Optimization of radiotherapy treatment plans using machine learning

Abstract

A technique is provided for solving a radiotherapy treatment planning optimization problem. The technique includes: receiving a radiotherapy treatment planning optimization problem; processing the radiotherapy treatment planning optimization problem using a machine learning model to estimate one or more optimization variables of the radiotherapy treatment planning optimization problem, wherein the machine learning model is trained to establish relationships between multiple parameters of the radiotherapy treatment planning optimization problem and the one or more optimization variables; and generating a solution to the radiotherapy treatment planning optimization problem based on the estimated one or more optimization variables of the radiotherapy treatment planning optimization problem.

Description

[0001] Cross-reference to related applications

[0002] This application claims the benefit of priority to U.S. Application Serial No. 16 / 512,972, filed July 16, 2019, which is incorporated herein by reference in its entirety. Technical Field

[0003] This disclosure generally relates to issues related to radiotherapy or radiotherapy optimization. Background Technology

[0004] Radiation therapy is used to treat cancers and other diseases in mammalian tissues (e.g., humans and animals). The direction and shape 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 surrounding healthy tissue (often referred to as organs at risk (OARs)). Treatment plans can be used to control radiation beam parameters, and radiation therapy devices deliver treatment by providing a spatially varying dose distribution to the patient.

[0005] Typically, for each patient, a radiotherapy treatment plan (“treatment plan”) can be created using optimization techniques based on clinical and dosimetric goals and constraints (e.g., maximum, minimum, and average doses to the tumor and critical organs). The treatment planning process may include using three-dimensional (3D) images of the patient to identify the target area (e.g., the tumor) and critical organs near the tumor. Creating a treatment plan can be a time-consuming process in which the planner attempts to adhere to various treatment goals or constraints, taking into account 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 OARs (Organizational Areas of Radiation) as the number of OARs increases (e.g., typically divided into 21 in head and neck treatments). OARs distant from the tumor may be easily protected from radiation, while OARs close to or overlapping with the target tumor may be difficult to protect from radiation.

[0006] Segmentation can be performed to identify the OAR and the area to be treated (e.g., the planned target volume (PTV)). After segmentation, a dose plan can be created for the patient, indicating the desired radiation dose to be received by one or more PTVs (e.g., targets) and / or one or more OARs. PTVs may have irregular volumes, and their size, shape, and location may be unique. A treatment plan can be calculated after optimizing a large number of planning parameters to ensure an adequate dose is delivered to the PTVs while delivering the lowest possible dose to surrounding healthy tissue. Therefore, a radiotherapy treatment plan can be determined by balancing the treatment of the tumor with efficient control of the dose to protect any OAR from harm. Typically, the quality of a radiotherapy plan can depend on the planner's level of experience. Further complications may arise due to variations in anatomy between patients. Summary of the Invention

[0007] In some embodiments, a computer-implemented method, a non-transitory computer-readable medium, and a system including a memory and a processor are provided for solving a radiotherapy treatment planning optimization problem by: receiving the radiotherapy treatment planning optimization problem by processor circuitry; processing the radiotherapy treatment planning optimization problem by the processor circuitry using a machine learning model to estimate one or more optimization variables of the radiotherapy treatment planning optimization problem, wherein the machine learning model is trained to establish relationships between a plurality of parameters of the radiotherapy treatment planning optimization problem and one or more optimization variables, wherein the machine learning model includes a deep neural network; and generating a solution to the radiotherapy treatment planning optimization problem by the processor circuitry based on the estimated one or more optimization variables of the radiotherapy treatment planning optimization problem, wherein the solution includes at least one of radiotherapy device parameters, fluence map, isocenter position, beam angle, or number of beam activations.

[0008] In some implementations, the machine learning model includes one or more intermediate estimated optimization variables used in subsequent iterations of the machine learning model.

[0009] In some implementations, the radiotherapy treatment planning optimization problem is handled by: using a machine learning model to process the radiotherapy treatment planning optimization problem to estimate one or more initial optimization variables; and using different learning or non-learning optimization processes to solve the radiotherapy treatment planning optimization problem starting from one or more initial optimization variables.

[0010] In some implementations, the non-learning optimization process includes at least one of the following: simplex method, interior point method, Newton's method, quasi-Newton method, Gauss-Newton method, Levenberg-Marquardt method, linear least squares method, gradient descent method, projected gradient method, conjugate gradient method, augmented Lagrangian method, Nelder-Mead method, branch and bound method, cutting plane method, simulated annealing or sequential quadratic programming method.

[0011] In some implementations, the multiple training radiotherapy treatment plan optimization problems include at least one of the following: optimization problems derived from previous radiotherapy treatment plans; or problems generated in combination.

[0012] In some implementations, the radiotherapy treatment planning optimization problem includes at least one of the following: patient volume or patient image, patient volume segmentation, dose kernel, dose-volume histogram constraint, or dose constraint.

[0013] In some implementations, the radiotherapy treatment planning optimization problem is a constrained optimization problem, and additionally, methods, non-transitory computer-readable media, and systems are provided for computer implementations that transform the radiotherapy treatment planning optimization problem into an unconstrained optimization problem based on a merit function; and machine learning models are used to process the transformed radiotherapy treatment planning optimization problem to estimate one or more optimization variables.

[0014] In some implementations, using a machine learning model to process a radiotherapy treatment planning optimization problem to estimate one or more optimization variables is performed by: selecting a first subset of constraints for the radiotherapy treatment planning optimization problem; performing a first iteration of processing the radiotherapy treatment planning optimization problem using a machine learning model based on the selected first subset of constraints to generate a first estimate of one or more optimization variables; selecting a second subset of constraints for the radiotherapy treatment planning optimization problem; and performing a second iteration of processing the radiotherapy treatment planning optimization problem using a machine learning model based on the selected second subset of constraints to generate a second estimate of one or more optimization variables.

[0015] In some implementations, performing one or more additional iterations involves selecting a subset of constraints and utilizing machine learning models to address the radiotherapy treatment planning optimization problem.

[0016] In some implementations, the first and second subsets of constraints are randomly selected.

[0017] In some embodiments, a computer-implemented method, a non-transitory computer-readable medium, and a system including a memory and a processor are provided for training a machine learning model to solve a radiotherapy treatment planning optimization problem by: receiving a plurality of training radiotherapy treatment planning optimization problems by processor circuitry; and training a machine learning model to generate estimates of one or more optimization variables of the radiotherapy treatment planning optimization problem by establishing relationships between parameters of the plurality of training radiotherapy treatment planning optimization problems and one or more optimization variables.

[0018] In some implementations, machine learning models are trained in a supervised manner based on multiple solutions to a training radiotherapy treatment plan optimization problem.

[0019] In some implementations, the machine learning model is trained iteratively in a supervised manner, wherein the intermediate output of a training iteration of the machine learning model, including intermediate estimates of optimization variables, is used in subsequent training iterations of the machine learning model.

[0020] In some implementations, the multiple training radiotherapy treatment plan optimization problems include constrained optimization problems, and the computer-implemented method, non-transitory computer-readable medium, and system further perform: transforming the multiple training radiotherapy treatment plan optimization problems into unconstrained optimization problems based on a value function; and training a machine learning model based on the transformed radiotherapy treatment plan optimization problem.

[0021] In some implementations, machine learning models are trained in an unsupervised manner.

[0022] The foregoing description is intended to provide an overview of the subject matter of this patent application. It is not intended to provide an exclusive or exhaustive description of the disclosure. Detailed descriptions are included to provide further information regarding this patent application. Attached Figure Description

[0023] In the accompanying drawings, which are not necessarily drawn to scale, the same reference numerals describe substantially similar parts throughout several views. The same reference numerals with different letter suffixes indicate different instances of substantially similar parts. The drawings illustrate, by way of example rather than limitation, the various embodiments discussed herein.

[0024] Figure 1 An exemplary radiotherapy system suitable for performing treatment plan generation processing is shown, based on some examples.

[0025] Figure 2A An exemplary image-guided radiotherapy device is shown, representing some examples of the present disclosure.

[0026] Figure 2BExamples of radiotherapy devices, such as the Gamma Knife, are shown in accordance with this disclosure.

[0027] Figure 3 Exemplary data streams for training and using machine learning techniques to solve the problem of optimizing radiotherapy treatment plans are shown, according to some examples of this disclosure.

[0028] Figures 4 to 6 A flowchart illustrating exemplary operations for training and using machine learning techniques to solve a radiotherapy treatment planning optimization problem, based on some examples of this disclosure, is shown. Detailed Implementation

[0029] This disclosure includes various techniques for generating radiotherapy treatment plans by using machine learning (ML) models to estimate one or more optimization variables in a radiotherapy treatment planning optimization problem, which can then be used to solve the problem. The technical benefits include reduced computational processing time for generating and solving the radiotherapy treatment plan optimization problem, along with improvements to processing, memory, and network resources used for these processes. These radiotherapy treatment plans can be applied to various medical treatment and diagnostic settings or radiotherapy treatment devices and equipment. Therefore, in addition to these technical benefits, this technology can also generate numerous significant medical treatment benefits, including improved accuracy of radiotherapy treatments and reduced exposure to unintended radiation.

[0030] Radiation therapy is one of the leading methods for treating cancer and is recommended for more than 50% of all cancer patients. Treatment plans are created through a complex design process involving mathematical optimization problems that capture the desired characteristics of dose delivery—typically requiring a sufficiently high dose to reach the target while minimizing the dose reaching healthy tissue. The overall structure of the optimization problem is the same for most forms of radiation therapy, including linear accelerator-based treatments (3D-CRT, IMRT, VMAT), proton therapy, Gamma Knife radiosurgery, and brachytherapy. The end result is the configuration of the radiation therapy device (e.g., control points) required to deliver the dose distribution.

[0031] Current planning software typically uses standard mathematical optimization methods to solve minimization problems. These methods can be slow, causing unnecessary waiting times for patients and clinicians. Future applications utilizing real-time imaging may even require real-time treatment planning, which cannot be performed using conventional optimization problem solvers.

[0032] The disclosed techniques address these challenges and improve the speed and efficiency of solving radiotherapy treatment planning optimization problems using ML models or parametric methods. Specifically, ML models are used to estimate one or more optimization variables for a given radiotherapy treatment planning optimization problem, simplifying the problem and allowing for faster solution using the estimated optimization variables. By increasing the speed of solving radiotherapy treatment planning optimization problems, the disclosed techniques enable real-time treatment planning and reduce patient and clinician wait times. Estimating optimization variables refers to generating estimates of one or more decision variables for a given optimization problem.

[0033] Specifically, the disclosed technique addresses a radiotherapy treatment planning optimization problem and utilizes a machine learning model to process the problem, estimating one or more optimization variables. The machine learning model is trained to establish relationships between the parameters of the multiple training radiotherapy treatment planning optimization problems and one or more optimization variables. A solution to the radiotherapy treatment planning optimization problem is generated based on the estimated optimization variables. Then, the solved radiotherapy treatment planning optimization problem can be used to determine and generate radiotherapy apparatus parameters (e.g., control points) for a given device.

[0034] Figure 1 An exemplary radiotherapy system 100 is illustrated, suitable for performing radiotherapy planning processing operations using one or more of the methods discussed herein. These radiotherapy planning processing operations are performed to enable the radiotherapy system 100 to deliver radiotherapy to a patient based on specific aspects of captured medical imaging data and therapy dose calculations or radiotherapy machine configuration parameters. Specifically, the following processing operations can be implemented as part of treatment processing logic 120. However, it will be understood that many variations and use cases of the following training models and treatment processing logic 120 can be provided, including data validation, visualization, and other medical assessment and diagnostic settings.

[0035] The radiotherapy system 100 includes a radiotherapy processing computing system 110, which hosts treatment processing logic 120. The radiotherapy processing computing system 110 can be connected to a network (not shown), and such a network can be connected to the Internet. For example, the network can connect the radiotherapy processing computing system 110 to one or more private and / or public medical information sources (e.g., radiology information systems (RIS), medical record systems (e.g., electronic medical record (EMR) / electronic health record (EHR) systems, oncology information systems (OIS)), one or more image data sources 150, image acquisition devices 170 (e.g., imaging modalities), treatment devices 180 (e.g., radiotherapy devices), and treatment data sources 160.

[0036] As an example, the radiotherapy processing computation system 110 can be configured to receive a subject's treatment objective (e.g., from one or more MR images) and generate a radiotherapy treatment plan by executing instructions or data from treatment processing logic 120 as part of the operation of generating a treatment plan that will be used by treatment device 180 and / or output on device 146. In an embodiment, treatment processing logic 120 solves an optimization problem to generate a radiotherapy treatment plan. Treatment processing logic 120 uses a trained ML model to solve the radiotherapy optimization problem, the trained ML model having been trained to establish relationships between parameters of a plurality of training radiotherapy treatment plan optimization problems and one or more optimization variables. In an example, treatment processing logic 120 classifies the received optimization problem into a specific type and then identifies a given ML model that has been trained to solve the optimization problem of that specific type. The given ML model is applied to the received optimization problem to estimate the optimization variables of the received optimization problem. Then, based on the estimated optimization variables, the optimization problem is solved using conventional optimization problem solvers (e.g., simplex method, interior point method, Newton's method, quasi-Newton method, Gauss-Newton method, Levenberg-Marquardt method, linear least squares method, gradient descent method, projected gradient method, conjugate gradient method, augmented Lagrange method, Nelder-Mead method, branch and bound method, cutting plane method, simulated annealing or sequential quadratic programming method).

[0037] The general radiotherapy treatment plan optimization problem can be defined as Equation 1:

[0038]

[0039] Where f: Let X be the objective function, x∈X be the decision variable, and... It is the set of feasible variables. Typically, the function f can be nonlinear, and the set Ω is nonconvex. An iterative scheme of some form is usually used to solve the optimization problem. For example, if f is smooth and convex, and Ω is convex, the projective gradient scheme can be used to solve equation (1) as follows:

[0040]

[0041] Where projΩ: X→X is the projection onto Ω. It is the step size, and X→X is the gradient. While these algorithms are generally provable to be convergent (e.g., given enough time (and the correct parameter choice), the algorithm will converge to a minimum), they are not always very fast and efficient. In fact, some algorithms may require hundreds or even thousands of iterations to achieve approximate convergence. Since each step can be computationally expensive, this can mean a runtime of minutes or even hours. According to the disclosed techniques, solving such optimization problems can be accelerated by using a trained ML model to estimate one or more optimization variables (e.g., x) and then applying conventional methods to solve the optimization problem. In some methods, the ML model provides a solution to the optimization problem within a deviation threshold of the expected or anticipated solution. In such cases, conventional methods may not be necessary because the ML model estimates the solution to the optimization problem.

[0042] In particular, the disclosed implementation improves the speed and efficiency of solving optimization problems using deep learning-based optimization. Specifically, it selects an optimization scheme that is fine-tuned for the type of problem at hand. The deep learning optimization method adds more parameters to the optimization problem in the form of a deep neural network, and then selects the parameters that best solve the optimization problem (e.g., those with a specified threshold deviation from the optimal solution).

[0043] The radiotherapy processing computing system 110 may include processing circuitry 112, memory 114, storage device 116, and other hardware and software operable components such as user interface 142, communication interface (not shown). Storage device 116 may store transient or non-transitory computer-executable instructions, such as operating systems, radiotherapy treatment plans, training data, software programs (e.g., image processing software, image or anatomical visualization software, artificial intelligence (AI) or ML implementations and algorithms such as those provided by deep learning models, ML models, and neural networks (NNs), and any other computer-executable instructions to be executed by processing circuitry 112.

[0044] In the example, processing circuitry 112 may include processing devices, such as one or more general-purpose processing devices like a microprocessor, central processing unit (CPU), graphics processing unit (GPU), accelerated processing unit (APU), etc. More specifically, processing circuitry 112 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. Processing circuitry 112 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.

[0045] As those skilled in the art will recognize, in some examples, the processing circuitry 112 may be a dedicated processor rather than a general-purpose processor. The processing circuitry 112 may include one or more known processing devices, such as those from Intel. TM Manufactured Pentium TM Core TM Xeon TM or The series of microprocessors and those from AMD TM Turion manufactured TM Athlon TM Sempron TM Opteron TM FX TM Phenom TM This refers to any microprocessor from the Sun Microsystems series or any processor from various processors manufactured by Sun Microsystems. The processing circuitry 112 may also include processors from sources such as Nvidia. TM Manufactured series and by Intel TM GMA and Iris manufactured TM Series or by AMD TM Radeon manufactured TM The series of GPUs' graphics processing units. The processing circuitry 112 may also include components such as those from Intel... TM Xeon Phi manufactured TMThe series of accelerated processing units. The disclosed embodiments 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 physical (circuit-based) or software-based processor, such as a multi-core design or multiple processors each having a multi-core design. Processing circuitry 112 can execute sequences of transient or non-transitory computer program instructions stored in memory 114 and accessed from storage device 116 to perform various operations, processes, and methods, which will be described in more detail below. It should be understood that any component of system 100 can be implemented individually and operate as a standalone device, and can be coupled to any other component of system 100 to perform the techniques described in this disclosure.

[0046] Memory 114 may include 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, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage devices, magnetic tape, other magnetic storage devices, or any other non-transitory medium that can be used to store information including images, training data, one or more ML models or technical parameters, data, or transient or non-transitory computer-executable instructions (e.g., stored in any format) that can be accessed by processing circuitry 112 or any other type of computer device. For example, computer program instructions can be accessed by processing circuitry 112, can be read from ROM or any other suitable memory location, and can be loaded into RAM for execution by processing circuitry 112.

[0047] Storage device 116 may constitute a drive unit including transient or non-transient machine-readable medium on which one or more transient or non-transient instruction sets and data structures (e.g., software) implemented or utilized by any one or more of the methods or functions described herein (including, in various examples, treatment processing logic 120 and user interface 142). During the execution of instructions by the radiotherapy processing computing system 110, the instructions may also reside wholly or at least partially in memory 114 and / or processing circuitry 112, wherein memory 114 and processing circuitry 112 also constitute transient or non-transient machine-readable medium.

[0048] Memory 114 and storage device 116 can constitute a non-transitory computer-readable medium. For example, memory 114 and storage device 116 can store or load transient or non-transitory instructions for one or more software applications onto the computer-readable medium. Software applications stored or loaded using memory 114 and storage device 116 can include, for example, operating systems for general-purpose computer systems and devices for software control. The radiotherapy processing computing system 110 can also operate various software programs including software code for implementing treatment processing logic 120 and user interface 142. Furthermore, memory 114 and storage device 116 can store or load the entire software application, a portion of the software application, or code or data associated with the software application that can be executed by processing circuitry 112. In other examples, memory 114 and storage device 116 can store, load, and manipulate one or more radiotherapy plans, imaging data, segmentation data, treatment visualizations, histograms or measurements, one or more AI model data (e.g., weights and parameters of the ML model of the disclosed embodiments), training data, labeling and mapping data, etc. It is anticipated that the software program can be stored not only on storage device 116 and memory 114, but also on removable computer media such as hard disk drive, computer disk, CD-ROM, DVD, Blu-ray DVD, USB flash drive, SD card, memory stick or any other suitable media; such software programs can also be transmitted or received over a network.

[0049] Although not depicted, the radiotherapy processing computing system 110 may include communication interfaces, network interface cards, and communication circuitry. Example communication interfaces 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., IEEE 802.11 / Wi-Fi adapters), telecommunications adapters (e.g., for communicating with 3G, 4G / LTE, and 5G networks, etc.). Such communication interfaces may include one or more digital and / or analog communication devices that allow the machine to communicate via a network with other machines and devices, such as remote components. The network 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, the network may be a LAN or WAN that may include other systems, including additional image processing computing systems or image-based components associated with medical imaging or radiotherapy operations.

[0050] In one example, the radiotherapy processing computing system 110 may obtain image data 152 from an image data source 150 (e.g., MR images) to be hosted on storage device 116 and memory 114. In yet another example, a software program may replace the function of patient images, such as a processed version of the image that emphasizes certain aspects of the image information or a symbolic distance function.

[0051] In this example, the radiotherapy processing computing system 110 may obtain image data 152 from image data source 150 or transmit image data 152 to image data source 150. In other examples, treatment data source 160 receives or updates planning data as a result of a treatment plan generated by treatment processing logic 120. Image data source 150 may also provide or host imaging data for use in treatment processing logic 120.

[0052] In the example, the computational system 110 can communicate with the treatment data source 160 and the input device 148 to generate: pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems; pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems and solutions for multiple training radiotherapy optimization problems; and pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems of a given type.

[0053] Processing circuitry 112 may be communicatively coupled to memory 114 and storage device 116, and processing circuitry 112 may be configured to execute computer-executable instructions stored thereon from memory 114 or storage device 116. Processing circuitry 112 may execute instructions to cause medical images from image data 152 to be received or acquired in memory 114 and processed using treatment processing logic 120 to generate a treatment plan. Specifically, treatment processing logic 120 receives an optimization problem derived based on the received medical images. Treatment processing logic implements a trained ML model, which is applied to the optimization problem to estimate one or more optimization variables of the optimization problem. Once the optimization variables are estimated, the received optimization problem is solved using the estimated optimization variables to generate a treatment plan.

[0054] Additionally, processing circuitry 112 can utilize software programs to generate intermediate data, such as updated parameters to be used by the NN model, machine learning model, treatment processing logic 120, or other aspects involved in the generation of treatment plans as discussed herein. Furthermore, such software programs can utilize the techniques further discussed herein, employing treatment processing logic 120 to generate new or updated treatment plan parameters for deployment to treatment data source 160 and / or presentation on output device 146. Processing circuitry 112 can then subsequently transmit the new or updated treatment plan parameters to treatment device 180 via a communication interface and network, in which a radiotherapy plan will be used to treat the patient with radiation via treatment device 180, consistent with the results of the trained ML model implemented by treatment processing logic 120 (e.g., according to the following combination). Figure 3 (Discussion processing).

[0055] In the examples herein, processing circuitry 112 may execute a software program that invokes treatment processing logic 120 to implement ML, deep learning, NN functions, and other aspects of artificial intelligence for generating treatment plans based on input radiotherapy medical information (e.g., CT images, MR images, and / or sCT images and / or dose information). For example, processing circuitry 112 may execute a software program that trains, analyzes, predicts, evaluates, and generates treatment plan parameters based on received radiotherapy medical information as discussed herein.

[0056] In the example, image data 152 may include one or more MRI images (e.g., 2D MRI, 3D MRI, 2D flow cytometry MRI, 4D MRI, 4D volumetric MRI, 4D imaging MRI, etc.), functional MRI images (e.g., fMRI, DCE-MRI, diffusion MRI), computed tomography (CT) images (e.g., 2D CT, 2D cone-beam CT, 3D CT, 3D CBCT, 4D CT, 4D CBCT), ultrasound images (e.g., 2D ultrasound, 3D ultrasound, 4D ultrasound), positron emission tomography (PET) images, X-ray images, fluorescence microscopy images, radiotherapy field images, single-photon emission computed tomography (SPECT) images, computer-generated composite images (e.g., pseudo-CT images), etc. Furthermore, image data 152 may also include or be associated with medical image processing data, such as training images, ground truth images, contour images, and dose images. In other examples, equivalent representations of anatomical regions can be represented in non-image formats (e.g., coordinates, mappings, etc.).

[0057] In the example, image data 152 can be received from image acquisition device 170, and this image data 152 can be stored in one or more image data sources 150 (e.g., Picture Archiving and Communication System (PACS), Vendor-Neutral Archive (VNA), medical record or information system, data warehouse, etc.). Therefore, image acquisition device 170 can include MRI imaging equipment, CT imaging equipment, PET imaging equipment, ultrasound imaging equipment, fluorescence microscope equipment, SPECT imaging equipment, integrated linear accelerator and MRI imaging equipment, CBCT imaging equipment, or other medical imaging equipment for acquiring medical images of a patient. Image data 152 can be received and stored in any data type or format type (e.g., in Medical Digital Imaging and Communication (DICOM) format) that can be used by image acquisition device 170 and radiotherapy processing computing system 110 to perform operations consistent with the disclosed embodiments. Furthermore, in some examples, the models discussed herein can be trained to process the raw image data format or its derivatives.

[0058] In the example, the image acquisition device 170 can be integrated with the treatment device 180 as a single device (e.g., an MRI apparatus combined with a linear accelerator, also referred to as an "MRI-Linac"). For example, such an MRI-Linac can be used to determine the location of a target organ or target tumor in a patient to accurately guide radiotherapy to a predetermined target according to a radiotherapy treatment plan. For example, a radiotherapy treatment plan can provide information about the specific radiation dose to be applied to each patient. The radiotherapy treatment plan may also include other radiotherapy information (including control points of the radiotherapy treatment device), such as treatment bed location, beam intensity, beam angle, dose-histogram-volume information, the number of radiation beams to be used during treatment, the dose per beam, etc.

[0059] The radiotherapy processing computing system 110 can communicate with an external database via a network to send / receive various types of data related to image processing and radiotherapy operations. For example, the external database may include machine data (including equipment constraints) providing information associated with the treatment device 180, image acquisition device 170, or other machines related to radiotherapy or medical procedures. Machine data information (e.g., control points) may include beam size, arc position, beam on and off duration, machine parameters, segments, multi-leaf collimator (MLC) configuration, gantry speed, MRI pulse sequences, etc. The external database may be a storage device and may be equipped with appropriate database management software programs. Furthermore, such a database or data source may include multiple devices or systems located in a centralized or distributed manner.

[0060] The radiotherapy processing computing system 110 can use one or more communication interfaces to collect and acquire data via a network and communicate with other systems. These communication interfaces are communicatively coupled to the processing circuitry 112 and the memory 114. For example, the communication interfaces can provide communication connections between the radiotherapy processing computing system 110 and components of the radiotherapy system (e.g., allowing data exchange with external devices). For example, in some examples, the communication interfaces may have appropriate interface circuitry with output device 146 or input device 148 to connect to a user interface 142, which may be a hardware keyboard, keypad, or touchscreen through which a user can input information into the radiotherapy system.

[0061] As an example, output device 146 may include a display device that outputs: a representation of user interface 142; and one or more aspects, visualizations, or representations of medical images, treatment plans, and the status of training, generation, verification, or implementation of such plans. Output device 146 may include one or more displays showing medical images, interface information, treatment plan parameters (e.g., contours, doses, beam angles, markings, graphs, etc.), treatment plans, targets, target localization and / or target tracking, or any user-related information. Input device 148 connected to user interface 142 may be a keyboard, keypad, touchscreen, or any type of device that the user can use with the radiotherapy system 100. Alternatively, features of output device 146, input device 148, and user interface 142 may be integrated into a device such as a smartphone or tablet (e.g., Apple). Lenovo Samsung In a single device (etc.).

[0062] Furthermore, any and all components of the radiotherapy system can be implemented as virtual machines (e.g., via virtualization platforms such as VMware, Hyper-V, etc.) or standalone devices. For example, a virtual machine can be software used 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 collectively serve as hardware. For example, the radiotherapy processing computing system 110, image data source 150, or similar components can be implemented as virtual machines or implemented within a cloud-based virtualization environment.

[0063] Image acquisition device 170 can be configured to acquire one or more images of a patient's anatomy for 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, etc.). In the example, image acquisition device 170 can acquire 2D slices of any orientation. For example, the orientation of a 2D slice can include sagittal orientation, oronal orientation, or axial orientation. Processing circuitry 112 can adjust one or more parameters (e.g., 2D slice thickness and / or orientation) to include the target organ and / or target tumor. In the example, 2D slices can be determined based on information such as 3D CBCT or CT or MRI volumes. When a patient is undergoing radiation therapy (e.g., when using treatment device 180), such 2D slices can be acquired by image acquisition device 170 "near real-time" (where "near real-time" means acquiring data in at least milliseconds or less).

[0064] The treatment processing logic 120 in the radiotherapy processing computation system 110 implements an ML model, which involves using a trained (learned) ML model. This ML model can be provided by a neural network (NN) trained as part of an NN model. One or more teacher ML models can be provided by different entities or at an off-site facility relative to the treatment processing logic 120, and can be accessed by issuing one or more queries to the off-site facility.

[0065] Supervised machine learning (ML) algorithms, models, or techniques can be summarized as function approximation. Training data consisting of some type of input-output pairs is obtained from, for example, an expert clinician or an existing optimization plan solver (e.g., one or more training optimization variables and training parameters for multiple training radiotherapy treatment plan optimization problems), and a function is "trained" to approximate this mapping. Some methods involve neural networks (NNs). Here, a set of parameterized functions Aθ is chosen, where θ is a set of parameters (e.g., convolution kernel and bias) chosen by minimizing the average error on the training data. If (x... m y m If ) represents an input-output pair, then the function can be formalized by solving a minimization problem such as Equation 2:

[0066]

[0067] Once the network has been trained (e.g., θ has been chosen), the function Aθ can be applied to any new input. For example, in the above setting of radiotherapy treatment planning optimization problem variables, a radiotherapy treatment planning optimization problem that has never been seen before can be fed into Aθ, and one or more radiotherapy treatment planning optimization problem variables that match the optimization problem solver will be estimated.

[0068] A simple neural network consists of an input layer, intermediate or hidden layers, and an output layer, each containing computational units or nodes. Nodes in one or more hidden layers have inputs from all the nodes in the input layer and are connected to all nodes in the output layer. Such a network is called "fully connected." Each node transmits a signal to the output node according to a non-linear function of the sum of its inputs. For a classifier, the number of nodes in the input layer is typically equal to the number of features for each object in the set of objects to be classified into a class, and the number of nodes in the output layer is equal to the number of classes. The network is trained by presenting the network with features of objects of known classes and adjusting the node weights using an algorithm called backpropagation to reduce training error. Thus, a trained network can classify new objects whose classes are unknown.

[0069] Neural networks have the ability to discover relationships between data and classes or regression values, and under certain conditions, they can model any function y = f(x), including nonlinear functions. In machine learning, it is assumed that both training and test data are generated through the same data generation process p. data Generate, where each {x i y i The samples are all identically and independently distributed (i.e., i ...

[0070] Figure 2A An exemplary image-guided radiotherapy apparatus 232 is shown, comprising: a radiation source such as an X-ray source or a linear accelerator, a bed 246, an imaging detector 244, and a radiotherapy output 234. The radiotherapy apparatus 232 can be configured to emit a radiotherapy beam 238 to provide treatment to a patient. The radiotherapy output 234 may include one or more attenuators or collimators (e.g., MLCs).

[0071] As an example, a patient can be placed in area 242 supported by treatment bed 246 to receive a radiotherapy dose according to a radiotherapy treatment plan. Radiotherapy output 234 can be mounted or attached to frame 236 or other mechanical support. When bed 246 is inserted into the treatment area, one or more chassis motors (not shown) can rotate frame 236 and radiotherapy output 234 about bed 246. In this example, frame 236 can rotate continuously about bed 246 when bed 246 is inserted into the treatment area. In another example, frame 236 can rotate to a predetermined position when bed 246 is inserted into the treatment area. For example, frame 236 can be configured to rotate treatment output 234 about an axis (“A”). Both the bed 246 and the radiotherapy output unit 234 can be moved independently to other locations around the patient, for example, by moving in a lateral direction (“T”), by moving in a side direction (“L”), or by rotating about one or more other axes, such as about a transverse axis (indicated as “R”). A controller communicatively connected to one or more actuators (not shown) controls the movement or rotation of the bed 246 according to the radiotherapy treatment plan to properly position the patient within or outside the radiotherapy beam 238. The fact that both the bed 246 and the gantry 236 can move independently of each other in multiple degrees of freedom allows the patient to be positioned such that the radiotherapy beam 238 can be precisely targeted at the tumor.

[0072] The coordinate system (including axes A, T, and L) may have an origin located at isocenter 240. Isocenter 240 may be defined as a location where the central axis of the radiotherapy beam 238 intersects the origin of the coordinate axes, for example, at a location where a prescribed radiation dose is delivered to the patient or within the patient's body. Alternatively, isocenter 240 may be defined as a location where, for various rotational positions of the radiotherapy output section 234 positioned by the gantry 236 about axis A, the central axis of the radiotherapy beam 238 intersects the patient.

[0073] The gantry 236 may also have an attached imaging detector 244. The imaging detector 244 is preferably located opposite the radiation source (output 234), and in this example, it may be located within the field of the therapy beam 238. The imaging detector 244 may be mounted on the gantry 236, preferably opposite the radiotherapy output 234, to maintain alignment with the radiotherapy beam 238. As the gantry 236 rotates, the imaging detector 244 rotates about a rotation axis. In this example, the imaging detector 244 may be a flat panel detector (e.g., a direct detector or a scintillator detector). In this way, the imaging detector 244 can be used to monitor the radiotherapy beam 238, or it can be used to image the patient's anatomy, such as through field imaging. The control circuitry of the radiotherapy device 232 may be integrated within or located away from the radiotherapy system 100.

[0074] In the illustrative example, one or more of the bed 246, therapy output 234, or gantry 236 can be automatically positioned, and the therapy output 234 can establish a therapy beam 238 according to a specified dose for a particular therapy delivery instance. The sequence of therapy delivery can be specified according to a radiotherapy treatment plan, for example, using one or more different orientations or positions of the gantry 236, bed 246, or therapy output 234. Therapy deliveries can occur sequentially, but can intersect on the patient or at the desired treatment site within the patient, for example, at isocenter 240. Thus, a prescribed cumulative dose of radiotherapy can be delivered to the therapy site while minimizing or avoiding damage to tissues near the therapy site.

[0075] therefore, Figure 2A Specifically, an example of a radiotherapy device 232 is shown, operable to provide radiotherapy treatment to a patient in accordance with or according to a radiotherapy treatment plan. The radiotherapy device 232 has 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, for example, located in a region laterally separated from the patient, and a platform supporting the patient can be used to align the center of the radiotherapy with a designated target site within the patient's body. In yet another example, the radiotherapy device can be a combination of a linear accelerator and an image acquisition device. As those skilled in the art will recognize, in some examples, the image acquisition device can be an MRI, X-ray, CT, CBCT, spiral CT, PET, SPECT, optical tomography, fluorescence imaging, ultrasound imaging, or radiotherapy field imaging equipment, etc.

[0076] Figure 2B A radiotherapy device 130 is shown, in which the Gamma Knife of this disclosure can be used. A patient 202 may wear a coordinate frame 220 to stabilize body parts (e.g., the head) of the patient undergoing surgery or radiotherapy. The coordinate frame 220 and patient positioning system 222 can establish a spatial coordinate system that can be used during patient imaging or during radiosurgery. The radiotherapy device 130 may include a protective housing 214 to enclose multiple radiation sources 212 for generating radiation beams (e.g., sub-beams) passing through a beam channel 216. The multiple beams may be configured to be focused at an isocenter 218 from different locations. While each individual radiation beam may have a relatively low intensity, the isocenter 218 can receive a relatively high level of radiation when multiple doses from different radiation beams accumulate at the isocenter 218. In some embodiments, the isocenter 218 may correspond to a target (e.g., a tumor) being treated or operated on.

[0077] As an example of implementation, the output element may include the dose of a voxel to be applied to a specific OAR. Furthermore, feature elements may be used to determine the output element. Feature elements may include the distance between a voxel in the OAR and the nearest boundary voxel in the target tumor. Therefore, a feature element may include a symbolic distance x, indicating the distance between a voxel in the OAR and the nearest boundary voxel in the target for radiotherapy. The output element may include a dose D in the voxel of the OAR, with x measured according to this dose D. In some other implementations, each training sample may correspond to a specific voxel in the target or OAR, such that multiple training samples within the training data correspond to the entire volume of the target or OAR and other anatomical portions subjected to radiotherapy.

[0078] Figure 3 An exemplary data flow for training and using a machine learning model to solve a radiotherapy treatment planning optimization problem is shown, according to some examples of this disclosure. The data flow includes training input 310, ML model (technical) training 330, and model usage 350.

[0079] Training input 310 includes model parameters 312 and training data 320. Training data 320 may include paired training datasets 322 (e.g., input-output training pairs) and constraints 326. Model parameters 312 store or provide the machine learning model. The parameters or coefficients of the corresponding model. During training, these parameters 312 are adapted based on the input-output training pairs of the training dataset 322. After adapting the parameters 312 (after training), the trained therapeutic model 360 uses these parameters to implement the trained machine learning model on a new dataset 370. The corresponding one in the list.

[0080] Training data 320 includes constraints 326, which may define physical constraints for a given radiotherapy apparatus. Paired training datasets 322 may include sets of input-output pairs, such as pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems; pairs of training parameters and one or more training optimization variables and solutions for multiple training radiotherapy treatment planning optimization problems; and pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems of a given type. Some components of training input 310 may be stored separately from other components in different one or more off-site facilities. In this implementation, each machine learning model can be trained on a specific type of optimization problem that has already been classified. Specifically, the classification problems in the training data 230 can be grouped according to type, and each specific type of training data is used to train the corresponding machine learning model. In this way, when a new optimization problem is encountered or received, the new optimization problem is classified to determine its type, and the corresponding ML model trained on that specific type can be accessed to estimate the optimization variables of the new optimization problem.

[0081] Machine learning model training 330 is based on a set of input-output pairs in a pairwise training dataset 322 to train one or more machine learning techniques. For example, model training 330 can train ML model parameters 312 by minimizing a first loss function based on the corresponding training parameters of a corresponding problem in a multiple training radiotherapy treatment plan optimization problem and one or more training optimization variables.

[0082] The result of minimizing the loss function of multiple training datasets is used to train, adapt, or optimize the model parameters 312 of the corresponding ML model. In this way, the ML model is trained to establish the relationship between the parameters of multiple training radiotherapy treatment planning optimization problems and one or more optimization variables.

[0083] In some implementations, an ML model is trained to estimate one or more optimization variables of a treatment planning optimization problem in closed-form. In such cases, one or more optimization variables can be used to directly solve a given radiotherapy treatment planning optimization problem. Specifically, a closed-form solution defines the mapping of f (the optimization problem function) to the optimal solution x. * Approximate neural network Λ θ .

[0084] In one implementation, the ML model is trained using supervised learning techniques to apply and provide closed-form solutions. Supervised learning techniques assume that, based on previous solutions to the optimization problem, the underlying principles are known. In this case, define a distance function d such that... This represents the distance to the solution. In this case, in order to train the ML model Λ θ Retrieve multiple training optimization problems previously solved for other patients (and / or including problems generated in combination) and their corresponding training parameters (e.g., optimization variables and solutions). Apply the ML model to the first batch of training optimization problems to estimate a given set of parameters (e.g., optimization variables and / or solutions). This batch of training optimization problems can be used to train the ML model using the same parameters, and can be in the range from one training optimization problem to all training problems. Compare the output or result of the ML model with the corresponding training parameters of the first batch of training optimization problems, and use the loss function l(θ, f) or L(θ) discussed below to calculate the deviation between the output or result and the corresponding training parameters of the first batch of training optimization problems. Based on this deviation, calculate the updated parameters of the ML model. Then apply the ML model with the updated parameters to a second batch of training optimization problems to again estimate a given set of parameters for comparison with the parameters previously determined for the second batch of training optimization problems. Update the parameters of the ML model again, and the iteration of this training process continues for a specified number of iterations or periods, or until a given convergence criterion has been met.

[0085] As mentioned in this disclosure, the “objective function” f can include any one or a combination of the following: the true objective function of the optimization problem; an extended objective function that is consistent with the true objective function on the feasible set but infinite outside the feasible set; a value function, which is a “more flexible” version of the extended objective function, wherein the value function is consistent with the true objective function on the feasible set and grows larger (but still finite) further away from the feasible set; and / or a relaxation of the objective function. For example, linear programming relaxation of integer programming problems eliminates integrity constraints and thus allows non-integer understanding. Lagrange relaxation of complex problems in combinatorial optimization penalizes violations of certain constraints, making it easier to solve the relaxation problem. Relaxation techniques complement or supplement branch and bound algorithms for combinatorial optimization; linear programming and Lagrange relaxation are used to obtain the boundaries in branch and bound algorithms for integer programming.

[0086] In one implementation, an ML model is trained using unsupervised learning techniques to apply and provide closed-form solutions, where the true solution (whether or not it is known) is not used. This is achieved by f(Λ) θ (f) Define a closed-form function value. In this case, to train the ML model Λ θRetrieve multiple training optimization problems for other patients (and / or including comprehensively generated problems). Apply the ML model to the first batch of training optimization problems to estimate a given set of parameters (e.g., optimization variables, efficient set of constraints (e.g., constraints efficient at the solution), and / or the solution). This batch of training optimization problems can be used to train the ML model using the same parameters, and can be used across all training optimization problems. Evaluate the output or outcome of the ML model using a loss function l(θ, f) or L(θ) discussed below to obtain feedback on the loss / utility of the current iteration. Based on this loss function, compute the updated parameters of the ML model. Then apply the ML model with the updated parameters to a second batch of training optimization problems to estimate the given set of parameters again. Update the parameters of the ML model again, and the iteration of this training process continues for a specified number of iterations or periods, or until a given convergence criterion has been met.

[0087] In some implementations, an ML model is trained to iteratively estimate one or more optimization variables for a radiotherapy treatment planning optimization problem. In such cases, the ML model estimates one or more intermediate optimization variables for one or more iterations used to solve the optimization problem. In some implementations, the intermediate optimization variables from one iteration of the optimization problem using the ML model are recursively processed by another iteration using the ML model. After a certain number of iterations or when a stopping criterion is met, the solution to the optimization problem can be output and used to generate a radiotherapy treatment plan. Alternatively, after the stopping criterion is met, other optimization problem-solving techniques can be used to further solve the optimization problem with the final estimated intermediate optimization variables. The iterative solution defines the initial guess x0 and the neural network Λ. θ Neural network Λ θ Map each iteration to the next x n+1 =Γ θ (f, x) n (x, ..., x0), where the objective is x n →x * .

[0088] In one implementation, the ML model is trained using supervised training techniques to apply and provide iterative solutions. In this case, a distance function d is defined such that the distance to the solution after n iterations is... In some cases, ML models are trained to estimate optimization variables such that the error is minimized after a given number of N iterations. In other cases, the ML model is trained to estimate the optimization variables such that the number of iterations required to reach a specified error is minimized. In other cases, the ML model is trained to estimate the optimization variables such that the weighted average function value over a given set of iterations is minimized, i.e., l(θ, f) = ∑ n w n d(x n x * In some cases, in order to train an iterative ML model Λ θ Retrieve multiple training optimization problems previously solved iteratively for other patients (and / or including problems generated in combination) (for a specified fixed number of iterations or already solved) along with their corresponding iterative training parameters (e.g., optimization variables and solutions for each iteration, or the final set of parameters corresponding to the solutions). Apply the ML model to the first batch of training optimization problems to estimate a given set of intermediate parameters (e.g., optimization variables and / or solutions). This batch of training optimization problems can be used to train the ML model using the same parameters, and can range from one training optimization problem to all training problems. Compare the output or result of the ML model with the corresponding training parameters of the first batch of training optimization problems, and calculate the deviation between the output or result and the corresponding training parameters of the first batch of training optimization problems using the loss function l(θ, f) or L(θ) discussed below. Based on this deviation, calculate the updated parameters of the ML model. Then apply the ML model with the updated parameters to a second batch of training optimization problems to again estimate a given set of parameters for comparison with the parameters previously determined for the second batch of training optimization problems. Update the parameters of the ML model again, and the iteration of this training process continues for a specified number of periods or until all training optimization problems have been processed.

[0089] In one implementation, an ML model is trained using unsupervised learning techniques to apply and provide iterative solutions, without using the actual solution (regardless of whether it is known). In such a case, to train the ML model Λ θ Retrieve multiple training optimization problems for other patients (and / or including comprehensively generated problems). Apply the ML model to the first batch of training optimization problems to estimate a given set of parameters (e.g., optimization variables, efficient set of constraints (e.g., constraints efficient at the solution), and / or the solution). Recursively and iteratively apply the ML model to the first batch of training optimization problems, such that intermediate optimization variables from one iteration of the optimization problem handled by the ML model are recursively handled by another iteration of the optimization problem handled by the ML model, until a certain number of iterations or when a stopping criterion is met. After the stopping criterion is met, the output or outcome of the ML model is evaluated using the loss function l(θ, f) or L(θ) discussed below to obtain feedback on the loss / utility of the current iteration. That is, for some The function value after a fixed number of iterations is l(θ, f) = f(x).N ). Calculate the weighted average function value l(θ, f) = ∑ n w n f(x n ), where wn is the weight. Based on this loss function, the updated parameters of the ML model are calculated. The ML model, along with the updated parameters, is then iteratively and recursively applied to a second batch of training optimization problems to iteratively estimate the given set of parameters again. The parameters of the ML model are updated again, and this training process iterates for a specified number of iterations or periods, or until a given convergence criterion has been met.

[0090] Specifically, based on a loss function, an ML model is trained in a supervised or unsupervised manner such that the set of intermediate parameters estimated by the ML model for a given optimization problem represents the solution to the given optimization problem after a specified fixed number of iterations, or after the problem has been fully solved. The ML model is trained until a stopping criterion is met (e.g., the maximum number of iterations has been reached, the objective value has decreased, the step size has been met, etc.) or when the solution is within a specified threshold error of the final solution to the given optimization problem, or within a specified threshold error of the solution after a specified number of iterations. In this way, the trained ML model can be applied to new optimization problems to estimate one or more optimization variables for the new optimization problem. In some cases, the new optimization problem can be solved iteratively and recursively using the estimated optimization variables by recursively applying the ML model for multiple iterations. In some cases, the number of iterations required to recursively solve the new optimization problem is less than the total number of iterations required to solve the optimization problem using conventional techniques with the initial optimization variables. In some cases, other optimization solving techniques can be used to solve new optimization problems with estimated optimization variables generated by applying the ML model.

[0091] In some cases, the trained ML model depends on the classification of the optimization problem in the training set. For example, if the objective function f is parameterized by images (e.g., dosing plans and tomographic images), the closed-form ML model can be a U-Net that can be used to predict the optimal decision variables. For the ML model used to provide iterative solutions, a learning scheme for projected gradients can be used, as given below: in X→X is a deep neural network and θ=(θ1,...,θ) N ) are their parameters.

[0092] Loss functions are defined for specific optimization problems, but the actual problem to be solved is unknown. Therefore, an optimization solver cannot solve only a single specific optimization problem. Instead, the solver needs to handle a series of related problems. Therefore, the loss function should depend only on the parameter choice L(θ). For this purpose, assume that f is randomly drawn from a set of possible optimization problems with assigned probability distributions.

[0093] In each training machine learning model (sometimes called Λ) θ After that, new data 370 can be received, including one or more patient input parameters (e.g., a radiotherapy treatment planning optimization problem). Trained machine learning techniques can then be used. The new data 370 is applied to generate a result 380 that includes one or more estimated optimization variables for a radiotherapy treatment planning optimization problem. Then, the generated estimated optimization variables for the radiotherapy treatment planning optimization problem are used to solve the received optimization problem, for example, by using methods such as: simplex method, interior point method, Newton's method, quasi-Newton method, Gauss-Newton method, Levenberg-Marquardt method, linear least squares method, gradient descent method, projected gradient method, conjugate gradient method, augmented Lagrange method, Nelder-Mead method, branch and bound method, cutting plane method, simulated annealing, and / or sequential quadratic programming.

[0094] In some implementations, the estimated optimization variables provided by the ML model can be used to solve a new radiotherapy treatment planning optimization problem (which includes at least one of the following: patient volume or an image of the patient, segmentation of the patient volume, dose kernel, dose-volume histogram constraint, or dose constraint). The solution to the optimization problem includes at least one of the following: radiotherapy device parameters, fluence map, irradiation location, or beam on-time. In some cases, to further simplify solving a constrained new radiotherapy treatment planning optimization problem (e.g., an optimization problem with certain constraints), the new radiotherapy treatment planning optimization problem can first be transformed into an unconstrained optimization problem. This can be done before, during, or after the ML model estimates the optimization variables for the new radiotherapy treatment planning optimization problem. To represent the constrained radiotherapy treatment planning optimization problem as an unconstrained optimization problem, a value function can be used. If the optimization problem is transformed into an unconstrained optimization problem before the ML model estimates the optimization variables for the new optimization problem, and then these optimization variables are used to solve the new optimization problem.

[0095] As an example, the general radiotherapy treatment planning optimization problem can usually be written as:

[0096]

[0097] Obey c i (x) = 0, i ∈ ε

[0098] c i (x)≥0, i∈I

[0099] Where f is the objective function, x is the decision variable, and {ci} is the set of functions, where ε and I are the sets of indices corresponding to equality and inequality constraints, respectively.

[0100] Example value functions for nonlinear programming problems include the l1 penalty function:

[0101]

[0102] The positive scalar μ is the penalty parameter, which determines the weight assigned to the constraint satisfaction relative to the goal. The l1 penalty function is an example of an exact value function, which means that if μ is large enough, any local solution to the optimization problem is a local minimum of φ(x; μ). Therefore, the disclosed technique can utilize suitable algorithms for unconstrained optimization on unconstrained problems: A fixed but limited number of iterations are performed, and the result is used as an initial guess for a general solver used for constrained optimization. Specifically, after applying the value function to the radiotherapy treatment plan optimization problem to transform it from a constrained optimization problem to an unconstrained optimization problem, a less complex unconstrained optimization problem solver can be used. After an appropriate number of iterations with the unconstrained optimization problem solver, the unconstrained optimization problem can be processed by the constrained optimization problem solver.

[0103] In some cases, after an appropriate number of iterations with an unconstrained optimization problem solver, an ML model is applied to an unconstrained optimization problem to estimate one or more radiotherapy treatment plan optimization variables. These estimated radiotherapy treatment plan optimization variables can then be applied to a radiotherapy treatment plan optimization problem that can be solved by a constrained optimization problem solver.

[0104] The objective function of the radiotherapy treatment planning optimization problem can be decomposed into separate parts. For example, some parts include different objectives related to the degree to which the tumor is targeted, while others focus on not harming healthy tissue. Similarly, the set of feasible variables in the radiotherapy treatment planning optimization problem can be decomposed. For example, the on-time of each beam may need to be positive. This leads to an optimization problem that can be defined as:

[0105]

[0106] obey

[0107] In these cases, iterative optimization algorithms can consider this structure to accelerate the optimization process. For example, in randomized constraint projection, projection is performed only on a subset of all constraints in each iteration:

[0108]

[0109]

[0110] Specifically, to further accelerate the solution of the radiotherapy treatment planning optimization problem, a first set of constraints can be (e.g., randomly) selected, and the radiotherapy treatment planning optimization problem is solved based on the selected first set of constraints in the first iteration. Subsequently, in one or more further iterations, a second set of constraints can be (e.g., randomly) selected, and the radiotherapy treatment planning optimization problem is solved based on the selected second set of constraints in one or more further iterations. In some cases, at each iteration, an ML model suitable for providing the iterative solution or intermediate optimization variables can be applied as follows:

[0111]

[0112]

[0113] As an example, a first ML model can be trained to provide intermediate iterative optimization problem variables (e.g., solutions) for a given radiotherapy treatment planning optimization problem considering a specific set of constraints (e.g., a specific number of constraints) after the first iteration (e.g., one iteration solving the optimization problem). In such a case, in the first iteration, the ML model can be applied to a new radiotherapy treatment planning optimization problem for which a random set of constraints (corresponding to a certain number of constraints) is selected. The ML model provides intermediate iterative optimization problem variables (e.g., intermediate solutions or updated values ​​of the optimization problem decision variables) for the new radiotherapy treatment planning optimization problem. In the second iteration, a second set of constraints for the radiotherapy treatment planning optimization problem can be selected, and the first or second ML model can be applied to the radiotherapy treatment planning optimization problem using the second set of constraints to provide updated intermediate iterative optimization problem variables for the new radiotherapy treatment planning optimization problem. In some cases, a second ML model can be trained to provide intermediate iterative optimization problem variables (e.g., intermediate solutions or updated values ​​of optimization problem decision variables) for a given radiotherapy treatment planning optimization problem considering a specific set of constraints (e.g., a specific number of constraints) after a second number of iterations (e.g., two iterations solving the optimization problem). The updated intermediate iterative optimization problem variables correspond to the radiotherapy treatment planning optimization problem variables generated after performing two iterations. This process can be repeated with any number of additional iterations n.

[0114] In some implementations, using a machine learning model to process a radiotherapy treatment planning optimization problem to estimate one or more optimization variables includes: selecting a first subset of constraints for the radiotherapy treatment planning optimization problem; performing a first iteration of processing the radiotherapy treatment planning optimization problem using the machine learning model based on the selected first subset of constraints to generate a first estimate of one or more optimization variables; selecting a second subset of constraints for the radiotherapy treatment planning optimization problem; and performing a second iteration of processing the radiotherapy treatment planning optimization problem using the machine learning model based on the selected second subset of constraints to generate a second estimate of one or more optimization variables. In some cases, the first and second subsets of constraints are selected randomly.

[0115] Figure 4 This is a flowchart illustrating example operation of treatment processing logic 120 during execution of process 400 according to an exemplary embodiment. Process 400 may be contained in computer-readable instructions executable by one or more processors, such that the operation of process 400 can be performed partially or wholly by functional components of treatment processing logic 120; therefore, process 400 is described below by way of example. However, in other embodiments, at least some of the operation of process 400 may be deployed on various other hardware configurations. Therefore, process 400 is not intended to be limited to treatment processing logic 120 and may be implemented wholly or partially by any other components. Some or all of the operation of process 400 may be parallel, unordered, or completely omitted.

[0116] At operation 410, treatment processing logic 120 receives training data. For example, treatment processing logic 120 receives pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems; pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems and solutions for multiple training radiotherapy treatment optimization problems; and pairs of training parameters and one or more training optimization variables for multiple training radiotherapy treatment planning optimization problems of a given type.

[0117] At operation 420, treatment processing logic 120 receives constraints for training.

[0118] At operation 430, treatment processing logic 120 performs model training. For example, treatment processing logic 120 can train ML model parameters 312 by minimizing a first loss function based on corresponding training parameters for a corresponding problem in a radiotherapy treatment planning optimization problem and one or more training optimization variables. Figure 3In this way, ML models are trained to establish relationships between parameters of multiple training radiotherapy treatment planning optimization problems and one or more optimization variables. Training can be performed in a supervised or unsupervised manner, and models can be generated in closed-form or iterative ways.

[0119] At operation 440, treatment processing logic 120 outputs the trained model. For example, the trained model can be output and stored in memory, or the model's parameters can be presented to the clinician on a display device.

[0120] At operation 450, treatment processing logic 120 utilizes a trained model to generate results. For example, in training each machine learning model... (sometimes called Λ) θ After that, new data 370 can be received, including one or more patient input parameters (e.g., a radiotherapy treatment planning optimization problem). Trained machine learning techniques It can be applied to new data 370 to generate results 380 that include one or more estimated optimization variables for a radiotherapy treatment planning optimization problem. The estimated optimization variables provided by the ML model can be used to solve the new radiotherapy treatment planning optimization problem (including at least one of patient volume or patient image, patient volume segmentation, dose kernel, dose-volume histogram constraint, or dose constraint).

[0121] Figure 5 This is a flowchart illustrating example operations of treatment processing logic 120 during execution of process 500 according to an exemplary embodiment. Process 500 may be embodied in computer-readable instructions executable by one or more processors, such that the operations of process 500 can be performed, partially or entirely, by functional components of treatment processing logic 120; therefore, process 500 is described below by way of example. However, in other embodiments, at least some of the operations of process 500 may be deployed on various other hardware configurations. Therefore, process 500 is not intended to be limited to treatment processing logic 120 and may be implemented wholly or partially by any other components. Some or all of the operations of process 500 may be parallel, unordered, or completely omitted.

[0122] At operation 510, the treatment processing logic 120 receives the radiotherapy treatment plan optimization problem.

[0123] At operation 520, treatment processing logic 120 uses a machine learning model to process a radiotherapy treatment planning optimization problem to estimate one or more optimization variables of the radiotherapy treatment planning optimization problem, wherein the machine learning model is trained to establish the relationship between multiple parameters of the radiotherapy treatment planning optimization problem and one or more optimization variables.

[0124] At operation 530, treatment processing logic 120 generates a solution to the radiotherapy treatment planning optimization problem based on one or more optimization variables of the estimated radiotherapy treatment planning optimization problem.

[0125] Figure 6 This is a flowchart illustrating example operation of treatment processing logic 120 during execution of process 600 according to an exemplary embodiment. Process 600 may be contained in computer-readable instructions executable by one or more processors, such that the operation of process 600 can be performed partially or wholly by functional components of treatment processing logic 120; therefore, process 600 is described below by way of example. However, in other embodiments, at least some of the operation of process 600 may be deployed on various other hardware configurations. Therefore, process 600 is not intended to be limited to treatment processing logic 120 and may be implemented wholly or partially by any other components. The operation of process 600 may be some or all parallel, unordered, or completely omitted.

[0126] At operation 610, treatment processing logic 120 receives multiple training radiotherapy treatment plan optimization problems.

[0127] At operation 620, treatment processing logic 120 trains a machine learning model to generate estimates of one or more optimization variables for the radiotherapy treatment planning optimization problem by establishing relationships between the parameters of multiple training radiotherapy treatment planning optimization problems and one or more optimization variables.

[0128] As previously discussed, various electronic computing systems or devices can implement one or more of the methods or functional operations discussed herein. In one or more embodiments, the radiotherapy processing computing system 110 can be configured, adapted, or used to control or operate the image-guided radiotherapy device 232, perform, or implement operations from... Figure 3The training or prediction operation, the operation of the trained treatment model 360, the execution or implementation of the flowcharts for processes 400 to 600, or the execution of any one or more of the other methods discussed herein (e.g., as part of treatment processing logic 120). In various embodiments, such an electronic computing system or device operates as a standalone device or can be connected (e.g., networked) to other machines. For example, such a computing system or device 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 features of the computing system or device may be implemented by a personal computer (PC), tablet PC, personal digital assistant (PDA), cellular phone, web device, or any machine capable of executing instructions (sequentially or otherwise) specifying actions to be performed by that machine.

[0129] As described above, the functions discussed above can be implemented by storing instructions, logic, or other information on a machine-readable medium. Although machine-readable media may have been described with reference to a single medium in various examples, the term "machine-readable medium" can include a single medium or multiple media (e.g., a centralized or distributed database and / or associated cache and server) storing one or more transient or non-transient instructions or data structures. The term "machine-readable medium" should also be considered to include any tangible medium capable of storing, encoding, or carrying transient or non-transient instructions for use by a machine and to enable the machine to perform any one or more methods of this disclosure, or data structures utilized by or associated with such instructions.

[0130] 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 present disclosure also contemplates examples that provide only those elements shown or described. Furthermore, the present disclosure contemplates examples using any combination or arrangement of those elements (or one or more aspects of those elements) shown or described with respect to a particular example (or one or more aspects of that particular example) or with respect to other examples shown or described herein (or one or more aspects of those other examples).

[0131] All publications, patents, and patent documents referenced herein are incorporated in their entirety by reference as if they were incorporated separately by reference. In the event of any inconsistency between the usage in this document and that of the incorporated references, the usage in the incorporated (one or more) references shall be considered supplementary to the usage in this document; in the event of any inconsistency, the usage in this document shall prevail.

[0132] In this document, when introducing various aspects of this disclosure or elements in its embodiments, the terms "a", "an", "the", and "described" 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, unless otherwise indicated, the term "or" is used to mean a non-exclusive "or", such that "A or B" includes "A but not B", "B but not A", and "A and B".

[0133] 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 other elements may be present in addition to 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.

[0134] This disclosure also relates to computing systems adapted, configured, or operated to perform the operations described herein. Such systems may be specifically built for a desired purpose, or they may include general-purpose computers that can be selectively started or reconfigured by computer programs (e.g., instructions, code, etc.) stored in the 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 a particular operation will run or be performed before, simultaneously with, or after another operation.

[0135] 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. Having described aspects of this disclosure in detail, it will be apparent that modifications and variations are possible without departing from the scope of the aspects of this disclosure as defined in the appended claims. Since various changes can be made to the above-described structures, products, and methods without departing from the scope of the various aspects of this disclosure, all content contained in the foregoing description and shown in the accompanying drawings is intended to be illustrative rather than restrictive.

[0136] The examples described herein can be implemented in a wide variety of embodiments. For example, one embodiment includes a computing device comprising processing hardware (e.g., a processor or other processing circuitry) and memory hardware (e.g., a storage device or volatile memory), the memory hardware including instructions implemented thereon such that, when executed by the processing hardware, the instructions cause the computing device to implement, perform, or coordinate electronic operations for these technologies and system configurations. Another embodiment discussed herein includes a computer program product, for example, which may be implemented by a machine-readable medium or other storage device, providing transient or non-transitory instructions for implementing, performing, or coordinating electronic operations for these technologies and system configurations. Another embodiment discussed herein includes a method capable of operating on the processing hardware of a computing device to implement, perform, or coordinate electronic operations for these technologies and system configurations.

[0137] In other embodiments, the logic, commands, or transient or non-transient instructions for implementing aspects of the above-described electronic operations can be provided in distributed or centralized computing systems—including any number of formal factors relating to computing systems such as desktop or notebook personal computers, mobile devices such as tablets, netbooks and smartphones, client terminals, and server-hosted machine instances, etc. Another embodiment discussed herein includes incorporating the techniques discussed herein into other forms, including other forms of programming logic, hardware configurations, or dedicated components or modules, including means of devices having the functionality to perform such techniques. Various algorithms for implementing the functionality of such techniques may include some or all of the sequences of the above-described electronic operations or other aspects depicted in the accompanying drawings and the detailed embodiments described above.

[0138] The above description is intended to be illustrative and not restrictive. For example, the examples (or one or more aspects of the examples) described above 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, types, and example parameters, functions, and implementations of the materials 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, together with the full scope of their equivalents.

[0139] Furthermore, in the specific embodiments described above, various features may be combined to organize the present disclosure. This should not be construed as meaning that all unclaimed disclosed features are necessary for any claim. Rather, the subject matter of the invention may lie in fewer than all features of a particular disclosed embodiment. Therefore, the appended claims are hereby incorporated into the specific embodiments, wherein each claim exists as an independent embodiment. The scope of this disclosure should be determined with reference to the appended claims together with the full scope of their equivalents.

Claims

1. A method for solving a radiotherapy treatment planning optimization problem, the method comprising: receiving, by a processor circuit, a radiotherapy treatment planning optimization problem; classifying, by the processor circuit, the radiotherapy treatment planning optimization problem as a particular type of optimization problem; generating, by the processor circuit, training data by applying different numbers of iterations of a same particular optimization process to a first training optimization problem to output different solutions of the first training optimization problem, a first set of the training data comprising a first solution of the different solutions resulting from applying the same particular optimization process a first number of iterations to the first training optimization problem, a second set of the training data comprising a second solution of the different solutions resulting from applying the same particular optimization process a second number of iterations to the first training optimization problem; identifying, by the processor circuit, a machine learning model of a plurality of machine learning models based on the particular type of the radiotherapy treatment planning optimization problem, the machine learning model trained to solve the particular type of radiotherapy treatment planning optimization problem comprising an objective function, a set of decision variables, and a set of constraints, a first machine learning model of the plurality of machine learning models trained based on the first set of the training data, a second machine learning model of the plurality of machine learning models trained based on the second set of the training data; processing, by the processor circuit, the radiotherapy treatment planning optimization problem with the machine learning model to estimate one or more optimization variables of the radiotherapy treatment planning optimization problem, wherein the machine learning model is trained to establish a relationship between parameters of a plurality of training radiotherapy treatment planning optimization problems and the one or more optimization variables, wherein the machine learning model comprises a deep neural network; and generating, by the processor circuit, a solution of the radiotherapy treatment planning optimization problem based on the estimated one or more optimization variables of the radiotherapy treatment planning optimization problem, wherein the solution of the radiotherapy treatment planning optimization problem comprises at least one of: a radiotherapy device parameter, a fluence map, an isocenter position, a beam angle, or a beam on time.

2. The method of claim 1, wherein, The machine learning model comprises one or more intermediate estimated optimization variables used in subsequent iterations of the machine learning model.

3. The method of claim 1, wherein, Processing the radiotherapy treatment planning optimization problem comprises: processing the radiotherapy treatment planning optimization problem with the machine learning model to estimate one or more initial optimization variables; and solving the radiotherapy treatment planning optimization problem starting from the one or more initial optimization variables using a different learning or non-learning optimization process.

4. The method of claim 3, wherein, The non-learning optimization process includes at least one of: a simplex method, an interior point method, a Newton method, a quasi-Newton method, a Gauss-Newton method, a Levenberg-Marquardt method, a linear least squares method, a gradient descent method, a projected gradient method, a conjugate gradient method, an augmented Lagrangian method, a Nelder-Mead method, a branch and bound method, a cutting plane method, a simulated annealing, or a sequential quadratic programming.

5. The method of claim 1, wherein, The plurality of training radiotherapy treatment plan optimization problems includes at least one of: optimization problems derived from prior radiotherapy treatment plans; or synthetically generated problems.

6. The method of claim 1, wherein, The radiotherapy treatment plan optimization problem includes at least one of: an image of a patient volume or the patient, a segmentation of the patient volume, a dose kernel, a dose volume histogram constraint, or a dose constraint.

7. The method of claim 1, wherein, The radiotherapy treatment plan optimization problem is a constrained optimization problem, further including: transforming the radiotherapy treatment plan optimization problem into an unconstrained optimization problem based on a value function; and processing the transformed radiotherapy treatment plan optimization problem with the machine learning model to estimate the one or more optimization variables.

8. The method of claim 1, wherein, Processing the radiotherapy treatment plan optimization problem with the machine learning model to estimate the one or more optimization variables includes: selecting a first subset of constraints of the radiotherapy treatment plan optimization problem; performing a first iteration of processing the radiotherapy treatment plan optimization problem with the machine learning model based on the selected first subset of constraints to generate a first estimate of the one or more optimization variables; selecting a second subset of constraints of the radiotherapy treatment plan optimization problem; and performing a second iteration of processing the radiotherapy treatment plan optimization problem with the machine learning model based on the selected second subset of constraints to generate a second estimate of the one or more optimization variables.

9. The method of claim 8, further comprising performing one or more additional iterations that include selecting a subset of constraints and processing the radiotherapy treatment plan optimization problem with the machine learning model.

10. The method of claim 9, wherein, The first subset of constraints and the second subset of constraints are selected randomly.

11. A method for training a machine learning model to solve radiotherapy treatment plan optimization problems, the method comprising: receiving, by a processor circuit, a plurality of training radiotherapy treatment plan optimization problems; classifying, by the processor circuit, the radiotherapy treatment plan optimization problems into a particular type of optimization problem; generating training data by applying a same particular optimization process a different number of iterations to a first training optimization problem to output different solutions of the first training optimization problem, a first set of the training data including a first solution of the different solutions resulting from applying the same particular optimization process a first number of iterations to the first training optimization problem, a second set of the training data including a second solution of the different solutions resulting from applying the same particular optimization process a second number of iterations to the first training optimization problem; based on the particular type of the radiotherapy treatment planning optimization problem, identifying a machine learning model of a plurality of machine learning models trained to solve the particular type of radiotherapy treatment planning optimization problem, the particular type of radiotherapy treatment planning optimization problem including an objective function, a set of decision variables, and a set of constraints, a first machine learning model of the plurality of machine learning models trained based on a first set of the training data, a second machine learning model of the plurality of machine learning models trained based on a second set of the training data; and training the machine learning model to generate estimates of one or more optimization variables of the radiotherapy treatment planning optimization problem by establishing a relationship between parameters of the plurality of training radiotherapy treatment planning optimization problems and the one or more optimization variables.

12. The method of claim 11, wherein, training the machine learning model in a supervised manner based on a plurality of solutions of the plurality of training radiotherapy treatment planning optimization problems.

13. The method of claim 12, wherein, training the machine learning model iteratively in the supervised manner, wherein an intermediate output of a training iteration of the machine learning model that includes intermediate estimated optimization variables is used in a subsequent training iteration of the machine learning model.

14. The method of claim 11, wherein, the plurality of training radiotherapy treatment planning optimization problems include a constrained optimization problem, the method further comprising: transforming the plurality of training radiotherapy treatment planning optimization problems into unconstrained optimization problems based on a value function; and training the machine learning model based on the transformed radiotherapy treatment planning optimization problems.

15. The method of claim 14, wherein, training the machine learning model in an unsupervised manner.

16. A non-transitory computer-readable medium comprising non-transitory computer-readable instructions, the computer-readable instructions comprising instructions for performing operations comprising: receiving a radiotherapy treatment planning optimization problem; classifying, by a processor circuit, the radiotherapy treatment planning optimization problem as a particular type of optimization problem; generating training data by applying a same particular optimization process a different number of iterations to a first training optimization problem to output different solutions of the first training optimization problem, a first set of the training data including a first solution of the different solutions resulting from applying the same particular optimization process a first number of iterations to the first training optimization problem, a second set of the training data including a second solution of the different solutions resulting from applying the same particular optimization process a second number of iterations to the first training optimization problem; identifying, based on the particular type of the radiotherapy treatment plan optimization problem, a machine learning model of a plurality of machine learning models trained to solve the particular type of radiotherapy treatment plan optimization problem, the particular type of radiotherapy treatment plan optimization problem including an objective function, a set of decision variables, and a set of constraints, a first machine learning model of the plurality of machine learning models trained based on a first set of training data, a second machine learning model of the plurality of machine learning models trained based on a second set of training data; processing the radiotherapy treatment plan optimization problem with the machine learning model to estimate one or more optimization variables of the radiotherapy treatment plan optimization problem, wherein the machine learning model is trained to establish a relationship between parameters of a plurality of training radiotherapy treatment plan optimization problems and the one or more optimization variables, wherein the machine learning model comprises a deep neural network; and generating a solution to the radiotherapy treatment plan optimization problem based on the estimated one or more optimization variables of the radiotherapy treatment plan optimization problem, wherein the solution to the radiotherapy treatment plan optimization problem comprises at least one of: a radiotherapy device parameter, a fluence map, an isocenter position, a beam angle, or a beam on time.

17. The non-transitory computer-readable medium of claim 16, wherein, the machine learning model comprises one or more intermediate estimated optimization variables used in subsequent iterations of the machine learning model.

18. The non-transitory computer-readable medium of claim 17, wherein, the radiotherapy treatment plan optimization problem comprises at least one of: a patient volume or an image of a patient, a segmentation of the patient volume, a dose kernel, a dose volume histogram constraint, or a dose constraint, and wherein the solution to the radiotherapy treatment plan optimization problem comprises at least one of: a radiotherapy device parameter, a fluence map, an irradiation position, or a beam on time.

19. A system comprising: a memory to store instructions; and one or more processors to execute the instructions stored in the memory to perform operations comprising: receiving a radiotherapy treatment plan optimization problem; classifying the radiotherapy treatment plan optimization problem as a particular type of optimization problem; generating training data by applying different numbers of iterations of a same particular optimization process to a first training optimization problem to output different solutions to the first training optimization problem, a first set of the training data comprising a first solution of the different solutions resulting from applying the same particular optimization process to the first training optimization problem for a first number of iterations, a second set of the training data comprising a second solution of the different solutions resulting from applying the same particular optimization process to the first training optimization problem for a second number of iterations; identifying, based on the particular type of the radiotherapy treatment plan optimization problem, a machine learning model of a plurality of machine learning models trained to solve the particular type of radiotherapy treatment plan optimization problem, the particular type of radiotherapy treatment plan optimization problem including an objective function, a set of decision variables, and a set of constraints, a first machine learning model of the plurality of machine learning models trained based on a first set of training data, a second machine learning model of the plurality of machine learning models trained based on a second set of training data; processing the radiotherapy treatment plan optimization problem with the machine learning model to estimate one or more optimization variables of the radiotherapy treatment plan optimization problem, wherein the machine learning model is trained to establish a relationship between parameters of a plurality of training radiotherapy treatment plan optimization problems and the one or more optimization variables, wherein the machine learning model comprises a deep neural network; and generating a solution to the radiotherapy treatment plan optimization problem based on the estimated one or more optimization variables of the radiotherapy treatment plan optimization problem, wherein the solution to the radiotherapy treatment plan optimization problem comprises at least one of: a radiotherapy device parameter, a fluence map, an isocenter position, a beam angle, or a beam on time.

20. The system of claim 19, wherein, the machine learning model comprises one or more intermediate estimated optimization variables used in subsequent iterations of the machine learning model.

21. The system of claim 19, wherein, the radiotherapy treatment plan optimization problem comprises at least one of: a patient volume or an image of a patient, a segmentation of the patient volume, a dose kernel, a dose volume histogram constraint, or a dose constraint, and wherein the solution to the radiotherapy treatment plan optimization problem comprises at least one of: a radiotherapy device parameter, a fluence map, an irradiation position, or a beam on time.

Images