Compressed radiation therapy treatment plan optimization problem

Abstract

Techniques for solving radiation therapy treatment planning optimization problems are provided. The techniques include receiving a first radiation therapy treatment planning optimization problem having a first set of parameters, processing the first set of parameters to estimate a second set of parameters for a second radiation therapy treatment planning optimization problem, generating a solution to the second radiation therapy treatment planning optimization problem based on the estimated second set of parameters, and generating a radiation therapy treatment plan based on the solution to the second radiation therapy treatment planning optimization problem.

Description

[0001] CLAIM OF PRIORITY

[0002] This application claims the benefit of priority to U.S. Application Serial No. 16 / 512,962, filed July 16, 2019, the entire contents of which are incorporated herein by reference. TECHNICAL FIELD

[0003] The present disclosure relates generally to radiotherapy or radiotherapy optimization problems. BACKGROUND

[0004] Radiotherapy is used to treat cancer and other diseases in mammalian (e.g., human and animal) tissue. The direction and shape of the radiation beam should be accurately controlled to ensure that the tumor receives the prescribed radiation, and the placement of the beam should be such that damage to surrounding healthy tissue, often referred to as organs at risk (OARs), is minimized. A treatment plan can be used to control the radiation beam parameters, and a radiotherapy device implements the treatment by delivering a spatially varying dose distribution to the patient.

[0005] Conventionally, for each patient, an optimization technique based on clinical and dose determination objectives and constraints (e.g., maximum, minimum, and mean dose to tumor and critical organs) can be used to create a radiotherapy treatment plan (“treatment plan”). The treatment planning process can include using three-dimensional (3D) images of the patient to identify a target region (e.g., a tumor) and to identify critical organs in the vicinity of the tumor. Creating a treatment plan can be a time-consuming process in which a planner attempts to adhere to treatment objectives or constraints (e.g., dose volume histogram (DVH) objectives) in light of respective importance (e.g., weights) of the treatment objectives or constraints to produce a clinically acceptable treatment plan. This task can be a time-consuming, trial-and-error process that is complicated by the various OARs, as the complexity of the process increases as the number of OARs increases (e.g., typically 21 are segmented in head and neck treatments). OARs that are far away from the tumor can be easily spared from radiation, while OARs that are close to or overlap with the target tumor can be difficult to spare.

[0006] Segmentation can be performed to identify OARs and regions to be treated (e.g., a planning target volume (PTV)). After segmentation, a dose plan can be created for the patient, indicating the desired amount of radiation to be received by the PTV (e.g., a target) and / or OARs. The PTV can have an irregular volume, and its size, shape, and location can be unique. A treatment plan can be computed after optimizing a large number of plan parameters to ensure that sufficient dose is provided to the PTV while providing as low a dose as possible to surrounding healthy tissue. Thus, a radiotherapy treatment plan can be determined by balancing the effective control of the dose to protect any OARs versus treating the tumor. Generally, the quality of a radiotherapy treatment plan can depend on the experience level of the planner. Further complications can arise from anatomical differences between patients. SUMMARY

[0007] In some implementations, a computer-implemented method, a non-transitory computer-readable medium, and a system including a memory and a processor are provided for receiving, by processor circuitry, a first radiotherapy treatment plan optimization problem having a first set of parameters; processing, by the processor circuitry, the first set of parameters to estimate a second set of parameters of a second radiotherapy treatment plan optimization problem; generating, by the processor circuitry, a solution to the second radiotherapy treatment plan optimization problem based on the estimated second set of parameters; and generating, by the processing circuitry, a radiotherapy treatment plan based on the solution to the second radiotherapy treatment plan optimization problem.

[0008] In some implementations, the first set of parameters and the second set of parameters include at least one of optimization variables, an objective function, or a set of constraints.

[0009] In some implementations, processing the first set of parameters includes applying a machine learning model to the first set of parameters to estimate the second set of parameters, wherein the machine learning model is trained to establish a relationship between the first set of parameters and the second set of parameters of a plurality of training radiotherapy treatment plan optimization problems.

[0010] In some implementations, a first optimization variable in the first set of parameters is excluded from the second set of parameters.

[0011] In some implementations, the second set of parameters is smaller than the first set of parameters.

[0012] In some implementations, processing the first set of parameters includes applying at least one of a non-linear functional relationship or a statistical model to the first set of parameters, wherein the statistical model includes modeling the first set of parameters as a random variable that depends on the second set of parameters, and an objective function or a constraint of the second radiotherapy treatment plan optimization problem is derived based on a measure of central tendency of an objective function or a constraint of the first radiotherapy treatment plan optimization problem.

[0013] In some implementations, the second radiation therapy treatment plan optimization problem includes fewer constraints than the first radiation therapy treatment plan optimization problem.

[0014] In some implementations, the matrix in the second radiation therapy treatment plan optimization problem is sparse or structured to correspond to the matrix in the first radiation therapy treatment plan optimization problem, and wherein generating the solution includes processing the second radiation therapy treatment plan optimization problem using an optimization procedure.

[0015] In some implementations, the solution to the second radiation therapy treatment plan optimization problem is invariant or equivalent for permutations of the first and second subsets of the first parameter set.

[0016] In some implementations, the second radiation therapy treatment plan optimization problem has a fixed size relative to the size of the first radiation therapy treatment plan optimization problem.

[0017] In some implementations, the solution to the second radiation therapy treatment plan optimization problem approximates the solution to the first radiation therapy treatment plan optimization problem.

[0018] In some implementations, the first radiation therapy treatment plan optimization problem includes at least one of a linear programming problem or a quadratic programming problem, and wherein the first parameter set and the second parameter set are related by a Gaussian process.

[0019] In some implementations, there is provided a computer-implemented method, a non-transitory computer-readable medium, and a system including a memory and a processor, for: computing a second parameter of a second parameter set based on a point estimate corresponding to a Gaussian distribution of a first parameter of a first parameter set; and replacing the first parameter of the first parameter set with the second parameter to generate a second radiation therapy treatment plan optimization problem, wherein the second radiation therapy treatment plan optimization problem includes the other linear programming problem or the quadratic programming problem.

[0020] In some implementations, there is provided a computer-implemented method, a non-transitory computer-readable medium, and a system including a memory and a processor, for marginalizing the first parameter set on a Gaussian distribution to estimate a second parameter set.

[0021] In some implementations, there is provided a computer-implemented method, a non-transitory computer-readable medium, and a system including a memory and a processor, for mapping a solution to the second radiation therapy treatment plan optimization problem onto a space of the first parameter set of the first radiation therapy treatment plan optimization problem.

[0022] In some implementations, mapping the solution onto the space of the first parameter set includes applying a machine learning model to the solution to estimate a further solution to the first radiation therapy treatment plan optimization problem.

[0023] In some implementations, the first radiotherapy treatment plan optimization problem and the second radiotherapy treatment plan optimization problem are processed by an optimization process until a predetermined criterion is satisfied.

[0024] In some implementations, 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 plan optimization problem by: receiving, by the processor circuitry, a plurality of first training radiotherapy treatment plan optimization problems, each of the plurality of first training radiotherapy treatment plan optimization problems including a first parameter set; and training, by the processor circuitry, the machine learning model to generate an estimate of a second parameter set of a second radiotherapy treatment plan optimization problem by establishing a relationship between the first parameter sets of the plurality of first training radiotherapy treatment plan optimization problems and the second parameter set of the second radiotherapy treatment plan optimization problem.

[0025] In some implementations, the machine learning model is trained in a supervised manner by: obtaining a first training data pair including a given parameter set of the first parameter sets and a corresponding solution of a given problem of the plurality of first training radiotherapy treatment plan optimization problems; processing the given parameter set of the first parameter sets with the machine learning model to generate an estimate of a second parameter set of a second radiotherapy treatment plan optimization problem; solving the second radiotherapy treatment plan optimization problem based on the estimate of the second parameter set to generate a given solution of the second radiotherapy treatment plan optimization problem; computing a deviation between at least one of: (1) the given solution and the corresponding solution or (2) a decoded version of the given solution in a space of the given problem of the plurality of first training radiotherapy treatment plan optimization problems and the corresponding solution; and updating parameters of the machine learning model based on the computed deviation.

[0026] In some implementations, the machine learning model is trained in an unsupervised manner by: obtaining a given parameter set of the first parameter sets; processing the given parameter set of the first parameter sets with the machine learning model to generate an estimate of a second parameter set of a second radiotherapy treatment plan optimization problem; solving the second radiotherapy treatment plan optimization problem based on the estimate of the second parameter set to generate a given solution of the second radiotherapy treatment plan optimization problem; computing a metric of the given solution; and updating parameters of the machine learning model based on the computed metric.

[0027] In some implementations, the metric includes a statistical metric representing a likelihood of observing the training data or an utility of a treatment plan corresponding to the given solution.

[0028] The above summary is intended to provide an overview of the subject matter of the present patent application. It is not intended to provide an exclusive or exhaustive explanation of the disclosure. The DETAILED DESCRIPTION is included to provide further information about the present patent application. BRIEF DESCRIPTION OF DRAWINGS

[0029] In the drawings, which are not necessarily drawn to scale, like numerals describe substantially similar components throughout the several views. Like numerals having different letter suffixes

[0030] Figure 1 An exemplary radiation therapy system adapted to perform a treatment plan generation process is shown in accordance with some examples.

[0031] Figure 2A An exemplary image-guided radiation therapy device is shown in accordance with some examples of the present disclosure.

[0032] Figure 2B A radiation therapy device, Gamma Knife, is shown in accordance with some examples of the present disclosure.

[0033] Figure 3 An exemplary data flow for training and using machine learning techniques to estimate parameters of a second radiation therapy treatment plan optimization problem from parameters of a first radiation therapy treatment plan optimization problem is shown in accordance with some examples of the present disclosure.

[0034] Figures 4 to 6 A flowchart of exemplary operations for training and using machine learning techniques to estimate parameters of a second radiation therapy treatment plan optimization problem from parameters of a first radiation therapy treatment plan optimization problem is shown in accordance with some examples of the present disclosure. DETAILED DESCRIPTION

[0035] The present disclosure includes various techniques for generating a radiation therapy treatment plan by simplifying a first radiation therapy treatment plan optimization problem to a second radiation therapy optimization problem that is less complex than the first radiation therapy treatment plan optimization problem. Specifically, parameters of the first radiation therapy treatment plan optimization problem are processed to estimate parameters of the second radiation therapy treatment plan optimization problem, and a solution to the second radiation therapy treatment plan optimization problem is generated and used to provide the radiation therapy treatment plan. In some implementations, the parameters of the first radiation therapy treatment plan optimization problem are processed by a machine learning model to estimate the parameters of the second radiation therapy treatment plan optimization problem. In some implementations, the parameters of the first radiation therapy treatment plan optimization problem are modeled by a statistical function or a non-linear function relationship to estimate the parameters of the second radiation therapy treatment plan optimization problem.

[0036] Technical benefits include reduced computational processing time for generating radiotherapy treatment plans and solving radiotherapy treatment plan optimization problems, and concomitant improvements in processing, memory, and network resources used for generating radiotherapy treatment plans and solving radiotherapy treatment plan optimization problems. These radiotherapy treatment plans can be applicable to a variety of medical treatment and diagnostic settings or radiotherapy treatment equipment and devices. Thus, in addition to these technical benefits, the present technology can also yield a number of apparent medical treatment benefits (including improved accuracy of radiotherapy treatments, reduced exposure to unintended radiation, etc.).

[0037] Radiotherapy is one of the primary methods of treating cancer and is recommended for more than 50% of cancer patients. Treatment plans are created through a complex design process involving a mathematical optimization problem that captures the desired characteristics of the dose delivery - typically a sufficiently high dose to reach the target while minimizing the dose to healthy tissue. The overall structure of the optimization problem is the same for most forms of radiotherapy - including linac-based treatments (3D-CRT, IMRT, VMAT), proton therapy, Gamma Knife radiosurgery, and brachytherapy. The end result is the radiotherapy device configuration (e.g., control points) needed to deliver the dose distribution.

[0038] Current planning software typically uses standard mathematical optimization methods to solve the minimization problem. These methods can be slow, resulting in unnecessary waiting for patients and clinicians. Future applications that utilize real-time imaging can even require real-time treatment planning, which cannot be performed using conventional optimization problem solvers.

[0039] The disclosed technology addresses these challenges and improves the speed and efficiency of solving radiotherapy treatment plan optimization problems by utilizing ML models, functional relationships, and / or statistical models to transform a complex first radiotherapy treatment plan optimization problem having a first set of optimization problem parameters (e.g., optimization variables, objective function, and / or constraint set) to a less complex second radiotherapy treatment plan optimization problem having a second set of optimization problem parameters (optimization variables, objective function, and / or constraint set). In some implementations, the ML model is used to process the parameters of the first radiotherapy treatment plan optimization problem and estimate the parameters of the second radiotherapy treatment plan optimization problem, which simplifies the radiotherapy treatment plan optimization problem and allows it to be solved more quickly. In such cases, the second radiotherapy treatment plan optimization problem can have a less complex optimization function, fewer constraints, and / or fewer decision variables than the first radiotherapy treatment plan optimization problem, making it less complex to solve. In some implementations, the parameters of the first radiotherapy treatment plan optimization problem are functionally and / or statistically modeled to generate the parameters of the second radiotherapy treatment plan optimization problem. In such cases, the optimization function of the second radiotherapy treatment plan optimization problem can have the same structure and form as the first radiotherapy treatment plan optimization problem, but can have fewer decision variables and / or constraints, making it less complex to solve.

[0040] In some cases, the solution to the second radiotherapy treatment plan optimization problem estimates the solution to the first radiotherapy treatment plan optimization problem and can be used directly to generate a radiotherapy treatment plan. In some cases, the solution to the second radiotherapy treatment plan optimization problem can be decoded to solve the first radiotherapy treatment plan optimization problem to generate a radiotherapy treatment plan. In some implementations, the first and / or second radiotherapy treatment plan optimization problems are solved using the processes discussed in commonly-assigned U.S. Patent Application No. 16 / 512,972 to Jens et al., filed concurrently herewith, entitled “optimization of radiotherapy treatment plans using machine learning” (Attorney Docket No. 4186.122US1), which is hereby incorporated by reference. According to the disclosed technology, by improving the speed of solving radiotherapy treatment plan optimization problems, the disclosed technology can enable real-time treatment planning and reduce patient and clinician wait times.

[0041] In particular, the disclosed technology receives a first radiotherapy treatment plan optimization problem having a first parameter set and processes the first parameter set to estimate a second parameter set of a second radiotherapy treatment plan optimization problem. The disclosed technology generates a solution to the second radiotherapy treatment plan optimization problem based on the estimated second parameter set and generates a radiotherapy treatment plan based on the solution to the second radiotherapy treatment plan optimization problem. Machine settings, e.g., control points, of a radiotherapy treatment device can then be determined and generated by the solution to the second radiotherapy treatment plan optimization problem and / or based on the estimated solution to the first radiotherapy treatment plan optimization problem corresponding to the solution to the second radiotherapy treatment plan optimization problem.

[0042] As referred to throughout this disclosure, the "first optimization problem" is the original optimization problem for a given radiotherapy treatment plan that needs to be simplified. The "second optimization problem" is a simplified version of the first optimization problem that is solved to generate the given radiotherapy treatment plan information.

[0043] Figure 1 An exemplary radiotherapy system 100 adapted to perform radiotherapy plan processing operations using one or more of the methods discussed herein is shown. These radiotherapy plan processing operations are performed to enable the radiotherapy system 100 to provide radiotherapy to a patient based on captured medical imaging data and particular aspects of therapy dose calculations or radiotherapy machine configuration parameters. In particular, the following processing operations can be implemented as part of the treatment processing logic 120. However, it will be understood that many variations and use cases of the following trained models and treatment processing logic 120 can be provided, including data validation, visualization, and other medical assessment and diagnostic settings.

[0044] The radiotherapy system 100 includes a radiotherapy processing computing system 110 that hosts the 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 with one or more private and / or public medical information sources (e.g., a radiology information system (RIS), a medical records system (e.g., an electronic medical record (EMR) / electronic health record (HER) system), an oncology information system (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.

[0045] As an example, the radiation therapy treatment computing system 110 can be configured to receive a treatment target of a subject (e.g., from one or more MR images) and generate a radiation therapy treatment plan by executing instructions or data of the treatment processing logic 120 as part of an operation to generate a treatment plan to be used by the treatment device 180 and / or for output on the device 146. In implementations, the treatment processing logic 120 solves an optimization problem to generate the radiation therapy treatment plan. The treatment processing logic 120 solves the radiation therapy optimization problem by simplifying a first radiation therapy treatment plan optimization problem to a second radiation therapy optimization problem that is less complex than the first radiation therapy treatment plan optimization problem. In some cases, first optimization variables in the first set of parameters are not included within the second set of parameters. In some cases, the second set of parameters is smaller than the first set of parameters. In some cases, the second radiation therapy treatment plan optimization problem includes fewer constraints than the first radiation therapy treatment plan optimization problem. In some cases, the second radiation therapy treatment plan optimization problem has a fixed size relative to the size of the first radiation therapy treatment plan optimization problem.

[0046] In some cases, the second radiation therapy treatment plan optimization problem includes parameters defined or selected according to a class of optimization problems for which there are more efficient algorithms to solve than the parameters of the first optimization problem. For example, the second optimization problem can be convex while the first optimization problem is non-convex. As another example, the second optimization problem can be unconstrained while the first optimization problem is constrained. As another example, the second optimization problem can be a semi-definite programming problem while the first optimization problem is a general non-linear problem. As another example, the second optimization problem can be defined or solved by linear programming while the first optimization problem is defined or solved by quadratic programming. In some cases, the second optimization problem can be less complex than the first optimization problem due to the second optimization problem enabling efficient linear algebra routines (e.g., sparse linear algebra or certain matrix decompositions).

[0047] In some implementations, the parameters of the first radiation therapy treatment plan optimization problem are processed to estimate the parameters of the second radiation therapy treatment plan optimization problem, and a solution to the second radiation therapy treatment plan optimization problem is generated and used to provide a radiation therapy treatment plan. In some implementations, the parameters of the first radiation therapy treatment plan optimization problem are processed by a machine learning model to estimate the parameters of the second radiation therapy treatment plan optimization problem. In some implementations, the parameters of the first radiation therapy treatment plan optimization problem are modeled by a statistical function or a non-linear function relationship to estimate the parameters of the second radiation therapy treatment plan optimization problem.

[0048] In treatment planning, clinical decision makers strive to overcome the physical limitations of radiotherapy devices to find an acceptable compromise for the patient. In most cases, this interaction is mediated by optimization problems. The overall structure of these optimization problems is identical for most forms of radiotherapy - including linac-based treatments (3D-CRT, IMRT, VMAT), proton therapy, Gamma Knife radiosurgery, and brachytherapy. The invention subject matter presented here applies to all of these.

[0049] The biological effect of a radiotherapy treatment depends on the absorbed dose d. Radiotherapy devices differ in the way they deliver dose, but the relationship can generally (at least approximately) be modeled as a linear function d(x) = Φx, where Φ is a dose influence matrix that maps from decision variables x to dose. Typically, at least some of the objective function f and constraints ci are formulated in terms of dose only. The physical meaning of the decision variables depends on the treatment modality, e.g., it is radiation time in Gamma Knife radiosurgery and brachytherapy, and it is fluence in linac-based treatments. Typically, both the elements in the dose influence matrix and the decision variables are non-negative.

[0050] The dose influence matrix Φ is typically static during optimization, meaning that it can be pre-computed and stored. Thus, the typical optimization problem in radiotherapy has the form defined by Equation 1:

[0051]

[0052] where Cxand Cdrepresent the sets of machine constraints and dose constraints, respectively. However, the associated constraint functions {ci} depend on the application. The functions f and ci can be both non-linear and non-convex (or even integer-valued). Two examples of complex optimization problems come from isocenter selection in Gamma Knife radiosurgery or arc sequencing in VMAT planning. In such cases, the optimization problem can be solved directly (e.g., direct aperture optimization in VMAT planning) or indirectly (e.g., via a series of increasingly difficult sub-problems). Regardless, the most complex optimization problem typically dominates the total run time.

[0053] For some generic feasible set X, the generic radiotherapy treatment planning optimization problem can be expressed according to Equation 2:

[0054]

[0055] The generic radiotherapy treatment planning optimization problem can be more explicitly written as Equation 3:

[0056]

[0057] where f is an objective function, x is an optimization variable, and {ci} is a set of functions, where e and I are index sets corresponding to equality and inequality constraints, respectively. According to the disclosed technology, solving such optimization problems is made less complex by estimating one or more parameters (e.g., x, f, and / or {ci}) of a less complex second optimization problem using parameters of the original optimization problem, and then applying a conventional method to solve the less complex optimization problem.

[0058] The original radiotherapy treatment planning optimization problem is difficult, computationally inefficient, and / or takes an unacceptable amount of time to solve due to the original radiotherapy treatment planning optimization problem being large or due to it having a complex structure (e.g., non-differentiable, non-convex, or mixed integer). According to some embodiments, the complexity of solving the original radiotherapy treatment planning optimization problem is simplified to reduce the computational resources and computation time required to solve the problem. One such approach includes approximating the original radiotherapy treatment planning optimization problem with another radiotherapy treatment planning optimization problem that has the same structure but fewer decision variables and / or constraints. Another approach includes approximating the original radiotherapy treatment planning optimization problem with another radiotherapy treatment planning optimization problem that has a simpler structure (e.g., a convex relaxation of a non-convex problem). In both cases, parameters of the original radiotherapy treatment planning optimization problem are used to estimate parameters of the other radiotherapy treatment planning optimization problem that is less complex to solve. In some embodiments, the parameters are estimated according to a statistical relationship between the parameters and / or a functional relationship between the parameters.

[0059] In some embodiments, a machine learning model is utilized to generate parameters of a less complex radiotherapy treatment planning optimization problem from parameters of the original radiotherapy treatment planning optimization problem. In such cases, the machine learning model is used to express a variable-size parameter of the radiotherapy treatment planning optimization problem as a fixed-size representation.

[0060] The radiotherapy treatment planning computing system 110 can include processing circuitry 112, memory 114, storage 116, and other hardware- and software-operable components such as a user interface 142, a communication interface (not shown), and the like. The storage 116 can store transitory or non-transitory computer-executable instructions, such as an operating system, a radiotherapy treatment plan, training data, software programs (e.g., image processing software, image or anatomical visualization software, artificial intelligence (AI) or ML implementations and algorithms provided, e.g., by deep learning models, ML models, and neural networks (NNs)), and any other computer-executable instructions to be executed by the processing circuitry 112.

[0061] In the example, the processing circuitry system 112 may include processing devices, such as one or more general-purpose processing devices such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), etc. More specifically, the processing circuitry system 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. The processing circuitry system 112 may also be implemented by one or more special-purpose processing devices such as an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), a system-on-a-chip (SoC), etc.

[0062] 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 are from AMD. TM Turion manufactured TM Athlon TM Sempron TM Opteron TM FX TM Phenom TM The microprocessor can be from any series of microprocessors, or any processor from various processors manufactured by Sun Microsystems. The processing circuitry system 112 may also include processors from, for example, those from Nvidia. TM Manufactured Series, by Intel TM GMA and Iris manufactured TM Series, or by AMD TM Radeon manufactured TM The series of GPUs' graphics processing units. The processing circuitry system 112 may also include components such as those from Intel... TM Xeon Phi manufactured TMa series of accelerated processing units. The disclosed implementations are not limited to any particular type of processor configured to otherwise meet the computational demands 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 (e.g., a multi-core design or multiple processors each having a multi-core design). The processing circuitry 112 can execute sequences of the transient or non-transient computer program instructions stored in memory 114, and accessed from storage 116, to perform various operations, processes, and methods that will be explained in greater detail below. It should be understood that any of the components in system 100 can be implemented separately and operate as standalone devices, and can be coupled to any of the other components in system 100 to perform the techniques described in this disclosure.

[0063] Memory 114 can include read-only memory (ROM), phase change random access memory (PRAM), static random access memory (SRAM), flash, random access memory (RAM), dynamic random access memory (DRAM), such as synchronous DRAM (SDRAM), electrically erasable programmable read-only memory (EEPROM), static storage (e.g., flash memory, flash drive, static random access memory), and other types of random access memory, cache memory, registers, compact discs read only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, 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 transitory or non-transitory computer-readable instructions (e.g., stored in any format) accessible to processing circuitry 112, or any other type of computer- accessible medium. For instance, the computer program instructions can be accessed from the ROM, or any other suitable memory location, read by processing circuitry 112, and loaded into the RAM for execution by processing circuitry 112.

[0064] Storage 116 can constitute a drive unit including a transitory or non-transitory machine-readable medium on which is stored one or more sets of transitory or non-transitory instructions and data structures (e.g., software) embodying or utilized by any one or more of the methodologies or functions described herein, including treatment processing logic 120 and user interface 142, in various examples. The instructions can also reside, completely or at least partially, within memory 114 and / or processing circuitry 112 during execution thereof by radiotherapy treatment computing system 110, memory 114 and processing circuitry 112 also constituting transitory or non-transitory machine-readable media.

[0065] The memory 114 and storage 116 can constitute non-transitory computer- readable media. For example, the memory 114 and storage 116 can store or load transitory or non-transitory instructions for one or more software applications on the computer-readable media. The software applications stored or loaded with the memory 114 and storage 116 can include, for example, operating systems for general computer systems as well as for software-controlled devices. The radiotherapy treatment computing system 110 can also operate various software programs including software code for implementing the treatment processing logic 120 and the user interface 142. Further, the memory 114 and storage 116 can store or load entire software applications, portions of software applications, or code or data associated with software applications that are executable by the processing circuitry 112. In another example, the memory 114 and storage 116 can store, load, and manipulate one or more radiotherapy treatment plans, imaging data, segmentation data, treatment visualizations, histograms or measurements, one or more AI model data (e.g., weights and parameters of the ML models of the disclosed embodiments), training data, labels and mapping data, etc. It is contemplated that software programs can be stored not only on the storage 116 and memory 114, but also on removable computer media such as hard drives, computer disks, CD-ROMs, DVDs, Blu-ray DVDs, USB flash drives, SD cards, memory sticks, or any other suitable media; such software programs can also be transmitted or received over a network.

[0066] Although not depicted, the radiotherapy treatment computing system 110 can include communication interfaces, network interface cards, and communication circuitry. Example communication interfaces can include, for example, network adapters, cable connectors, serial connectors, USB connectors, parallel connectors, high-speed data transfer adapters (e.g., fiber optic, USB 3.0, thunderbolt, etc.), wireless network adapters (e.g., IEEE 802.11 / Wi-Fi adapters), telecommunication adapters (e.g., to communicate with 3G, 4G / LTE, and 5G networks, etc.), etc. Such communication interfaces can include one or more digital and / or analog communication devices that allow the machine to communicate with other machines and devices, e.g., remotely located components, via a network. The network can provide the functionality of a local area network (LAN), a wireless network, a 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 can be a LAN or WAN that can include other systems, including additional image processing computing systems or image-based components associated with medical imaging or radiotherapy operations.

[0067] In an example, the radiotherapy treatment computing system 110 can obtain image data 152 from an image data source 150 (e.g., MR images) to host on the storage 116 and the memory 114. In yet another example, a software program can replace the function of patient images, such as a processed version of the image that emphasizes some aspect of the image information or a signed distance function.

[0068] In an example, the radiotherapy treatment computing system 110 can obtain image data 152 from or transmit image data 152 to the image data source 150. In a further example, the treatment data source 160 receives or updates plan data as a result of a treatment plan generated by the treatment processing logic 120. The image data source 150 can also provide or host imaging data for use in the treatment processing logic 120.

[0069] In an example, the computing system 110 can communicate with the treatment data source 160 and the input device 148 to generate one or more training solutions to a first training radiotherapy treatment plan optimization problem and pairs of multiple training parameters to the first training radiotherapy treatment plan optimization problem.

[0070] The processing circuitry 112 can be communicatively coupled to the memory 114 and the storage 116, and the processing circuitry 112 can be configured to execute computer-executable instructions from the memory 114 or the storage 116 stored thereon. The processing circuitry 112 can execute the instructions to cause medical images from the image data 152 to be received or obtained in the memory 114 and processed using the treatment processing logic 120 to generate a treatment plan. In particular, the treatment processing logic 120 receives an optimization problem derived based on the received medical images. The treatment processing logic processes parameters of the received optimization problem to estimate parameters of a second, less complex optimization problem. Once the parameters are estimated, the second optimization problem is solved and used to generate the treatment plan.

[0071] Additionally, processing circuitry 112 can utilize software programs to generate intermediate data, such as updated parameters to be used by, for example, a neural network (NN) model, a machine learning model, treatment processing logic 120, or other aspects related to the generation of a treatment plan as discussed herein. Moreover, such software programs can utilize treatment processing logic 120 to produce new or updated treatment plan parameters using techniques discussed further herein for deployment to treatment data source 160 and / or rendering on output device 146. Subsequently, processing circuitry 112 can then transmit the new or updated treatment plan parameters to treatment device 180 via a communication interface and network, where the radiation therapy plan will be used to treat a 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 processes discussed below in connection with FIGS. 3-5). Figure 3

[0072] In examples herein, processing circuitry 112 can execute such software programs that invoke treatment processing logic 120 to implement ML, deep learning, NN, and other aspects of artificial intelligence for generating treatment plans from input radiation therapy medical information (e.g., CT images, MR images, and / or sCT images and / or dose information). For example, processing circuitry 112 can execute such software programs that train, analyze, predict, evaluate, and generate treatment plan parameters from received radiation therapy medical information as discussed herein.

[0073] In examples, image data 152 can include one or more MRI images (e.g., 2D MRI, 3D MRI, 2D flow MRI, 4D MRI, 4D volume MRI, 4D cine 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, fluoroscopy images, radiotherapy portal images, single photon emission computed tomography (SPECT) images, computer-generated synthetic images (e.g., pseudo-CT images), etc. Moreover, image data 152 can also include or be associated with medical image processing data (e.g., 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, maps, etc.). ​

[0074] In examples, image data 152 can be received from image acquisition devices 170 and stored in one or more of image data sources 150 (e.g., a picture archiving and communication system (PACS), a vendor neutral archive (VNA), a medical records or information system, a data warehouse, etc.). Thus, image acquisition devices 170 can include MRI imaging devices, CT imaging devices, PET imaging devices, ultrasound imaging devices, fluoroscopy devices, SPECT imaging devices, integrated linear accelerator and MRI imaging devices, CBCT imaging devices, or other medical imaging devices for obtaining medical images of patients. Image data 152 can be received and stored in any data type or any format type (e.g., in a digital imaging and communications in medicine (DICOM) format) that image acquisition devices 170 and radiation therapy processing computing system 110 can use to perform operations consistent with the disclosed implementations. Moreover, in some examples, the models discussed herein can be trained to process raw image data formats or derivatives thereof.

[0075] In examples, image acquisition devices 170 can be integrated with therapy devices 180 as a single device (e.g., an MRI device combined with a linear accelerator, also referred to as an “MRI-Linac”). Such an MRI-Linac can be used, for example, to determine the location of a target organ or a target tumor in a patient to accurately direct radiation therapy to a predetermined target according to a radiation therapy treatment plan. For example, a radiation therapy treatment plan can provide information about a specific radiation dose to be applied to each patient. The radiation therapy treatment plan can also include other radiation therapy information including control points for the radiation therapy device, such as bed position, beam intensity, beam angle, dose-histogram-volume information, number of radiation beams to be used during therapy, dose per beam, etc.

[0076] Radiation therapy processing computing system 110 can communicate with external databases over a network to send / receive a plurality of various types of data related to image processing and radiation therapy operations. For example, the external databases can include machine data (including device constraints) that provide information associated with therapy devices 180, image acquisition devices 170, or other machines related to radiation therapy or medical procedures. Machine data information (e.g., control points) can include beam size, arc placement, beam on and off duration, machine parameters, segments, multi-leaf collimator (MLC) configuration, gantry speed, MRI pulse sequence, etc. The external databases can be storage devices and can be equipped with appropriate database management software programs. Moreover, such databases or data sources can include multiple devices or systems located in a centralized or distributed manner.

[0077] The radiotherapy treatment computing system 110 can collect and obtain data and communicate with other systems via networks using one or more communication interfaces communicatively coupled to the processing circuitry 112 and the memory 114. For example, the communication interfaces can provide a communication connection between the radiotherapy treatment computing system 110 and radiotherapy system components (e.g., allowing data to be exchanged with external devices). For example, in some examples, the communication interfaces can have appropriate interface circuitry relative to the output device 146 or the input device 148 to connect to the user interface 142, which can be a hardware keyboard, keypad, or touchscreen through which a user can input information into the radiotherapy system 100.

[0078] As an example, the output device 146 can include a display device that outputs: a representation of the user interface 142; and one or more aspects, visualizations, or representations of medical images, treatment plans, and a status of training, generation, validation, or implementation of such plans. The output device 146 can include one or more display screens that display medical images, interface information, treatment plan parameters (e.g., contours, dose, beam angles, markers, maps, etc.), treatment plans, targets, localization of targets, and / or tracking of targets, or any information relevant to a user. The input device 148 connected to the user interface 142 can be a keyboard, keypad, touchscreen, or any type of device that a user can use for the radiotherapy system 100. Alternatively, features of the output device 146, the input device 148, and the user interface 142 can be integrated into a single device such as a smartphone or tablet computer (e.g., Apple iPhone®, Apple iPad®, Lenovo ThinkPad®, Samsung Galaxy®, etc.). Lenovo Samsung etc.).

[0079] Further, any and all components of the radiotherapy system 100 can be implemented as virtual machines (e.g., via a VMWare, Hyper-V, etc. virtualization platform) or standalone devices. For example, a virtual machine can be software that acts as hardware. Thus, 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 act as hardware. For example, the radiotherapy treatment computing system 110, the image data source 150, etc. components can be implemented as virtual machines or implemented within a cloud-based virtualization environment.

[0080] The image acquisition device 170 can be configured to acquire one or more images of the patient’s anatomy for a region of interest (e.g., a target organ, a target tumor, or both). Each image, which is typically a 2D image or slice, can include one or more parameters (e.g., 2D slice thickness, orientation, and location, etc.). In an example, the image acquisition device 170 can acquire 2D slices in any orientation. For example, the orientation of the 2D slices can include a sagittal orientation, a coronal orientation, or an axial orientation. The processing circuitry 112 can adjust one or more parameters (e.g., thickness and / or orientation of the 2D slices) to include the target organ and / or the target tumor. In an example, the 2D slices can be determined from information such as a 3D CBCT or CT or MRI volume. Such 2D slices can be acquired by the image acquisition device 170 “near real-time” (where “near real-time” means acquiring data in at least a few milliseconds or less) as the patient is undergoing radiotherapy treatment, e.g., when using the treatment device 180.

[0081] The treatment processing logic 120 in the radiotherapy treatment computing system 110 implements an ML model that involves using a trained (learned) ML model. The ML model can be provided by a NN trained as part of the NN model. One or more teacher ML models can be provided by a different entity or at an off-site facility related to the treatment processing logic 120 and can be accessed by issuing one or more queries to the off-site facility.

[0082] Machine learning (ML) algorithms or ML models or techniques can be summarized as function approximation. Training data consisting of pairs of inputs and outputs of some type (e.g., training parameters and one or more training optimization variables of a plurality of training radiotherapy treatment plan optimization problems) are acquired from, e.g., expert clinicians or previous optimization plan solvers and functions, and a function is “trained” to approximate the mapping. Some methods involve NNs. In these methods, a set of parameterized functions A0is chosen, where 0 is a set of parameters (e.g., convolution kernels and biases) chosen by minimizing the average error of the training data. If the input-output pairs are represented by (x m ,y m ), then the function can be formalized by solving, e.g., the following minimization problem:

[0083]

[0084] Once the network is trained (e.g., theta is chosen), the function A0can be applied to any new input. For example, in the above setting of a radiotherapy treatment planning optimization problem parameter, parameters of a radiotherapy treatment planning optimization problem that has never been seen before can be input to A0and the parameters of another, less complex radiotherapy treatment planning optimization problem are estimated.

[0085] A simple NN is composed of an input layer, an intermediate or hidden layer, and an output layer, each containing computational units or nodes. The hidden layer nodes have inputs from all input layer nodes 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 nodes according to a non-linear function of the sum of its inputs. For a classifier, the number of input layer nodes is typically equal to the number of features of each of the set of objects being classified into classes, and the number of output layer nodes is equal to the number of classes. The network is trained by presenting it with features of objects of known classes and adjusting the node weights to reduce training error through an algorithm called backpropagation. Thus, a trained network can classify new objects whose classes are unknown.

[0086] Neural networks have the ability to discover relationships between data and classes or regression values, and under certain conditions can model any function y = f(x) including non-linear functions. In ML, it is assumed that both training and test data are generated by the same data generation process p data generated, where each {x i , y i} sample is identically and independently distributed (i.i.d.). In ML, the goal is to minimize training error and to make the difference between training and test error as small as possible. If the training error is too large, underfitting occurs; when the training-test error gap is too large, overfitting occurs. Both types of performance deficiencies are related to model capacity: large capacity can fit the training data very well but lead to overfitting, while small capacity can lead to underfitting.

[0087] Figure 2A An exemplary image-guided radiotherapy device 242 is shown, which includes 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 device 232 can be configured to emit a radiotherapy beam 238 to provide treatment to a patient. The radiotherapy output 234 can include one or more attenuators or collimators (e.g., MLC).

[0088] As an example, a patient can be placed in the region 242 supported by the treatment couch 246 to receive a radiation therapy dose according to a radiation therapy treatment plan. The radiation therapy output 234 can be mounted or attached to the gantry 236 or other mechanical support. One or more gantry motors (not shown) can rotate the gantry 236 and the radiation therapy output 234 about the couch 246 as the couch 246 is inserted into the treatment region. In an example, the gantry 236 is capable of continuous rotation about the couch 246 as the couch 246 is inserted into the treatment region. In another example, the gantry 236 can be rotated to a predetermined position as the couch 246 is inserted into the treatment region. For example, the gantry 236 can be configured to rotate the therapy output 234 about an axis (“A”). Both the couch 246 and the radiation therapy output 234 are capable of independent movement to other positions around the patient, for example, capable of movement in a transverse direction (“T”), capable of movement in a lateral direction (“L”), or rotation about one or more other axes, for example, rotation about a transverse axis (indicated as “R”). A controller communicatively connected to one or more actuators (not shown) can control movement or rotation of the couch 246 to properly position the patient in or out of the radiation therapy beam 238 according to a radiation therapy treatment plan. Both the couch 246 and the gantry 236 are capable of independent movement from one another in multiple degrees of freedom, which enables the patient to be positioned so that the radiation therapy beam 238 can accurately target a tumor.

[0089] The coordinate system (including axis A, axis T, and axis L) can have an origin at the isocenter 240. The isocenter 240 can be defined as a location at which a central axis of the radiation therapy beam 238 intersects the origin of the coordinate axes, for example, to deliver a prescribed radiation dose to a location on the patient or a location within the patient. Alternatively, the isocenter 240 can be defined as a location at which the central axis of the radiation therapy beam 238 intersects the patient for various rotational positions of the radiation therapy output 234 about axis A as positioned by the gantry 236.

[0090] The gantry 236 can also have an imaging detector 244 attached. The imaging detector 244 is preferably located opposite the radiation source (output 234), and in examples, the imaging detector 244 can be located in the field of the therapy beam 238. The imaging detector 244 can be mounted on the gantry 236, preferably opposite the radiation therapy output 234, to remain in alignment with the radiation therapy beam 238. As the gantry 236 rotates, the imaging detector 244 rotates about the rotation axis. In examples, the imaging detector 244 can 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 radiation therapy beam 238, or the imaging detector 244 can be used to image the patient’s anatomy, such as portal imaging. Control circuitry of the radiation therapy device 232 can be integrated within the radiation therapy system 100 or remote from the radiation therapy system 100.

[0091] In illustrative examples, one or more of the couch 246, the therapy output 234, or the gantry 236 can be automatically positioned, and the therapy output 234 can establish the therapy beam 238 according to a specified dose for a particular therapy delivery instance. A sequence of therapy delivery can be specified according to one or more different orientations or positions of the gantry 236, the couch 246, or the therapy output 234, for example, in accordance with a radiation therapy treatment plan. The therapy delivery can occur sequentially, but can intersect at a desired treatment site on or within the patient, for example, at the isocenter 240. In this way, a cumulative dose of a prescription of radiation therapy can be delivered to the treatment site, while damage to tissue near the therapy site can be reduced or avoided.

[0092] Thus, Figure 2A Specific examples of radiation therapy devices 232 are shown that are operable to provide radiation therapy treatment to a patient in accordance with or consistent with a radiation therapy plan, with a configuration in which a radiation therapy output is rotatable about a central axis (e.g., axis “A”). Other radiation therapy output configurations can be used. For example, the radiation therapy output can be mounted to a robotic arm or manipulator with multiple degrees of freedom. In another example, the therapy output can be fixed, such as in a region laterally separated from the patient, and a platform supporting the patient can be used to align a radiation therapy isocenter with a specified target site within the patient. In yet another example, the radiation therapy device can be a combination of a linear accelerator and an image acquisition device. As will be appreciated by those of ordinary skill in the art, in some examples, the image acquisition device can be an MRI, X-ray, CT, CBCT, helical CT, PET, SPECT, optical tomography, fluorescence imaging, ultrasound imaging, or a radiation therapy portal imaging device, among others.

[0093] Figure 2BA Gamma Knife, in which the present disclosure can be used, is shown. A patient 202 can wear a coordinate frame 220 to keep a body part of the patient (e.g., a head) stable while undergoing surgery or radiation therapy. The coordinate frame 220 and patient positioning system 222 can establish a spatial coordinate system that can be used when imaging the patient or during a radiation procedure. The radiation therapy device 130 can include a protective housing 214 to enclose a plurality of radiation sources 212 used to generate radiation beams (e.g., sub-beams) that pass through beam channels 216. The plurality of beams can be configured to focus on an isocenter 218 from different locations. While each individual radiation beam can have a relatively low intensity, the isocenter 218 can receive a relatively high level of radiation when the multiple doses from the different radiation beams are accumulated at the isocenter 218. In certain implementations, the isocenter 218 can correspond to a target (e.g., a tumor) that is being operated on or treated.

[0094] According to some implementations, a class of optimization problems can be parameterized by an input parameter u e U, which can correspond to, for example, an objective function f and constraint functions ci. As an example, any linear programming problem can be equivalently expressed in standard form, where the objective function is determined by a cost vector w, and the constraint functions are determined by a constraint matrix A and a constraint vector b (but details depend on the convention used), so in some implementations the input parameter u = {w, A, b}.

[0095] A given selection of the parameter u defines an optimization problem, e.g., by a mapping F(u) = {f, X}, where f is the objective function, and X is the set of feasible points. Assuming there exists a point that satisfies the constraints (e.g., X is not empty), a (local) solution to the optimization problem is defined as a point set x* e X such that for all feasible x in a neighborhood of x*, f(x*) < f(x). The solution can not be unique. According to some implementations, solving the optimization problem includes finding at least one point x* that is a solution to the optimization problem or within a specified tolerance of a solution to the optimization problem.

[0096] A loss function can be used to quantify how far a point x is from being a solution to the first optimization problem. One example, if x and x* are points in a normed vector space and x* is unique, the relevance is to use the induced metric l(x, x * ) = ||x - x * ||, where the norm ||.|| can be a p-norm. Alternatively, x can be projected onto the feasible set X before the loss is evaluated. An alternative is to use the objective function value of the first optimization problem directly as the loss function l(x). A variation is to use the objective function value (or the current iteration in the loss) of the first optimization problem after it has been processed by the optimization algorithm (e.g., for a fixed number of iterations or until convergence) - in this case, the loss quantifies how good the local solution found after starting from x is.

[0097] To illustrate the disclosed implementation method, T is used to encode the process of solving the optimization problem according to Equation 4:

[0098]

[0099] In this way, the original (first) optimization problem can be compactly formulated as searching for point x*(u) in the image of T°F(u) according to Equation 5:

[0100]

[0101] As mentioned above, the original (first) optimization problem can be difficult and complex to solve and use due to its variable size and ordering, requiring considerable effort to solve. Therefore, according to some implementations, the second, more convenient, simplified, and less complex optimization problem is formulated by the treatment processing logic 120, for example, according to Equation 6, using a mapping G(u) = {g, Z} different from F(u):

[0102]

[0103] The mapping G(u) = {g, Z} defining the relationship between the parameters of the second optimization problem and the parameters of the original optimization problem can be generated or provided by an ML model, a nonlinear functional relationship, and / or a statistical model. Specifically, the behavior of the mapping G, the encoder henc (which maps the parameters of the first optimization problem to the parameters of the second optimization problem), and / or the decoder hdec (which maps the parameters of the second optimization problem to the first optimization problem) can be controlled by the treatment processing logic 120 retrieving or generating a parameter set θ. In some implementations, some of these parameters θ are calculated or estimated by the treatment processing logic 120 based on a nonlinear functional relationship or a statistical model. The statistical model models the first parameter set as random variables dependent on the second parameter set and derives the objective function or constraints of the second radiotherapy treatment planning optimization problem based on the constraints or a measure of the central tendency of the objective function of the first radiotherapy treatment planning optimization problem.

[0104] In some implementations, such as combining Figure 3 As shown and described, some of these parameters θ can be selected by the treatment processing logic 120 and estimated using an ML model trained on a dataset of other treatment plans. In such a case, if both G and hdec are parameterized, they can be jointly trained by the treatment processing logic 120. Training can be performed in a supervised manner, for example, using the pair of M... The data is constructed, and the goal is to choose θ to minimize the expected loss on the dataset. Equation 7 shows how to compute the θ parameter in a supervised manner:

[0105]

[0106] When trained in an unsupervised manner, e.g., with data consisting of M problem instances The corresponding representation is provided according to equation 8:

[0107]

[0108] In such cases, the purpose of the training can also be to maximize a statistical measure of, e.g., the likelihood of observing the training data, or to maximize an indicator of the utility of a treatment plan generated based on the training data.

[0109] In some implementations, the solution z* of the second optimization problem can itself be used, such that a treatment plan can be generated by the treatment processing logic 120 based on the solution z*. The process of defining the second optimization problem can be designed to maximize the utility of z* in generating a treatment plan. In some implementations, z* (the solution of the second optimization problem) can be decoded by the treatment processing logic 120 to map the solution to the same space as the parameters (e.g., optimization variables) of the original first optimization problem. In some implementations, the decoding is performed by the treatment processing logic 120 using a function hdec(z). In some cases, the function hdec is generated by the treatment processing logic 120 such that when the solution of the second optimization problem is applied to the function hdec(z*), the result approximately solves the original first optimization problem. Similarly, an encoder can be used to map parameters (e.g., optimization variables) from the first optimization problem to the second optimization problem.

[0110] Figure 3 An exemplary data flow for training and using a machine learning model to estimate parameters of a second radiation therapy treatment plan optimization problem from parameters of a first radiation therapy treatment plan optimization problem is shown in accordance with some examples of the present disclosure. The data flow includes training input 310, ML model (technique) training 330, and model use 350.

[0111] The training input 310 includes model parameters 312 and training data 320, which can include pairs of training data sets 322 (e.g., input-output training pairs) and constraints 326. The model parameters 312 store or provide the parameters or coefficients of the machine learning model During training, these parameters 312 are adapted based on the input-output training pairs of the training data sets 322. After the parameters 312 are adapted (after training), the parameters are used by the trained treatment model 360 to implement the trained machine learning model

[0112] The training data 320 includes constraints 326, which can define physical constraints of a given radiation therapy device or constraints of the first or second optimization problems. The paired training data sets 322 can include a set of input-output pairs, such as pairs of solutions of the first plurality of first radiation therapy treatment plan optimization problems and parameters of the first plurality of first radiation therapy treatment plan optimization problems. Some components of the training input 310 can be stored separately from other components at different off-premise facility or facilities.

[0113] The machine learning model training 330 trains one or more machine learning techniques based on the set of input-output pairs of the paired training data sets 322 For example, the model training 330 can train the ML model parameters 312 by minimizing a first loss function based on one or more solutions of the first plurality of first radiation therapy treatment plan optimization problems and corresponding training parameters of the first plurality of first radiation therapy treatment plan optimization problems.

[0114] The result of minimizing the loss function of the plurality of training data sets trains, adjusts, or optimizes the model parameters 312 of the corresponding ML model. In this way, the ML model is trained to establish a relationship between the parameters of the second radiation therapy treatment plan optimization problem and the parameters of the plurality of first training radiation therapy treatment plan optimization problems.

[0115] As an example, the ML model can be trained according to a supervised learning technique. In such a case, to train the ML model Λ θ, a plurality of training optimization problems (and / or include synthetically generated problems) previously solved for other patients are retrieved along with their corresponding training parameters (e.g., optimization variables and solutions). Specifically, a first training batch of data is obtained that includes a batch of parameters of a first training radiation therapy treatment plan optimization problem and a corresponding batch of solutions to the first training radiation therapy treatment plan optimization problem. As another example, the first training batch of data includes a batch of parameters of a plurality of first training radiation therapy treatment plan optimization problems and a corresponding batch of solutions to the plurality of first training radiation therapy treatment plan optimization problems. This batch of training data can be used to train the ML model with the same parameters of the ML model and can range from one parameter set of the first training radiation therapy treatment plan optimization problem and the corresponding solution to the first training radiation therapy treatment plan optimization problem to all training problems and solutions. The given batch of parameters is processed using the ML model to generate an estimate of the parameters of the second radiation therapy treatment plan optimization problem. For example, the ML model is applied by the treatment processing logic 120 to the first training batch of data to compute the mapping G(u) = {g, Z}. The second radiation therapy treatment plan optimization problem is solved by the treatment processing logic 120 based on the estimate of the parameters of the second radiation therapy treatment plan optimization problem to generate a given solution to the second radiation therapy treatment plan optimization problem. For example, the second radiation therapy treatment plan optimization problem is solved to compute A deviation is computed between at least one of (1) the given solution z* and the corresponding solution x* or (2) a decoded version of the given solution hdec(z*) and the corresponding solution x* in the space of the first training radiation therapy treatment plan optimization problems of the given batch. The parameters of the machine learning model are updated based on the computed deviation.

[0116] The ML model is then applied by the treatment processing logic 120 with the updated parameters to another batch of training data to again estimate the given parameter set of the second radiation therapy treatment plan optimization problem in a similar manner as the first batch to compute a deviation and update the parameters of the ML model. The parameters of the ML model are again updated and the iteration of the training process continues for a specified number of iterations or epochs or until a given convergence criterion is met.

[0117] The ML model can be trained by the treatment processing logic 120 according to an unsupervised learning technique. The unsupervised learning technique assumes that the true solution is unknown. In such a case, to train the ML model Λ θ, training data comprising a given batch of parameters of the first training radiotherapy treatment plan optimization problem is obtained. The batch of parameters of the first training radiotherapy treatment plan optimization problem can be used to train the ML model with the same parameters of the ML model, and can range from one parameter set of the first training radiotherapy treatment plan optimization problem to all parameters of the first training radiotherapy treatment plan optimization problem. The obtained given batch of parameters is processed by the treatment processing logic 120 with the ML model to generate an estimate of the parameters of the second radiotherapy treatment plan optimization problem. The second radiotherapy treatment plan optimization problem is solved by the treatment processing logic 120 based on the estimate of the parameters of the second radiotherapy treatment plan optimization problem to generate a given solution to the second radiotherapy treatment plan optimization problem. For example, the second radiotherapy treatment plan optimization problem is solved to compute a metric of the given solution z*. Based on the metric, updated parameters of the ML model are computed by the treatment processing logic 120.

[0118] The ML model is then applied by the treatment processing logic 120 with the updated parameters to another batch of parameters of the first training radiotherapy treatment plan optimization problem to again estimate a given parameter set of the second radiotherapy treatment plan optimization problem, and generate a solution to the second radiotherapy treatment plan optimization problem. The parameters of the ML model are again updated, and the iteration of the training process continues for a specified number of epochs, or until all batch parameters of the first training radiotherapy treatment plan optimization problem are processed.

[0119] After the machine learning model is trained (sometimes referred to as Λ θ ), new data 370 comprising one or more patient input parameters (e.g., a radiotherapy treatment plan optimization problem) can be received. The trained machine learning technique is applied to the new data 370 to generate a generated result 380 comprising one or more estimated parameters of a second radiotherapy treatment plan optimization problem. The generated one or more estimated parameters of the second radiotherapy treatment plan optimization problem are then used by the treatment processing logic 120 to solve the second optimization problem, for example, by using a simplex method, an interior point method, or a sequential quadratic programming. The solution to the second optimization problem can then be used directly by the treatment processing logic 120 to generate a radiotherapy treatment plan and / or decoded onto the space of the first optimization problem to estimate a solution to the first optimization problem.

[0120] Optimization problems often have two symmetrical properties, including commutativity of permutable variables and invariance of permutable constraints. Commutativity of permutable variables means that reordering the optimization variables gives a corresponding reordering of the solution (without affecting the optimal value of the objective function). Invariance of permutable constraints means that if the feasible set is described as a set of equalities and / or inequalities, reordering them has no effect on the solution set (but can change the behavior of the optimization algorithm).

[0121] The class of permutation-invariant functions can be characterized such that for some suitable transformations p and f, any permutation-invariant function h(Y) operating on a set Y can be decomposed in the form according to equation 9:

[0122] h(Y) = p(∑ y∈Y f(y)) (9) This means that permutation-invariant functions can be learned by the therapy processing logic 120 using this expression, e.g., by learning the transformations p and f with an ML algorithm such as a neural network. In some cases, a permutation-commutative neural network can be provided.

[0123] According to some embodiments, the mapping G(u) between the parameters of the first optimization problem and the parameters of the second optimization problem can be defined by the therapy processing logic 120 as a permutation-invariant or commutative neural network (ML model). Alternatively, the second optimization problem can be processed by the therapy processing logic 120 based on permutation-invariance or commutativity. If the result of processing the second optimization is a permutation-invariant quantity, the parameters of the second optimization problem can be decoded by the therapy processing logic 120 according to the original ordering of the parameters of the first optimization problem to reverse the permutation. In this way, the decoded parameters recover the original ordering. The symmetry requirement can be limited to some subset of the optimization variables or constraints (e.g., emissions in the case of a gamma knife plan).

[0124] To address the size variability of the first optimization problem, a fusion step involving a quasi-arithmetic mean according to equation 10 can be utilized to process the input:

[0125]

[0126] where p is a reversible function that can be specified a priori or learned, e.g., using a reversible neural network. The quasi-arithmetic mean can also be permutation invariant. In another implementation, some parameters between the first and second optimization problems can be identified as correlated by a base stochastic process. By Kolmogorov extension theorem, any finite set of random variables from a stochastic process satisfies (i) commutativity, which is essentially the same as permutation invariance, and (ii) consistency, which means that after marginalizing out a subset, the remaining random variables still belong to the same stochastic process. Due to the consistency property, the optimization variables or parameters of the first optimization problem can be reduced by the therapy processing logic 120 to a set of parameter variables with fixed size.

[0127] In some cases, the mapping function G(u) between the first and second optimization problems is created by the therapy processing logic 120 such that the second optimization problem has fixed size and ordering relative to the first optimization problem. In such cases, an ML model (e.g., a neural network) is trained by the therapy processing logic 120 to predict the solution of the second optimization problem. In this way, any parameters can be trained in an end-to-end manner if G and hdec are differentiable.

[0128] In some implementations, the original (first) optimization problem variables are encoded by the therapy processing logic 120 into a lower dimensional representation z = henc(x). In such implementations, the second optimization problem is defined by g(z) = f(hdec(z)) and Z = henc(X). In such cases, such that the solution of the second optimization problem is defined according to equation 11:

[0129]

[0130] The encoder and decoder can be different and can be parameterized by the therapy processing logic 120 using neural networks that can be trained to reconstruct the input (as explained above in connection with Figure 3 In some implementations, the cycle-consistency constraint is used by the therapy processing logic 120 in training. In some implementations, the decoder is the inverse of the encoder, and in such cases, the encoder and decoder can be learned using reversible neural networks.

[0131] In some implementations, the first optimization problem is a quadratic programming problem. Such a problem can be written in general form by equation 12:

[0132]

[0133] where Qx is a symmetric matrix (not necessarily positive definite), px is a vector, and Ax is a matrix. x* can be an undefined prior - which is determined by the optimization problem. Given the index variable r, a base stochastic process y(r) can be provided such that x = y(rx). The stochastic process evaluated at rz is defined by z = y(rz). From the stochastic process, a less complex z compared to x is selected such that the first optimization problem can be represented by the second optimization problem. y(r) can be, for example, a Gaussian process, where y ~ GP(μ(r), k(r; r')), where μ(r) is a mean function and k(r, r') is a covariance function. The Gaussian process assumption requires

[0134]

[0135] where μx= μ(rx), kxz= k(rx, rz), and so on. Given z = y(rz) at some selected locations rz, x can be estimated using such conditions, for example, using Gaussian process regression. The conditional distribution is also a Gaussian distribution as shown in Equations 14-16:

[0136] x|z ~ N(μ x|z ,∑ x|z ) (14)

[0137]

[0138]

[0139] This can be used to show two different ways of formulating the second optimization problem (in the form of z) based on the first optimization problem (in the form of x). In the special case of a quadratic program where x and z are correlated through a base Gaussian process, the two approximations coincide. The first method for reformulating the parameters of the first optimization problem by the treatment processing logic 120 with the parameters of the second optimization problem is to use a point-wise mapping, where x in the first optimization problem is replaced with the point estimate μx|z. The second method involves marginalizing all expressions of x by the treatment processing logic 120 with the conditional distribution.

[0140] In some implementations, the parameters of the second optimization problem (in the form of z) are reformulated by the treatment processing logic 120 based on the parameters of the first optimization problem (in the form of x) based on the point estimate μx|z. In such a case, the second optimization problem (which is also a quadratic optimization problem) is defined by Equation 17 by replacing the parameters (e.g., optimization variable x) in the first optimization problem defined by Equation 12 with the point estimate μx|z from Equation 15 (ignoring the constant term):

[0141]

[0142] According to the second optimization problem, the optimization variable of the first optimization problem is defined by z instead of x. In some cases, x can be represented by a fixed and much smaller set of points z. In some cases, z can be chosen so that the involved matrices have some particular structure that is computationally superior to the parameters of the first optimization problem. In some implementations, instead of replacing z and x, the equation constraints can be used by the treatment processing logic 120.

[0143] In some implementations, the parameters of the second optimization problem (in the form of z) are reformulated by the treatment processing logic 120 based on the parameters of the first optimization problem (in x) by marginalizing all expressions of x with the conditional distribution P(x|z). For simplicity, the convex case is described when Qx is positive definite, but the indefinite case can also be treated in a similar way. If Qx is positive definite, it has a Cholesky decomposition Qx = LTL. From the general rule for linear transformations of normally distributed variables s ~ N(μ,∑), it follows that z = Ls is distributed according to equation 18:

[0144] z ~ N(Lμ, L∑LT) (18)

[0145] Furthermore, from the definition of the covariance, E[ssT] =∑+ μμT. Combining this expression with the covariance of z, the conditional expectation on the objective function is defined by equations 19-23:

[0146]

[0147] The first term Tr(Q∑x|z) is constant and thus irrelevant for the optimization. From equation (18), it follows that and E[Axx|z] = Axμx|z. In summary, for the convex quadratic optimization problem, the marginalization over the conditional distribution is equivalent to replacing x with the point estimate μx|z. The resulting approximate second optimization problem is given in equation 17. The indefinite case can be handled by using eigenvalue decomposition and splitting the diagonal matrix into positive and negative parts. By linearity, the result is the difference of two positive definite quadratic forms each of which can be handled as described above.

[0148] As an example, a scenario is described that considers dose planning for gamma knife radiosurgery. A reasonably accurate model for the first optimization problem for the dose deposited by a gamma knife at least some distance from the surface is defined according to equation 24:

[0149]

[0150] where is the (translationally invariant) dose rate kernel that gives the dose rate at position r centered at position r' (for each of the 24 machine configurations), i.e. p(r') scales the dose rate by the density at r' and w(r') is the radiation time of the 24 machine configurations at location r'. Current gamma knife treatments use a discrete set of isocenter locations ri, i = 1,..., n, which means that where, is the 24 configured radiation time at location ri.

[0151] According to the disclosed embodiments, the parameters of the first optimization problem (e.g., w) can be approximated by w ~ GP(μ(r'), k(r', r")) where the mean function μ(r) and the covariance function k(r, r') are suitably parameterized functions. A useful property of Gaussian processes is that they are closed under linear function transformations L[w]. In particular, the dose calculation in equation 24 is such a transformation. Thus, the dose can be expressed by a second optimization problem described by a Gaussian process according to equations 25-27:

[0152] d ~ GP(L[μ], L2[k]) (25)

[0153]

[0154]

[0155] In this way, the first optimization problem can be approximated by the treatment processing logic 120 by marginalizing or point-estimating the second optimization problem.

[0156] In one example, the first optimization problem can be reformulated by the treatment processing logic 120 as a second optimization problem by subsampling certain parameters of the first optimization problem. Specifically, consider a one-dimensional dose planning problem where the delivered dose is a weighted combination of basis functions of the form defined by equation 28:

[0157]

[0158] This means that the delivered dose is where n "isocenter" locations r' i are chosen randomly within the target and the prescribed dose is 1 within the target and 0 outside the target. The goal is to determine the best settings of the non-negative weights w. A reasonable way to formulate this optimization problem is as a non-negative least squares problem of equation 29:

[0159]

[0160] The disclosed embodiments simplify the first optimization problem of equation 29 using a second optimization problem defined by equation 30 by subsampling the parameters w of the first optimization problem:

[0161]

[0162] That is, the second optimization problem is defined by sampling a subset s of indices with corresponding weights ws, and is solved only with respect to these (the others are effectively set to zero). The corresponding (decoded) variables are w = {wsif i∈s, otherwise 0} according to the first optimization problem.

[0163] As another example of reducing the number of optimization variables, the first optimization problem according to equation 31 is a smooth, constrained, non-convex problem:

[0164]

[0165] Such a form of the optimization problem can occur multiple times with different values of the input parameters u = {a > 0, b > 0, β ∈ [0, π], R > 0}. The second optimization problem can be generated by the therapy processing logic 120 based on the input parameters u. For example, the therapy processing logic 120 can switch the first optimization problem decision variables to be expressed in polar coordinates, so that it is solved only along radial lines, the angle of which is predicted by the ML model. In such a case, the original decision variables can be expressed in polar coordinates according to equation 32:

[0166] x(r, φ) = r cos φ

[0167] y(r, φ) = y sin φ (32)

[0168] The variable φ ∈ [0, π] can be predicted by the therapy processing logic 120 based on a trained ML model defined as φ = fθ(u) and kept fixed during optimization.

[0169] Since the number and order of the input parameters is fixed in this example, a two-layer, fully connected neural network (multilayer perceptron) can be used to predict φ. Furthermore, the constraint y >= 0 is automatically satisfied from φ ∈ [0, π] and r > 0. Thus, the second optimization problem can be expressed by the therapy processing logic 120 as a smooth, constrained, non-convex problem according to equation 33:

[0170]

[0171] As shown, the parameters x and y of the first optimization problem are encoded into the parameters r cos φ and r sin φ of the simpler second optimization problem using a trained ML model, where φ is predicted by the trained ML model.

[0172] The therapy processing logic 120 can solve the second optimization problem for a given input parameter set u to obtain a solution r* = r*(u), which can be decoded into the decision variables of the first optimization problem using equation 32 together with the predicted φ.

[0173] In some implementations, the parameters θ of the ML model φ = fθ(u) that predicts φ can be trained by the therapy processing logic 120 by approximately solving the first optimization problem using the decoded solution of the decoding by encouraging the decoded solution. Specifically, the loss function for training the ML model can be defined by Equation 34:

[0174]

[0175] This particular choice of the loss function that only utilizes the input parameters u and does not require knowledge of any solution corresponds to the case of unsupervised training. The ML model can be applied to a first set of training parameters to estimate the value of φ. The estimated value of φ can be input to Equation 33 to provide a second optimization problem. Equation 33 is then solved using an optimization technique to obtain a value of the parameters r of the second optimization problem. The obtained value of the parameters r is input to Equation 34 (e.g., using the solution of the second optimization problem with the estimated value of φ) to compute the loss.

[0176] During supervised training, the parameters r of the second optimization problem are decoded onto the space of the first optimization problem (e.g., the parameters r are used to generate the x and y parameter values) according to Equation 32. The decoded solution of the second optimization problem is compared to a batch of training solutions of the first optimization problem that were obtained to compute the bias. This batch of training solutions of the first optimization problem can be used to train the ML model with the same parameters of the ML model, and can range from one training solution to all training solutions. When a stopping criterion is satisfied, the training of the ML model is complete. In some implementations, the stopping criterion includes whether a maximum number of iterations has been reached, whether the loss is within a threshold amount, or whether the parameter updates are less than a threshold amount. If not, the parameters of the ML model are updated and used to generate another solution of the second optimization problem and corresponding solution of the first optimization problem based on another batch of training parameters and solutions of the first optimization problem. The parameters of the ML model are updated, and the iteration of this training process continues for a specified number of iterations or epochs, or until a given convergence criterion or stopping criterion is satisfied.

[0177] During unsupervised training of the ML model, the training of the ML model is complete when a stopping criterion is satisfied. In some implementations, the stopping criterion includes whether a maximum number of iterations has been reached, whether the loss is within a threshold amount, or whether the parameter updates are less than a threshold amount. If not, the parameters of the ML model are updated and used to generate another solution of the second optimization problem based on another batch of training parameters of the first optimization problem.

[0178] The therapy processing logic 120 can know exactly what probability distribution the input parameters u follow, but this can not matter because training typically minimizes the empirical risk. That is, the average of the loss function over a fixed training dataset is expressed according to Equation 35:

[0179]

[0180] As another example of mapping a variable-size problem to a fixed representation, a first optimization problem with a different number and ordering of decision variables can be mapped to a fixed-size representation and expressed as a second optimization problem. This can be done by assuming that the decision variables belong to a stochastic process (e.g., a Gaussian process). The first optimization problem can be formulated as a one-dimensional dose planning problem, where the delivered dose is a weighted combination of basis functions in the form expressed in Equation 36:

[0181]

[0182] The input variable u = {r'1,..., r'n} is a variable-size set (e.g., n is not constant) corresponding to isocenter locations within a target tumor located within a patient's body, which extends from r = -2 to r = 2. The delivered dose can be expressed according to Equation 37:

[0183]

[0184] For simplicity, it can be assumed that the desired dose is 1 within the target tumor and 0 outside the target tumor (|r| > 2). The goal is to determine the best setting of the non-negative weights w. Thus, the first optimization problem can be expressed as a non-negative least squares problem, as shown in Equation 38:

[0185]

[0186] If the dose is evaluated on a grid, the integral becomes a sum. For this example, the grid can be restricted to r e {-3, -2, -1, 0, 1, 2, 3}, resulting in the objective function of the first optimization problem being expressed according to Equation 39:

[0187]

[0188] where the last expression assumes an implicit summation over r. Thus, the first optimization problem can be rewritten as a quadratic optimization problem according to Equation 40:

[0189]

[0190] In some implementations, the first optimization problem defined by Equation 40 can be simplified to a second optimization problem by sampling a subset s of the indices with the corresponding weights ws. The second optimization problem can then be solved only according to these sampled values, with the remaining being 0. In this way, the decoded variables of the first optimization problem can be expressed as w = {wsif i e s, otherwise 0}.

[0191] In some implementations, instead of sampling, the second optimization problem can be defined by applying a Gaussian process model to the weights, according to some implementations. Specifically, a plurality of synthetic center-of-equilibria positions u* can be used, which can or can not include elements of the observed center-of-equilibria positions. These synthetic center-of-equilibria positions can be selected according to some heuristic or by training based on the observed set of problem instances. In some implementations, the center-of-equilibria positions can be defined as u* = {-2, 0, 2}. Assuming that the weights can be modeled as random samples from a Gaussian process with zero mean and a covariance function given by spatial proximity according to an exponential kernel, the following equations 41 and 42 can be expressed:

[0192] μ(r) = 0 (41)

[0193]

[0194] where Δ is a hyperparameter (length scale of relevance) that can be selected based on some heuristic or training. Equations 41 and 42 can be used as a function that maps the parameters or variables of the first optimization problem to the parameters and variables of the second optimization problem.

[0195] The optimization variables of the first optimization problem are x = {w-1, w0, w1, w2}, while the optimization variables of the second optimization problem are z = {w-2, w0, w2}. The treatment processing logic 120 can reformulate the first optimization problem as a second optimization problem in the form of z by replacing x in the first optimization problem with point estimates based on z and using equations 41 and 42 to optimize with respect to z instead of x. The conditional mean of the Gaussian process can be used as a point estimate according to equation 43:

[0196]

[0197] To clean up the notation, one can introduce to restate the second optimization problem according to equation 44:

[0198]

[0199] Figure 4is a flowchart showing example operations of the treatment processing logic 120 in performing the process 400 according to example embodiments. The process 400 can be implemented in computer readable instructions for execution by one or more processors, such that the operations of the process 400 can be performed, in part or in whole, by functional components of the treatment processing logic 120; thus, the process 400 is described below by way of example with reference to the treatment processing logic 120. However, in other embodiments, at least some of the operations of the process 400 can be deployed on various other hardware configurations. The process 400 is therefore not intended to be limited to the treatment processing logic 120 and can be implemented in whole or in part by any other component. Some or all of the operations of the process 400 can be performed in parallel, out of order, or omitted entirely.

[0200] At operation 410, the treatment processing logic 120 receives training data. For example, the treatment processing logic 120 receives pairs of training parameters and one or more training optimization variables for a plurality of training radiotherapy treatment plan optimization problems; pairs of training parameters and one or more training optimization variables and solutions to a plurality of training radiotherapy treatment optimization problems; and pairs of training parameters and one or more training optimization variables for a plurality of training radiotherapy treatment plan optimization problems of a given type.

[0201] At operation 420, the treatment processing logic 120 receives constraints for training.

[0202] At operation 430, the treatment processing logic 120 performs training of the model. For example, the treatment processing logic 120 can train the ML model parameters 312 by minimizing a loss function based on the one or more solutions to the first batch of first radiotherapy treatment plan optimization problems and the corresponding training parameters of the first batch of first radiotherapy treatment plan optimization problems. Figure 3 In this way, the ML model is trained to establish a relationship between the parameters of the second radiotherapy treatment plan optimization problem and the parameters of the plurality of first training radiotherapy treatment plan optimization problems. The training can be performed in a supervised manner or an unsupervised manner.

[0203] At operation 440, the treatment processing logic 120 outputs the trained model. For example, the trained model can be output and stored in memory, or the parameters of the model can be presented to a clinician on a display device.

[0204] At operation 450, the treatment processing logic 120 utilizes the trained model to generate a result. For example, after each machine learning model (sometimes referred to as Λ θ ) is trained, new data 370 can be received that includes one or more patient input parameters for a first radiotherapy treatment plan optimization problem. The trained machine learning technique The new data 370 can be applied to generate the results 380 including one or more parameters of the second optimization problem.

[0205] Figure 5 is a flowchart illustrating example operations of the treatment processing logic 120 in performing the process 500, according to an example implementation. The process 500 can be implemented in computer-readable instructions for execution by one or more processors, such that the operations of the process 500 can be performed, in part or in whole, by functional components of the treatment processing logic 120; thus, the process 500 is described below by way of example with reference to the treatment processing logic 120. However, in other implementations, at least some of the operations of the process 500 can be deployed on various other hardware configurations. The process 500 is therefore not intended to be limited to the treatment processing logic 120 and can be implemented in whole or in part by any other component. Some or all of the operations of the process 500 can be performed in parallel, out of order, or omitted entirely.

[0206] At operation 510, the treatment processing logic 120 receives a first radiation therapy treatment plan optimization problem having a first set of parameters.

[0207] At operation 520, the treatment processing logic 120 processes the first set of parameters to estimate a second set of parameters of a second radiation therapy treatment plan optimization problem.

[0208] At operation 530, the treatment processing logic 120 generates a solution to the second radiation therapy treatment plan optimization problem based on the estimated second set of parameters.

[0209] At operation 540, the treatment processing logic 120 generates a radiation therapy treatment plan based on the solution to the second radiation therapy treatment plan optimization problem.

[0210] Figure 6 is a flowchart illustrating example operations of the treatment processing logic 120 in performing the process 600, according to an example implementation. The process 600 can be implemented in computer-readable instructions for execution by one or more processors, such that the operations of the process 600 can be performed, in part or in whole, by functional components of the treatment processing logic 120; thus, the process 600 is described below by way of example with reference to the treatment processing logic 120. However, in other implementations, at least some of the operations of the process 600 can be deployed on various other hardware configurations. The process 600 is therefore not intended to be limited to the treatment processing logic 120 and can be implemented in whole or in part by any other component. Some or all of the operations of the process 600 can be performed in parallel, out of order, or omitted entirely.

[0211] At operation 610, the treatment processing logic 120 receives a plurality of first training radiotherapy treatment plan optimization problems, each of the plurality of first training radiotherapy treatment plan optimization problems including a first set of parameters.

[0212] At operation 620, the treatment processing logic 120 trains a machine learning model to generate an estimate of a second set of parameters of a second radiotherapy treatment plan optimization problem by establishing a relationship between the first set of parameters of the plurality of first training radiotherapy treatment plan optimization problems and the second set of parameters of the second radiotherapy treatment plan optimization problem.

[0213] As previously discussed, various electronic computing systems or devices can implement one or more of the method or functional operations as discussed herein. In one or more implementations, the radiotherapy processing computing system 110 can be configured, adapted, or employed to control or operate the image-guided radiotherapy device 202, perform or implement training or prediction operations from Figure 3 the trained treatment model 360, perform or implement the operations of the flowcharts of the processes 400-600, or perform any one or more of the other methods discussed herein (e.g., as part of the treatment processing logic 120). In various implementations, such electronic computing systems or devices operate as standalone devices or can be connected (e.g., networked) to other machines. For example, such a computing system or device can operate in the capacity of a server or a client machine in server-client network environments, or it can operate as a peer machine in peer-to-peer (or distributed) network environments. The features of the computing system or device can be implemented by a personal computer (PC), a tablet PC, a personal digital assistant (PDA), a cellular telephone, a web appliance, or any machine capable of executing instructions (sequential or otherwise) that specify actions to be taken by that machine. Further, while a single machine can perform various functions as described, it is contemplated that multiple machines can perform the same or similar functions, and that the functions can be distributed among the machines.

[0214] As also indicated above, the functionality discussed above can be implemented by instructions, logic or other information storage on a machine readable medium. While it can have been described in a variety of examples that a machine readable medium can comprise a single medium, the term “machine readable medium” should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) storing one or more transitory or non-transitory instructions or data structures. The term “machine readable medium” shall also be taken to include any tangible medium that is capable of storing, encoding or carrying the transitory or non-transitory instructions or data structures used in the execution of the methodologies of the present disclosure by a machine, and that causes the machine to perform any one or more of the methodologies of the present disclosure, or that is otherwise utilized or associated with such

[0215] The above detailed description includes references to the accompanying drawings, which form a part of the detailed description. The drawings show, by way of illustration, specific embodiments in which the disclosure can be practiced. These embodiments are also referred to as “examples.” Such examples can include elements in addition to those shown or described. However, the present disclosure also contemplates examples in which only those elements shown or described are provided. Moreover, the present disclosure also contemplates examples using any combination or permutation of those elements shown or described (or one or more aspects thereof), either with respect to a particular example (or one or more aspects thereof), or with respect to other examples (or one or more aspects thereof) shown or described herein.

[0216] All publications, patents, and patent documents referred to in this document are incorporated by reference herein in their entirety, as though individually incorporated by reference. In the event of inconsistent usages between this document and those documents so incorporated by reference, the usage in the incorporated reference(s) should be considered supplementary to that of this document; for irreconcilable inconsistencies, the usage in this document prevails.

[0217] In this document, the terms “a,” “an,” “the,” and “said” are used in reference to an element of the disclosure that precedes them, whether explicit or implicit, and encompasses one or more than one, independent of any other instances or usages of “at least one” or “one or more.” In this document, the term “or” is used in its nonexclusive sense, such that “A or B” includes “A but not B,” “B but not A,” and “A and B,” unless otherwise indicated.

[0218] In the appended claims, the terms “including,” “containing,” and “having” are used as the plain-English equivalents of the respective terms “comprising,” and “wherein.” Also, in the following claims, the terms “first,” “second,” and “third,” etc. are used merely as labels, and are not intended to impose numerical requirements on their objects.

[0219] The present disclosure also relates to computing systems adapted, configured or operated to perform the operations herein. This system can be specially constructed for the required purposes or it can comprise a general purpose computer selectively activated or reconfigured by a computer program stored in the computer. The order of execution or performance of the operations in embodiments of the disclosure illustrated and described herein is not essential, unless otherwise specified. That is, embodiments of the present disclosure can perform the operations outlined in any order, unless otherwise specified, and embodiments of the present disclosure can include additional or fewer operations than those disclosed herein. For example, it is contemplated that executing or performing a particular operation before, concurrently with, or after another operation is within the scope of aspects of the present disclosure. The above description is that of current embodiments of the disclosure. Various alterations and changes can be made without departing from the spirit and scope of aspects of the disclosure. It therefore is understood that the application is not limited to the exact details of construction, products and methods described herein and that all changes and modifications encompassed by the scope of the claims are meant to be within the scope of aspects of the present disclosure.

[0220] In light of the above, it will be seen that the several objects of the disclosure are achieved and other advantageous results attained. Having described aspects of the disclosure in detail, it will be apparent to those skilled in the art that modifications and variations are possible without departing from the aspects of the disclosure defined in the appended claims. Since various changes could be made in the above constructions, products and methods without departing from the scope of aspects of the disclosure, it is intended that all matter contained in the above description and shown in the accompanying drawings shall be interpreted as illustrative and not in a limiting sense.

[0221] Examples described herein can be implemented in a wide variety of embodiments. For example, one embodiment includes a computing device that includes processing hardware (e.g., a processor or other processing circuitry) and memory hardware (e.g., a storage device or volatile memory) that includes instructions embodied thereon such that the instructions, when executed by the processing hardware, cause the computing device to implement, perform, or coordinate the electronic operations for these techniques and system configurations. Another embodiment discussed herein includes a computer program product, for example, embodied in a machine-readable medium or other storage device, that provides transitory or non-transitory instructions for implementing, performing, or coordinating the electronic operations for these techniques and system configurations. Another embodiment discussed herein includes a method that is operable on processing hardware of a computing device to implement, perform, or coordinate the electronic operations for these techniques and system configurations.

[0222] In other implementations, the logic, commands, or transitory or non-transitory instructions implementing aspects of the above-described electronic operations can be provided in a distributed or centralized computing system, including any number of form factors for computing systems such as desktop or notebook personal computers, mobile devices such as tablets, netbooks, and smartphones, client terminals, and server-hosted machine instances. Another implementation discussed herein includes incorporating the technology discussed herein into other forms, including into other forms of programming logic, hardware configurations, or specialized components or modules, including devices with individual pieces of equipment performing the functions of such technology. Individual algorithms for implementing the functions of such technology can include sequences of some or all of the above-described electronic operations or other aspects depicted in the drawings and detailed description above.

[0223] The above description is intended to be illustrative, and not restrictive. For example, the above-described examples (or one or more aspects of the examples) can be used in combination with each other. In addition, many modifications can be made to adapt a particular situation or material to the teachings of the disclosure without departing from the scope thereof. While the dimensions, types, and example parameters, functions, and implementations of the materials described herein are intended to define the parameters of the disclosure, they are in no way limiting, but rather exemplary implementations. Many other implementations will be apparent to those of ordinary skill in the art in view of the foregoing description. The scope of the disclosure should therefore be determined with reference to the appended claims along with the full scope of equivalents to which such claims are entitled.

[0224] Further, in the detailed description of implementations herein, various features can be combined together in order to organize this disclosure. This should not be interpreted as a requirement to claim such combinations in any claims that can issue from this disclosure. Rather, inventive subject matter can lie in fewer than all features of a particular disclosed implementation. Accordingly, the attached claims, which follow this detailed description, are hereby expressly incorporated into this detailed description, with each claim acting as a separate implementation independent of all other claims. The scope of the disclosure should be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.

Claims

1. A method for solving a radiotherapy treatment planning optimization problem, the method comprising: receiving, by processor circuitry, a complex first radiotherapy treatment planning optimization problem having a first set of parameters; converting, by the processor circuitry, the complex first radiotherapy treatment planning optimization problem to a second radiotherapy treatment planning optimization problem that is less complex than the first radiotherapy treatment planning optimization problem having a second set of parameters, wherein the first set of parameters and the second set of parameters comprise at least one of optimization variables, an objective function, or a set of constraints, wherein the converting comprises processing, by the processor circuitry, the first set of parameters to estimate the second set of parameters, wherein processing the first set of parameters comprises applying a machine learning model to the first set of parameters to estimate the second set of parameters, wherein the machine learning model is trained to establish a relationship between the first set of parameters and the second set of parameters for a plurality of training radiotherapy treatment planning optimization problems; generating, by the processor circuitry, a solution to the second radiotherapy treatment planning optimization problem based on the estimated second set of parameters; and generating, by the processor circuitry, a radiotherapy treatment plan based on the solution to the second radiotherapy treatment planning optimization problem.

2. The method of claim 1, wherein, a first optimization variable in the first set of parameters is excluded from the second set of parameters.

3. The method of claim 1, wherein, the second set of parameters is smaller than the first set of parameters.

4. The method of claim 1, wherein, the second radiotherapy treatment planning optimization problem includes fewer constraints than the first radiotherapy treatment planning optimization problem.

5. The method of claim 1, wherein, a matrix in the second radiotherapy treatment planning optimization problem is sparse or structured to correspond to a matrix in the first radiotherapy treatment planning optimization problem, and wherein generating the solution comprises processing the second radiotherapy treatment planning optimization problem using an optimization procedure.

6. The method of claim 1, wherein, the solution to the second radiotherapy treatment planning optimization problem is equivariant to permutations of a first subset and a second subset of the first set of parameters.

7. The method of claim 1, wherein, the solution to the second radiotherapy treatment planning optimization problem is invariant to permutations of a first subset and a second subset of the first set of parameters.

8. The method of claim 1, wherein, the second radiotherapy treatment planning optimization problem has a fixed size relative to a size of the first radiotherapy treatment planning optimization problem.

9. The method of claim 1, wherein, the solution to the second radiotherapy treatment planning optimization problem approximates a solution to the first radiotherapy treatment planning optimization problem.

10. The method of claim 1, wherein, the first radiotherapy treatment planning optimization problem comprises at least one of a linear programming problem or a quadratic programming problem, and wherein the first set of parameters and the second set of parameters are related by a Gaussian process.

11. The method of claim 10, further comprising: computing a second parameter of the second set of parameters based on a point estimate corresponding to a Gaussian distribution of a first parameter in the first set of parameters; and using the second parameter in place of the first parameter in the first set of parameters to generate the second radiotherapy treatment planning optimization problem, wherein the second radiotherapy treatment planning optimization problem comprises other linear programming or quadratic programming problems.

12. The method of claim 1, further comprising marginalizing the first parameter set on a Gaussian distribution to estimate the second parameter set.

13. The method of claim 1, further comprising mapping a solution to the second radiation therapy treatment planning optimization problem onto a space of the first parameter set of the first radiation therapy treatment planning optimization problem.

14. The method of claim 13, wherein, Mapping the solution onto the space of the first parameter set comprises applying a machine learning model to the solution to estimate a further solution to the first radiation therapy treatment planning optimization problem.

15. The method of claim 1, wherein, The first radiation therapy treatment planning optimization problem and the second radiation therapy treatment planning optimization problem are processed by an optimization process until a predetermined criterion is satisfied.

16. A method for solving a radiation therapy treatment planning optimization problem, the method comprising: receiving, by processor circuitry, a complex first radiation therapy treatment planning optimization problem having a first parameter set; converting the complex first radiation therapy treatment planning optimization problem to a second radiation therapy treatment planning optimization problem that is less complex than the first radiation therapy treatment planning optimization problem having a second parameter set, wherein the first parameter set and the second parameter set comprise at least one of optimization variables, an objective function, or a set of constraints, wherein the converting comprises processing, by the processor circuitry, the first parameter set to estimate the second parameter set, wherein processing the first parameter set comprises applying at least one of a non-linear function relationship or a statistical model to the first parameter set, wherein the statistical model comprises modeling the first parameter set as a random variable that depends on the second parameter set and deriving an objective function or a constraint of the second radiation therapy treatment planning optimization problem based on a measure of central tendency of the objective function or the set of constraints of the first radiation therapy treatment planning optimization problem; generating, by the processor circuitry, a solution to the second radiation therapy treatment planning optimization problem based on the estimated second parameter set; and generating, by the processor circuitry, a radiation therapy treatment plan based on the solution to the second radiation therapy treatment planning optimization problem.

17. A method for training a machine learning model to solve a radiation therapy treatment planning optimization problem, the method comprising: receiving, by processor circuitry, a plurality of complex first training radiation therapy treatment planning optimization problems, each of the plurality of first training radiation therapy treatment planning optimization problems comprising a first parameter set; and converting the plurality of complex first radiation therapy treatment planning optimization problems to a second radiation therapy treatment planning optimization problem that is less complex than the plurality of complex first radiation therapy treatment planning optimization problems having a second parameter set, wherein the first parameter set and the second parameter set comprise at least one of optimization variables, an objective function, or a set of constraints, wherein the converting comprises training, by the processor circuitry, the machine learning model to generate estimates of the second parameter set by establishing a relationship between the first parameter set of the plurality of first training radiation therapy treatment planning optimization problems and the second parameter set of the second radiation therapy treatment planning optimization problem.

18. The method of claim 17, wherein, The machine learning model is trained in a supervised manner by: obtaining a first training data pair comprising a given parameter set of the first parameter sets and a corresponding solution of a given problem of the plurality of first training radiotherapy treatment plan optimization problems; processing the given parameter set of the first parameter sets with the machine learning model to generate an estimate of the second parameter set of the second radiotherapy treatment plan optimization problem; solving the second radiotherapy treatment plan optimization problem based on the estimate of the second parameter set to generate a given solution of the second radiotherapy treatment plan optimization problem; computing a deviation between at least one of: (1) the given solution and the corresponding solution or (2) a decoded version of the given solution in a space of the given problem of the plurality of first training radiotherapy treatment plan optimization problems and the corresponding solution; and updating parameters of the machine learning model based on the computed deviation.

19. The method of claim 17, wherein, The machine learning model is trained in an unsupervised manner by: obtaining a given parameter set of the first parameter sets; processing the given parameter set of the first parameter sets with the machine learning model to generate an estimate of the second parameter set of the second radiotherapy treatment plan optimization problem; solving the second radiotherapy treatment plan optimization problem based on the estimate of the second parameter set to generate a given solution of the second radiotherapy treatment plan optimization problem; computing a metric of the given solution; and updating parameters of the machine learning model based on the computed metric.

20. The method of claim 19, wherein, The metric comprises a statistical metric representing a likelihood of observing training data or an utility of a treatment plan corresponding to the given solution.

21. A non-transitory computer-readable medium comprising non-transitory computer- readable instructions, the computer-readable instructions comprising instructions for performing operations of the method of any of claims 1-20.

22. A system comprising: a memory to store instructions; and one or more processors to execute the instructions stored in the memory to perform operations of the method of any of claims 1-20.

Images