Machine learning optimization of dose maps for radiotherapy

Abstract

Systems and methods for generating fluence maps for radiation therapy treatment plans predicted using machine learning are disclosed. The systems and methods include identifying image data indicative of a target dose region in an anatomical structure of a subject and treatment constraints for organ at risk regions; generating anatomical projection images representing views of the anatomical structure according to respective beam angles; using a trained neural network model to generate a computer simulated fluence map representation based on the anatomical projection images, wherein the fluence map is indicative of a fluence distribution of a radiation therapy treatment at each of the beam angles.

Description

[0001] Priority requirements

[0002] This application claims the benefit of priority to U.S. Application Serial No. 16 / 948,486, filed on September 21, 2020, which is incorporated herein by reference in its entirety. Technical Field

[0003] Implementations of this disclosure generally relate to determining planning parameters for radiotherapy performed by a radiotherapy treatment system. Specifically, this disclosure relates to using machine learning techniques to determine the fluence map used in the treatment planning of a radiotherapy system. Background Technology

[0004] Radiation therapy (or "radiotherapy") can be used to treat cancer or other diseases in mammalian tissues (e.g., humans and animals). One such radiation therapy technique uses a gamma knife, where the patient is irradiated with a large amount of low-intensity gamma rays focused precisely on the target (e.g., the tumor) at high intensity. Another such radiation therapy technique uses a linear accelerator (linac), which irradiates the tumor with high-energy particles (e.g., electrons, protons, ions, high-energy photons, etc.). The arrangement and dosage of the radiation beam must be precisely controlled to ensure that the tumor receives the prescribed radiation, and the beam arrangement should minimize damage to the surrounding healthy tissue, often referred to as the organ at risk (OAR). The radiation is called "prescribed" because the physician delivers a predefined amount of radiation to the tumor and surrounding organs, similar to a prescription for a drug. Typically, ionizing radiation, in the form of a collimated beam, is directed from an external radiation source at the patient.

[0005] Specified or selectable beam energies can be used, such as those for delivering diagnostic or therapeutic energy levels. Modulation of the beam can be provided by one or more attenuators or collimators (e.g., multi-leaf collimators (MLCs)). The intensity and shape of the beam can be adjusted by collimation to avoid damaging adjacent healthy tissue by aligning the projected beam with the contours of the target tissue (e.g., OAR).

[0006] Treatment planning can involve using three-dimensional (3D) images of the patient to identify target regions (e.g., tumors) and key organs near the tumor. Creating a treatment plan can be a time-consuming process in which planners attempt to adhere to various treatment goals or constraints (e.g., dose-volume histograms (DVH), overlap-volume histograms (OVH)) and consider their respective importance (e.g., weights) to produce a clinically acceptable treatment plan. This task can be a time-consuming trial-and-error process, complicated by the variety of organ-area-associated (OARs), as the complexity increases with the number of OARs (e.g., a dozen or more for head and neck treatments). OARs located far from the tumor may be easily protected from radiation, while those close to or overlapping with the target tumor may be more vulnerable.

[0007] Traditionally, an initial treatment plan can be generated “offline” for each patient. Treatment plans can be well-developed before radiotherapy delivery, for example, using one or more medical imaging techniques. Imaging information can include images, for example, from X-rays, computed tomography (CT), magnetic resonance imaging (MR), positron emission tomography (PET), single-photon emission computed tomography (SPECT), or ultrasound. Healthcare providers (e.g., physicians) can use 3D imaging information indicating the patient's anatomy to identify one or more target tumors and adjacent organs of interest (OARs). Healthcare providers can use manual techniques to delineate target tumors to receive a prescribed radiation dose, and similarly, they can delineate nearby tissues, such as organs, at risk of damage from radiotherapy. Alternatively or additionally, automated tools (e.g., ABAS provided by Elekta AB of Sweden) can be used to aid in the identification or delineation of target tumors and organs at risk. Numerical optimization techniques can then be used to create a radiotherapy treatment plan (“treatment plan”) that minimizes an objective function that includes clinical and dosimetric goals and constraints (e.g., maximum, minimum, and partial radiation doses to a portion of the tumor volume (“95% target should receive not less than 100% of the prescribed dose”), and similar measures for critical organs). The optimized plan includes numerical parameters specifying the direction, cross-sectional shape, and intensity of each radiation beam.

[0008] The treatment plan can then be executed by positioning the patient in the treatment machine and delivering prescribed radiotherapy guided by optimized planning parameters. The radiotherapy treatment plan may include dose “gradations” to provide a sequence of radiotherapy over predetermined time periods (e.g., 30 to 45 fractions per day), where each treatment comprises a designated fraction of the total prescribed dose.

[0009] As part of the treatment planning process for radiotherapy dosage, the fluence is determined and evaluated. Fluence is the density of radioactive photons or particles perpendicular to the beam direction, while dose is related to the energy released in the material when the photons or particles interact with the atoms of the material. Therefore, dose depends on both the fluence and the physics of radio-matter interactions. Important planning is carried out as part of determining the treatment plan and the fluence and dose for a specific patient. Summary of the Invention

[0010] In some embodiments, methods, systems, and computer-readable media are provided for generating optimized fluence maps or sets of fluence maps to be used as part of one or more radiotherapy treatment plans. The methods, systems, and computer-readable media can be configured to perform operations including: acquiring image data corresponding to a subject receiving radiotherapy treatment, the image data indicating one or more target dose regions and one or more organ-at-risk regions in the subject's anatomy; generating anatomical projection images from the image data, each anatomical projection image providing a view of the subject according to a corresponding beam angle of the radiotherapy treatment; and using a trained neural network model to generate estimated fluence maps based on the anatomical projection images, each of the estimated fluence maps indicating the fluence distribution of the radiotherapy treatment at the corresponding beam angle. In these and other configurations, such neural network models can be trained with corresponding pairs of anatomical projection images and fluence maps to produce estimated fluence maps.

[0011] In some implementations, each of the estimated fluence maps is a two-dimensional array of unit beam weights perpendicular to the corresponding beam direction, and the beam angle of the radiotherapy treatment corresponds to the gantry angle of the radiotherapy machine. Furthermore, obtaining a set of three-dimensional image data corresponding to the subject may include obtaining and projecting image data for each gantry angle of the radiotherapy machine, such that each generated anatomical projection image represents a view of the subject's anatomy according to a given gantry angle used to provide treatment with a given radiotherapy beam.

[0012] In some implementations, the generated estimated fluence maps are used during operation to calculate and optimize the radiation dose in a radiotherapy treatment plan, such as for radiotherapy treatments providing volumetric intensity-modulated rotational therapy (VMAT) performed by a radiotherapy machine, where multiple radiotherapy beams are shaped to achieve a modulated dose for a target region from multiple beam angles, thereby delivering a prescribed radiation dose. For example, the workflow for radiotherapy planning may include: using a neural network model to generate a set of estimated fluence maps; performing numerical optimization using the estimated fluence maps as input to optimization, where optimization incorporates radiotherapy treatment constraints; and generating a Pareto-optimal fluence plan to be used in the radiotherapy treatment plan for the subject. Such a Pareto-optimal fluence plan can be used to generate a set of initial control points corresponding to each of the multiple radiotherapy beams using arc sorting, and then performing direct aperture optimization to generate a set of final control points corresponding to each of the multiple radiotherapy beams. In addition, this set of final control points can be used to perform radiotherapy treatment because it controls the position of the multi-leaf collimator (MLC) blades of the radiotherapy treatment machine at a given gantry angle corresponding to a given beam angle.

[0013] Other aspects of generating, recognizing, and optimizing fluence maps are disclosed, including the use of specific neural network training arrangements. For example, in a test or validation setting, fluence maps generated from a neural network model in response to an input set of anatomical structure projection images can be compared with fluence maps generated from another source. Specific model training aspects, including Generative Adversarial Networks (GANs), Conditional Generative Adversarial Networks (cGANs), and Cycle-Consistent Generative Adversarial Networks (CycleGANs), are also disclosed.

[0014] The above overview is intended to provide an overview of the subject matter of this patent application. It is not intended to provide an exclusive or exhaustive explanation of the subject matter of the invention. Detailed descriptions are included to provide further information regarding this patent application. Attached Figure Description

[0015] In the accompanying drawings, which are not necessarily drawn to scale, similar reference numerals describe substantially similar parts throughout several views. Similar reference numerals with different letter suffixes indicate different instances of substantially similar parts. The accompanying drawings illustrate, by way of example rather than limitation, the various embodiments discussed in this document.

[0016] Figure 1 An exemplary radiotherapy system based on some examples is shown.

[0017] Figure 2A and Figure 2B The diagram shows elliptical projections and exemplary prostate target anatomy structures, based on some examples.

[0018] Figure 3A An exemplary radiotherapy system, according to some examples, may include a radiotherapy output component configured to provide a therapeutic beam.

[0019] Figure 3B Exemplary systems, including combined radiotherapy systems and imaging systems such as cone-beam computed tomography (CBCT) imaging systems, are shown according to some examples.

[0020] Figure 4 A partial cross-sectional view is shown of an exemplary system, including a combination of a radiotherapy system and an imaging system such as a magnetic resonance (MR) imaging (MRI) system, according to some examples.

[0021] Figure 5 An exemplary gamma knife radiotherapy system is shown according to some examples.

[0022] Figure 6A and Figure 6B The differences between exemplary MRI images and their corresponding CT images, based on some examples, are described respectively.

[0023] Figure 7 Exemplary collimator configurations for shaping, guiding, or modulating the intensity of a radiotherapy beam are shown, based on some examples.

[0024] Figure 8 The data flow and process for radiotherapy planning are shown based on some examples.

[0025] Figure 9 An example of optimizing operations based on some examples of influx graphs is shown.

[0026] Figure 10 Examples of anatomical projection images and fluence maps at specific angles of a radiotherapy beam are shown, based on some examples.

[0027] Figure 11 Examples of anatomical projections and radiotherapy constraints at multiple angles during radiotherapy treatment are shown, based on several examples.

[0028] Figure 12 Examples of fluence map projections at multiple angles during radiotherapy treatment are shown, based on some examples.

[0029] Figure 13 Paired anatomical projections, radiotherapy constraints, and fluence map projections at multiple angles in radiotherapy treatment are shown, based on some examples.

[0030] Figure 14The deep learning process for training a model to predict flux maps from projected image data and flux map data is illustrated using some examples.

[0031] Figure 15A and Figure 15B Schematic diagrams of generative deep convolutional neural networks and discriminative deep convolutional neural networks used in predicting injection map representations, based on some examples, are depicted respectively.

[0032] Figure 16A and Figure 16B Schematic diagrams of generative adversarial networks and recurrent consistency generative adversarial networks, respectively, are depicted based on some examples for training generative models for predicting injection graph representations.

[0033] Figure 17 and Figure 18 The diagram illustrates the corresponding data flow for training and using machine learning models suitable for generating simulated injection graphs, based on some examples.

[0034] Figure 19 The methods shown are for generating fluence maps for use in radiotherapy treatment planning and for generating machine parameters for delivering radiotherapy treatment plans, based on some examples.

[0035] Figure 20 An exemplary block diagram is shown of a machine that can implement one or more of the methods discussed herein. Detailed Implementation

[0036] In the following detailed description, reference is made to the accompanying drawings, which form part of the detailed description, and specific embodiments are illustrated by way of examples that allow the practice of this disclosure. These embodiments, also referred to herein as “examples,” are described in sufficient detail to enable those skilled in the art to practice this disclosure, and it should be understood that embodiments may be combined or other embodiments may be utilized and structural, logical, and electrical changes may be made without departing from the scope of this disclosure. Therefore, the following detailed description should not be considered limiting, and the scope of this disclosure is defined by the appended claims and their equivalents.

[0037] Intensity-modulated radiotherapy (IMRT) and volumetric modulated arctherapy (VMAT) have become the standard of care in modern cancer radiotherapy. Creating an individualized IMRT or VMAT treatment plan for a patient is typically an iterative process that weighs the target dose against the risk of adverse effects from oral radiation therapy (OAR) and adjusts plan constraints whose impact on plan quality metrics and dose distribution can be difficult to predict. In fact, the very order in which plan constraints are adjusted can lead to dose variability. The quality of a treatment plan often depends on the planner's subjective judgment, which is influenced by his / her experience and skill. Even the most skilled planner cannot guarantee that their plan will be close to the best possible outcome, or whether a little or a lot of effort will produce a significantly better plan.

[0038] This disclosure includes various techniques for improving and enhancing radiotherapy treatment by generating fluence map values ​​as part of a model-driven fluence map optimization (FMO) process during radiotherapy planning. The model may include a trained machine learning model, such as an artificial neural network model, trained to produce (predict) fluence map values ​​from a computer-modeled, image-based representation. These fluence map values ​​can then be used to plan and implement radiotherapy machine parameters, including the planning and optimization of control points that control the operation of the radiotherapy machine to deliver radiotherapy with treatment to the patient's depicted anatomy.

[0039] The technical benefits of these technologies include reduced radiotherapy treatment plan creation time, improved quality of generated radiotherapy treatment plans, and the ability to evaluate less data or user input to produce higher quality fluence map designs. These technical benefits can lead to numerous significant medical benefits, including improved accuracy of radiotherapy treatments and reduced exposure to unintended radiation. The disclosed technologies are applicable to a wide range of medical and diagnostic settings or radiotherapy treatment devices and apparatuses, including those using IMRT and VMAT treatment plans.

[0040] FMO is typically performed as a numerical computation to generate a 3D dose distribution covering the target while attempting to minimize the dose's impact on nearby OARs. As will be understood, the optimal fluence map and the final 3D dose distribution generated using the fluence map are generally referred to as the "plan," even though the fluence 3D dose distribution must be resampled and transformed to suit the characteristics of linear accelerators and multi-leaf collimators (MLCs) to become a clinically deliverable treatment plan. Such changes can include arc segmentation and aperture optimization operations, as well as other aspects of transformation or modification, as referenced below. Figure 8Further discussion is needed. However, for simplicity, the term "plan" as used below generally refers to the planned radiation dose derived from fluence plot optimization and the results of the training model suitable for generating the fluence plot.

[0041] Delivering the correct fluence to achieve the desired dose in tissue involves an anteroposterior tomography procedure in which a 2D array of appropriately weighted unit beams is guided through a linear accelerator MLC from numerous angles around the target. In effect, the fluence map at each beam angle is a 2D array of eye-view projection images of the beams spanning the target. Each element of the fluence map is a real-valued weight proportional to the expected dose in the target. VMAT radiotherapy can have 100 or more beams, where the total number of unit beam weights equals 10. 5 Or more.

[0042] FMO performs exhaustive optimization of target and OAR constraints based on thousands of small unit beams aiming at the target from many directions, along with the weights of the unit beams and physical parameters describing the material fluence distribution of each unit beam. This high-dimensional optimization typically starts with default initial values ​​for the parameters, without considering the anatomy of a particular patient. Among other techniques, the creation and training of anatomically relevant models of FMO parameters are discussed below, allowing computations to be initialized to values ​​closer to the ideal final values ​​of the parameters, thus reducing the time required to produce satisfactory fluence maps. Furthermore, such anatomically relevant models of FMO parameters can be used for fluence map validation or verification and integrated in various ways for radiotherapy planning.

[0043] The following paragraphs provide an overview of the implementation of an example radiotherapy system and treatment planning (see reference). Figures 2A to 7 This includes the use of computing systems and hardware implementations (see reference). Figure 1 and Figure 20 The following paragraphs also provide specific considerations for optimizing the injection plot (see reference). Figures 8 to 9 ) and the representation of the injection volume diagram relative to the projection of the patient's anatomy (see reference) Figures 10 to 13 The discussion continues. Finally, a discussion of machine learning techniques (see [reference]). Figures 14 to 16B ) is provided as a method for training and using machine learning models. Figures 17 to 19 ).

[0044] Figure 1A radiotherapy system 100 for delivering radiotherapy to a patient is shown. The radiotherapy system 100 includes an image processing unit 112. The image processing unit 112 can be connected to a network 120. The network 120 can be connected to the Internet 122. The network 120 can connect the image processing unit 112 to one or more of the following: a database 124, a hospital database 126, an oncology information system (OIS) 128, a radiotherapy unit 130, an image acquisition unit 132, a display unit 134, and a user interface 136. The image processing unit 112 can be configured to generate a radiotherapy treatment plan 142 and plan-related data to be used by the radiotherapy unit 130.

[0045] Image processing apparatus 112 may include memory device 116, image processor 114, and communication interface 118. Memory device 116 may store computer-executable instructions, such as operating system 143, radiotherapy treatment plans 142 (e.g., original treatment plans, adjusted treatment plans, etc.), software programs 144 (e.g., executable implementations of artificial intelligence, deep learning neural networks, radiotherapy treatment planning software), and any other computer-executable instructions to be executed by processor 114. In an example, software program 144 may convert a medical image of one format (e.g., MRI) into another format (e.g., CT) by generating synthetic images such as pseudo-CT images. For example, software program 144 may include an image processing program for training a predictive model to convert a medical image 146 of one modality (e.g., MRI image) into a synthetic image of a different modality (e.g., pseudo-CT image); alternatively, the image processing program may convert a CT image into an MRI image. In another example, software program 144 can register a patient image (e.g., a CT image or MR image) with the patient's dose distribution (also represented as an image), such that corresponding image voxels and dose voxels are appropriately correlated via a network. In yet another example, software program 144 can replace the functionality of the patient image, such as a signature distance function or a processed version of the image that emphasizes some aspect of the image information. Such functionality may emphasize the edges or differences in voxel texture, or any other structural aspect useful for neural network learning. In another example, software program 144 can replace the functionality of the dose distribution that emphasizes some aspect of the dose information. Such functionality may emphasize steep gradients around the target or any other structural aspect useful for neural network learning. Memory device 116 can store data, including medical images 146, patient data 145, and other data required to create and implement at least one radiotherapy treatment plan 142, or data associated with at least one plan.

[0046] In yet another example, software program 144 can generate projection images for a collection of two-dimensional (2D) CT or MR images and / or 3D CT or MR images depicting anatomical structures (e.g., one or more targets and one or more OARs), the projection images representing different views of the anatomical structures according to one or more beam angles used to deliver radiotherapy, the beam angles potentially corresponding to corresponding gantry angles of the radiotherapy apparatus. For example, software program 144 can process a collection of CT or MR images and create a stack of projection images depicting different views of the anatomical structures depicted from various angles of the radiotherapy beams in the CT or MR images as part of the fluence data for generating a radiotherapy treatment plan. For example, one projection image might represent a view of the anatomical structure from 0 degrees of the gantry, a second projection image might represent a view of the anatomical structure from 45 degrees of the gantry, and a third projection image might represent a view of the anatomical structure from 90 degrees of the gantry, where individual radiotherapy beams are located at each angle. In other examples, each projected image may represent a view of an anatomy based on a specific beam angle, corresponding to the position of the radiotherapy beam at the corresponding angle on the gantry.

[0047] exist Figure 2A The diagram schematically illustrates a projected view of a simple ellipse 202. Figure 2A In this view, the views are oriented relative to the center of the ellipse and capture the shape and extent of the ellipse 202 as seen from each angle (e.g., 0 degrees represented by view 203, 45 degrees represented by view 204, and 90 degrees represented by view 205). For example, when viewed at a 0-degree angle relative to the y-axis 206 of the ellipse 202, the view of the ellipse 202 is projected as view 203. For example, when viewed at a 45-degree angle relative to the y-axis 206 of the ellipse 202, the view of the ellipse 202 is projected as view 204. For example, when viewed at a 90-degree angle relative to the y-axis 206 of the ellipse 202, the view of the ellipse 202 is projected as view 205.

[0048] exist Figure 2B The image shows the projection of male pelvic anatomy relative to a set of original 3D CT images 201. Selected organs at risk and target organs are outlined in the 3D CT images 201, and voxel code values ​​are assigned to their respective anatomical structures based on their type. Projected images 250 at selected angles (0 degrees, 45 degrees, and 90 degrees) around the central axis of the 3D CT images 201 can be obtained using the orthogonal projection capabilities of reconstruction procedures (e.g., cone-beam CT reconstruction procedures). Projected images can also be calculated by directly recreating the geometry of the projected view from ray tracing or by means such as Fourier reconstruction used in computed tomography.

[0049] In the example, a projected image can be calculated by tracing the path of light as pixels in the image plane and simulating the effect of it encountering a virtual object. In some implementations, the projected image is generated by tracing the path from a hypothetical eye (the beam's eye view or MLC view) to each pixel in the virtual screen and calculating the color of objects visible through it. Other tomographic reconstruction techniques can be used to generate projected images from views of anatomical structures depicted in 3D CT images 201.

[0050] For example, a group (or set) of 3D CT images 201 can be used to generate one or more views of the anatomical structures (e.g., bladder, prostate, seminal vesicle, rectum, first target, and second target) depicted in the 3D CT images 201. These views can be from the perspective of the radiotherapy beam (e.g., provided by the gantry of the radiotherapy device), and for simplicity, refer to... Figure 2B These views are measured in degrees relative to the y-axis of the 3D CT image 201 and based on the distance between the anatomical structures depicted in the image and the MLC. Specifically, the first view 210 represents the projection of the 3D CT image 201 when viewed from the gantry at 0 degrees relative to the y-axis and at a given distance from the anatomical structures depicted in the 3D CT image 201; the second view 220 represents the projection of the 3D CT image 201 when viewed from the gantry at 45 degrees relative to the y-axis and at a given distance from the anatomical structures depicted in the 3D CT image 201; and the third view 230 represents the projection of the 3D CT image 201 when viewed from the gantry at 90 degrees relative to the y-axis. Any other views may be provided, such as different views at each of the 360 ​​degrees around the anatomical structures depicted in the 3D CT image 201.

[0051] Return to reference Figure 1 In yet another example, software program 144 can use the machine learning techniques discussed herein to generate graphical representations (referred to differently as fluence map representations, fluence map images, or “fluence maps”) of fluence map data at various radiotherapy beam and gantry angles. Specifically, software program 144 can optimize information from these fluence map representations with machine learning assistance for fluence map optimization. Such fluence map data is ultimately used to generate and refine a set of control points for controlling the radiotherapy apparatus to produce a radiotherapy beam. These control points can represent machine parameters such as beam intensity, gantry angle relative to the patient position, and blade position of the MLC to deliver the dose specified by the fluence map representation.

[0052] In another example, software program 144 stores treatment planning software that includes a trained machine learning model, such as a generative model trained from a generative adversarial network (GAN), a conditional generative adversarial network (cGAN), or a cycle-consistent generative adversarial network (CycleGAN), to generate or estimate a fluence map image representation at a given radiotherapy beam angle based on inputs to a model of a projected image of an anatomical structure representing a view of the anatomical structure from a given angle and treatment constraints (e.g., target dose and organs at risk) in such an anatomical structure. Software program 144 may also store functions for optimizing or accepting further optimization of the fluence map data, and for converting or calculating the fluence map into machine parameters or control points for a given type of radiotherapy machine (e.g., to output a beam from an MLC to achieve a fluence map using the MLC blade position). Therefore, the treatment planning software can perform multiple calculations to adapt the beam shape and intensity and gantry angle of each radiotherapy beam to the radiotherapy treatment constraints, and calculate control points for a given radiotherapy device to achieve that beam shape and intensity in the subject patient.

[0053] In addition to the memory device 116 storing the software program 144, it is conceivable that the software program 144 may be stored on a removable computer medium, such as a hard disk drive, computer disk, CD-ROM, DVD, HD, Blu-ray DVD, USB flash drive, SD card, memory stick or any other suitable medium; and the software program 144 may be executed by the image processor 114 when it is downloaded to the image processing device 112.

[0054] Processor 114 may be communicatively coupled to memory device 116, and processor 114 may be configured to execute computer-executable instructions stored on memory device 116. Processor 114 may send or receive medical images 146 to memory device 116. For example, processor 114 may receive medical images 146 from image acquisition device 132 via communication interface 118 and network 120 for storage in memory device 116. Processor 114 may also send medical images 146 stored in memory device 116 to network 120 via communication interface 118 for storage in database 124 or hospital database 126.

[0055] Furthermore, processor 114 can utilize software program 144 (e.g., treatment planning software) along with medical images 146 and patient data 145 to create a radiotherapy treatment plan 142. Medical images 146 may include information such as imaging data associated with patient anatomical regions, organs, or the amount of segmented data of interest. Patient data 145 may include information such as: (1) functional organ modeling data (e.g., sequential vs. parallel organs, appropriate dose response models, etc.); (2) radiation dose data (e.g., DVH information); or (3) other clinical information about the patient and treatment process (e.g., other surgeries, chemotherapy, previous radiotherapy, etc.).

[0056] Additionally, processor 114 can utilize software programs to generate intermediate data, such as updated parameters to be used by machine learning models, such as neural network models; or to generate intermediate 2D or 3D images, which can then be stored in memory device 116. Processor 114 can then transmit an executable radiotherapy treatment plan 142 to radiotherapy apparatus 130 via communication interface 118 to network 120, in which the radiotherapy plan will be used to treat a patient with radiotherapy. Furthermore, processor 114 can execute software program 144 to perform functions such as image transformation, image segmentation, deep learning, neural networks, and artificial intelligence. For example, processor 114 can execute software program 144 to train medical images or to outline medical images; such software program 144, when executed, can train boundary detectors or utilize a shape dictionary.

[0057] Processor 114 may be a processing device, including one or more general-purpose processing devices, such as a microprocessor, central processing unit (CPU), graphics processing unit (GPU), accelerated processing unit (APU), etc. More specifically, processor 114 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor implementing other instruction sets, or a processor implementing a combination of instruction sets. Processor 114 may also be implemented by one or more special-purpose processing devices such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), system-on-a-chip (SoCs), etc. As those skilled in the art will understand, in some examples, processor 114 may be a special-purpose processor rather than a general-purpose processor. Processor 114 may include one or more known processing devices, such as those from Intel. TM Manufactured Pentium TM Core TM XeonTM or The series of microprocessors are from AMD. TM Turion manufactured TM Athlon TM Sempron TM Opteron TM FX TM Phenom TM The processor 114 may be any of the microprocessors in the series or any of the various processors manufactured by Sun Microsystems. The processor 114 may also include a graphics processing unit, such as one from Nvidia. TM Manufactured Series, by Intel TM GMA and Iris manufactured TM series or by AMD TM Radeon manufactured TM The series of GPUs. Processor 114 may also include accelerated processing units, such as those from Intel. TM Xeon Phi manufactured TM The examples disclosed are not limited to any type of processor otherwise configured to meet the computational needs of identifying, analyzing, maintaining, generating and / or providing large amounts of data or manipulating such data to perform the methods disclosed herein. Additionally, the term "processor" can include more than one processor (e.g., a multi-core design or multiple processors each having a multi-core design). Processor 114 can execute a sequence of computer program instructions stored in memory device 116 to perform various operations, processes, and methods, which will be described in more detail below.

[0058] The memory device 116 can store medical images 146. In some examples, the medical image 146 may include one or more MRI images (e.g., 2D MRI, 3D MRI, 2D flow cytometry MRI, four-dimensional (4D) MRI, 4D volumetric MRI, 4D cine MRI, projection images, fluence map representation images, graphic aperture images, pairing information between projection images and fluence map representation images, and pairing information between projection images and graphic aperture images, etc.), functional MRI images (e.g., fMRI, DCE-MRI, diffusion MRI), CT images (e.g., 2D CT, cone-beam CT, 3D CT, 4D CT), ultrasound images (e.g., 2D ultrasound, 3D ultrasound, 4D ultrasound), one or more projection images representing views of anatomical structures depicted in MRI, synthetic CT (pseudo-CT) and / or CT images at different angles of the gantry relative to the patient axis, PET images, X-ray images, fluoroscopic images, radiotherapy portal images, SPECT images, computer-generated synthetic images (e.g., pseudo-CT images), aperture images, graphic aperture image representations of MLC blade positions at different gantry angles, etc. Furthermore, medical image 146 may also include medical image data, such as training images, contour images, and dose images. In this example, medical image 146 can be received from image acquisition device 132. Therefore, image acquisition device 132 may include an MRI imaging device, a CT imaging device, a PET imaging device, an ultrasound imaging device, a fluoroscopy device, a SPECT imaging device, an integrated linear accelerator and MRI imaging device, or other medical imaging devices for acquiring medical images of a patient. Image processing device 112 may use any data type or any format type to perform operations conforming to the disclosed examples to receive and store medical image 146.

[0059] Memory device 116 may be a non-transitory computer-readable medium, such as read-only memory (ROM), phase-change random access memory (PRAM), static random access memory (SRAM), flash memory, random access memory (RAM), dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM), electrically erasable programmable read-only memory (EEPROM), static memory (e.g., flash memory, flash disk, static random access memory) and other types of random access memory, cache memory, registers, CD-ROM, DVD or other optical storage devices, magnetic tape cassette, other magnetic storage devices, or any other non-transitory medium that can be used to store information including images, data, or computer-executable instructions (e.g., stored in any format) that can be accessed by processor 114 or any other type of computer device. Computer program instructions may be accessed by processor 114, read from ROM or any other suitable memory location, and loaded into RAM for execution by processor 114. For example, memory device 116 may store one or more software applications. The software applications stored in memory device 116 may include, for example, an operating system 143 for a public computer system and for a software-controlled device. Furthermore, memory device 116 may store the entire software application or only a portion of a software application that can be executed by processor 114. For example, memory device 116 may store one or more radiotherapy treatment plans 142.

[0060] Image processing device 112 can communicate with network 120 via communication interface 118, which can be communicatively coupled to processor 114 and memory device 116. Communication interface 118 can provide communication connectivity between image processing device 112 and components of radiotherapy system 100 (e.g., allowing data exchange with external devices). For example, in some examples, communication interface 118 may have a suitable interface circuitry to connect to user interface 136, which may be a hardware keyboard, keypad, or touchscreen through which a user can input information into radiotherapy system 100.

[0061] Communication interface 118 may include, for example, network adapters, cable connectors, serial connectors, USB connectors, parallel connectors, high-speed data transmission adapters (e.g., fiber optic, USB 3.0, Thunderbolt, etc.), wireless network adapters (e.g., Wi-Fi adapters), telecommunications adapters (e.g., 3G, 4G / LTE, etc.), etc. Communication interface 118 may include one or more digital communication devices and / or analog communication devices that allow image processing device 112 to communicate with other machines and devices (e.g., remotely located components) via network 120.

[0062] Network 120 may provide the functionality of a local area network (LAN), wireless network, cloud computing environment (e.g., Software as a Service, Platform as a Service, Infrastructure as a Service, etc.), client-server, wide area network (WAN), etc. For example, network 120 may be a LAN or WAN, and may include other systems S1 (138), S2 (140), and S3 (141). Systems S1, S2, and S3 may be the same as image processing device 112, or they may be different systems. In some examples, one or more systems in network 120 may form a distributed computing / simulation environment that collaboratively performs the examples described herein. In some examples, one or more systems S1, S2, and S3 may include a CT scanner that acquires CT images (e.g., medical image 146). Additionally, network 120 may be connected to the Internet 122 to communicate with servers and clients residing remotely on the Internet.

[0063] Therefore, network 120 can allow data transmission between image processing device 112 and multiple different other systems and devices such as OIS 128, radiotherapy device 130, and image acquisition device 132. Furthermore, data generated by OIS 128 and / or image acquisition device 132 can be stored in memory device 116, database 124, and / or hospital database 126. As needed, data can be sent / received via communication interface 118 through network 120 for access by processor 114.

[0064] Image processing device 112 can communicate with database 124 via network 120 to send / receive various types of data stored in database 124. For example, database 124 may include machine data (control points) that include information associated with radiotherapy device 130, image acquisition device 132, or other machines related to radiotherapy. Machine data information may include control points such as radiation beam size, arc placement, beam on and off duration, machine parameters, segmentation, MLC configuration, gantry speed, MRI pulse sequence, etc. Database 124 may be a storage device and may be equipped with appropriate database management software. Those skilled in the art will understand that database 124 may include multiple devices located in a centralized or distributed manner.

[0065] In some examples, database 124 may include processor-readable storage media (not shown). While the processor-readable storage media in the examples may be a single medium, the term "processor-readable storage media" should be considered to include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) that store one or more sets of computer-executable instructions or data. The term "processor-readable storage media" should also be considered to include any medium capable of storing or encoding instruction sets that are executed by a processor and cause the processor to perform any or more methods of the present disclosure. Therefore, the term "processor-readable storage media" should be considered to include, but is not limited to, solid-state memory, optical media, and magnetic media. For example, a processor-readable storage media may be one or more volatile, non-transitory, or non-volatile tangible computer-readable media.

[0066] Image processor 114 can communicate with database 124 to read images into memory device 116 or store images from memory device 116 into database 124. For example, database 124 can be configured to store multiple images received from image acquisition device 132 (e.g., 3D MRI, 4D MRI, 2D MRI slice images, CT images, 2D fluorescence fluoroscopy images, X-ray images, raw data from MR or CT scans, Medical Digital Imaging and Communication (DICOM) data, projection images, graphic aperture images, etc.). Database 124 can store data to be used by image processor 114 when executing software program 144 or when creating radiotherapy treatment plan 142. Database 124 can store data generated by trained machine learning patterns, such as neural networks, including network parameters constituting the model learned by the network and the resulting prediction data. Image processing device 112 can receive imaging data such as medical images 146 (e.g., 2D MRI slice images, CT images, 2D fluorescence fluoroscopy images, X-ray images, 3D MRI images, 4D MRI images, projection images, graphic aperture images, etc.) from database 124, radiotherapy device 130 (e.g., MRI-linac), and / or image acquisition device 132 to generate radiotherapy treatment plans 142.

[0067] In this example, the radiotherapy system 100 may include an image acquisition device 132 that can acquire medical images of the patient (e.g., MRI images, 3D MRI, 2D flow cytometry MRI, 4D volumetric MRI, CT images, cone-beam CT, PET images, functional MRI images (e.g., fMRI, DCE-MRI, and diffusion MRI), X-ray images, fluoroscopy images, ultrasound images, radiotherapy field images, SPECT images, etc.). The image acquisition device 132 may be, for example, an MRI imaging device, a CT imaging device, a PET imaging device, an ultrasound device, a fluoroscopy device, a SPECT imaging device, or any other suitable medical imaging device for acquiring one or more medical images of the patient. The images acquired by the image acquisition device 132 may be stored as imaging data and / or test data in a database 124. By way of example, the images acquired by the image acquisition device 132 may also be stored as medical images 146 by the image processing device 112 in a memory device 116.

[0068] In this example, the image acquisition device 132 may be integrated with the radiotherapy device 130 as a single device (e.g., an MRI-linac). Such an MRI-linac can be used, for example, to determine the location of a target organ or target tumor within a patient's body, so as to accurately guide radiotherapy to the predetermined target according to the radiotherapy treatment plan 142.

[0069] Image acquisition device 132 can be configured to acquire one or more images of a patient's anatomy targeting a region of interest (e.g., a target organ, a target tumor, or both). Each image—typically a 2D image or slice—can include one or more parameters (e.g., 2D slice thickness, orientation, and location). In this example, image acquisition device 132 can acquire 2D slices in any orientation. For example, the orientation of a 2D slice can include sagittal orientation, coronal orientation, or axial orientation. Processor 114 can adjust one or more parameters, such as the thickness and / or orientation of the 2D slices, to include the target organ and / or target tumor. In this example, 2D slices can be determined based on information such as 3D MRI volume. For example, in the case of using radiotherapy device 130, such 2D slices can be acquired by image acquisition device 132 "in real time" while the patient is undergoing radiotherapy treatment, where "in real time" means acquiring data in at least milliseconds or less.

[0070] Image processing device 112 can generate and store radiotherapy treatment plans 142 for one or more patients. Radiotherapy treatment plans 142 can provide information about the specific radiation dose to be applied to each patient. Radiotherapy treatment plans 142 may also include other radiotherapy information, such as control points, including beam angle, gantry angle, beam intensity, dose histogram volume information, the number of radiation beams to be used during treatment, and the dose per beam.

[0071] Image processor 114 can be used with software program 144, such as treatment planning software (e.g., manufactured by Elekta AB in Stockholm, Sweden). To generate a radiotherapy treatment plan 142, the image processor 114 can communicate with an image acquisition device 132 (e.g., a CT device, MRI device, PET device, X-ray device, ultrasound device, etc.) to access images of the patient and delineate targets such as tumors. In some examples, it may be necessary to delineate one or more OARs, such as the tumor periphery or healthy tissue adjacent to the tumor. Therefore, when an OAR is close to the target tumor, segmentation of the OAR can be performed. Additionally, if the target tumor is close to an OAR (e.g., the prostate near the bladder and rectum), by segmenting the OAR from the tumor, the radiotherapy system 100 can study not only the dose distribution in the target but also the dose distribution in the OAR.

[0072] To delineate a target organ or tumor from an OAR, medical images of a patient undergoing radiotherapy, such as MRI, CT, PET, fMRI, X-ray, ultrasound, radiotherapy field images, SPECT images, etc., can be non-invasively acquired via image acquisition device 132 to reveal the internal structure of the body part. Based on information from the medical images, the 3D structure of the relevant anatomical portion can be obtained. Furthermore, during treatment planning, numerous parameters can be considered to achieve a balance between effective treatment of the target tumor (e.g., ensuring the target tumor receives a sufficient radiation dose for effective treatment) and low radiation exposure to the OAR (e.g., ensuring the OAR receives the lowest possible radiation dose). Other parameters that can be considered include the location of the target organ and tumor, the location of the OAR, and the movement of the target relative to the OAR. For example, a 3D structure can be obtained by contouring the target within each 2D layer or slice of the MRI or CT image, or by contouring the OAR and combining the contours of each 2D layer or slice. It can be done manually (e.g., by a physician, dosimeter, or healthcare professional using a program such as MONACO manufactured by Elekta AB in Stockholm, Sweden). TM ) or automatically (e.g., using programs such as ABAS, an Atlas-based automatic segmentation software manufactured by Elekta AB in Stockholm, Sweden). TM and subsequent automatic segmentation software product ADMIRE TM Generate outlines. In some examples, the 3D structure of the target tumor or OAR can be automatically generated using treatment planning software.

[0073] After the target tumor and one or more OARs have been located and mapped, a dosimeter, physician, or healthcare professional can determine the radiation dose to be applied to the target tumor and any maximum dose that can be received by OARs located near the tumor (e.g., left and right parotid glands, optic nerve, eye, lens, inner ear, spinal cord, brainstem, etc.). After the radiation dose has been determined for each anatomical structure (e.g., target tumor, OAR), processes known as inverse planning can be performed to determine one or more treatment planning parameters that will achieve the desired radiation dose distribution. Examples of treatment planning parameters include volume mapping parameters (e.g., those defining the target volume, contour-sensitive structures, etc.), the edges around the target tumor and OARs, beam angle selection, collimator settings, and beam-on time. During reverse planning, the physician can define dose constraint parameters that set limits on how much radiation an OAR can receive (e.g., defining a full dose to the tumor target and a zero dose to any OAR; defining 95% of the dose to the target tumor; defining ≤45 Gy, ≤55 Gy, and <54 Gy to the spinal cord, brainstem, and optic nerve structures, respectively). The result of reverse planning can form a radiotherapy treatment plan 142 that can be stored in memory device 116 or database 124. Some of these treatment parameters can be related. For example, adjusting one parameter (e.g., weighting for different objects, such as increasing the dose to the target tumor) in an attempt to change the treatment plan may affect at least one other parameter, which in turn may lead to the development of different treatment plans. Therefore, image processing device 112 can generate a customized radiotherapy treatment plan 142 with these parameters so that radiotherapy device 130 can deliver radiotherapy treatment to the patient.

[0074] Additionally, the radiotherapy system 100 may include a display device 134 and a user interface 136. The display device 134 may include one or more displays showing medical images, interface information, treatment planning parameters (e.g., projected images, graphic aperture images, contours, dose, beam angle, etc.), treatment plans, targets, target location and / or target tracking, or any related information to the user. The user interface 136 may be a keyboard, keypad, touchscreen, or any type of device through which the user can input information into the radiotherapy system 100. Alternatively, the display device 134 and user interface 136 may be integrated into a tablet computer (e.g., an Apple device). Lenovo Samsung In devices such as (etc.).

[0075] Furthermore, any and all components of the radiotherapy system 100 can be implemented as virtual machines (e.g., VMware, Hyper-V, etc.). For example, a virtual machine can be software acting as hardware. Therefore, a virtual machine can include at least one or more virtual processors, one or more virtual memories, and one or more virtual communication interfaces that together act as hardware. For example, image processing device 112, OIS 128, and image acquisition device 132 can be implemented as virtual machines. Given the available processing power, memory, and computing power, the entire radiotherapy system 100 can be implemented as a virtual machine.

[0076] Figure 3A A radiotherapy apparatus 302 is shown, which may include a radiation source such as an X-ray source or linear accelerator, a bed 316, an imaging detector 314, and a radiotherapy output 304. The radiotherapy apparatus 302 can be configured to emit a radiation beam 308 to provide treatment to a patient. The radiotherapy output 304 may include one or more attenuators or collimators, such as those shown below. Figure 7 The MLC is described in the illustrative example.

[0077] Return to reference Figure 3A The patient can be positioned in area 312 and supported by treatment bed 316 to receive a radiation therapy dose according to a radiotherapy treatment plan. Radiation therapy output 304 can be mounted or attached to frame 306 or other mechanical support. When bed 316 is inserted into the treatment area, one or more chassis motors (not shown) can rotate frame 306 and radiation therapy output 304 about bed 316. In one example, frame 306 can rotate continuously about bed 316 when bed 316 is inserted into the treatment area. In another example, frame 306 can rotate to a predetermined position when bed 316 is inserted into the treatment area. For example, frame 306 can be configured to rotate radiation therapy output 304 about an axis (“A”). Both the bed 316 and the radiotherapy output unit 304 can be moved independently to other locations around the patient, for example, by moving laterally (“T”), laterally (“L”), or by rotation about one or more other axes, such as rotation about the transverse axis (denoted as “R”). A controller communicatively connected to one or more actuators (not shown) can control the movement or rotation of the bed 316 to properly position the patient inside or outside the radiation beam 308 according to the radiotherapy treatment plan. Both the bed 316 and the gantry 306 can move independently of each other with multiple degrees of freedom, which allows the patient to be positioned so that the radiation beam 308 can be precisely targeted at the tumor. The MLC can be integrated and included within the gantry 306 to deliver a radiation beam 308 of a specific shape.

[0078] Figure 3AThe coordinate system shown (including axes A, T, and L) may have an origin located at isocenter 310. The isocenter may be defined as the location where the central axis of the radiation beam 308 intersects the origin of the coordinate axes for delivering a prescribed radiation dose to or within the patient. Alternatively, isocenter 310 may be defined as a location where, for various rotational positions of the radiotherapy output 304 positioned by the gantry 306 about axis A, the central axis of the radiation beam 308 intersects the patient. As discussed herein, gantry angles correspond to the position of the gantry 306 relative to axis A, although any other axis or combination of axes may be referenced and used to determine the gantry angles.

[0079] The gantry 306 may also have an attached imaging detector 314. The imaging detector 314 is preferably located opposite the radiation source, and in this example, the imaging detector 314 may be located within the field of the radiation beam 308.

[0080] Imaging detector 314 may be mounted on gantry 306 (preferably opposite radiotherapy output 304) to maintain alignment with treatment beam 308. Imaging detector 314 rotates about a rotation axis as gantry 306 rotates. In this example, imaging detector 314 may be a flat panel detector (e.g., a direct detector or a scintillator detector). In this way, imaging detector 314 may be used to monitor radiation beam 308, or it may be used to image the patient's anatomy, such as field imaging. The control circuitry of radiotherapy apparatus 302 may be integrated within or remote from radiotherapy system 100.

[0081] In the illustrative example, one or more of the bed 316, therapy output 304, or gantry 306 can be automatically positioned, and the therapy output 304 can establish a radiation beam 308 according to a specified dose for a particular treatment delivery instance. The treatment delivery sequence can be specified according to a radiotherapy treatment plan (e.g., using one or more different orientations or positions of the gantry 306, bed 316, or therapy output 304). Treatment deliveries can occur sequentially, but can cross over on or within the patient at the desired treatment site (e.g., at isocenter 310). This allows a prescribed cumulative dose of radiotherapy to be delivered to the treatment site while minimizing or avoiding damage to tissues near the treatment site.

[0082] Figure 3BA radiotherapy apparatus 302 is shown, which may include a combined linear accelerator and imaging system, such as a CT imaging system. The radiotherapy apparatus 302 may include an MLC (not shown). The CT imaging system may include an imaging X-ray source 318, such as one providing X-ray energy in the kiloelectron volt (keV) energy range. The imaging X-ray source 318 may provide a fan-shaped and / or cone-shaped radiation beam 308 directed at an imaging detector 322, such as a flat panel detector. The radiotherapy apparatus 302 may be similar to the one described above. Figure 3A The described system includes, for example, a radiotherapy output 304, a stand 306, a bed 316, and another imaging detector 314 (e.g., a flat panel detector). An X-ray source 318 can provide a relatively low-energy X-ray diagnostic beam for imaging.

[0083] exist Figure 3B In the illustrative example, the radiotherapy output 304 and the X-ray source 318 can be mounted on the same rotating stage 306, rotated 90 degrees apart from each other. In another example, two or more X-ray sources can be mounted along the periphery of the stage 306, each with its own detector arrangement to provide diagnostic imaging from multiple angles simultaneously. Similarly, multiple radiotherapy outputs 304 can be provided.

[0084] Figure 4 A radiotherapy system 400 is depicted, which may include a combined radiotherapy device 302 and an imaging system such as a magnetic resonance (MR) imaging system conforming to the disclosed examples (e.g., MR-linac known in the art). As shown, system 300 may include a bed 316, an image acquisition device 420, and a radiation delivery device 430. System 300 delivers radiotherapy to a patient according to a radiotherapy treatment plan. In some examples, image acquisition device 420 may correspond to... Figure 1 The original image of the first modality can be obtained (e.g., Figure 6A The MRI image shown) or the target image of the second modality (e.g., Figure 6B Image acquisition device 132 (the CT image shown).

[0085] Bed 316 can support a patient (not shown) during the treatment phase. In some implementations, bed 316 can move along a horizontal translation axis (labeled "I"), allowing bed 316 to move a patient lying on bed 316 into and / or out of system 400. Bed 316 can also rotate about a vertical rotation axis transverse to the center of the translation axis. To allow such movement or rotation, bed 316 may have motors (not shown) that enable bed 316 to move in various directions and rotate along various axes. A controller (not shown) can control these movements or rotations to properly position the patient according to the treatment plan.

[0086] In some examples, the image acquisition device 420 may include an MRI machine for acquiring 2D or 3D MRI images of a patient before, during, and / or after a treatment phase. The image acquisition device 420 may include a magnet 421 for generating a main magnetic field for magnetic resonance imaging. The magnetic field lines generated by the operation of the magnet 421 may extend substantially parallel to the central translation axis I. The magnet 421 may include one or more coils having an axis extending parallel to the translation axis I. In some examples, one or more coils of the magnet 421 may be spaced apart such that the central window 423 of the magnet 421 is free of coils. In other examples, the coils in the magnet 421 may be thin enough or have a reduced density such that the coils are substantially transmissive to radiation of wavelengths generated by the radiotherapy device 430. The image acquisition device 420 may also include one or more shielding coils that can generate approximately equal amplitude and opposite polarity magnetic fields outside the magnet 421 to eliminate or reduce any magnetic field outside the magnet 421. As described below, the radiation source 431 of the radiation delivery device 430 can be positioned in a region where the magnetic field is at least eliminated to the first order or reduced.

[0087] The image acquisition device 420 may further include two gradient coils 425 and 426, which can generate a gradient magnetic field superimposed on the main magnetic field. The coils 425 and 426 can generate a gradient in the resulting magnetic field, which enables spatial encoding of protons so that their positions can be determined. The gradient coils 425 and 426 can be positioned about a common central axis with the magnet 421 and can be displaced along this central axis. This displacement can create a gap or window between the coils 425 and 426. In an example where the magnet 421 may also include a central window 423 between the coils, the two windows can be aligned with each other.

[0088] In some examples, the image acquisition device 420 may be an imaging device other than MRI, such as X-ray, CT, CBCT, spiral CT, PET, SPECT, optical computed tomography, fluorescence imaging, ultrasound imaging, radiotherapy field imaging devices, etc. As will be appreciated by those skilled in the art, the above description of the image acquisition device 420 relates to certain examples and is not intended to be limiting.

[0089] The radiation delivery device 430 may include a radiation source 431 such as an X-ray source or a linear accelerator and an MLC 432 (hereinafter referred to as...). Figure 7(Shown in more detail below). The radiation delivery device 430 can be mounted on a chassis 435. When the bed 316 is inserted into the treatment area, one or more chassis motors (not shown) can rotate the chassis 435 around the bed 316. In the example, the chassis 435 can rotate continuously around the bed 316 when the bed 316 is inserted into the treatment area. The chassis 435 may also have an attached radiation detector (not shown), which is preferably located opposite the radiation source 431 and wherein the axis of rotation of the chassis 435 is positioned between the radiation source 431 and the detector. Furthermore, the device 430 may include a control circuitry (not shown) for controlling one or more of, for example, the bed 316, the image acquisition device 420, and the radiotherapy device 430. The control circuitry of the radiation delivery device 430 may be integrated within or remote from the system 400.

[0090] During the radiotherapy treatment phase, the patient can be positioned on bed 316. System 400 can then move bed 316 to the treatment area defined by magnet 421, coils 425 and 426, and chassis 435. The control circuitry system can then control radiation source 431, MLC 432, and chassis motor to deliver radiation to the patient through the window between coils 425 and 426 according to the radiotherapy treatment plan.

[0091] Figure 3A , Figure 3B and Figure 4 An example of a radiotherapy device configured to deliver radiotherapy treatment to a patient is generally illustrated, including a configuration in which the radiotherapy output can rotate about a central axis (e.g., axis "A"). Other radiotherapy output configurations can be used. For example, the radiotherapy output can be mounted to a robotic arm or manipulator with multiple degrees of freedom. In yet another example, the therapy output can be fixed (e.g., located in an area laterally separated from the patient), and a platform supporting the patient can be used to align the center point of the radiotherapy with a designated target site within the patient's body.

[0092] Figure 5 An example of another type of radiation therapy device 530 (e.g., the Leksell Gamma Knife) is shown. Figure 5As shown, during the radiotherapy treatment phase, patient 502 may wear a coordinate frame 520 to stabilize body parts (e.g., the head) of the patient undergoing surgery or radiotherapy. The coordinate frame 520 and patient positioning system 522 can establish a spatial coordinate system that can be used during patient imaging or during radiotherapy. The radiotherapy device 530 may include a protective housing 514 to enclose multiple radiation sources 512. The radiation sources 512 may generate multiple radiation beams (e.g., beamlets) passing through a beam channel 516. The multiple radiation beams may be configured to be focused at an isocenter 310 from different directions. While each individual radiation beam may have a relatively low intensity, the isocenter 310 can receive a relatively high level of radiation when multiple doses from the different radiation beams accumulate at the isocenter 310. In some examples, the isocenter 310 may correspond to a target in the surgery or treatment, such as a tumor.

[0093] As discussed above, by Figure 3A , Figure 3B and Figure 4 The described radiotherapy apparatus includes an MLC for shaping, guiding, or modulating the intensity of a radiotherapy beam to a designated target site within the patient's body. Figure 7 MLC 432 is shown, which includes blades 732A to 732J that can be automatically positioned to define an aperture approximating a cross-section or projection of a tumor 740. Blades 732A to 732J allow modulation of the radiotherapy beam. Depending on the radiotherapy plan, blades 732A to 732J can be made of a material designated for attenuating or blocking the radiation beam in areas other than the aperture. For example, blades 732A to 732J may include a metal plate (e.g., comprising tungsten) wherein the long axis of the plate is oriented parallel to the beam direction and has ends orthogonally oriented to the beam direction (e.g., ...). Figure 2A (As shown in the diagram). The “state” of the MLC 432 can be adaptively adjusted during the course of radiotherapy treatment to establish a treatment beam that better approximates the shape or location of the tumor 740 or other target sites. This is in contrast to the use of a static collimator configuration, or to the use of an MLC configuration determined specifically using “offline” treatment planning techniques. A radiotherapy technique that uses the MLC 432 to produce a specified radiation dose distribution to a tumor or a specific region within a tumor can be referred to as IMRT. The resulting beam shape obtained using the output of the MLC 432 is represented as a graphic aperture image. That is, a given graphic aperture image is generated to represent the appearance (beam shape) and intensity of the beam after passing through and being output by the MLC 432.

[0094] IMRT involves irradiating the patient at a small number of fixed gantry angles; while VMAT typically involves irradiating the patient from 100 or more gantry angles. Specifically, using a VMAT radiotherapy device, a linear accelerator rotating around the patient continuously irradiates the patient with a beam continuously shaped by an aperture generated by an MLC, achieving modulated coverage of the target at each angle with a prescribed radiation dose. VMAT has become popular because it accurately targets the patient while minimizing the dose to adjacent OARs, and VMAT treatment generally takes less time than IMRT treatment.

[0095] Creating personalized plans for each patient using IMRT or VMAT is challenging. Treatment planning systems model the physics of radiation dose, but they offer little guidance to planners on how to modify treatment parameters to achieve a high-quality plan. Changing planning variables often yields non-intuitive results, and treatment planning systems fail to inform planners whether a little or a lot of effort is required to advance the current plan to a clinically usable one. Automated multi-criteria optimization reduces planning uncertainty through automated, exhaustive numerical optimization that satisfies objective-OAR constraints, but this approach is time-consuming and often fails to produce deliverable plans.

[0096] Creating a radiation therapy plan typically involves applying multiple processes to address treatment planning considerations. Figure 8 The data flow for three typical phases defined through the VMAT program is shown: Flux Map Optimization (FMO) 820, Arc Sequencing 840, and Direct Aperture Optimization 860. (See also...) Figure 8 As shown, patient image structure 810, such as image data received from CT, MRI, or similar imaging modalities, is received as input for the treatment plan. Through a process of fluence map optimization 820, a fluence map 830 is identified and created. For a VMAT plan, the fluence map 830 represents the ideal target dose coverage that must be replicated by constructing segments (MLC aperture and monitor unit weights) with a set of linear accelerator gantry angles.

[0097] Specifically, fluence map 830 provides a model of the ideal 3D dose distribution for radiotherapy treatment and is constructed during fluence map optimization (FMO) 820. FMO is a hierarchical, multi-criteria numerical optimization that models the target radiation by irradiating the target with a number of small X-ray unit beams constrained by the target dose and OAR. The resulting fluence map 830 represents a 2D array of unit beam weights that map radiation onto the target beam eye view; thus, when planning VMAT treatment, a fluence map exists for each VMAT beam at each of 100 or more angular settings surrounding the linear accelerator gantry. Since fluence is the density of rays passing through a unit surface perpendicular to the beam direction, while dose is the energy released in the irradiated material, the final 3D dose covering the target is specified by the set of 2D fluence maps.

[0098] The 3D dose represented in the fluence diagram 830 generated by the FMO 820 does not include sufficient information about how the machine can deliver radiation to achieve that distribution. Therefore, an initial set of linear accelerator / MLC weighted apertures (one set per gantry angle; also known as control points) must be created by iteratively modeling the 3D dose through a series of MLC apertures at different gantry angles and appropriate intensities or weights. These initial control points 850 are generated by arc sorting 840, where the aperture and parameters of the resulting (initial control points 85) depend on the specific patient's anatomy and target geometry.

[0099] Even after generating numerous control points 850, further refinement of the aperture and weights is frequently involved, with control points occasionally being added or subtracted. Refinement is necessary because the 3D dose distribution obtained from arc sorting 840 is reduced relative to the original optimal infusion map 830, and some refinement of the aperture always improves the quality of the resulting plan. The process of optimizing the aperture of these control points is called direct aperture optimization 860, where the resulting refined aperture and weights (final control points 870) depend on the specific patient's anatomy and target geometry.

[0100] In each of operations 820, 840, and 860, the achievable solution corresponds to the minimum of an objective function in a high-dimensional space, which may have many minima and requires lengthy numerical optimization. In each case, the objective function describes a mapping or relationship between patient anatomy and the dose distribution or the set of parameters of a linear accelerator / MLC machine. The following technique discusses a mechanism by which the generation of fluence maps 830 and the fluence map optimization process 820 can be optimized from modeling. Specifically, the optimization of the fluence map can be performed by generating the fluence map using a probabilistic model (e.g., a learned model trained via machine learning techniques).

[0101] Probabilistic modeling of fluency graphs based on models learned from clinical planning populations can provide... Figure 8 The FMO operation discussed here offers two significant benefits. One benefit of using a probabilistic model is the accelerated search for solutions. Using a trained probabilistic model, a fluence map that approximates the true solution can be inferred from the structure of a new patient (e.g., fluence map 830). This approximation of the solution can be used as a starting point for numerical optimization and yields the correct solution in a shorter time compared to starting from a point with less information (e.g., final control point 870). Another benefit of using a probabilistic model involves reliably obtaining higher quality results using approximations than by starting with less information. For example, in some settings, the inferred fluence map can be used as a lower bound on the expected optimization quality of the fluence map.

[0102] FMO is a high-dimensional optimization that typically uses default initial values ​​for the problem parameters without considering the specific anatomy of the patient. Performing FMO to produce Pareto optimal results means that the resulting dose distribution satisfies both the objective and OAR constraints, such that no constraint can be improved without reducing another constraint. Achieving this high level of accuracy requires exhaustive calculations of the unit beam weights in the beam-normal fluence plot, adjusting them under a sorted list of constraints. Therefore, current forms of FMO calculations are time-consuming, typically taking 10 to 20 minutes for typical prostate cases and longer for more complex head / neck treatments. Thus, computation time generally scales with the complexity of the plan and the number of constraints. And this only completes the first step in a three-step process.

[0103] In various examples, generative machine learning models are well-suited for performing FMO to generate fluence map values ​​from input imaging data. Therefore, the following examples identify ways to create and train anatomically relevant models of FMO parameters from machine learning models. By using trained machine learning models, FMO parameter computations can be initialized to values ​​closer to the final parameters, thereby reducing the time required to compute the optimal set of fluence maps.

[0104] Additionally, in various examples, the following techniques utilize machine learning models to perform FMO as part of a measurement, validation, or verification process or procedure. For instance, results from machine learning models can be used to provide an independent measure of the performance of an FMO procedure beyond the specific physical and clinical information involved in the patient. Therefore, machine learning models can provide or evaluate FMO data for use as a benchmark in future radiotherapy planning.

[0105] As a more detailed overview, the following outlines the FMO process implemented using a probabilistic machine learning model, which is executed in relation to a VMAT radiotherapy plan. As discussed above, in VMAT treatment, multiple beams are directed at the target, and the cross-sectional shape of each beam is related to the target's apparent direction from that direction. Figure 1Consistent with, or consistent with, a set of segments that together provide variable or modulated intensity modes. Each beam is discretized into unit beams occupying elements of a virtual rectangular grid in a plane perpendicular to the beam. The dose is a linear function of the unit beam intensity or fluence, expressed by the following equation:

[0106]

[0107] Where, d i (b) is from the source with strength b j The dose of unit beam j deposited in voxel i, and the vector of n unit beam weights is b = (b1, ..., bn). n ) T D ij It is the dose deposition matrix.

[0108] The FMO problem is solved through multi-criteria optimization. For example, Romeijn et al. (“A unifying framework for multi-criteria fluence map optimization models”, Phys Med Biol (49), 1991–2013, 2004) provided the following FMO model formula:

[0109] (P):

[0110]

[0111] …

[0112]

[0113] (Equation 2)

[0114] Where F(d(b)) is its dose objective function that minimizes the listed constraints, and where the specific objective G(b) is affected by the dose constraint C(b), and L is the number of constraints. The objective F(d(b)) minimizes the difference between the calculated dose and the specified dose P(b):

[0115]

[0116] The summation occurs over all voxels. Under this setup, solutions to the constrained objective can be achieved in several ways. Pareto optimal solutions to multi-criteria problems can be generated, possessing the property that any criterion value can only be improved if at least one other criterion value deteriorates, and all members of the Pareto optimal solution family lie on the Pareto boundary in the solution space. By varying the relative weights of the constraints, the planner can move along the optimal plan to explore the trade-off between the target dose and organ protection.

[0117] The iCycle or mCycle implementation demonstrates an example of FMO operation; iCycle was originally implemented by researchers at Erasmus University Medical Center (Rotterdam, Netherlands) and reimplemented as mCycle by Elekta Ltd. (Stockholm, Sweden; and St. Charles, Missouri, USA). iCycle and mCycle FMO perform beam angle and beam profile (flux map) optimization based on a wish list with priority objectives and constraints. Starting with an empty plan (no beams selected), the optimal beam orientation is selected from a predefined set of input directions. Iteration i begins by selecting a candidate beam for the i-th orientation to add to the plan. All orientations not yet selected are evaluated one by one, and for each of them, the beam profile optimization for the i-th trial beam and all previously selected beams is solved. Each iteration optimizes to satisfy all wish list objectives and their constraints.

[0118] Figure 9 An example of FMO operation is illustrated, providing a comparison between a conventional FMO process (e.g., performed by iCycle or mCycle) and FMO optimization using machine learning modeling performed with the various examples discussed in this paper. As shown, conventional FMO iterates through beam angle selection and profile optimization. Each iteration begins with selecting candidate beam orientations to add to the plan 910, followed by multi-criteria optimization 920 of the beam profiles of the new beam and all previously selected beams. Candidate beams with the best scores representing the optimal orientation 940 are added to the plan. The first optimization phase completes when additional beams identified in the search for new orientations 930 fail to sufficiently improve the optimization score. If possible, a second optimization phase 960 is performed to further refine the objective. This iterative process of beams and profiles results in a Pareto-optimal plan 980 with respect to the desired list of objectives and constraints.

[0119] In contrast, the machine learning modeling FMO technique discussed below begins with estimates of beam orientation and fluence profile 950 learned from the clinical planning population. This "planning estimate" proceeds directly to the second optimization phase 960 to refine the target against the wish list. This avoids the time-consuming accumulation of the search performed in the first optimization phase and achieves shorter plan creation time because the machine learning estimate begins to approach Pareto optimality in the parameter space compared to conventional FMO initialization parameters.

[0120] For each iteration of beam addition, the target and OAR objective are guided by a wish list, which provides a priority schedule for hard constraints and optimization objectives similar to those in Equation (2) above. For example, the wish list for prostate radiotherapy may include hard constraints (e.g., specifying maximum radiation limits for the planned target volume (PTV) region, anatomical structures such as the rectum, bladder, etc.) and objectives (priority target treatment for the PTV region, anatomical structures such as the rectum, bladder, etc.). In the example, the wish list objectives are optimized one at a time using multi-criteria optimization. The highest priority objective is the prescribed dose for the objective and the dose limit for the most vulnerable OAR. Lower priority objectives provide additional constraints to achieve the highest possible therapeutic value. At each step, the previous optimization serves as the initial point and previous constraint for subsequent constraint optimization. Furthermore, hard constraints may be defined as having the highest priority and including hard limits (that cannot be exceeded) for the target volume (PTV) and the primary OAR. Objectives are the objectives achieved through repeated numerical optimization of the fluence plot (if possible).

[0121] In the iCycle and mCycle implementations, the target dose is optimized by minimizing the logarithmic probability of tumor control (LTCP).

[0122]

[0123] This penalty dose is insufficient, but allows the PTV to be affected by nearby OARs. In this equation, V is the set of voxels including the target PTV, and d P This is the prescribed dose. α is a cell sensitivity parameter—a higher α results in fewer target voxels with a lower dose, and therefore a higher proportion of voxels receiving 95% of the prescribed dose (good PTV coverage). Generalized equivalent uniform dose (gEUD) is the second useful dose function applied to OAR:

[0124]

[0125] Here, V is the set of voxels in the relevant organ, and α is a parameter that modulates the dose delivered to that organ.

[0126] Multi-criteria optimization of the wish list actually occurs in two phases. In the first phase, the optimization proceeds from the first objective on the wish list to the last objective, minimizing each objective within its constraints. After each objective is minimized, and based on its result, the constraints of that objective become new constraints for subsequent lower-priority objectives. Adding newly realized constraints ensures that lower-priority optimizations do not degrade any higher-priority objectives. Therefore, lower-priority objectives have more constraints than higher-priority objectives. At the end of the first phase, each objective with a constrained objective has reached a value equal to that objective (even if further minimization is possible), or a value higher than its objective if minimization constraints prevent the optimization from reaching that objective.

[0127] In the second phase, all objectives are re-minimized to their maximum extent. This means that, apart from the LTCP objective, the first-phase objectives, which could have been further minimized, are now minimized to the maximum extent allowed by the relevant constraint set. Minimization of the LTCP objective stops at a defined sufficient value to leave more room for minimizing lower-priority objectives and to avoid unnecessarily increasing the dose.

[0128] The resulting FMO plan (e.g., Pareto optimal plan 980) is essentially a three-dimensional array of physical doses within a coordinate frame of the target anatomy. Given the anatomy and an array with fixed beam directions, the resulting optimal dose distribution is also fixed because many optimization layers are deterministic, at least for the numerical precision of the computer. Therefore, the set of parameters constraining the optimal 3D dose is a set of optimal beam directions and optimal fluence maps, one for each beam. Furthermore, if the beams are constrained to a fixed set of angles, the fluence map will individually constrain the 3D dose distribution.

[0129] Figure 10 An example arrangement is depicted (provided in the 3D CT image set 1001) of the original patient image, a projected image 1010 of the target anatomy at a 90° gantry angle, and a corresponding fluence map image 1020 of the planned fluence at the same gantry angle. The projected image 1010 of the target anatomy specifically represents the target planning target area (PTV) for treatment and the maximum OAR as part of the prostate radiotherapy plan. The fluence map image 1020 specifically represents a 2D array of unit beam weights corresponding to the rays passing through the fluence map pixels aimed at the target. The projected image 1010 and fluence map image 1020 of the target anatomy are shown for a hypothetical beam at a 90° gantry angle; however, it should be understood that a VMAT treatment plan can have 100 to 150 beams at different angles, each with its own view of the target anatomy and fluence map.

[0130] In the example, various data formatting techniques are applied within the machine learning model to train and infer injection data for analysis. Figure 10 The provided image representations are projection images and fluence map images. In practice, anatomical structures and fluence or dose data exist as 3D linear arrays. However, FMO results correspond to an idealized 3D dose distribution corresponding to the optimal fluence map. Using the techniques discussed below, a 2D fluence map representation can be generated in a geometric plane perpendicular to the beam (in coplanar therapy), which is a beam eye-view projection related to the linear accelerator and patient coordinate systems and similar to the patient's anatomy. Therefore, the anatomical structures and fluence map can be represented as planar projections in cylindrical coordinates and used to train a machine learning model for FMO from which inferences are derived.

[0131] Figure 11 First, the creation of multiple anatomical projections (1110, 1120, 1130) from the 3D volume of CT image data is depicted. Equivalent techniques can be used to generate projections for MR images; therefore, it is understood that the following references to CT image data are provided for illustrative purposes rather than for limitation. Figure 11 The depicted 3D CT image 1101, relative to this anatomical structure, depicts multiple projections of the male pelvic organs, providing views 1110, 1120, and 1130 at 0 degrees, 45 degrees, and 90 degrees respectively (previously regarding...). Figure 2A (Introduction). The patient's position is head-first supine, with the patient's head extending beyond the top of the projection. The organs at risk (bladder, rectum), target organs (prostate, seminal vesicles), and their enclosed target volumes (target 1, target 2) are depicted (contour depiction), and a constant density value is assigned to each organ voxel. The density is summed for voxels across two or more structures.

[0132] For example, the forward projection capabilities of the RTK cone-beam CT reconstruction toolkit, an open-source cone-beam CT reconstruction toolkit based on the Insight Toolkit (ITK), can be used to obtain projected images of the anatomical structure around the central axis of a 3D CT volume of 1100 and at assigned density. In these views, the bladder is positioned in front of the seminal vesicles at 0° (the bladder is closest to the observer), and rotated to the left in the next two views.

[0133] Figure 12 The following describes the relationship with Figure 11The views depicted are a set of equivalent geometric fluence maps. As depicted, a 2D fluence map is generated as a dose distribution from a radiotherapy beam (e.g., as depicted in 2D dose distribution projection 1201) by converting the FMO dose distribution volume 1200 to a threshold volume 1202, and then the threshold dose intensity (as shown in 1205) is converted into a projection map-image (e.g., in 2D fluence map projections 1210, 1220, and 1230). For example, the orthogonal projection capabilities of the RTK cone-beam CT reconstruction toolkit and the ITK toolkit can be used to create a projection from a 3D ideal fluence distribution. The resulting projection view corresponds to Figure 11 The views shown are: volume view 1210 corresponding to 0-degree projection view 1110; volume view 1220 corresponding to 45-degree projection view 1120; and volume view 1230 corresponding to 90-degree projection view 1130.

[0134] Figure 13 Each corresponding row further depicts a set of 2D anatomical projections 1310, corresponding 2D plethysmograms 1320, and superimposed 2D plethysmograms 1330, which show plethysmograms superimposed on the anatomical projections. Figure 13 Each column further depicts these projections and diagrams at linear accelerator gantry angles of 90° (arrangement 1340), 120° (arrangement 1350), 150° (arrangement 1360), and 180° (arrangement 1370), respectively. By using projection transformations, 3D voxel data can be accurately represented in a format compatible with the geometry of the treatment.

[0135] In various examples, these projections can be used to train machine learning models to generate predictions of fluence maps or planning parameters. Specifically, the following approach discusses trained machine learning models that predict fluence maps or planning parameters only given images of a new patient and relevant radiotherapy anatomy such as OAR and treatment goals.

[0136] In the example, a probabilistic planning model learned from a population of existing best-practice injection plans is used for prediction. New patient data combined with the model enables the prediction of injection plans that serve as a starting point for direct aperture optimization. Among other benefits, this reduces the time required to refine the plan to achieve clinical quality.

[0137] The probabilistic model can be constructed as follows. Anatomical data is represented as a random variable X, and dosing plan information is represented as a random variable Y. Bayes' rule states that the probability p(Y|X) of predicting plan Y given patient X is proportional to the conditional probability p(X|Y) of observing patient X given training plan Y and the prior probability p(Y) of the training plan, or:

[0138] p(Y|X)∝p(X|Y)p(Y)

[0139] (Equation 6)

[0140] Bayesian inference predicts new patient X * Plan Y * The conditional probability p(Y) is derived from the posterior distribution p(Y|X) during training. * |X * In practice, the new anatomical structure X * The input is fed into a trained network, which then generates a prediction plan Y based on the stored model p(Y|X). * The estimate.

[0141] Figure 14 A schematic diagram depicts the deep learning process used to train a model to predict fluence maps. In the example, the training data includes paired 3D imaging projections 1410 and stacked 3D fluence maps 1420 from the same data source (e.g., the same patient). Training generates an estimated Y that can be inferred from these projections. * The model f(X;Θ) is 1430. The estimate itself is a 3D data volume 1440, which has the same size and shape as the input anatomy and fluence data volume. The estimate can be converted into a function set of fluence and used as a way to accelerate the hot start of FMO.

[0142] The posterior model p(Y|X) is constructed by training a convolutional neural network with pairs of known data (anatomical structure, plan; X, Y) in optimization, which minimizes the network loss function and simultaneously determines the values ​​of the network layer parameters Θ. These neural network parameter values ​​embed the posterior model p(Y|X) as p(Y|X; Θ). Once trained, the network can infer plans for new anatomical structures using the Bayesian analysis described above. The performance of the neural network is established by comparing the inferred plans from test patients not used for training with the original clinical plans from those same test patients—the better the neural network, the smaller the discrepancies between the plan sets.

[0143] In various examples, machine learning models of various forms can be implemented using artificial neural networks (NNs). In its simplest implementation, an NN consists of an input layer, intermediate or hidden layers, and an output layer. Each layer consists of nodes connected to more than one input node and connected to one or more output nodes. Each node outputs its input x = (x1, ..., x2). n The function y ~ σ(w) of the sum of ) TThe input layer is defined as σ(x + β), where w is a vector of input node weights, β is the layer bias, and the nonlinear function σ is typically a sigmoid function. The parameters Θ = (w, β) are the implementation of a model learned to represent the relationship Y = f(X; Θ). The number of input layer nodes is typically equal to the number of features for each object in the set of objects classified into categories, while the number of output layer nodes is equal to the number of categories. For regression, the output layer typically has a single node that transmits estimates or probabilities of the parameters.

[0144] The network is trained by presenting the object's category or known object features (parameter values) to the network and adjusting the node weights w and biases β to reduce training error by working backward from the output layer to the input layer (an algorithm called backpropagation). The training error is the normed difference ||yf(x)|| between the true answer y and the inferred estimate f(x) at any stage of training. The trained network then computes the node output σ(w) at each layer by passing the data forward from the input layer to the output layer. T x+β) is used to perform inference (classification or regression).

[0145] Neural networks possess the ability to discover general relationships between data and categories or regression values, including nonlinear functions of arbitrary complexity. This is relevant to problems such as radiotherapy dose prediction, treatment machine parameter prediction, or planning modeling, because the shape or volume overlap relationships of targets and organs captured in dose-volume histograms and overlap-volume histograms are highly nonlinear and have been shown to correlate with dose distribution shape and planning quality.

[0146] Modern deep convolutional neural networks (CNNs) have far more layers than earlier NNs—they can consist of dozens or hundreds of layers, each containing thousands to hundreds of thousands of nodes, arranged in complex geometries. Furthermore, convolutional layers are isomorphically mapped to images or any other data, which can be represented as multidimensional arrays, and can learn features embedded in the data without any prior pre-regulation or feature design. For example, a convolutional layer can locate edges in an image or temporal / pitch features in a sound stream, while subsequent layers discover larger structures comprising these primitives. In the past six years, some CNNs have approached human-level performance on standard image classification tests—correctly classifying images into thousands of classes from databases of millions of images.

[0147] The CNN is trained to learn a general mapping f: X→Y between data in the source domain X and the target domain Y, respectively. Examples of X include images of patient anatomy or functions of anatomical structures that convey structural information. Examples of Y may include graphs of radiation dose or delivery dose, or graphs of machine parameters superimposed on the target anatomical structure X. Figure 14As shown, a CNN can be trained using paired, matched known X and Y data. The CNN learns the anatomical structure and network parameters Θ = (θ1, ..., θ2). n ) T The mapping or function between the two, f(X; Θ), where θ i ={w i ,β i} represents the parameters of the i-th layer. Training makes the mapping f and the true or reference plan parameters... loss function on minimize

[0148]

[0149] The first term makes the network estimate the objective f(X; Θ) and the reference property... The first term minimizes the difference between the two terms, while the second term minimizes the variation in the value of Θ. The subscripts K and L specify the norm. The L2 norm (K, L=2) is globally convex but produces a fuzzy estimate of Y, while the L1 norm (K, L=1) encourages a clearer estimate. Network performance generally determines which combination of norms is useful.

[0150] Figure 15A A schematic diagram of the U-Net deep convolutional neural network (CNN) is depicted. Specifically, the schematic diagram depicts a U-Net deep CNN model suitable for generating fluence map representations in a generative apparatus to provide a generative model suitable for the techniques discussed herein. A pair of input images (top image) representing a target anatomical structure constraint and a radiotherapy treatment X-ray fluence representation corresponding to the target anatomical structure (bottom image) are shown, provided in the input training set 1510 to train the network. The output is a predicted fluence map representation 1540 inferred for the target image. The input training set 1510 may include separate pairs of input images projected from a 3D anatomical structure imaging volume and a 3D fluence volume; these separate pairs of input images may include separate images projected at the relevant beam angles used for treatment with a radiotherapy machine. The output dataset provided in the fluence map representation 1540 may include a representation that may include a separate output image or a 3D fluence volume.

[0151] U-Net CNN creates a scaled version of the input data array on the encoding side using max pooling, and then recombines the scaled data with the learned features on the encoding side at an increased scale using transposed convolutions for high-performance inference. The black rectangular blocks represent combinations of convolutional / batch normalization / rectified linear unit (ReLU) layers; two or more are used at each scale level. The vertical dimension of the block corresponds to the image scale (S), while the horizontal dimension is proportional to the number of convolutional filters (F) at that scale. Equation 7 above is the typical U-Net loss function.

[0152] Figure 15A The model shown depicts an arrangement suitable for generating an output dataset (output fluence map representation images 1540) based on an input training set 1510 (e.g., pairs of anatomical structure images and fluence map representation images). The name derives from the "U" configuration, and this form of CNN model is known to produce pixel-level classification or regression results. In some cases, the first path to the CNN model includes one or more deformable offset layers and one or more convolutional layers, which include convolution, batch normalization, and activations such as the Corrected Linear Unit (ReLU) or one of its variants.

[0153] The left side of the model operation ("encoding" operation 1520) learns the right side ("decoding" operation 1530) to reconstruct a set of features for the output. U-Net has 1550 gate levels consisting of conv / BN / ReLU (convolution / batch normalization / rectified linear unit) blocks, and each block has skip connections for residual learning. The block size is... Figure 15A The input image is represented by the numbers "S" and "F"; the number of feature layers is equal to F. The output of each block is a pattern of feature responses in an array of the same size as the image.

[0154] As the encoding path proceeds, the block size decreases by 1 / 2 or 2 at each level. -1 The feature size is increased by a factor of 2 as is customary. The decoding side of the network scales back from S / 2n, while adding to the feature content from the left at the same level; this is copy / connect data communication. The difference between the output image and the training version of that image drives the generator network weight adjustment via backpropagation. For inference or testing, by using the model, the input will be a single projected image of the radiotherapy treatment constraint or a set of multiple projected images (e.g., at different beam or gantry angles), and the output will be a graphical fluence map representation of image 1540 (e.g., one or more graphical images corresponding to different beam or gantry angles).

[0155] Figure 15A The representation of the model specifically illustrates the training and prediction of the generative model, which is suitable for performing regression rather than classification. Figure 15B An exemplary CNN model suitable for discriminating synthetic fluence map representations according to this disclosure is shown. As used herein, a “synthetic” image refers to an image generated by the model, and therefore “synthetic” is used interchangeably with the terms “estimate,” “computer simulation,” or “computer generation.” Figure 15BThe discriminator network shown may include several blocks of convolutional layers configured with a stride of 2, batch normalization layers and ReLU layers, as well as separate pooling layers. At the end of the network, there will be one or more fully connected layers to form 2D patches for discrimination purposes. Figure 15B The discriminator shown can be a patch-based discriminator configured to receive an input synthetic fluence map representation image (e.g., from...). Figure 15A The generator shown in the image classifies the image as real or fake and provides the classification as the output detection result 1570.

[0156] In the examples, specific types of CNNs and generative adversarial networks (GANs) can be used to generate cost-based FMO modeling techniques to predict dose planning parameters (dose maps) based on radiotherapy treatment constraints from new patient anatomy. In another example, a recurrent consistency GAN can be used to predict dose planning parameters based on new patient anatomy. An overview of the relevant GAN techniques is provided below.

[0157] Generative Adversarial Networks (GANs) are generative models (generating probability distributions) that learn the mapping from a random noise vector Z to an output image y as G: z→y. Conditional Adversarial Networks (CANs) learn the mapping from an observed image x and random noise Z as G: {x, z}→y. Both adversarial networks consist of two networks: a discriminator (D) and a generator (G). The generator G is trained to produce outputs that cannot be distinguished from "real" or actual training images by the adversarially trained discriminator D. The adversarially trained discriminator D is trained to be as accurate as possible in detecting the output of G or "fake" images.

[0158] Conditional GANs differ from unconditional GANs because both the discriminator and generator inferences are conditioned on example images of type X discussed above. The conditional GAN ​​loss function is expressed as:

[0159]

[0160]

[0161] In this context, relative to the adversary D, which attempts to maximize this loss, G attempts to minimize this loss, or...

[0162]

[0163] Additionally, it is desirable for the generator G to minimize the difference between the training estimates and the actual training images.

[0164]

[0165] Therefore, the complete loss is the λ-weighted sum of the two losses:

[0166]

[0167] In the example, the generator in a conditional GAN ​​can be a U-Net.

[0168] Consistent with the examples in this disclosure, treatment modeling methods, systems, devices, and / or processes based on such models comprise two phases: training a generative model using discriminator / generator pairs in a GAN; and using a generator trained with the GAN to make predictions using the generative model. Various examples involving GANs and CycleGANs for generating fluence map representations of images are discussed in detail in the following examples. It should be understood that other variations and combinations of deep learning models and other neural network processing methods can also be implemented using this technique. Furthermore, although this example is discussed with reference to images and image data, it should be understood that the following networks and GANs can operate using other non-image data representations and formats.

[0169] Figure 16A The data flow for training and using a GAN adapted to generate fluence planning parameters (fluence map representation) from a received set of projected images representing views of the anatomical structures of a subject's images is shown. For example, a generator model 1660 trained to generate this model... Figure 16A The generator model 1632 can be trained to achieve as Figure 1 The processing functions are provided by a portion of the image processor 114 in the radiotherapy system 100.

[0170] Therefore, the GAN model uses 1650 (predicted or inferred) data streams in Figure 16A The process is described as follows: providing new patient data 1670 (e.g., a projected image of radiotherapy treatment constraints in a view representing the anatomical structures of an input image of a new patient) to a trained generator model 1660, and using the trained generator model 1660 to generate a prediction or estimate of the generator output (image) 1634 (e.g., a synthetic graphic fluence representation image corresponding to the input projected image representing the anatomical structures of the subject image). The projected image can be generated from one or more CT or MR images representing the anatomical structures from a given beam position (e.g., at an angle on the gantry) or other defined position.

[0171] GANs consist of two networks: a generator network (e.g., generator model 1632), trained to perform classification or regression; and a discriminator network (e.g., discriminator model 1640), which samples the output distribution of the generator network (e.g., generator output (image) 1634) or a training injection map representation image from training image 1623 and determines whether the sample is the same as or different from the real test distribution. The goal of this network system is to drive the generator network to learn the real model as accurately as possible, such that the discriminator network can only determine the correct origin of a generator sample with a 50% probability, thus achieving a balance with the generator network. The discriminator has access to the real-world data, but the generator only accesses the training data through the detector's response to the generator output.

[0172] Figure 16A The data stream also illustrates the reception of training input 1610, which includes model parameters 1612 and training data 1620 (where such training images 1623 include a set of projected images and conditions or constraints 1626, which represent different views of the anatomical structures of the paired subject-patient imaging data corresponding to real-world graphical fluence maps of the patient imaging data at different views). These conditions or constraints 1626 (e.g., one or more radiotherapy treatment target areas, one or more organ-at-risk areas, etc.) may be indicated directly in the anatomical structure images themselves (e.g., as shown in projected image 1010) or provided or extracted as a separate dataset. Training input 1610 is provided to GAN model training 1630 to produce a trained generator model 1660 used in GAN model usage 1650.

[0173] As part of the training of the GAN model 1630, the generator model 1632 represents the subject image pair 1622 (in Figure 16AThe anatomical structure (also depicted as 1623) is trained on real training fluence map representations and corresponding training projected images to generate and map segment pairs in the CNN. In this way, generator model 1632 is trained to generate computer-simulated (estimated or synthesized) images of fluence map representations and fluence values ​​based on the input map as generator output (image) 1634. Discriminator model 1640 determines whether one or more simulated fluence map representation images come from training data (e.g., training fluence map representation images or real fluence map representation images) or from the generator (e.g., estimated or synthesized fluence map representation images), as passed between generator model 1632 and discriminator model 1640. Discriminator output 1636 is a decision from discriminator model 1640 indicating whether the received image is a simulated image or a real image, and is used to train generator model 1632. In some cases, the discriminator is used to train generator model 1632 on the generated images, and generator model 1632 is further trained based on cycle consistency loss information. This training process leads to backpropagation of weight adjustments 1638 and 1642 to improve the generator model 1632 and the discriminator model 1640.

[0174] During the training of generator model 1632, a set of training data can be selected from patient images (indicating radiotherapy treatment constraints) and expected outcomes (flux map representations). The selected training data may include at least one projected image of a patient anatomy representing a view of the patient's anatomy at a given beam / gantry angle, and a corresponding training flux map representation image or a true flux map representation image at that given beam / gantry angle. The selected training data may include multiple projected images of a patient's anatomy representing views of the same patient's anatomy from multiple equidistant or non-equidistant angles (e.g., at bench angles, such as from 0 degrees, from 15 degrees, from 45 degrees, from 60 degrees, from 75 degrees, from 90 degrees, from 105 degrees, from 120 degrees, from 135 degrees, from 150 degrees, from 165 degrees, from 180 degrees, from 195 degrees, from 210 degrees, from 225 degrees, from 240 degrees, from 255 degrees, from 270 degrees, from 285 degrees, from 300 degrees, from 315 degrees, from 330 degrees, from 345 degrees, and / or from 360 degrees), as well as corresponding training fluency map representations and / or machine parameter data at those different equidistant or non-equidistant bench angles.

[0175] Therefore, in this example, the data preparation for training the GAN model 1630 requires fluence map representation images paired with projected images representing views of the anatomical structures of the subject images (these can be referred to as training projected images representing views of the anatomical structures of the subject images at various beam / gantry angles). That is, the training data includes paired sets of fluence map representation images with corresponding projected images at the same gantry angle. In this example, the raw data includes paired projected images representing views of the subject's anatomical structures at various beam / gantry angles, and corresponding fluence map representations of the fluence at the respective beam / gantry angles, which can be registered and resampled to a common coordinate system to produce paired anatomical structure derived images. The training data can include multiple such pairs of images from multiple patients at any number of different beam / gantry angles. In some cases, the training data can include 360 ​​pairs of projected images and fluence map representation images, one pair of projected images and fluence map representation images for each angle of the gantry for each training patient.

[0176] The expected results may include estimates of fluence results or synthetic graphical fluence map representations, which can be further optimized and converted into control points. Such control points can be transformed and optimized to generate beam shapes at corresponding beam / gantry angles to define the delivery of radiotherapy to the patient. Based on the fluence map specifications, control points or machine parameters may include at least one beam / gantry angle, at least one multi-leaf collimator blade position, and at least one aperture weight or intensity.

[0177] In detail, in a GAN model, the generator (e.g., generator model 1632) learns the distribution p on the data x. G (x), from the distribution p Z The noise input (z) begins as the generator learns the mapping G(z; θ). G ):p Z (z)→p G (x), where G represents a layer with layer weights and bias parameters θ. G Differentiable functions of a neural network. Discriminator D(x; θ) D (For example, discriminator model 1640) maps the generator output to binary scalars {true, false} if the generator output comes from the actual data distribution p. data (x) is then considered true, and if the generator output comes from the generator distribution p G If (x) is true, then D(x) is false. In other words, D(x) is true if x comes from p. data (x) instead of coming from p GThe probability of (x). In another example, paired training data can be used, where, for example, Y is conditional on X (dependent on X). In this case, the GAN generator mapping is G(y|x;θ) from the data domain X and the domain Y. G X→Y means that in the data domain X, data x∈X represents a projected image of an anatomical structure, and in the domain Y, data y∈Y represents the fluence map representation value corresponding to x. Here, the estimation of the fluence map representation value is conditioned on its projection. Another difference from direct GANs is that the projected image x is the generator input, rather than random noise z. For this example, the discriminator setup is the same as above. In general, the generator model 1632 and the discriminator model 1640 are in a recurrent data stream, with the results of one fed into the other. The discriminator takes either the training image or the generated image, and its output is used to adjust both the discriminator weights and guide the training of the generator network.

[0178] Another useful extension of GANs is the following combination Figure 16B The CycleGAN described. Figure 16B The training and use of CycleGAN 1631 for generating a set of fluence map representation images (e.g., a set of synthesized or estimated fluence map representation images projected at the angle of a radiotherapy beam) from a received set of projected images (e.g., a set of projected images of anatomical structures indicating radiotherapy treatment constraints projected at the angle of a radiotherapy beam) are illustrated according to some examples of this disclosure. CycleGAN 1631 includes a first generator model 1635, a second generator model 1637, a first discriminator model 1639A, and a second discriminator model 1639B. The first generator model 1635 includes deformable offset layers and convolutional blocks, and the second generator model 1637 includes deformable offset layers and convolutional blocks. Both models 1635 and 1637 may each be implementations of generator model 1632 (e.g., in...). Figure 16A In this context, CycleGAN1631 can be divided into two parts: a first part 1633A and a second part 1633B. The first discriminator model 1639A and the second discriminator model 1639B can each be an implementation of the discriminator model 1640 (e.g., as a classification DCNN model).

[0179] The convolutional blocks of each generator model 1635 and 1637 can be trained together or separately from the training of other generator and discriminator models. Specifically, the convolutional blocks of generator models 1635 and 1637 are trained to obtain the correct weights to perform their functions. Deformable offset layers can be trained individually to coordinate offsetting, resampling, and interpolation. Deformable offset layers can be trained together or separately from the training of generator and discriminator models. These offset layers modify the original regular sampling grid from the upper convolutional blocks, introduce coordinate offsets, and resample the image using interpolation. Alternatively or additionally, deformable offset layers can be implemented using spatial transformers, other types of convolutional layers, and / or any other modules that can store deformable structural information of the image. The number of offset layers in the deformable offset layers can vary based on image size, the number of downsampling convolutional layers, and other factors.

[0180] In the example, in the first part 1633A, the first generator model 1635 can be trained to receive a set of training projection images 1623A (which may include one of the anatomical projection images and image pairs 1622) and generate a corresponding set of first synthetic femtograph representation images as a first generation result 1636A. The first generator model 1635 is referred to as G. proj2fluence .

[0181] The first generated result 1636A can be provided to the first discriminator model 1639A. The first discriminator model 1639A can classify the set of synthetic fluence map representation images into a set of training images with real fluence map representations or a set of training images with simulated fluence map representations, and provide this classification as a detection result 1644A. The first generated result 1636A and the detection result 1644A can be fed back to the first generator model 1635 and the first discriminator model 1639A to adjust the weights implemented by the first generator model 1635 and the first discriminator model 1639A. For example, the first generated result 1636A (e.g., the set of fluence map representation images generated by the first generator model 1635) and the detection result 1644A can be used to calculate the adversarial loss.

[0182] The first generated result 1636A (e.g., a set of images representing a synthetic fluence map) can also be simultaneously provided to the second generator model 1637. The second generator model 1637 can receive the first generated result 1636A and generate a corresponding set of simulated anatomical projection images as output. This set of simulated anatomical projection images can be referred to as the set of cyclic anatomical projection images 1641 and can be used to calculate a cyclic loss to adjust the weights of the first generator model 1635 / second generator model 1637. The second generator model 1637 that generates the set of the first cyclic anatomical projection images 1641 is referred to as G. fluence2proj .

[0183] In the example, in the second part 1633B, the second generator model 1637 can be trained to receive a set of real training injection map representation images 1623B (which may include one of the image pairs 1622) and generate a corresponding set of synthetic anatomical projection images (a set of synthetic or simulated anatomical projection images) as a first generation result 1636B. The second generator model 1637 that generates the first generation result 1636B is the same generator used in the first part 1633A.

[0184] The first generated result 1636B can be provided to the second discriminator model 1639B. The second discriminator model 1639B can classify the set of synthetic anatomical projection images into a set of real anatomical projection training images or a set of simulated anatomical projection training images, and provide this classification as a detection result 1644B. ​​The first generated result 1636B and the detection result 1644B can be fed back to the second generator model 1637 and the second discriminator model 1639B to adjust the weights implemented by the second generator model 1637 and the second discriminator model 1639B. For example, the first generated result 1636B (e.g., the set of synthetic anatomical projection images generated by the second generator model 1637) and the detection result 1644B can be used to calculate an adversarial loss.

[0185] The first generated result 1636B (e.g., a set of synthetic anatomical projection images) can also be simultaneously provided to the first generator model 1635. The first generator model 1635 can receive the first generated result 1636B and generate a corresponding cyclic injection map representation image 1643 as output. The cyclic injection map representation image 1643 can be used to calculate the cyclic loss to adjust the first generator model 1635 /

[0186] The weights of the second generator model 1637. The first generator model 1635 representing the generated cyclic invoke graph of image 1643 is the same generator used in the first part 1633A, and the second generator model 1637 representing the generated cyclic invoke graph of image 1643 is the same generator used in the first part 1633A.

[0187] In some examples, "adversarial loss" can explain the classification loss of the first discriminator model 1639A and the second discriminator model 1639B. The first discriminator model 1639A and the second discriminator model 1639B can classify whether a synthetic image has a distribution similar to that of a real image. For the cycle consistency loss, losses are computed between each set of real projected images and the set of cyclic projected images, and between each pair of real femtograph representation images and cyclic femtograph representation images. For example, a first loss can be computed between the set of projected training images 1623A and the set of cyclic projected images 1641, and a second loss can be computed between the set of real training femtograph representation images 1623B and the set of cyclic femtograph representation images 1643. Both the set of cyclic projected images 1641 and the set of cyclic femtograph representation images 1643 can be obtained by performing forward and backward loops. Each set of real projected images 1623A and the set of cyclic projected images 1641 can reside in the same domain of projected image sets, and each pair of real training injection map representation images 1623B and cyclic injection map representation images 1643 can reside in the same domain of graph injection map representation images. CycleGAN 1631 can accordingly generate synthetic injection map representation images (set of injection map representation images), a set of synthetic projected images, a set of cyclic projected images 1641, and a set of cyclic injection map representation images 1643, relying on the entire pool (or multiple) of real or real projected training images 1623A and the entire pool (or multiple) of real training injection map representation images 1623B. Based on "adversarial loss" and "cyclic consistency loss", CycleGAN 1631 can generate sharp synthetic injection map representation images with image resolution similar to that of real injection map representation images.

[0188] In some examples, (e.g., of a radiotherapy system 100) the processor can apply image registration to register training images of the real fluence map representation to a set of training projected images. This can create a one-to-one correspondence between the projected images at different angles (e.g., beam angle, gantry angle, etc.) in the training data and the fluence map representation image at each of those different angles. This relationship can be referred to as paired or matched projected images and fluence map representation images.

[0189] In some implementations, CycleGAN 1631 can be implemented as generating a set of fluency map representations based on an objective function that includes adversarial and cycle consistency loss terms. The CycleGAN network has two independent adversarial losses. Similar to conditional GANs, the mapping G: X→Y and its associated discriminator D... y The loss is given by the following representation:

[0190]

[0191] For CycleGAN, the forward cycle consistency of the network is x→G(x)→F(G(x))≈x and the backward cycle consistency is y→F(y)→G(F(y))≈y. The adversarial loss in the network is captured in the cycle consistency loss (e.g., using the first generator model 1635 / second generator model 1637 and the first discriminator model 1639A / second discriminator model 1639B) as the L1 norm.

[0192]

[0193] Furthermore, when inputting real samples from the target domain Y, the identification loss regularizes the generator to approximate an identity mapping:

[0194]

[0195] Therefore, the total loss function of a cycle-consistent GAN is

[0196]

[0197] Where D X It is the first discriminator model that determines whether an image is a set of true fluence map representations or a set of synthetic fluence map representations. D Y It is a second discriminator model that determines whether an image is a set of real projected images or a set of synthetic projected images.

[0198] The preceding examples provide examples of how GANs, conditional GANs, or CycleGANs can be trained specifically from image data in 2D or 3D image slices across multiple parallel or sequential paths, based on pairs of sets of images represented by injection maps and sets of projected images. It should be understood that GANs, conditional GANs, or CycleGANs can process other forms of image data (e.g., 3D or other multidimensional images) or include representations of such data in non-image formats. Furthermore, although only grayscale (including black and white) images are depicted in the accompanying figures, it should be understood that other image formats and image data types can be generated and / or processed by GANs.

[0199] Figure 17 An example flowchart 1700 is shown for a method of training a neural network model, such as a model to be trained to generate a flux graph using the techniques and constraints discussed above. It is evident that additional operations or variations in the sequence of operations can be implemented within this method.

[0200] At operation 1710, an operation is performed to obtain a training anatomical structure projection image, and at operation 1720, an operation is performed to obtain a training fluence map projection image. In this training scenario of flowchart 1700, pairs of anatomical structure projection images and fluence maps can be obtained from multiple human subjects, such that each pair of corresponding projection images and fluence maps is provided from the same human subject. Furthermore, corresponding pairs of anatomical structure projection images and fluence maps for training the neural network model can be obtained for each beam angle of the radiotherapy treatment machine. Operations 1710 and 1720 may also include other aspects of identifying, extracting, projecting, and modifying the projection images and fluence maps, as presented above.

[0201] Flowchart 1700 proceeds to operation 1730 to perform and supervise the training of the model. In various examples, the model is implemented as a neural network, and the neural network model can be a generative model of a generative adversarial network (GAN). Such a GAN may include at least one generative model and at least one discriminative model, wherein the at least one generative model and at least one discriminative model correspond to a corresponding generative convolutional neural network and a discriminative convolutional neural network. In yet another example, the GAN includes a conditional adversarial network (cGAN) or a cycle-consistent generative adversarial network (CycleGAN). In operations 1740 to 1750, discussed below, operations for GAN training are performed; in operations 1755 to 1775, further operations for CycleGAN training are performed.

[0202] In the example, the operations for GAN training (operations 1740 to 1750) include using the GAN to train a generative model using a discriminative model. For example, this could include using adversarial training between the discriminative and generative models to establish neural network parameter values ​​for learning values ​​through the generative and discriminative models. Such adversarial training could include: training the generative model to generate a first estimated fluence map at a first beam angle, representing a view of the subject's anatomy trained according to the first beam angle (operation 1740); training the discriminative model to classify the first estimated fluence map as an estimated or true training fluence map projection image (operation 1745); and using the output of the generative model to train the discriminative model and using the output of the discriminative model to train the generative model (operation 1750).

[0203] In the example, the operations (operations 1755 to 1775) used for CycleGAN training include two sets of models in the GAN setup: a first discriminative model and a first generative model (trained with operations 1755 and 1760 corresponding to operations 1740 and 1745), and a second generative model and a second discriminative model (trained with operations 1765 and 1770). Specifically, the second generative model is trained to process a given fluence map at a given beam angle as input, based on a given pair of anatomical projection images and fluence maps; and to generate an estimated anatomical projection image as output, the estimated anatomical projection image representing a view of the subject's anatomy according to the given beam angle. The second discriminative model is trained to classify the estimated anatomical projection image as either an estimated or a true anatomical projection image.

[0204] In the example, adversarial training for the first part of CycleGAN (the first generative model and the first discriminative model, corresponding to operations 1755 and 1760) includes: obtaining a set of training anatomical projection images representing different views of the patient's anatomy from previous treatments, the set of training anatomical projection images being paired with a set of training invoke maps corresponding to each of the different views, each of the training invoke maps being aligned with a corresponding one of the training anatomical projection images; inputting the set of training anatomical projection images into the first generative model; and outputting an estimated set of invoke maps from the first generative model; inputting the estimated set of invoke maps into the first discriminative model, and classifying the estimated set of invoke maps into a set of estimated or true invoke maps using the first discriminative model; and inputting the set of invoke maps into the second generative model, and generating an estimated set of anatomical projection images to compute a cycle consistency loss. In the example, adversarial training for the second part of CycleGAN (the second generative model and the second discriminative model, corresponding to operations 1765 and 1770) includes: inputting a set of training injection maps corresponding to each of the different views into the second generative model, and outputting a set of generated anatomical structure projection images from the second generative model; inputting the set of generated anatomical structure projection images into the second discriminative model, and classifying the set of generated anatomical structure projection images into a set of estimated or true anatomical structure projection images; and inputting the set of anatomical structure projection images into the first generative model to generate a set of estimated injection maps to compute the cycle consistency loss. Based on this adversarial training, the cycle consistency loss can be computed and considered to improve the training of the first generative model and the overall CycleGAN (operation 1775).

[0205] Flowchart 1700 ends at operation 1780, providing a trained generative model (from GAN, CycleGAN, or other training setups) for use with projected images of the patient's anatomy and for use in the radiotherapy planning process described herein.

[0206] Figure 18 An example of the method in flowchart 1800 is shown, which is used to determine a fluence map using a trained neural network model based on the techniques discussed above. Additional operations or variations in the sequence of operations can be implemented within this method, particularly when implemented as part of a radiotherapy plan or treatment procedure.

[0207] Flowchart 1800 begins with operation 1810 to obtain a set of three-dimensional image data corresponding to the subject undergoing radiotherapy treatment. This can be performed, for example, by obtaining data from an imaging modality (e.g., CT, MRI) used to image the subject. Flowchart 1800 continues with operation 1820 to obtain the subject's radiotherapy treatment constraints. This can be defined, for example, as part of the radiotherapy treatment planning process, by defining target dose regions and organ-at-risk regions. Flowchart 1800 continues with operation 1830 to generate three-dimensional image data indicating the radiotherapy treatment constraints (e.g., target dose regions, organ-at-risk regions). In this example, the input data provided by a trained model (mentioned in operation 1860 below) is image data indicating one or more target dose regions and one or more organ-at-risk regions of the subject's anatomy.

[0208] Flowchart 1800 continues to operate to perform a forward projection on the three-dimensional image data (at operation 1840) and generate projected images of the subject's anatomy for each radiotherapy beam angle (at operation 1850). In the example, each anatomy projection image provides a view of the subject according to the corresponding beam angle of the radiotherapy treatment, such as the angle corresponding to each gantry angle used by the radiotherapy machine.

[0209] Flowchart 1800 continues at operation 1860, using a trained neural network model (e.g., trained according to the method indicated in flowchart 1700) to generate one or more fluence maps. In the example, the neural network model is trained using corresponding pairs of anatomical projection images and fluence maps provided from multiple human subjects (e.g., as shown in reference). Figure 17 (As discussed). Flowchart 1800 ends at operation 1870, producing a three-dimensional fluence map representation, provided, for example, by generating multiple two-dimensional fluence maps generated by a trained model. In the example, the generated (estimated) fluence maps and model training are provided for each radiotherapy beam angle used in radiotherapy treatment.

[0210] Figure 19 A flowchart 1900 is provided illustrating the general example operation of a processing system (e.g., image processing device 112 or other computing system) for coordinating radiotherapy treatment and planning methods according to various examples. As discussed above, additional operations or variations in the order of operations can be implemented in this method, particularly when implemented as part of a radiotherapy plan or treatment operation.

[0211] At operation 1910, the method begins by acquiring three-dimensional image data, including radiation therapy constraints, corresponding to a subject undergoing radiotherapy treatment. As described above, such image data can indicate one or more target dose regions and one or more organ-at-risk regions within the subject's anatomy, and such image data can be converted or generated into a projection for further processing.

[0212] At operation 1920, a trained neural network model is used to generate an estimated fluence map representation (fluence map). For example, each element in the estimated fluence map can indicate the fluence distribution of a radiotherapy treatment at a corresponding beam angle. In a specific example, for the beam angle of a radiotherapy treatment corresponding to the gantry angle of the radiotherapy machine, each element in the estimated fluence map is a two-dimensional array of unit beam weights perpendicular to the corresponding beam direction.

[0213] At operation 1930, the estimated fluence map represents the optimal fluence distribution. For example, such optimization could include performing numerical optimization, where the estimated fluence map is provided as input to the optimization, which incorporates radiotherapy treatment constraints to produce a Pareto-optimal fluence plan used in the subject's radiotherapy treatment plan.

[0214] At operation 1940, a set of initial control points is generated for the radiotherapy beam based on the flux distribution. In this example, the set of initial control points can be generated by performing arc sorting based on a Pareto-optimal flux plan to generate a set of initial control points corresponding to each of the multiple radiotherapy beams. At operation 1950, a set of final control points is generated for the radiotherapy beam based on the initial control points. In this example, the set of final control points can be generated by performing direct aperture optimization to generate a set of final control points corresponding to each of the multiple radiotherapy beams.

[0215] At operation 1960, radiotherapy is delivered using radiotherapy beams based on the final control point. In this example, the radiotherapy is provided as volumetric intensity-modulated rotational therapy (VMAT) performed by a radiotherapy machine, and multiple radiotherapy beams are shaped to achieve modulated coverage of the target dose region from multiple beam angles to deliver the prescribed radiation dose. It is evident that other methods and optimizations of radiotherapy can also be used.

[0216] Figure 20 A block diagram of an example machine 2000 is shown, on which one or more of the methods discussed herein may be implemented. In one or more examples, one or more items of the image processing device 112 may be implemented by the machine 2000. In alternative examples, the machine 2000 operates as a standalone device or may be connected (e.g., networked) to other machines. In one or more examples, the image processing device 112 may include one or more items of the machine 2000. In a networked deployment, the machine 2000 may operate as a server or client machine in a server-client network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The machine may be a personal computer (PC), server, tablet computer, smartphone, network device, edge computing device, network router, switch, or bridge, or any machine capable of (sequentially or otherwise) executing instructions specifying actions to be taken by that machine. Furthermore, although only a single machine is shown, the term "machine" should also be understood to include any set of machines that individually or jointly execute a set (or more sets) of instructions to perform any one or more of the methods discussed herein.

[0217] Example machine 2000 includes processing circuitry systems or processors 2002 (e.g., CPU, graphics processing unit (GPU), ASIC, circuitry systems (e.g., one or more transistors, resistors, capacitors, inductors, diodes, logic gates, multiplexers, buffers, modulators, demodulators, radios (e.g., transmitting or receiving radios or transceivers)), sensors 2021 (e.g., transducers that convert one form of energy (e.g., light, heat, electricity, mechanical, or other energy) into another form of energy), and combinations thereof), main memory 2004, and static memory 2006, which communicate with each other via bus 2008. Machine 2000 (e.g., computer system) may also include a video display device 2010 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)). Machine 2000 also includes an alphanumeric input device 2012 (e.g., a keyboard), a user interface (UI) navigation device 2014 (e.g., a mouse), a disk drive or mass storage unit 2016, a signal generation device 2018 (e.g., a speaker), and a network interface device 2020.

[0218] The disk drive unit 2016 includes a machine-readable medium 2022 on which one or more sets of instructions and data structures (e.g., software) 2024 are stored, which embody or be utilized by any one or more of the methods or functions described herein. During execution of the instructions 2024 by the machine 2000, the instructions 2024 may also reside wholly or at least partially in main memory 2004 and / or in processor 2002, which also constitute the machine-readable medium.

[0219] As shown, machine 2000 includes an output controller 2028. The output controller 2028 manages data streams to and from machine 2000. The output controller 2028 is sometimes referred to as a device controller, and the software that interacts directly with the output controller 2028 is referred to as a device driver.

[0220] Although machine-readable medium 2022 is shown as a single medium in the example, the term "machine-readable medium" can include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) storing one or more instructions or data structures. The term "machine-readable medium" should also be considered to include any tangible medium capable of storing, encoding, or carrying instructions or data structures capable of storing, encoding, or carrying any one or more methods by which the instructions are executed by a machine and cause the machine to perform the contents of this disclosure, and the data structures are utilized by or associated with such instructions. Therefore, the term "machine-readable medium" should be considered to include, but is not limited to, solid-state memory, as well as optical and magnetic media. Specific examples of machine-readable media include non-volatile memory, which by way of example includes: semiconductor storage devices, such as erasable programmable read-only memory (EPROM), EEPROM, and flash memory devices; magnetic disks, such as internal hard disks and removable hard disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0221] Instructions 2024 can also be sent or received via a communication network 2026 using a transmission medium. Instructions 2024 can be sent using a network interface device 2020 and any of many well-known transmission protocols (e.g., HTTP). Examples of communication networks include LANs, WANs, the Internet, mobile phone networks, ordinary old-style telephone (POTS) networks, and wireless data networks (e.g., Wi-Fi and 4G / 5G data networks). The term "transmission medium" should be considered to include any intangible medium capable of storing, encoding, or carrying instructions executed by a machine and comprising digital or analog communication signals, or other intangible media facilitating communication by such software.

[0222] As used in this article, “communicatively coupled between” means that entities on any of the couplings must communicate through items between them, and that these entities cannot communicate with each other without communicating through those items.

[0223] Additional notes

[0224] The above detailed description includes reference to the accompanying drawings, which form part of the detailed description. The drawings illustrate, by way of illustration and not limitation, specific embodiments in which the present disclosure may be practiced. These embodiments are also referred to herein as “examples.” Such examples may include elements other than those shown or described. However, the inventors also contemplate providing examples that only include those elements shown or described. Furthermore, the inventors contemplate examples of any combination or substitution of those elements (or one or more aspects thereof) using those elements, either with respect to a particular example (or one or more aspects thereof) or with respect to other examples (or one or more aspects thereof) shown or described herein.

[0225] All publications, patents, and patent documents referenced in this document are incorporated herein by reference in their entirety as if they were individually incorporated by reference. In the event of any inconsistency between the usage in this document and those incorporated by reference, the usage in one or more incorporated references shall be considered supplementary to the usage in this document; in the event of any inconsistency, the usage in this document shall prevail.

[0226] In this document, when describing various aspects of this disclosure or elements in its embodiments, the terms “a,” “an,” “the,” and “the” are used, as is common in patent literature, to include one or more elements, independent of any other instance or usage of “at least one” or “one or more.” In this document, the term “or” is used to indicate non-exclusivity, or such that, unless otherwise stated, “A or B” includes “A but not B,” “B but not A,” and “A and B.”

[0227] In the appended claims, the terms “including” and “in which” are used as common English equivalents to the corresponding terms “comprising” and “wherein”. Furthermore, in the appended claims, the terms “comprising,” “including,” and “having” are intended to be open-ended, meaning that there may be other elements besides those listed, such that anything following such terms in the claims (e.g., “comprising,” “including,” “having”) is still considered to fall within the scope of the claims. Additionally, in the appended claims, the terms “first,” “second,” and “third,” etc., are used merely as designations and are not intended to impose numerical requirements on their objects.

[0228] Embodiments of this disclosure can be implemented using computer-executable instructions. Computer-executable instructions (e.g., software code) can be organized into one or more computer-executable components or modules. Various aspects of this disclosure can be implemented using any number of such components or modules and any organization of such components or modules. For example, various aspects of this disclosure are not limited to the specific computer-executable instructions or specific components or modules shown in the accompanying drawings and described herein. Other embodiments of this disclosure may include different computer-executable instructions or components having more or fewer functions than those shown and described herein.

[0229] The method examples described herein (e.g., operations and functions) may be implemented at least in part by a machine or computer (e.g., implemented as software code or instructions). Some examples may include a computer-readable or machine-readable medium encoded with instructions operable to configure electronic devices to perform the methods described in the examples above. Implementations of such methods may include software code such as microcode, assembly language code, high-level language code, etc. (e.g., “source code”). Such software code may include computer-readable instructions (e.g., “object” or “executable code”) for performing various methods. Software code may form part of a computer program product. Software implementations of the embodiments described herein may be provided via an article of art on which code or instructions are stored, or via a method of operating a communication interface to transmit data via a communication interface (e.g., wirelessly, via the Internet, via satellite communication, etc.).

[0230] Furthermore, software code may be tangibly stored on one or more volatile or non-volatile computer-readable storage media during execution or at other times. These computer-readable storage media may include any means of storing information in a form accessible by a machine (e.g., computing device, electronic system, etc.), such as, but not limited to, floppy disks, hard disks, removable disks, any form of disk storage media, CD-ROMs, magneto-optical disks, removable optical disks (e.g., compact discs and digital video discs), flash memory devices, magnetic tape cassettes, memory cards or memory sticks (e.g., secure digital cards), RAM (e.g., CMOS RAM), recordable / non-recordable media (e.g., read-only memory (ROM)), EPROMs, EEPROMs, or any type of media suitable for storing electronic instructions. Such computer-readable storage media are coupled to a computer system bus for access by the processor and other parts of the OIS.

[0231] In this implementation, the computer-readable storage medium may have encoded a data structure for treatment planning, wherein the treatment plan may be adaptive. The data structure for the computer-readable storage medium may be at least one of the following: Medical Digital Imaging and Communication (DICOM) format, extended DICOM format, XML format, etc. DICOM is an international communication standard that defines a format for transmitting medical image-related data between various types of medical devices. DICOM RT refers to a radiotherapy-specific communication standard.

[0232] In various embodiments of this disclosure, the methods for creating components or modules can be implemented in software, hardware, or a combination thereof. For example, the methods provided by the various embodiments of this disclosure can be implemented in software using standard programming languages ​​such as, for example, C, C++, Java, Python, and combinations thereof. As used herein, the terms “software” and “firmware” are interchangeable and include any computer program stored in memory for execution by a computer.

[0233] A communication interface includes any mechanism that interfaces with hardwired media, wireless media, optical media, etc., to communicate with another device, such as a memory bus interface, processor bus interface, Internet connection, disk controller, etc. A communication interface can be configured by providing configuration parameters and / or sending signals to prepare it to provide data signals describing the software content. A communication interface can be accessed via one or more commands or signals sent to it.

[0234] This disclosure also relates to systems for performing the operations described herein. Such systems may be specifically constructed for a desired purpose or may include general-purpose computers that are selectively activated or reconfigured by computer programs stored in a computer. Unless otherwise stated, the order in which operations are run or performed in the embodiments of this disclosure shown and described herein is not mandatory. That is, operations may be performed in any order unless otherwise stated, and embodiments of this disclosure may include additional or fewer operations compared to those disclosed herein. For example, it is contemplated that running or performing a particular operation before, simultaneously with, or after another operation falls within the scope of various aspects of this disclosure.

[0235] In view of the foregoing, it will be seen that several objectives of this disclosure have been achieved and other advantageous results have been obtained. Various aspects of this disclosure have been described in detail, and it will be apparent that modifications and variations are possible without departing from the scope of the various aspects of this disclosure as defined in the appended claims. Since various changes can be made to the above-described constructions, products, and methods without departing from the scope of the various aspects of this disclosure, it is intended that all content contained in the foregoing description and shown in the accompanying drawings be interpreted as illustrative and not restrictive.

[0236] The above description is intended to be illustrative and not restrictive. For example, the examples above (or one or more aspects thereof) may be used in combination with each other. Furthermore, many modifications may be made to adapt a particular situation or material to the teachings of this disclosure without departing from the scope of this disclosure. While the dimensions and types of materials and coatings described herein are intended to define the parameters of this disclosure, they are by no means limiting but rather exemplary embodiments. Many other embodiments will become apparent to those skilled in the art upon review of the above description. Therefore, the scope of this disclosure should be determined by reference to the appended claims and the full scope of their equivalents.

[0237] Furthermore, in the above specific embodiments, various features may be combined to simplify this disclosure. This should not be construed as an intention that any unclaimed disclosed feature is necessary for any of the claims. Rather, the subject matter of the invention may lie in fewer than all features of a particular disclosed embodiment. Therefore, the appended claims are thus incorporated into the specific embodiments, wherein each claim is an independent embodiment. The scope of this disclosure should be determined with reference to the appended claims and the full scope of equivalents to which such claims are conferred. Moreover, the limitations of the appended claims are not written in the form of means plus function, and are not intended to be interpreted based on paragraph 6 of 35 U.S.SC §112, unless and until the limitation of such a claim is expressly expressed using the phrase “means for…” followed by a functional statement without further structural details.

[0238] An abstract is provided to allow readers to quickly determine the substance of the technical disclosure. It should be understood at the time of submission that it will not be used to interpret or limit the scope or meaning of the claims.

Claims

1. A computer-implemented method for generating a dose map used in a radiotherapy treatment plan, the method comprising: Obtain a set of three-dimensional image data corresponding to a subject receiving radiotherapy, the image data indicating one or more target dose regions and one or more organ-at-risk regions in the subject's anatomy; Based on the image data, anatomical projection images are generated, each anatomical projection image providing a view of the subject according to the corresponding beam angle of the radiotherapy treatment; as well as A trained neural network model is used to generate estimated fluence maps based on the anatomical structure projection images, each of the estimated fluence maps indicating the fluence distribution of the radiotherapy treatment at a corresponding beam angle, wherein the neural network model is trained with corresponding pairs of anatomical structure projection images and fluence maps. Each of the estimated fluence diagrams represents a unit beam weight perpendicular to the corresponding beam direction, and the beam angle of the radiotherapy treatment corresponds to the gantry angle of the radiotherapy treatment machine.

2. The method of claim 1, wherein, Each of the estimated flux diagrams uses a two-dimensional array to represent the unit beam weights perpendicular to the corresponding beam direction.

3. The method of claim 2, wherein, Obtaining the set of three-dimensional image data corresponding to the subject includes obtaining image data for each gantry angle of the radiotherapy treatment machine, and wherein each generated anatomical projection image represents a view of the subject's anatomy according to a given gantry angle for providing treatment with a given radiotherapy beam.

4. The method according to claim 1, further comprising: The estimated fluence map is used to determine the radiation dose in the radiotherapy treatment plan, wherein the radiotherapy treatment includes volumetric intensity-modulated rotational therapy (VMAT) performed by a radiotherapy machine, wherein multiple radiotherapy beams are shaped to achieve a modulated dose for a target region from multiple beam angles, thereby delivering a prescribed radiation dose.

5. The method of claim 1, wherein, Each anatomical structure projection image is generated by orthogonally projecting the three-dimensional image data set from corresponding angles of multiple beam angles.

6. The method of claim 1, wherein, The neural network model is trained using paired anatomical projection images and fluence maps from multiple human subjects, wherein each individual pair is provided from the same human subject, and wherein the neural network model is trained using operations including: Multiple sets of training anatomical projection images are obtained, each set of training anatomical projection images indicating one or more target dose regions and one or more organ-at-risk regions in the anatomical structure of the corresponding subject; Obtain multiple sets of training fluence maps corresponding to the projected images of the training anatomy, each set of training fluence maps indicating the fluence distribution of the corresponding subject; and The neural network model is trained based on the training anatomical structure projection image corresponding to the training injection map.

7. The method according to claim 6, wherein, For each beam angle of the radiotherapy treatment machine, corresponding pairs of anatomical projection images and fluence maps are obtained for training the neural network model.

8. The method according to claim 6, wherein, The neural network model is a generative adversarial network (GAN) model comprising at least one generative model and at least one discriminative model, wherein the at least one generative model and the at least one discriminative model correspond to a corresponding generative convolutional neural network and a discriminative convolutional neural network.

9. The method according to claim 8, wherein, The GAN includes Conditional Adversarial Network (cGAN) or CycleGAN.

10. The method according to claim 8, wherein, The GAN is configured to train the generative model using a discriminative model. The neural network parameter values ​​learned by the generative model and the discriminative model are established using adversarial training between the discriminative model and the generative model. The adversarial training includes: The generative model is trained to generate a first estimated fluence map at a first beam angle based on a projected image, the projected image representing a view of the anatomical structure of the training subject according to the first beam angle; and The discrimination model is trained to classify the first estimated injection map as either an estimated or a true training injection map projection image; and The output of the generative model is used to train the discriminative model, and the output of the discriminative model is used to train the generative model.

11. The method according to claim 8, wherein, The GAN is a CycleGAN, a generative adversarial network that includes the generative model and the discriminative model. The generative model is a first generative model, and the discriminative model is a first discriminative model. The CycleGAN further includes: The second generative model, which is trained to: Based on a given pair of anatomical projection images and fluence maps, a given fluence map at a given beam angle is processed as input; and Generate an estimated anatomical projection image as output, the estimated anatomical projection image representing a view of the subject's anatomy based on the given beam angle; and A second discriminant model is trained to classify the estimated anatomical projection image as either an estimated or a true anatomical projection image.

12. The method according to claim 11, wherein, The CycleGAN includes a first part that trains the first generative model, the first part being trained to: A set of training anatomical projection images representing different views of the patient's anatomy obtained from previous treatments, the set of training anatomical projection images being paired with a set of training inflection maps corresponding to each of the different views, each of the training inflection maps being aligned with a corresponding one of the training anatomical projection images. The set of training anatomical structure projection images is input into the first generative model, and the set of estimated fluence maps is output from the first generative model. The estimated set of injection volume graphs is input into the first discriminant model, and the first discriminant model is used to classify the estimated set of injection volume graphs into a set of estimated or true injection volume graphs. as well as The estimated set of fluence maps is input into the second generative model, and a set of estimated anatomical projection images is generated to calculate the cycle consistency loss.

13. The method according to claim 12, wherein, The CycleGAN includes a second part, which is trained to: The set of training injection maps corresponding to each of the different views is input into the second generation model, and the set of generated anatomical structure projection images is output from the second generation model. The generated set of anatomical structure projection images is input into the second discrimination model, and the generated set of anatomical structure projection images is classified into an estimated set of anatomical structure projection images or a set of real anatomical structure projection images. as well as The set of anatomical structure projection images is input into the first generative model to generate a set of estimated fluency maps to calculate the cycle consistency loss.

14. The method according to claim 12, wherein: The cycle consistency loss is generated based on the following: a) Compare the generated set of anatomical structure projection images with the training set of anatomical structure projection images, and b) Compare the generated set of injection maps with the set of training injection maps; Wherein, the first generative model is trained to minimize a first loss term, which represents the expected difference between multiple estimated injection maps and corresponding pairs of training injection maps; and The second generative model is trained to minimize a second loss term, which represents the expected difference between multiple estimated anatomical structure projection images and corresponding pairs of trained anatomical structure projection images.

15. The method according to claim 1, further comprising: The neural network model is used to generate a set of estimated injection maps; as well as The estimated fluence map is used as input to perform numerical optimization, wherein the optimization incorporates radiotherapy treatment constraints to produce a Pareto-optimal fluence plan for use in the subject's radiotherapy treatment plan.

16. The method of claim 15, further comprising: Arc sorting is performed based on the Pareto optimal injection plan to generate a set of initial control points corresponding to each of the multiple radiotherapy beams; as well as Perform direct aperture optimization to generate a set of final control points corresponding to each of the plurality of radiotherapy beams.

17. The method according to claim 16, wherein, The radiotherapy treatment includes volumetric intensity-modulated rotational therapy (VMAT) performed by a radiotherapy machine, wherein the arc sequencing is performed based on the Pareto optimal dose plan, such that the plurality of radiotherapy beams are shaped to achieve modulated coverage of the target dose region from a plurality of beam angles to deliver a prescribed radiation dose.

18. The method of claim 16, further comprising: The set of final control points is used to generate data for controlling radiotherapy treatment, wherein the set of final control points is used to control the position of the multi-leaf collimator (MLC) blades of the radiotherapy treatment machine at a given gantry angle corresponding to a given beam angle.

19. The method according to claim 1, further comprising: The fluence map generated from the neural network model in response to a set of input images of anatomical structures is compared with a fluence map generated from another source.

20. A computer system for generating a fluence map used in a radiotherapy treatment plan, the computer system comprising: One or more memory devices for storing a set of three-dimensional image data corresponding to a subject receiving radiotherapy, the image data indicating one or more target dose regions and one or more organ-at-risk regions in the subject's anatomy; as well as One or more processors, said processors being configured to perform the method according to any one of claims 1 to 19.

21. A computer-readable storage medium comprising computer-readable instructions for generating a fluence map used in a radiotherapy treatment plan, wherein, When executed by a computer system, the instructions cause the computer system to perform the method according to any one of claims 1 to 19.

Images