the Internet. One or more elements of the computing system may operate in the capacity of a server or a client machine in a client-server network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. One or more elements of the computing system may be a personal computer (PC), a tablet computer, a set-top box (STB), a Personal Digital Assistant (PDA), a cellular telephone, a web appliance, a server, a network router, switch or bridge, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that machine. The computing system 910 includes controller circuitry 911 and a memory 913 (e.g., read-only memory (ROM), flash memory, dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM) or Rambus DRAM (RDRAM), etc.). The memory 913 may comprise a static memory (e.g., flash memory, static random access memory (SRAM), etc.), and / or a secondary memory (e.g., a data storage device), which communicate with each other via a bus (not shown). Memory 913 may be used to store or buffer projection data until required for image processing. Controller circuitry 911 represents one or more general-purpose processors such as a microprocessor, central processing unit, accelerated processing units, or the like. More particularly, the controller circuitry 911 may comprise a complex instruction set computing (CISC) microprocessor, reduced instruction set computing (RISC) microprocessor, very long instruction word (VLIW) microprocessor, processor implementing other instruction sets, or processors implementing a combination of instruction sets. Controller circuitry 911 may also include one or more special-purpose processing devices such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), network processor, or the like. One or more processors of the controller circuitry may have a multicore design. Controller circuitry 911 is configured to execute the processing logic for performing the operations and steps discussed herein. The computing system 910 may further include a network interface circuitry 915. The computing system 910 may be communicatively coupled to an input device 920 and / or an output device 930, via input / output circuitry 916. In some implementations, the input device 920 and / or the output device 930 may be elements of the computing system 910. The input device 920 may include an alphanumeric input device (e.g., a keyboard or touchscreen), a cursor control device (e.g., a mouse or touchscreen), an audio device such as a microphone, and / or a haptic input device. The output device 930 may include an audio device such as a speaker, a video display unit (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)), and / or a haptic output device. In some implementations, the input device 920 and the output device 930 may be provided as a single device, or as separate devices. In some implementations, the computing system 910 may comprise image processing circuitry 914. Image processing circuitry 914 may be configured to process image data 970 (e.g., images, imaging data, projections, projection data), such as medical images obtained from one or more imaging data sources, a treatment device 950 and / or an image acquisition device 940. Image processing circuitry 914 may be configured to process, or pre-process, image data 970. For example, image processing circuitry 914 may convert received image data into a particular format, size, resolution or the like. Image processing circuitry 914 may be configured to perform methods described herein for correcting scatter artifacts. Image processing circuitry 914 may be configured to perform image reconstruction. In some implementations, image processing circuitry 914 may be combined with controller circuitry 911. In some implementations, the radiotherapy system 900 may further comprise an image acquisition device 940 and / or a treatment device 950. The image acquisition device 940 and the treatment device 950 may be provided as a single device. In some implementations, treatment device 950 is configured to perform imaging, forexample in addition to providingtreatment and / or during treatment. Image acquisition device 940 may be configured to perform positron emission tomography (PET), computed tomography (CT), magnetic resonance imaging (MRI), single positron emission computed tomography (SPECT), X-ray, and the like. Image acquisition device 940 may be configured to output image data 970, which may be accessed by computing system 910. Treatment device 950 may be configured to output treatment data 960, which may be accessed by computing system 910. Treatment data 960 may be obtained from an internal data source (e.g., from memory 913) or from an external data source, such as treatment device 950 or an external database. The various methods described above may be implemented by a computer program. The computer program may include computer code (e.g., instructions) arranged to instruct a computer to perform the functions of one or more of the various methods described above. For example, the steps of the methods described in relation to any of Figures 1 and 2 may be performed by the computer code. The steps of the methods described above may be performed in any suitable order. The computer program and / or the code for performing such methods may be provided to an apparatus, such as a computer, on one or more computer readable media or, more generally, a computer program product. The computer readable media may be transitory or non-transitory. The one or more computer readable media could be, for example, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, or a propagation medium for data transmission, for example for downloading the code over the Internet. Alternatively, the one or more computer readable media could take the form of one or more physical computer readable media such as semiconductor or solid-state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disc, and an optical disk, such as a CD-ROM, CD-R / W or DVD. The instructions may also reside, completely or at least partially, within the memory 913 and / or within the controller circuitry 911 during execution thereof by the computing system 910, the memory 913 and the controller circuitry 911 also constituting computer-readable storage media. A “hardware component” is a tangible (e.g. non-transitory) physical component (e.g. a set of one or more processors) capable of performing certain operations and may be configured or arranged in a certain physical manner. A hardware component may include dedicated circuitry or logic that is permanently configured to perform certain operations. A hardware component may comprise a special-purpose processor, such as an FPGA or an ASIC. A hardware component may also include programmable logic or circuitry that is temporarily configured by software to perform certain operations. In addition, the modules and components may be implemented as firmware or functional circuitry within hardware devices. Further, the modules and components may be implemented in any combination of hardware devices and software components, or only in software (e.g. code stored or otherwise embodied in a machine-readable medium or in a transmission medium). Unless specifically stated otherwise, as apparent from the following discussion, it is appreciated that throughout the description, discussions utilizing terms such as “receiving”, “determining”, “displaying”, “enabling”, “applying”, “inputting”, “obtaining”, “training”, or the like, refer to the actions and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system memories or registers or other such information storage, transmission or display devices. While certain embodiments have been described, these embodiments have been presented by way of example only and are not intended to limit the scope of the inventions. Indeed, the novel methods and apparatuses described herein may be embodied in a variety of other forms; furthermore, various omissions, substitutionsand changes in the form of methods and apparatus described herein maybe made.

Claims

1. A computer-implemented method of training a machine learning model to predict a correction volume for correcting scatter in reconstructed medical images, the method comprising:obtaining an input training dataset comprising a plurality of pairs of reconstructed volumes, each pair comprising: (i) a first reconstructed volume comprising scatter artifacts, and (ii) a second reconstructed volume comprising fewer scatter artifacts than the first reconstructed volume, wherein the second reconstructed volume corresponds to the first reconstructed volume;training the machine learning model based on the input training dataset, the trained machine learning model configured to accept a reconstructed volume and output a correction volume which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume, wherein values of the correction volume are constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts.

2. The method of claim 1, wherein the first reconstructed volume is obtained by adding scatter artifacts to the second reconstructed volume.

3. The method of claim 2, wherein adding scatter artifacts comprises using Monte-Carlo techniques.

4. The method of claim 2 or 3, wherein adding scatter artifacts comprises randomly adding smooth dark and / or light patches to the second reconstructed volume.

5. The method of any preceding claim, wherein the plurality of pairs of reconstructed volumes comprise simulated reconstructed volumes.

6. A computer-implemented method of correcting scatter in reconstructed medical images, the method comprising:obtaining a reconstructed volume;inputting the reconstructed volume into a trained machine learning model which is trained to output a correction volume, which, when applied to the reconstructed volume, obtains a scatter-corrected version of the reconstructed volume, wherein values of the correction volumeare constrained such that the correction volume is configured to only alter the reconstructed volume in a way that is consistent with correcting scatter artifacts;applying the correction volume to the reconstructed volume to obtain a final corrected volume.

7. The method of claim 6, wherein the trained machine learning model is trained in accordance with any of claims 1 to 5.

8. The method of claim 6 or 7, wherein applying the correction volume comprises multiplying the reconstructed volume by the correction volume and / or adding the correction volume to the reconstructed volume.

9. The method of any of claims 6 to 8, further comprising displaying the final corrected volume to a user.

10. The method of any of claims 6 to 9, wherein the reconstructed volume has already undergone a first scatter correction and the inputting and applying steps comprise a second scatter correction of the reconstructed volume.

11. The method of any of claims 6 to 9, wherein the reconstructed volume has not been previously corrected for scatter.

12. The method of any preceding claim, wherein values of the correction volume are constrained to be smoothly varying across the correction volume.

13. The method of claim 12, wherein constraining the values of the correction volume to be smoothly varying across the correction volume comprises enforcing a maximum spatial gradient across the correction volume.

14. The method of claim 12 or 13, wherein constraining the values of the correction volume to be smoothly varying across the correction volume comprises predicting parameters of gaussians in 3D space and calculating a weighted sum of the gaussians to determine the correction volume.

15. The method of any preceding claim, wherein values of the correction volume are constrained such that they do not produce negative values when applied to the reconstructed volume.

16. The method of any preceding claim, wherein values of the correction volume are constrained such that they do not change values of the reconstructed volume by more than a predetermined amount, and optionally wherein the predetermined amount is 1%, 3%, 5%, 10%, 15% or 20%.

17. The method of any preceding claim, wherein values of the correction volume are constrained such that a number of regions within the correction volume having increasing values is at or below a predetermined number, and optionally wherein the predetermined number is 1,2, 3, 4 or 5.

18. The method of any preceding claim, wherein values of the correction volume are constrained such that a number of regions within the correction volume having decreasing values is at or belowa predetermined number, and optionally wherein the predetermined number is 1,2, 3,4 or 5.

19. The method of any preceding claim, wherein the reconstructed volume comprises a cone beam computed tomography, CBCT, reconstructed volume.

20. A computer program product comprising a computer readable medium, the computer readable medium having computer readable code embodied therein, the computer readable code beingconfigured such that, on execution by a suitable computeror processor, the computer or processor is caused to perform a method as claimed in any one of the preceding claims.

21. A training node for training a machine learning model to predict a correction volume for correcting scatter in reconstructed medical images, the training node comprising processing circuitry configured to cause the training node to carry out a method as claimed in any one of claims 1 to 5 or anyone of claims 12 to 19 when dependent on anyone of claims 1 to 5.

22. A processing node for correcting scatter in reconstructed medical images, the processing node comprising processing circuitry configured to cause the processing node to carry out a method as claimed in any one of claims 6 to 11 or any one of claims 12 to 19 when dependent on any one of claims 6 to 11.

23. A radiotherapy treatment apparatus comprising at least one of:a training node as claimed in claim 21;a processing node as claimed in claim 22.s

Images