Apparatus and method for facilitating optimization of patient-specific radiation treatment plan using at least one multi-blade collimator

Abstract

Devices and methods are provided that facilitate optimizing a radiation treatment plan for a particular patient using at least one multi-blade collimator. A control circuit configured for the graphics processing unit optimizes the blade movement of the at least one multi-blade collimator, and wherein the optimized blade movement is configured for highly parallel optimization opportunities. By means of one method, the second control circuit is at least partially used to configure the optimization of the blade movement in such a way as a highly parallel optimization opportunity. The second control circuit itself may be configured as a central processing unit (different from, for example, the graphics processing unit described above).

Description

Technical Field

[0001] These teachings generally relate to treating a planned target volume of a patient with energy according to an energy-based treatment plan, and more particularly to optimizing the energy-based treatment plan. Background Art

[0002] The use of energy to treat medical conditions comprises a known area of ​​prior art effort. For example, radiation therapy comprises an important component of many treatment plans for reducing or eliminating unwanted tumors. Unfortunately, the energy applied does not inherently distinguish between unwanted material and adjacent tissues, organs, etc., which are desirable or even critical for the continued survival of the patient. As a result, energy such as radiation is typically applied in a carefully administered manner to at least attempt to limit the energy to a given target volume. So-called radiation treatment plans typically function in the aforementioned manner.

[0003] A radiation treatment plan typically includes specified values ​​for each of the various treatment platform parameters during each of a plurality of sequential fields. A treatment plan for a radiation treatment session is typically generated automatically by a so-called optimization process. As used herein, "optimization" will be understood to refer to improving a candidate treatment plan without necessarily ensuring that the result of the optimization is actually a singular optimal solution. Such optimization typically includes automatically adjusting one or more physical treatment parameters (typically while complying with one or more corresponding constraints on these aspects), and mathematically calculating possible corresponding treatment outcomes (such as dose levels) to identify a given set of treatment parameters that represents a good compromise between a desired treatment outcome and avoiding undesirable side effects.

[0004] The foregoing may include optimizing the adjustment and use of one or more multi-leaf collimators. Unfortunately, optimizing the movement and use of the multi-leaf collimators typically comprises a significant portion of the overall optimization process and may contribute significantly to the total length of time required to create an optimized radiation treatment plan. For example, some optimization algorithms include recurring substeps in which the optimizer attempts to find a leaf sequence (i.e., a set of control points) that matches a given target annotation as closely as possible while complying with given leaf movement constraints. In the context of a serially computed optimization method, the latter leaf sequence optimization may consume a disproportionate amount of time compared to the overall optimization sequence. Summary of the invention

[0005] In general, these various embodiments are used to facilitate optimization of a radiation treatment plan for a particular patient using at least one multi-leaf collimator. In general, these teachings provide for optimizing the leaf movement of at least one multi-leaf collimator using a control circuit configured as a graphics processing unit, and wherein the optimized leaf movement is configured as a highly parallel optimization opportunity. By one approach, a second control circuit is at least partially used to configure the optimization of the leaf movement as a highly parallel optimization opportunity. The second control circuit itself can be configured as a central processing unit (different from, for example, the graphics processing unit described above). So configured, the second control circuit can assume the overall task of optimizing the radiation treatment plan, while being appropriately informed by the graphics processing unit of the optimization information for the leaf sequence.

[0006] By one approach, the graphics processing unit can be configured to optimize leaf movement according to at least one metaheuristic method. By one approach, the metaheuristic method can include a differential evolution method. If desired, the latter can include using differential evolution-specific parameters including an iteration count of about 100 iterations and a population size of about 50 agents.

[0007] By another approach, in lieu of or in combination with the foregoing approach, the graphics processing unit may be configured to optimize leaf movement according to at least one parallel explicit integration method. If desired, the at least one parallel explicit integration method may include: determining a force acting on each leaf of the at least one multi-leaf collimator based at least in part on the determined fluence under each of the leaves and based on constraints between the leaves in adjacent control points (i.e., a "pseudo" force that adapts the leaf to the fluence, rather than a force that physically moves the leaf); and then updating the velocity and position of the leaf based on the force. By one approach, when the force is so determined, the fluence may be assumed to be continuous.

[0008] When optimization is performed using both of the aforementioned methods, these teachings will be adapted, if desired, to optimize blade movement first according to at least one metaheuristic method and then subsequently according to at least one parallel explicit integration method.

[0009] So configured, blade optimization may be accomplished via a highly parallelized approach that is easily implemented on and exploits the processing power of a graphics processing unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above needs are met, at least in part, by providing an apparatus and method for facilitating optimization of a radiation treatment plan for a specific patient using at least one multi-leaf collimator, as described in the following detailed description, particularly when studied in conjunction with the accompanying drawings, in which:

[0011] Figure 1 including block diagrams as configured in accordance with various embodiments of these teachings;

[0012] Figure 2 including flow charts as configured according to various embodiments of these teachings;

[0013] Figure 3 including diagrams as configured according to various embodiments of these teachings;

[0014] Figure 4 includes screen shots as configured according to various embodiments of these teachings;

[0015] Figure 5 including pixel maps and charts as configured according to various embodiments of these teachings; and

[0016] Figure 6 Included are schematic representations of various embodiments as configured in accordance with these teachings.

[0017] The elements in the drawings are shown for simplicity and clarity, and do not have to be drawn to scale. For example, the size and / or relative positioning of some of the elements in the drawings may be exaggerated relative to other elements to help improve the understanding of the various embodiments of this teaching. In addition, common but well-known elements that are useful or necessary in commercially feasible embodiments are usually not depicted, so as to facilitate the less obstructed viewing of these various embodiments of this teaching. Certain actions and / or steps can be described or depicted in a specific order of occurrence, and those skilled in the art will understand that this specificity about the order is not actually required. The terms and expressions used herein have the common technical meanings of such terms and expressions as given by the technicians of the above-mentioned technical field, unless different specific meanings are set forth in addition herein. Unless otherwise specifically noted, the word "or" should be interpreted as having a disjunctive structure rather than a conjunctive structure when used herein. DETAILED DESCRIPTION

[0018] These and other benefits may become more apparent upon a thorough review and study of the following detailed description. Referring now to the drawings, and in particular to the Figure 1 , an illustrative device 100 that is compatible with many of these teachings will first be presented.

[0019] In this particular example, the enabling device 100 includes a first control circuit 101. As a "circuit," the first control circuit 101 thus includes a structure that includes at least one (and typically many) conductive paths (such as paths including conductive metals such as copper or silver) that transmit power in an orderly manner, which path(s) will typically also include corresponding electrical components (including both passive (such as resistors and capacitors) and active (such as any of a variety of semiconductor-based devices) as appropriate) to allow the circuit to implement the control aspects of these teachings.

[0020] The first control circuit 101 may include a fixed-purpose hardwired hardware platform (including but not limited to an application-specific integrated circuit (ASIC) (which is customized by design for a specific purpose rather than an integrated circuit intended for general use), a field programmable gate array (FPGA), etc.), or may include a partially or fully programmable hardware platform (including but not limited to a microcontroller, a microprocessor, etc.). These architectural options for such structures are well known and understood in the art and do not require further description here. The first control circuit 101 is configured to perform one or more of the steps, actions, and / or functions described herein (e.g., by using corresponding programming as will be familiar to those skilled in the art).

[0021] For purposes of illustration, the description will assume that the first control circuit is configured as a central control unit (CPU). The central processing unit comprises the main hardware components of a computer system and is responsible for executing instructions and performing calculations. The CPU interprets and executes sequences of instructions from one or more software programs to perform tasks and manipulate data. The CPU typically includes a plurality of components, including: a control unit (which manages the execution of instructions by fetching instructions from the computer's memory, decoding those instructions to understand their meaning, and then coordinating various other components of the CPU to execute the instructions), an arithmetic logic unit (ALU) (which is responsible for performing arithmetic calculations (such as addition, subtraction, multiplication, and division) and logical operations (such as AND, OR, NOT) as specified by the aforementioned instructions; the ALU is typically the core computing component of the CPU), registers (which are small, high-speed memory units located within the CPU that store data and intermediate results during processing and allow the CPU to quickly access data without having to retrieve data from a slower main memory (such as random access memory)), cache memory (which is a bit A small, high-speed memory located on or near the CPU and which stores frequently used data and instructions to reduce the time the CPU needs to access the contents compared to slower main memory; caches come in different levels (L1, L2, L3) of varying sizes and speeds), a clock (which provides a clock signal that synchronizes the activities of the CPU, the clock signal typically comprising a steady rhythm of pulses, each pulse representing a discrete unit of time (clock cycle); the speed of the clock (typically measured in Hertz) determines how many instructions the CPU can execute per second), and at least one pipeline in at least one modern CPU, within which instructions are broken down into smaller stages, and multiple instructions can be in different stages of processing at the same time to improve throughput and more efficiently utilize available resources.

[0022] The first control circuit 101 is operably coupled to a memory 102. The memory 102 may be integrated into the first control circuit 101, or may be physically separate (in whole or in part) from the first control circuit 101, as desired. The memory 102 may also be local to the control circuit 101 (where, for example, the two share a common circuit board, chassis, power supply, and / or housing), or may be partially or fully remote to the first control circuit 101 (where, for example, the memory 102 is physically located in another facility, city, or even country than the first control circuit 101).

[0023] In addition to information as described herein (such as optimization information for a particular patient and information about a particular radiation treatment platform), the memory 102 can also be used to, for example, non-transitory store computer instructions that, when executed by the first control circuit 101, cause the first control circuit 101 to behave as described herein. (As used herein, such references to "non-transitory" will be understood to refer to the non-transitory state of the stored content (and thus exclude when the stored content merely constitutes a signal or wave), rather than to the volatility of the storage medium itself, and thus include both non-volatile memory (such as read-only memory (ROM)) and volatile memory (such as dynamic random access memory (DRAM)).

[0024] By an optional approach, the first control circuit 101 is also operatively coupled to a user interface 103. The user interface 103 may include any of a variety of user input mechanisms (such as but not limited to a keyboard and keypad, a cursor control device, a touch-sensitive display, a voice recognition interface, a gesture recognition interface, etc.) and / or user output mechanisms (such as but not limited to a visual display, an audio transducer, a printer, etc.) to facilitate receiving information and / or instructions from a user and / or providing information to a user.

[0025] If necessary, the first control circuit 101 may also be operably coupled to a network interface (not shown). The first control circuit 101 configured in this manner may communicate with other elements (both other elements within the device 100 and other elements outside the device 100) via the network interface. Network interfaces, including both wireless and non-wireless platforms, are well known in the art and do not need to be specifically described in detail here.

[0026] By one approach, some or all of any desired patient-related imaging information may be obtained by a computed tomography device 106 and / or other imaging device 107 as is known in the art.

[0027] In this illustrative example, the first control circuit 101 is configured to ultimately output an optimized energy-based treatment plan (such as, for example, the optimized radiation treatment plan 113). The energy-based treatment plan typically includes specified values ​​for each of the various treatment platform parameters during each of the plurality of sequential exposure fields. In this case, the energy-based treatment plan is generated by an optimization process, examples of which are further provided herein.

[0028] By one approach, the first control circuit 101 may be operably coupled to an energy-based treatment platform 114 configured to deliver therapeutic energy 112 to a corresponding patient 104 according to an optimized energy-based treatment plan 113, the corresponding patient 104 having at least one treatment volume 105 and also having one or more organs at risk (at least one of the organs at risk). Figure 1 These teachings are generally applicable to use with any of a variety of energy-based treatment platforms / devices. In a typical application setting, the energy-based treatment platform 114 will include an energy source, such as a radiation source 115 of ionizing radiation 116.

[0029] By one approach, the radiation source 115 can be selectively moved via the gantry along an arcuate pathway (where the pathway at least to some extent encompasses the patient itself during the administration of the treatment). The arcuate pathway can include a complete or nearly complete circle as desired. By one approach, the control circuit 101 controls the movement of the radiation source 115 along the arcuate pathway and can accordingly control when the radiation source 115 starts moving, stops moving, speeds up, slows down, and / or the speed at which the radiation source 115 travels along the arcuate pathway.

[0030] As an illustrative example, radiation source 115 may include, for example, a radio frequency (RF) based linear particle accelerator (linac based) x-ray source. A linac is a particle accelerator that greatly increases the kinetic energy of charged subatomic particles or ions by subjecting the charged particles to a series of oscillating electric potentials along a linear beamline, which can be used to produce ionizing radiation (e.g., x-rays) 116 and high energy electrons.

[0031] As desired, a typical energy-based treatment platform 114 may also include: one or more support devices 110 (such as a couch) to support the patient 104 during the treatment session, one or more patient fixation devices 111, a gantry or other movable mechanism that allows selective movement of the radiation source 115, and one or more energy shaping devices that provide selective energy shaping and / or energy modulation (for example, a beam shaping device 117 such as a jaw, a multi-leaf collimator, etc.).

[0032] In a typical application setting, it is contemplated herein that the patient support device 110 can be selectively controlled by the first control circuit 101 to move in any direction (i.e., any X, Y, or Z direction) during an energy-based treatment session. Because the aforementioned elements and systems are well known in the art, further detailed description of these aspects is not provided herein unless otherwise relevant to the description.

[0033] In this illustrative example, the device 100 also includes a second control circuit 118, which is operably coupled to and interacts with the first control circuit 101 described above. In this example, the second control circuit 118 is configured as a graphics processing unit (GPU). A GPU is a dedicated hardware component that is generally designed for rendering and processing graphics and visual data. Unlike CPUs, which are more general and handle a wide range of tasks, GPUs are specifically optimized for parallel processing tasks that are generally well served in graphics-intensive applications such as games, three-dimensional rendering, video editing, and scientific simulations. A number of features and components of a typical GPU include:

[0034] Parallel architecture: GPUs contain a large number of smaller processing units called cores or stream processors. These cores work on different parts of a task simultaneously, enabling the GPU to perform many calculations at once.

[0035] CUDA / OpenCL: Many modern GPUs support programming languages ​​such as CUDA (Unified Compute Device Architecture) for GPUs manufactured by NVIDIA and OpenCL (Open Computing Language) for various GPU brands. These programming models allow developers to apply the parallel processing power of GPUs to non-graphics tasks, such as scientific simulations and machine learning.

[0036] Memory: GPUs typically have their own dedicated high-speed memory, often called video memory or VRAM. This high-speed memory can be used, for example, to store textures, frame buffers, and other graphics data required for graphics rendering. The fast access speeds of VRAM support fast rendering of high-resolution graphics without bottlenecking the overall performance of the system.

[0037] API (Application Programming Interface): GPUs are usually controlled using APIs such as DirectX (for Windows) and OpenGL (cross-platform). These APIs provide a standardized way for software, such as in the first control circuit 101 described above, to communicate with and utilize the capabilities of the GPU.

[0038] By one approach, this second control circuit 118 configured as a GPU may be built into the same chip as the first control circuit 101 used as a CPU, in which case the two control circuits may share at least some system memory if desired. However, dedicated GPUs (i.e., circuits with their own separate chips and their own on-board memory) tend to provide significantly better performance for many parallel processing intensive tasks.

[0039] In this illustrative example, the GPU is not used (at least not primarily) for graphics processing. Instead, the present teachings utilize the parallel architecture of the GPU to accelerate calculations for optimization of the multi-leaf collimator as described herein, which would be time-consuming on the first control circuit (which is configured as a CPU).

[0040] Reference now Figure 2 , a process 200 will be described, which may be performed, for example, in conjunction with the above-described application arrangement (and more particularly via the above-described second control circuit 118). In general, the process 200 is used to facilitate optimizing leaf movement of one or more multi-leaf collimators as part of optimizing a radiation treatment plan for a particular patient using a particular radiation treatment platform, utilizing therapeutic radiation.

[0041] At block 201, the process 200 optimizes the movement of leaves of at least one multi-leaf collimator that includes a portion of the beam shaping device 117 described above. A multi-leaf collimator includes a plurality of individual portions (referred to as "leaves") formed of a high atomic number material (such as tungsten) that can be independently moved into and out of the path of a radiation therapy beam in order to selectively block (and thereby shape) the beam. Typically, the leaves of a multi-leaf collimator are organized in pairs, with the paired leaves aligned co-linearly relative to each other, and the paired leaves can be selectively moved toward and away from each other. A typical multi-leaf collimator has many such leaf pairs, typically more than twenty, fifty, or even one hundred such pairs.

[0042] In the case where the second control circuit 118 is configured as a GPU, the GPU undertakes the optimization task as a highly parallel optimization opportunity.

[0043] By an optional method, the optimization is performed under the overall control of the above-mentioned first control circuit 101 (which is configured as a CPU in this example). So configured, when and according to the needs of the first control circuit 101, the second control circuit 118 undertakes the appropriate calculations for optimizing the blade movement, and the second control circuit 118 uses the information provided by the first control circuit 101 and provides its optimization results to the first control circuit 101.

[0044] These teachings will be applicable to using any of a variety of methods to so utilize the second control circuit 118 (configured as a GPU). As an example, the GPU can be configured to optimize blade movement according to at least one of the at least one metaheuristic method 202 and the at least one parallel explicit integration method 203. More specifically, the GPU can employ one or the other of these two methods, or both methods if desired. When using both methods, these teachings will be applicable to, for example, first optimizing according to the at least one metaheuristic method 202, and then subsequently optimizing according to the at least one parallel explicit integration method (in that order).

[0045] By one approach, when using metaheuristic method 202, metaheuristic method 202 may include, at least in part, a differential evolution method. If desired, the differential evolution method may in turn include using differential evolution-specific parameters including an iteration count of about 100 iterations and a population size of about 50 agents. (As used in the preceding sentence, the word "about" will be understood to mean a range of plus or minus a selected threshold, such as 5 percent, 10 percent, 15 percent, or 25 percent, as desired.)

[0046] By one approach, when the parallel explicit integration method 203 is used, the parallel explicit integration method 203 may include, at least in part, determining pseudo forces corresponding to each leaf of the one or more multi-leaf collimators based at least in part on the determined fluence at each of the leaves and based on constraints between the leaves in adjacent control points, and then updating the velocity and position of the leaves based on those forces. By one approach, when those forces are so determined, the fluence may be assumed to be continuous. (Of course, fluence represents the radiation flux integrated over time and comprises a fundamental metric in dosimetry (i.e., the measurement and calculation of absorbed dose of ionizing radiation in matter and tissue).

[0047] Further details consistent with these teachings will now be presented. It will be understood that the specific details of these examples are intended for illustrative purposes and are not intended to imply any particular limitations regarding these teachings.

[0048] Metaheuristics

[0049] Metaheuristic methods can work on a set of solution candidates simultaneously, and thus simplify the core computational operations. Such methods require little or no assumptions about the underlying optimization problem, and can search very large solution spaces without getting stuck in local optima. Although metaheuristic methods may be much more computationally demanding than dedicated algorithms, applicants have determined that significantly increased parallelism can offset this disadvantage. Several metaheuristic algorithms are known in the art. The following example contemplates the use of differential evolution, and in particular box-standard differential evolution.

[0050] The box-standard differential evolution algorithm first randomly generates a population of agents (i.e., solution candidates): During an iteration, each agent x∈P is mutated by combining it with three other agents a,b,c∈P using the element-wise mutation formula:

[0051]

[0052] where F∈[0,2], r i ~U(0,1), and CR∈[0,1]. Then use the given cost function Each mutated agent y is evaluated, and if the mutated agent y is found to have a lower associated cost, the original agent x is replaced with the mutated agent y:

[0053]

[0054] The basic differential evolution algorithm is easy to implement and parallelize because only the cost function needs to know how the agents are encoded. However, the applicant has determined that through a method, four application area specific components can be carefully selected and / or implemented to make the differential evolution algorithm work efficiently. These components include differential evolution specific parameters (particularly iteration count, population size, differential weight F and crossover probability CR), cost function, initial population generation and optional post-processing.

[0055] In this example, the above four differential evolution specific parameters are selected experimentally. For performance reasons, both the iteration count and the population size are kept relatively small (e.g., about 100 iterations and 50 agents). The differential weight F, which is part of the mutation formula, is selected to be a relatively modest 0.15, but the optimal value may depend on how the population is initialized. The crossover probability CR, which defines the element-by-element mutation probability, is selected to be 0.9 (so that 90% of the elements are mutated).

[0056] In this example, target annotation volume fitting, MU smoothing, and leaf movement constraints (e.g., between adjacent sectors, adjacent control points, and adjacent leaves) are integrated into the cost function as separate objectives:

[0057] Cost(x)=w1Obj fitting (x)+w2Obj MU (x)+w3Obj sector (x)+w4Obj CP (x)+w5Obj leaf (x).

[0058] So configured, new objectives can be added without modifying the optimizer itself. Both the blade position and each control point monitoring unit can be encoded into the agent (i.e., the optimizer optimizes both the blade position and each control point monitoring unit). The cost calculation is performed in a highly parallel manner because sectors (where each sector includes multiple control points in this example, and where each sector has its own population), agents, blades, and fluence pixels can be processed independently of each other (excluding fast reduction operations).

[0059] The differential evolution algorithm can be sensitive to the way the initial population is generated. It is possible to start with a completely randomly generated initial population, but convergence to a good quality global solution may require thousands of iterations and very carefully tuned objective weights in this case. On the other hand, an initial population consisting of simple, consistent leaf openings does not contain enough variation and may therefore lead to poor quality local solutions.

[0060] Applicants have determined that in order to achieve adequate performance and satisfactory results, an initial population can be generated by adapting individual leaves to a target annotation in a greedy manner. (In the context of differential evolution, the term "greedy" refers to an algorithm that follows a problem-solving heuristic that makes a local optimal choice at each stage. In many problems, a greedy strategy does not produce an optimal solution, but a greedy heuristic can produce a local optimal solution that approximates the global optimal solution in a reasonable amount of time.) Initial monitoring units are randomly selected (from a distribution that matches the distribution of values ​​of the target annotation), the control points are traversed in two opposite orders, and the adaptation begins in two opposite directions.

[0061] By a method, and with reference to Figure 3, half of the agents can be initialized using the stacking method and half can be initialized using the sliding window method. In this illustrative example, four examples 301 are presented, which depict how leaves can be adapted to target annotations during initial population generation. The rows show two opposite control point orders (via the depicted numbering) and the columns show two opposite adaptation directions (via the depicted arrows).

[0062] The resulting initial population contains two agents that are close enough to an acceptable solution for the algorithm to converge quickly, and enough variation for the algorithm to converge to a global solution of good quality. The output of previous prior art iterations can also be included in the population if desired. This means that the solution can be iteratively refined and the iteration count can also be kept relatively small. That is, the initialization can also be easily extended with new types of initial agents when and because new types of initial agents become available.

[0063] In addition to mutating the agent, these teachings will also be suitable for post-processing the mutated agent (and the initial agent if necessary) before cost calculation. This method can be used to further enforce the leaf movement constraints (e.g., during the last few iterations). The leaf positions can be represented as a sequence of position offsets (between adjacent control points), and this sequence can be gradually pruned (using a tightening threshold) so that offsets that violate any leaf movement constraints are eliminated. In the general case, the back-transformed leaf sequence will likely be very different from the original leaf sequence, but since the movement constraints have been (almost) satisfied and the solution has been gradually refined, such an approach is performed as a post-processing step.

[0064] Figure 4 An illustrative output 401 based on the foregoing teachings for 100 iterations from a cold start with blade movement restrictions expressed is presented. From the upper left, these illustrations present the synthetic target annotations in the control points numbered 1 through numbered 7, the output fluence (for the 7 control points), and the blade opening and monitoring units (which may be presented using different colors via the user interface 103 described above, where appropriate). (In this example, the blades are assumed to be moving vertically.)

[0065] The prior art is typically configured to work on a single solution candidate that is gradually improved by mutating individual leaves using complex mutation formulas, while the present teachings provide for working on a set of solution candidates that are mutated concurrently using formulas that can be simple mutation formulas. So configured, these teachings are well suited for GPU acceleration. In addition, the foregoing teachings can be used to find a global optimal solution. In particular, the above method can handle several peaks and valleys in the direction of movement of the blade.

[0066] It will be appreciated that the foregoing teachings tend to be much simpler than known technical solutions that rely on complex mutation formulas for imposing movement restrictions, and that the foregoing teachings can be configured to formulate movement restrictions as penalty terms as part of the cost function. Penalty terms are easier to understand, implement, and experiment with. New types of objectives, initial agents, and post-processing steps can be easily added to the algorithm.

[0067] Parallel Explicit Integration Method

[0068] As described above, these teachings will be suitable for formulating the leaf sequencing problem in a form that can be solved by parallel explicit integration methods. Parallel explicit integration methods are known in the art and find use, for example, in molecular dynamics. In view of the foregoing, the applicant has determined that the positions of the leaves of a multi-leaf collimator can generally be analogized to atoms, the target annotation quantities can be analogized to external force fields, and the leaf movement constraints can be analogized to the interactions between atoms. Therefore, the iterative solution steps of the leaf positions can correspond to time steps in molecular dynamics (without simulating physical time).

[0069] In this example, the method has the following main loop.

[0070] Through the iterative loop {

[0071] Loop through leaves and control points #1{

[0072] Calculate the forces acting on the blade

[0073] }

[0074] Loop through leaves and control points #2

[0075] Update blade position and injection

[0076] }

[0077] }

[0078] In the foregoing example, the two sub-loops (1 and 2) are parallelized over all optimization variables (the number of control points multiplied by the number of leaves in the multi-leaf collimator), making the scheme suitable for processing by a graphics processing unit. The forces acting on the leaves may include, for example, forces derived from the fluence at each leaf and from constraints between leaves in adjacent control points.

[0079] To derive force from fluence, the fluence can be assumed to be continuous (as opposed to typical pixel images of fluence).The relevant fluence variable in this problem is the difference fluence, which is the difference between the target annotation fluence and the current fluence produced by the current leaf aperture.

[0080] Figure 5A continuous fluence approximation is shown. The force applied to the blade from the differential fluence is fluence*mu, where mu is the monitoring unit weight of the control point in the direction of blade opening at positive differential fluence and blade closing at negative differential fluence. The differential fluence may be stored as a pixel map 501, but the fluence values ​​according to the blade position may be continuously approximated via interpolation of pixel values, as illustrated by graph 502.

[0081] The leaf position constraints may include constraints between adjacent leaves in a multi-leaf collimator and leaf movement constraints between adjacent control points. Figure 6 The latter (blade movement constraint) is shown in . In this example, if the movement between control points CP(i-1) and CP(i) exceeds the free range, the movement is penalized by a rigid force (wherein a blade movement exceeding the free range distance x is opposed by a linear force g with rigidity k). An efficient implementation of blade movement constraints can be very beneficial in producing a blade sequence with optimal and controllable treatment delivery time.

[0082] After the forces have been calculated for each blade, the blade positions can be updated from solution step t to t+1 using the Symplectic (semi-explicit) Euler scheme:

[0083] Speed ​​update: v t+1 =v t +hg

[0084] Position update: x t+1 =x t +hv t+1

[0085] where g is the total force on the blade, and h is a small step size that can be limited by the stiffness of the system, which depends on the fluence amplitude and stiffness of the blade movement and position penalties. Including a velocity v in the solver can be used to speed up convergence, and indicates that the blades have inertia that helps them roll through flat areas and small local minima in the landscape of fluence and blade constraint penalties. If desired, a damping force proportional to the velocity can be included in sub-loop 1 mentioned above.

[0086] At the end of sub-cycle 2, the differential fluence may be updated with the new blade position. The iterative loop may be run until, for example, the blades reach a new equilibrium. If desired, a sufficient fixed number of iterations may be used to ensure sufficient convergence, or a convergence criterion based on, for example, kinetic energy (e.g., represented by the sum of squared velocities) may be used.

[0087] In a direct aperture optimization application setting, the foregoing method may be iteratively called along with other functionality for updating target annotations, with successive passes starting from the leaf positions resulting from the previous call.

[0088] So configured, leaf sequencing can be readily implemented on a graphics processing unit, allowing, for example, the development of faster GPU-accelerated direct aperture optimizers. Since the force and update sub-steps are completely separable, complex interactions between leaves and between leaves and fluence are more easily implemented. It will also be appreciated that these teachings can be readily applied when optimizing leaves in an application setting employing a dual-layer multi-leaf collimator with only small changes that may be beneficial.

[0089] The leaf movement optimization results thus obtained can be used in the overall optimization of the radiation therapy treatment plan.The resulting optimized radiation treatment plan 113 can then be used in the normal process of administering radiation to a patient using the corresponding radiation treatment platform 114.

[0090] Those skilled in the art will appreciate that various modifications, changes and combinations may be made to the above-described embodiments without departing from the scope of the present invention, and such modifications, changes and combinations should be deemed to be within the scope of the inventive concept.

Claims

1. A method for facilitating optimization of a radiation treatment plan for a specific patient using at least one multi-leaf collimator, the method comprising: A first control circuit configured as a graphics processing unit is used to optimize leaf movement of the at least one multi-leaf collimator, wherein optimizing the leaf movement is configured as a highly parallel optimization opportunity.

2. The method according to claim 1, further comprising: Optimizing the blade movement is configured as highly parallel optimization opportunities by a second control circuit. The method according to claim 2 , wherein the second control circuit is configured as a central processing unit.

4. The method of claim 1 , wherein the graphics processing unit is configured to optimize the blade movement according to at least one of: At least one metaheuristic; and / or At least one parallel explicit integration method.

5. The method of claim 4, wherein the graphics processing unit is configured to optimize the blade movement according to both of the following: at least one metaheuristic; and At least one parallel explicit integration method. 6 . The method of claim 5 , wherein the graphics processing unit is configured to optimize the blade movement according to: first optimizing according to the at least one metaheuristic method, and then optimizing according to the at least one parallel explicit integration method.

7. The method of claim 4, wherein the at least one metaheuristic method comprises, at least in part: Differential evolution method.

8. The method of claim 7, wherein the differential evolution method comprises using differential evolution-specific parameters, the differential evolution-specific parameters comprising an iteration count of about 100 iterations and a population size of about 50 agents.

9. The method of claim 4, wherein the at least one parallel explicit integration method comprises, at least in part: determining a force acting on each leaf of the at least one multi-leaf collimator based at least in part on the determined fluence under each of the leaves and based on constraints between the leaves in adjacent control points; as well as The velocity and position of the blade are updated based on the force.

10. The method of claim 9, wherein when determining the force, the fluence is assumed to be continuous.

11. An apparatus for facilitating optimization of a radiation treatment plan for a specific patient using at least one multi-leaf collimator, the apparatus comprising: A first control circuit configured as a graphics processing unit, the first control circuit configured to optimize leaf movements of the at least one multi-leaf collimator, wherein the graphics processing unit is configured to optimize the leaf movements as a highly parallel optimization opportunity.

12. The apparatus according to claim 11, further comprising: A second control circuit is operably coupled to the first control circuit and is configured to optimize the blade movement as a highly parallel optimization opportunity.

13. The apparatus according to claim 12, wherein the second control circuit is configured as a central processing unit.

14. The apparatus of claim 11, wherein the graphics processing unit is configured to optimize the blade movement according to at least one of: At least one metaheuristic; and / or At least one parallel explicit integration method.

15. The apparatus of claim 14, wherein the graphics processing unit is configured to optimize the blade movement according to both of the following: at least one metaheuristic; and At least one parallel explicit integration method.

16. The apparatus of claim 15, wherein the graphics processing unit is configured to optimize the leaf movement according to: first optimizing according to the at least one metaheuristic method, and then optimizing the leaf movement according to the at least one parallel explicit integration method.

17. The apparatus of claim 14, wherein the at least one metaheuristic method comprises, at least in part: Differential evolution method.

18. The apparatus of claim 17, wherein the differential evolution method comprises using differential evolution-specific parameters, the differential evolution-specific parameters comprising an iteration count of about 100 iterations and a population size of about 50 agents.

19. The apparatus of claim 14, wherein the at least one parallel explicit integration method comprises, at least in part: determining a force acting on each leaf of the at least one multi-leaf collimator based at least in part on the determined fluence under each of the leaves and based on constraints between the leaves in adjacent control points; as well as The velocity and position of the blade are updated based on the force.

20. The apparatus of claim 19, wherein when determining the force, the fluence is assumed to be continuous.

Images