Method and system for radiotherapy contouring

Abstract

A computer-implemented method for radiotherapy contouring, the method comprising: receiving a dosimetric constraint for a physiological structure of a patient; determining a contour delineating and cl

Description

Field of the Invention The present invention relates to methods and systems for radiotherapy contouring. More specifically, the present invention relates to a computer-implemented method for radiotherapy contouring, and data processing apparatuses, computer programs, and non-transitory computer-readable storage mediums configured to execute methods for radiotherapy contouring. Background of the Invention Radiation therapy or radiotherapy (RT) the use of ionising radiation to damage or destroy unhealthy cells in both humans and animals. The ionising radiation may be directed to tumours on the surface of the skin or deep inside the body. Common forms of ionising radiation include X-rays and charged particles. An example of an RT technique is Gamma Knife®, where a patient may be irradiated using a number of lower-intensity gamma rays that converge with higher intensity and high precision at a targeted region (e.g., a tumour). Another example of RT comprises use of a linear accelerator (“linac”), whereby a targeted region is irradiated by high-energy particles (e.g., electrons, high-energy photons, and the like). In another example, RT may be provided using a heavy charged particle accelerator (e.g., protons, carbon ions, and the like). The placement and dose of the radiation beam may be accurately controlled to provide a prescribed dose of radiation to the target region (e.g., the tumour) and to reduce damage to surrounding healthy tissue (known as organs at risk or OARs). An aspect of treatment planning concerns accurate localisation of physiological structures of patients. A treatment planning procedure may include using an image (two- or three-dimensional) of the patient to identify a target region and to identify critical organs near the target region. The target region (or area to be treated, e.g., a planned target volume, PTV), and surrounding region (or Organs at Risk, OARs) may be identified using contouring or segmentation. After contouring, a dose plan may be created for the patient indicating the desirable amount of radiation to be received by the target region and / or the surrounding region. The target region may have an irregular volume and may be distinctive in terms of its size, shape, and position. Contour delineation is an important process in treatment planning as it involves outlining tumours (or areas at risk of microscopic disease) as well as nearby organs at risk (OARs) to guide RT plans that seek to optimise tumour control and reduce radiation toxicity. However, variation in contour delineation among providers is common and can affect the resulting plan quality and patient outcomes. At present, manual contouring of OARs, primary tumour and involved nodal regions is perhaps the most laborious, but crucial, component of the clinical oncologist's role. Errors in this process can lead to underdosing of the tumour or increased toxicity, with consequential effects on survival and quality of life. A RT treatment plan (treatment plan, or simply plan) may be established using contours. An optimisation procedure may be based on clinical and dosimetric objectives and constraints. Examples of clinical and dosimetric objectives and constraints include maximum, minimum, and mean doses to target regions and surrounding regions (e.g., tumours and critical organs). Clinical and dosimetric objectives and constraints may be referred to as treatment planning objectives. Optimisation is usually carried out with respect to one or more treatment plan parameters to reduce beam-on time, improve dose uniformity, etc. In a practical example, multiple anatomical structures (target regions and / or surrounding regions) may be present. For example, in a head and neck treatment, there may be over 20 anatomical structures. For each structure, compliance with various treatment-planning objectives may be desired. Since most patients receive more than one fraction of radiation as part of a course of RT, and because the anatomy may change (deform) between these fractions, it is not straightforward to identify contours and determine doses to be delivered during the individual fractions so the physician can accurately gauge how the treatment is proceeding relative to the original intent as defined by the prescription. Contouring is one of the critical steps in the clinical workflow that determines the success of a RT and treatment plan generation. Traditionally, contouring is performed manually by experienced physicians or medical physicists. Recent development in artificial intelligence (Al) and pattern recognition algorithms has significantly helped to automate this process. For instance, Al-assisted auto-contouring has created a significant impact in the field of RT and treatment planning by replacing the tedious manual process of data curation. Ronneberger et al., proposed a U-net architecture in their seminal work, where the U-net-based model is, in general, used as the backbone for solving auto-contouring problems with Al. The Al assisted auto-contouring task is accomplished in two steps. A model is firstly trained in a supervised or a semi-supervised manner on a set of curated data. The trained model is then fed to an inference engine, which receives a set of un-curated patient image data and applies the pretrained model to produce a label map of each voxel (or pixel in 2D cases) within the data. The inference procedure assigns an organ-label to each voxel by minimising the misclassification error. Minimisation of the misclassification error is particularly suitable for use cases where the contouring is performed only for visual inspection by, for example, medical practitioners. In the field of RT, the results of contouring systems often act as inputs to treatment planning systems for further decision making. In such cases, supporting the planning system is potentially more important than visual accuracy of contours alone. Nevertheless, whether the contouring is performed manually or with the aid of a software tool, a manual step of review and approval is required to validate the contouring outcomes. In order to alleviate the subjective variance, the review process often comprises consensus of more than a single expert. Therefore, its tedious, expensive, and time consuming. Classical methods for the automatic quality assurances and error detection on contouring outcomes are primarily based on the statistical distribution of the geometric features (e.g., centroid, volume, and shape) and their relative position with the neighboring structures. Dong Joo Rhee et al. proposed Sorensen-Dice coefficients between the contours generated by two independent system (namely Al generated contours vs atlas-based contours) as a metric for detection of contouring failure. However, little to no effort has been observed by the inventors that considers the clinical constraints and dosimetric goals as a part of this process. It is therefore desirable to equip the medical practitioner with tools and techniques capable of reducing risks associated with poorly defined contours of physiological structures. Summary of the Invention The invention is defined in the independent claims. Further features are set out in the dependent claims. According to an aspect, there is a provided a computer-implemented method for radiotherapy contouring. The method includes a step of receiving one or more dosimetric constraints, such as a goal dose to be delivered for a physiological structure of a patient (and, if appropriate, structural details of the physiological structure, such as target regions and surrounding regions). The dosimetric constraint(s) may be a reference dose value or a defined dose value, which are reference objectives, representative of goals to be achieved by the RT process. The dose values may typically be provided in SI units of gray (Gy). The dose values may be mean dose values for the entire region in question. The dose values may be dose-based (e.g., indicating an absorbed dose to be delivered to the target region or surrounding region) or may be volume-based (e.g., indicating an absorbed dose to be delivered to a specified volume of the target region or surrounding region). The dose values may be input by a user or may be otherwise received by the computer implementing the method. The method then includes a step of determining a contour or image segment, which delineates (or identifies the border or boundary) of the physiological structure (or underlying structure thereof), and which classifies the physiological structure within voxels or pixels of image data of the patient. Determining a contour in this context may involve generation of a contour using image data or may involve acquisition of a pre-generated contour from external means. The contours may be manual contours (i.e., manually segmented by a user) or may be generated contours, or a combination thereof. The method then includes determining the probability that the contour misclassifies the physiological structure. The method includes a step of determining a misclassification cost for the physiological structure using the dosimetric constraint. The misclassification cost is a metric that quantifies the adverse effect of misclassification of the physiological structure. The method includes a step of calculating an expected cost metric (or confidence score) for the physiological structure using both the probability of misclassification of the physiological structure and the misclassification cost for the physiological structure. Contouring almost inevitably includes some uncertainty, regardless of contouring methodology, and these uncertainties are, in some cases, modelled or modellable. For example, deep-learning based segmentation techniques are able to provide probability distribution-based contours. Determining an expected cost metric based on misclassification probabilities and the cost of the misclassification, the techniques herein allow for physiological structure shapes to be optimised based on both probability distributions and the expected clinical impact. As a simple example, if a contoured or segmented voxel is considered to have a 49% probability of being spinal tissue and a 51% probability of being normal tissue, based on probability alone, this voxel would be classified as normal tissue. If the voxel is classified as normal tissue, a treatment plan optimisation procedure may seek to minimise exposure of this normal tissue to therapeutic radiation (or vice versa, depending on the objective of the treatment plan). However, it may well be preferable to classify the voxel as spinal tissue in spite of the lower probability of classification (or, equivalently, the higher probability misclassification), for instance in scenarios where exposure of normal tissue to therapeutic radiation is deemed to be an acceptable risk in the pursuit of treatment of spinal tissue. The selection makes the process less sensitive to the contouring misclassification. The techniques herein reduce the sensitivity of RT treatment planning to contouring misclassification, making treatments safer. The method may be used with any treatment planning system, including within an online adaptive context. The technique produces standard, easily-interpretable outputs, which are easy to comprehend and thus integrate well within common clinical practices. Optionally, the radiotherapy contouring method may be performed in respect of multiple physiological structures. That is, the method may include a step of receiving a plurality of dosimetric constraints for corresponding multiple physiological structures. The method may include steps of determining contours delineating and classifying the physiological structures; determining probabilities of misclassification of the physiological structures; determining misclassification costs for the physiological structures using the dosimetric constraints; and calculating expected cost metrics for the physiological structures using the probabilities of misclassification of the physiological structures and the misclassification costs for the physiological structures. Optionally, contours may be associated with multiple potential structures and the determined contour may have corresponding classification probabilities (or misclassification probabilities), where the classification probabilities correspond to a plurality of possible physiological structures (for instance, the contour may be associated with bone tissue and surrounding lung tissue). The misclassification cost for the physiological structure may then be determined for the plurality of possible physiological structures. The method may then include a step of outputting a final classification for the physiological structure based on a lowest expect cost metric. That is, the final classification for the physiological structure may be decided as the physiological structure from amongst the plurality of physiological structures that provides the lowest expected cost (i.e., an organ label on a least-cost criterion). Optionally, the dosimetric constraint may include a generic dosimetric constraint in addition to any treatment plan specific constraints. That is, in addition to treatment plan dosimetric constraints such as goal doses for specific physiological structures, the input may additionally include dosimetric constraints acquired from external libraries, such as libraries based on clinical studies. One example includes clinical study libraries, which set the limits of doses in different OARs (see Tables 1, S1 and S2 in “Dose-Volume Constraints fOr oRganS At risk In Radiotherapy (CORSAIR): An “All-in-One” Multicenter-Multidisciplinary Practical Summary”, Bisello, S. et al, which provide general dose-volume constraints for adult patients). Optionally, the RT contouring method may include a step of outputting the expected cost metric for the physiological structure via a graphical user interface (GUI). In this way, the method enables visualisation of potentially problematic regions, for further scrutiny by clinical professionals. The method may identify regions of particular concern, which may require more attention. For instance, the output may illuminate or otherwise emphasise physiological structures (or regions thereof) that are associated with a high probability of misclassification and / or associated with high misclassification cost (indicating that misclassification would be potentially clinically problematic). A GUI may be utilised to accept user input, for instance for the purposes of manual contouring. Optionally, determining the contour (or contours) may be achieved through acquisition of any or all of the following: a manual contour; an Al-based contour; a deformed structure; and a scripted structure. A deformed structure in this context refers to a structure for which the contours have been adapted from a previous computed tomography (CT) or magnetic resonance (MR) scan of the same patient using a deformation field, calculated by a deformable registration algorithm. A scripted structure in this context is automatically contoured using some non-AI algorithm, such as atlas-based contouring. Optionally, the RT contouring method may further include receiving volumetric image data (for instance, CT or MR data) of the patient. For example, some methods may convert received image data into a particular format, size, or resolution, suitable for contouring and defining target regions and surrounding regions. Determining the contour may then comprise passing the image data through a trained machine learning model, which is configured or trained to generate the contour. For example, the trained machine learning model may be (or include) a U-Net based convolutional neural network (CNN). Optionally, the method may be suitable and configured for RT contouring within an adaptive radiotherapy process. With online adaptive RT systems, for instance, contouring results are required to be fed to the planning system in a short timeframes with a little or no manual intervention. Optionally, the method may further include a step of preparing a treatment plan in accordance with the expected cost metric. For example, the final classification for the physical structures may be used within a treatment plan, which classification is determined using expected cost metrics. As an example, any optimisation procedure that accepts treatment plan parameters including a dosimetric constraints, such as reference dose value for a target region or a defined dose value for a surrounding region, which seeks to determine optimum (as far as practicable) parameter values may be suitable use in the development of treatment plans. For instance, the method may implement the optimisation procedure described in European patent application publication EP3681600A1. Optionally, the RT contouring method may include a step of outputting parameter values corresponding to the treatment plan. For instance, a set of optimisable parameter values may include a dose excess value, which is an amount of dosage violation that is considered acceptable. The optimisable parameters may include weights of beamlets. In a radiation beam may be divided into a number of beamlets where the contribution, at (hypothetical) unit fluence, may be determined; the weight of a beamlet is then a scaling factor, by which the unit fluence of the beamlet may be scaled to arrive at another fluence value. The optimisable parameters may include beam (or beamlet) angles, which may be the angle of a beam - relative to a reference point, such as a radiation head of a radiotherapy system, towards a target region. The optimisable parameters may include dose-histogram-volume information, which provides information related to the cumulative dose per volume fraction of target region or surrounding region. A “fraction” may be derived using a process of "fractioning," whereby a sequence of radiation therapy deliveries is provided over a predetermined period of time (e.g., 45 fractions), with each therapy delivery including a specified fraction of a total prescribed dose. The optimisable parameters may include a number of radiation beams. The optimisable parameters may include a dose per radiation beam. The optimisable parameters may include segment or control point shapes. The optimisable parameters may include segment or control point weights. These parameter values may correspond to instructions (or be included within an instruction set), to be executed by suitable RT hardware. In this way, the method may output parameter values (and, of course, configuration settings) for a radiotherapy system, so as to be able to deliver the radiation plan to the patient. Embodiments of another aspect include a data processing apparatus comprising a memory storing computer-readable instructions and a processor. The processor (or controller circuitry) is configured to execute the instructions to carry out the computer-implemented method for radiotherapy contouring. Embodiments of another aspect include a computer program comprising instructions, which, when executed by computer, causes the compute to execute the computer-implemented method for radiotherapy contouring. Embodiments of another aspect include a non-transitory computer-readable storage medium comprising instructions, which, when executed by a computer, causes the compute to execute the computer-implemented method for radiotherapy contouring. 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. Brief Description of the Drawings Reference is made, by way of example only, to the accompanying drawings in which: FIGURE 1 is a flow chart of a general method for radiotherapy contouring; FIGURE 2 is a schematic overview of a treatment planning workflow, in which contours determined in accordance with embodiments may be used; FIGURE 3 is a radiotherapy system, suitable for implementing a radiation treatment plan using contours determined according to embodiments; and FIGURE 4 is a radiotherapy device or apparatus, suitable for implementing a radiation treatment plan using contours determined according to embodiments. Detailed Description FIGURE 1 shows a general method for radiotherapy contouring. At step 102, the computer implementing the method receives a dosimetric constraint for a physiological structure of a patient. The dosimetric constraint may be, for instance, a dosage goal for a target region or a dosage limit for a surrounding region within the patient. At this point, in addition, the computer may also receive or access an Al model, eventually used for contouring, and volumetric images (e.g., CT or MR) of the patient. The dosimetric constraints may be defined for a treatment plan or they might be generated from a library, which is not specific to or tied to any individual treatment plan. At step 104, the computer determines a contour, which delineates the physiological structure. The contouring process may also classify the delineated physiological structure, e.g., assigning a label to identify the structure. This step may be performed using the above-mentioned Al contouring model, and may result in, in effect, a probability map for each class available in the input data. Alternatively, image data and contours may be input into an Al model configured to generate a probability map for all voxels, indicating the probability of each respective voxel being labelled correctly in the input data. At step 106, the computer determines a probability of misclassification of the physiological structure. Equivalently, this process may be seen as determination of a probability of classification of the physiological structure. This determination may be inherent within the preceding step 104 (that is, many Al contouring models by default output probabilistic contours). At step 108, the computer derives a misclassification cost for each structure (e.g., all structures identified as possible physiological structure candidates in the contouring process). This misclassification cost may be based on or derived from any clinical constraints and dosimetric goals of the treatment, which are known in advance for each oftheOARs. The misclassification cost indicates the clinical impact or cost of misclassification. The cost function brings the adverse effect of misclassification into consideration. At step 110, the computer calculates an expected cost metric for the physiological structure using the probability of misclassification of the physiological structure and the misclassification cost for the physiological structure. In this way, the misclassification probability map will be combined with misclassification cost to set a confidence score on the input structure set labels. The expected cost metric is a combined metric, which acts as an indicator to identify the voxels or regions that may require higher attention than others in a contour review process. While Al assisted treatment planning systems have created a significant impact in the field of RT and treatment planning by replacing the tedious manual process of data curation, the inventors have not observed any significant efforts that involves Al to speed up the review process of the contouring outcomes. To this end, the techniques herein speed up this process by incorporating consideration of the clinical impact of misclassification. The method of RT contouring simplifies any review process by combining two different aspects together: a probability contouring map; and the clinical impact or cost of any misclassification. The technique may be used to decide on organ or physiological structural labels based on the expected cost metric. Treatment plan optimisation systems need then only accept standard inputs (the structure sets) and do not need to be modified to further consider any contouring probability distribution. As a worked example, consider a scenario in which we have a voxel (V) that we want to classify into one of three available classes (physiological structures): “spine”; “bone”; and “muscle”. Using, for example, an Al contouring model, we may assign a certain probability for V for each of these available classes. A trivial probability assignment would be as V being spine: 0.4; V being muscle: 0.45; and V being bone: 0.05. This allocation of probabilities to the voxel (or indeed all voxels within a region of interest from image data) may be referred to as a probability map. The misclassification cost is the cost of incorrect labelling (e.g., the clinical cost of labelling V as “muscle” if it was in reality “spine” or “bone”), which is derived from the clinical constraints and dosimetric goals. TABLE 1 below exemplifies such a misclassification cost. The columns represent actual, ground-truth class labels and the rows represent assigned class labels. In this example, assigning “spine” to (ground truth) “spine” incurs no cost. However, assigning “spine” to “muscle” will include a cost of 50 arbitrary units. TABLE 1: Misclassification cost functions: Actual label Spine Muscle Bone Assigned label Spine 0 50 90 Muscle 20 0 20 Bone 20 30 0 Of course, the above allocation of probabilities and misclassification costs for the voxel V may be repeated for all voxels in the region of interest in the image data. In one example, the probabilities of classification (that is, 1 - the probability of misclassification) may be combined with the misclassification costs to determine an expected cost metric (or set thereof) for the voxel V as follows: Cost of assigning V as Spine: 0 x 0.4 + 20 x 0.45 + 20 x 0.05 = 0 + 9 + 1 = 10 Cost of assigning V as Muscle: 50 x 0.4 + 0 x 0.45 + 30 x 0.05 = 20 + 0 + 1.5 = 21.5 Cost of assigning V as Bone: 90 x 0.4 + 20 x 0.45 + 0 x 0.05 = 36 + 9 + 0 = 45 In this case, the voxel will be assigned the label “spine” as the expected cost metric for assignment of spine is the minimum. Without consideration of the misclassification cost, prior art techniques would assign the voxel the label “muscle” as - according to contouring - this is the most probable label. Mathematically, this example method of classification for the physiological structure, with organ label Lv for voxel V, may be expressed as follows: n Lv= arg min i e {i...n} Here, P is the probability vector (comprising classification / misclassification probabilities) for j labels, and M is the cost function matrix (comprising classification / misclassification costs) for j assigned labels and i actual labels. Of course, alternative mathematical functions for Lv classification may be implemented, including arg max functions where a high expected cost metric is desired. The above-described embodiment may be used to, in effect, generate confidence maps on contours for RT treatment plan with misclassification probability. Alternatively, embodiments may be utilised to generate regions of uncertainty in structure set RT treatment plans with misclassification probability. An Al model may be trained to identify areas of potential errors or inaccuracies within preexisting contours. The input data of such as system is a structure set, which may have been produced by manual contouring, deep learning based Al contouring, deformed structures, scripted structures, or any other method of structure generation. Example manual contouring errors include missing a slice in the spine, a deformation error may include misplaced bone near lungs (large lung deformation often causes deformation error), and scripting structuring errors include structures that might be unnaturally small or missing. In adaptive treatment, multiple methods are often used to generate all required contours or structures. However, there is limited time to evaluate all of the structures. This embodiment allows clinicians to review all structures generated by different methods at the same time. The implementing system receives the trained Al model and the contours (structures as input (without any volumetric medical image data) and dosimetric constraints or objectives associated with each structure. The system then derives a misclassification cost for each structure based on the clinical constraints including dosimetric goals. The system then feeds the contours and optionally any image data to the Al model so as to generate a probability map for each voxel to indicate the respective voxel being labelled correctly in the input data. The probability map may then be combined with the misclassification cost derived above, in order to set the expected cost metric or confidence score for each of the input structure set labels. The expected cost metric can then be used to accelerate the review process by indicating the regions or voxels of structures that require higher attention than others. In embodiments, Al-based techniques may be used to determine contours from patent image data. One example model suitable for image segmentation is an adaptation of that described in the work of Yang, G. et al. (Autosegmentation for thoracic radiation treatment planning: A grand challenge at AAPM 2017; Med. Phys. 45 (10), October 2018). This approach uses a deep-convolutional neural network (DCNN) for thoracic CT image segmentation. The DCNN model was modified from the U-Net architecture (see the work of Ronneberger, O. et al.) with 27 convolutional layers in total and with the sequential convolutional layers at each resolution level being combined into a residual block. To improve computation efficiency, two models are trained and applied in sequence. A fast 2.5D model with an input size of 5*360*360 voxels is trained to segment the lungs, the results of which were also used to automatically define a bounding box for the other structures. A 3D model with an input size of 32*192*192 voxels is trained and applied within the smaller ROI to get the final segmentation of the heart, the oesophagus, and the spinal cord. The models are implemented using the PyTorch package and trained from scratch using 36 training datasets provided for the Annual Meeting of American Association of Physicists in Medicine. These training datasets comprise sixty thoracic CT scans, provided by three different institutions. Training used an ADAM optimiser, with a learning rate of 0.002 (see the work of Kingma, D. P &Ba, J. L arXiv:1412.6980v9). Using a Titan X 12GB GPU, training from scratch over 3 days provided a hierarchical segmentation model, where the lung is used to constrain the locations of other structures. When trained, the model is capable of inference and testing using the same hardware in 30 seconds. The skilled reader will appreciate that other training datasets with a focus on other physiological structures may be used, other underlying model architectures may be used, and other hardware may be used. The resultant expected cost metric for the physiological structure (calculated using the probabilities of misclassification and the misclassification cost) may be used as an input into an adaptive RT procedure or adaptive RT optimisation procedure. As an example, the resultant expected cost metric may be used with Elekta’s Unity package. Elekta Al-assisted autocontouring has been shown to reduce online prostate contouring time to under two minutes, which reduces previous Unity prostate workflows by around 10 minutes. Such time savings help to further enhance treatment accuracy by lowering the risk of anatomical movement during session and improve patient comfort by reducing time on the treatment table. In online-adaptive case, the contouring results are required to be fed to the planning system in a rapid phase with a little or no manual intervention. Therefore, the robustness of the autocontouring system has a very high importance. The techniques herein generate contours by minimizing the cost of misclassification rather than, e.g., minimizing the misclassification error. An example clinical MR workflow, which incorporates online adaptive RT may proceed as illustrated in FIGURE 2. At 202, MR eligibility of the patient is verified and recorded at the time of consultation, at simulation (if using MR), and at every treatment session on the MR apparatus. At 204, the simulation process is supported by replicating the MR apparatus. MR-compatible patient positioning devices are provided. Indexing positions are recorded. A reference plan is generated (CT and / or MR reference data can be used) and provides a starting point for the adaptive workflow. Electron densities for each structure are predefined at this stage to permit dose calculation on the daily MR images. Additional defaults can be set to streamline the online adaptive process. At 206, image acquisition is performed, based on sequences typically predetermined in the patient’s plan of care, without the need for replicating the patient’s setup position. The MR images provide a wealth of information. At 208, as soon as 3D image acquisition has finished, live 2D motion monitoring images can be acquired continuously in up to three planes at any point in the adaptive and treatment delivery workflow. The anatomical structure(s) of choice is displayed on these images, allowing the user to assess shifts in anatomy and whether intervention is necessary. At 210, pretreatment 3D images are automatically registered to the reference image. Depending on the variations in anatomy that are visible and the clinical indication, the user can select the adaptive path to take. At 212, the treatment plan is adapted. Different clinical cases have different objectives, depending on the dose being delivered and whether anatomy of interest is subject to deformations. Consequently, Elekta Unity supports two adaptive workflows: “adapt to position”, where the reference dose is shifted to the daily target position, which is an efficient workflow in terms of time and expertise required in the online environment; “adapt to shape”, where the dose is adjusted to conform to the daily deformed anatomical structures, which is more resource-intense and will improve conformity when high doses per fraction are being delivered. Contours selected in accordance with the calculated expected cost metric according to embodiments may be used within this “adapt to shape” adaptive workflow, where the dose is adjusted to conform to the selected contours. During plan adaptation, advanced imaging protocols can be applied to collect real-time anatomical or biological information regarding treatment effects on the target region and surrounding tissue. At 214, once the user is satisfied that the adapted plan meets the specified criteria, the plan can be approved for delivery. Additional 3D MR images may be acquired for verification of patient position or for offline analysis. The reader is directed to the white paper “Elekta Unity for Magnetic Resonance Radiation Therapy (MR / RT)” from Brown, K. L., et al. for a further description and discussion of the Elekta Unity system. FIGURE 3 is a block diagram of an implementation of a radiotherapy system 300, suitable for executing methods for radiation treatment planning using contours determined according to embodiments. The example radiotherapy system 300 comprises a computing system 310 within which a set of instructions, for causing the computing system 310 to perform any one or more of the methods (or steps thereof) discussed herein, may be executed. The computing system 310 may implement a contouring system. The computing system 310 may also be referred to as a computer. In particular, the methods described herein may be implemented by a processor or controller circuitry 311 of the computing system 310. The computing system 310 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 any one 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 310 includes controller circuitry 311 and a memory 313 (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 313 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). Controller circuitry 311 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 311 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 311 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 311 is configured to execute the processing logic for performing the operations and steps discussed herein. The computing system 310 may further include a network interface circuitry 315. The computing system 310 may be communicatively coupled to an input device 320 and / or an output device 330, via input / output circuitry 316. In some implementations, the input device 320 and / or the output device 330 may be elements of the computing system 310. The input device 320 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 330 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 320 and the output device 330 may be provided as a single device, or as separate devices. In some implementations, the computing system 310 may comprise image processing circuitry 314. Image processing circuitry 314 may be configured to process image data 380 (e.g., images, or imaging data), such as medical images obtained from one or more imaging data sources, a treatment device 330 and / or an image acquisition device 340 . Image processing circuitry 314 may be configured to process, or pre-process, image data 380. For example, image processing circuitry 314 may convert received image data into a particular format, size, resolution or the like. In some implementations, image processing circuitry 314 may be combined with controller circuitry 311. In some implementations, the radiotherapy system 300 may further comprise an image acquisition device 340 and / or a treatment device 330. The image acquisition device 340 and the treatment device 330 may be provided as a single device. In some implementations, treatment device 330 is configured to perform imaging, for example in addition to providing treatment and / or during treatment. The treatment device 330 comprises the main radiation delivery components of the radiotherapy system. Image acquisition device 340 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 340 may be configured to output image data 380, which may be accessed by computing system 310. Treatment device 330 may be configured to output treatment data 360, which may be accessed by computing system 310. Computing system 310 may be configured to access or obtain treatment data 360, planning data 370 and / or image data 380. Treatment data 360 may be obtained from an internal data source (e.g., from memory 313) or from an external data source, such as treatment device 330 or an external database. Planning data 370 may be obtained from memory 313 and / or from an external source, such as a planning database. Planning data 370 may comprise information obtained from one or more of the image acquisition device 340 and the treatment device 330. 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 FIGURE 1 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 313 and / or within the controller circuitry 311 during execution thereof by the computing system 310, the memory 313 and the controller circuitry 311 also constituting computer-readable storage media. In an implementation, the modules, components and other features described herein may be implemented as discrete components or integrated in the functionality of hardware components such as ASICS, FPGAs, DSPs or similar devices. 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). FIGURE 4 depicts a radiotherapy apparatus, suitable for implementing radiation treatment plans utilising contours determined according to embodiments. The cross-section through radiotherapy apparatus 400 includes a radiation head 410 and a beam receiving apparatus 402, both of which are attached to a gantry 404. The radiation head 410 includes a radiation source 412, which emits a beam of radiation 406. The radiation head 410 also includes a beam shaping apparatus 418, which controls the size and shape of the radiation field associated with the beam. The beam receiving apparatus 402 is configured to receive radiation emitted from the radiation head 410, for the purpose of absorbing and / or measuring the beam of radiation. In the view shown, the radiation head 410 and the beam receiving apparatus 402 are positioned diametrically opposed to one another. The gantry 404 is rotatable, and supports the radiation head 410 and the beam receiving apparatus 402 such that they are rotatable around an axis of rotation 408, which may coincide with the patient longitudinal axis. The gantry provides rotation of the radiation head 410 and the beam receiving apparatus 402 in a plane perpendicular to the patient longitudinal axis (e.g., a sagittal plane). Three gantry directions Xg, Yg, Zg may be defined, where the Yg direction is perpendicular with gantry axis of rotation. The ZG direction extends from a point on the gantry corresponding to the radiation head, towards the axis of rotation of the gantry. Therefore, from the patient frame of reference, the Zg direction rotates around as the gantry rotates. Radiotherapy apparatus 400 also includes a support surface 420 on which a subject (or patient) is supported during radiotherapy treatment. The radiation head 410 is configured to rotate around the axis of rotation 408 such that the radiation head 410 directs radiation towards the subject from various angles around the subject in order to spread out the radiation dose received by healthy tissue to a larger region of healthy tissue while building up a prescribed dose of radiation at a target region. The radiotherapy apparatus 400 is configured to deliver a radiation beam towards a radiation isocentre, which is substantially located on the axis of rotation 408 at the centre of the gantry 404 regardless of the angle at which the radiation head 410 is placed. The rotatable gantry 404 and radiation head 410 are dimensioned so as to allow a central bore 422 to exist. The central bore 422 provides an opening, sufficient to allow a subject to be positioned therethrough without the possibility of being incidentally contacted by the radiation head 410 or other mechanical components as the gantry rotates the radiation head 410 about the subject. The radiation head 410 emits the radiation beam 406 along a beam axis 424 (or radiation axis or beam path), where the beam axis 424 is used to define the direction in which the radiation is emitted by the radiation head. The radiation beam 406 is incident on the beam receiving apparatus 402, which may include at least one of a beam stopper and a radiation detector. The beam receiving apparatus 402 is attached to the gantry 404 on a diametrically opposite side to the radiation head 410 to attenuate and / or detect a beam of radiation after the beam has passed through the subject. The radiation beam axis 424 may be defined as, for example, a centre of the radiation beam 406 or a point of maximum intensity. The beam shaping apparatus 418 delimits the spread of the radiation beam 406. The beam shaping apparatus 418 is configured to adjust the shape and / or size of a field of radiation produced by the radiation source. The beam shaping apparatus 418 does this by defining an aperture (also referred to as a window or an opening) of variable shape to collimate the radiation beam 406 to a chosen cross-sectional shape. In this example, the beam shaping apparatus 418 may be provided by a combination of a diaphragm and an MLC. Beam shaping apparatus 418 may also be referred to as a beam modifier. The radiotherapy apparatus 400 may be configured to deliver both coplanar and non-coplanar (also referred to as tilted) modes of radiotherapy treatment. In coplanar treatment, radiation is emitted in a plane which is perpendicular to the axis of rotation of the radiation head 410. In non-coplanar treatment, radiation is emitted at an angle which is not perpendicular to the axis of rotation. In order to deliver coplanar and non-coplanar treatment, the radiation head 410 may move between at least two positions, one in which the radiation is emitted in a plane which is perpendicular to the axis of rotation (coplanar configuration) and one in which radiation is emitted in a plane which is not perpendicular to the axis of rotation (non-coplanar configuration). In the coplanar configuration, the radiation head is positioned to rotate about a rotation axis and in a first plane. In the non-coplanar configuration, the radiation head is tilted with respect to the first plane such that a field of radiation produced by the radiation head is directed at an oblique angle relative to the first plane and the rotation axis. In the non-coplanar configuration, the radiation head is positioned to rotate in a respective second plane parallel to and displaced from the first plane. The radiation beam is emitted at an oblique angle with respect to the second plane, and therefore as the radiation head rotates the beam sweeps out a cone shape. The beam receiving apparatus 402 remains in the same place relative to the rotatable gantry when the radiotherapy apparatus is in both the coplanar and non-coplanar modes. Therefore, the beam receiving apparatus 402 is configured to rotate about the rotation axis in the same plane in both coplanar and non-coplanar modes. This may be the same plane as the plane in which the radiation head rotates. The beam shaping apparatus 410 is configured to reduce the spread of the field of radiation in the non-coplanar configuration in comparison to the coplanar configuration. The radiotherapy apparatus 400 includes a controller 430, which is programmed to control the radiation source 412, beam receiving apparatus 402 and the gantry 404. Controller 430 may perform functions or operations such as treatment planning, treatment execution, image acquisition, image processing, motion tracking, motion management, and / or other tasks involved in a radiotherapy process. Controller 430 is programmed to control features of apparatus 400 according to a radiotherapy treatment plan for irradiating a target region, also referred to as a target tissue, of a patient. The treatment plan includes information about a particular dose to be applied to a target tissue, as well as other parameters such as beam angles, dose-histogram-volume information, the number of radiation beams to be used during therapy, the dose per beam, and the like. Controller 430 is programmed to control various components of apparatus 400, such as gantry 404, radiation head 410, beam receiving apparatus 402, and support surface 420, according to the treatment plan. The treatment plan may be generated using contours determined according to embodiments. Hardware components of controller 430 may include one or more computers (e.g., general purpose computers, workstations, servers, terminals, portable / mobile devices, etc.); processors (e.g., central processing units (CPUs), graphics processing units (GPUs), microprocessors, digital signal processors (DSPs), field programmable gate arrays (FPGAs), special-purpose or specially-designed processors, etc.); memory / storage devices such as a memory (e.g., read-only memories (ROMs), random access memories (RAMs), flash memories, hard drives, optical disks, solid-state drives (SSDs), etc.); input devices (e.g., keyboards, mice, touch screens, mics, buttons, knobs, trackballs, levers, handles, joysticks, etc.); output devices (e.g., displays, printers, speakers, vibration devices, etc.); circuitries; printed circuit boards (PCBs); or other suitable hardware. Software components of controller 430 may include operation device software, application software, etc. The radiation head 410 may be connected to a head actuator 414, which is configured to actuate the radiation head 410, for example between a coplanar configuration and one or more non-coplanar configurations. This may involve translation and rotation of the radiation head 410 relative to the gantry. In some implementations, the head actuator may include a curved rail along which the radiation head 410 may be moved to adjust the position and angle of the radiation head 410. The controller 430 may control the configuration of the radiation head 430 via the head actuator 414. The beam shaping apparatus 418 includes a shaping actuator 416. The shaping actuator is configured to control the position of one or more elements in the beam shaping apparatus 418 in order to shape the radiation beam 406. In some implementations, the beam shaping apparatus 418 includes an MLC, and the shaping actuator 416 includes means for actuating leaves of the MLC. The beam shaping apparatus 418 may further comprise a diaphragm, and the shaping actuator 416 may include means for actuating blocks of the diaphragm. The controller 430 may control the beam shaping apparatus 418 via the shaping actuator 416. A treatment plan may comprise positioning information of beam shaping apparatus 418. The positioning information of beam shaping apparatus 418 may comprise information indicating a configuration of one or more elements of beam shaping apparatus 418, such as leaf configuration of an MLC of beam shaping apparatus 418, a configuration of a diaphragm of beam shaping apparatus 418, a configuration of an opening (e.g., window or aperture) of the MLC, and / or the like. Unless specifically stated otherwise, as apparent from the following discussion, it is appreciated that throughout the description, discussions utilizing terms such as “receiving”, “determining”, “comparing”, “enabling”, “maintaining,” “identifying”, “obtaining”, “accessing”, 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, substitutions and changes in the form of methods and apparatus described herein may be made.

Claims

1. A computer-implemented method for radiotherapy contouring, the method comprising: receiving a dosimetric constraint for a physiological structure of a patient;determining a contour delineating and classifying the physiological structure;determining a probability of misclassification of the physiological structure;determining a misclassification cost for the physiological structure using the dosimetric constraint; andcalculating an expected cost metric for the physiological structure using the probability of misclassification of the physiological structure and the misclassification cost for the physiological structure.

2. The method according to claim 1, comprising receiving dosimetric constraints for multiple physiological structures.

3. The method according to any preceding claim, wherein:the contour has corresponding classification probabilities, the classification probabilities corresponding to a plurality of possible physiological structures;the misclassification cost for the physiological structure is determined for the plurality of possible physiological structures; andthe method further comprises outputting a final classification for the physiological structure based on a lowest expect cost metric.

4. The method according to any preceding claim, wherein the dosimetric constraint includes a generic dosimetric constraint.

5. The method according to any preceding claim, further comprising outputting the expected cost metric for the physiological structure via a graphical user interface.

6. The method according to any preceding claim, wherein determining the contour comprises acquiring any or all of the following: a manual contour; an Al-based contour; a deformed structure; and a scripted structure.

7. The method according to any preceding claim, further comprising receiving image data of the patient, and wherein determining the contour comprises passing the image data through a trained machine learning model configured to generate the contour.

8. The method according to claim 7, wherein the trained machine learning model is a Linet based convolutional neural network.

9. The method according to any preceding claim, wherein the method is for radiotherapy contouring in an adaptive radiotherapy process.

10. The method according to any preceding claim, further comprising preparing a treatment plan in accordance with the expected cost metric.

11. The method of claim 10, further comprising outputting parameter values corresponding to the treatment plan, wherein the parameter values comprise one or more of: number of beams, beam angles, a dose per beam, beamlet weights, segment or control point shapes, segment or control point weights, dose-volume histogram information, and a dose excess value.

12. A data processing apparatus comprising a memory storing computer-executable instructions and a processor configured to execute the instructions to:receive a dosimetric constraint for a physiological structure of a patient;determine a contour delineating and classifying the physiological structure;determine a probability of misclassification of the physiological structure;determine a misclassification cost for the physiological structure using the dosimetric constraint; andcalculate an expected cost metric for the physiological structure using the probability of misclassification of the physiological structure and the misclassification cost for the physiological structure.

13. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to:receive a dosimetric constraint for a physiological structure of a patient;determine a contour delineating and classifying the physiological structure;determine a probability of misclassification of the physiological structure;determine a misclassification cost for the physiological structure using the dosimetric constraint; andcalculate an expected cost metric for the physiological structure using the probability of misclassification of the physiological structure and the misclassification cost for the physiological structure.

14. A non-transitory computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to:receive a dosimetric constraint for a physiological structure of a patient;determine a contour delineating and classifying the physiological structure;determine a probability of misclassification of the physiological structure;determine a misclassification cost for the physiological structure using the dosimetric constraint; andcalculate an expected cost metric for the physiological structure using the probability of misclassification of the physiological structure and the misclassification cost for the physiological structure.26

Images