Administration of therapeutic radiation using deep learning models to generate leaf sequences
Generate leaf sequences of multi-leaf collimators through deep learning models and reinforcement learning techniques, optimize the radiation therapy plan, solve the problem of inaccurate energy application in the existing technology, improve the therapeutic effect and adaptability, and reduce complexity and cost .
Patent Information
- Application Number
- CN202380031640.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-03-30
- Filing Date
- 2023-03-16
- Publication Date
- 2024-11-22
AI Technical Summary
Existing radiation therapy programs are difficult to effectively distinguish unwanted tissue from target tissue, resulting in inaccurate energy application and may cause side effects on important tissues.
Deep learning model, especially the neural network model trained by reinforcement learning, generate leaf sequences of multi-leaf collimators to optimize the radiation therapy plan, and use multiple agents to independently generate leaf sequences based on the patient's flux map and treatment. The target generates an optimized leaf sequence.
It improves the accuracy and adaptability of leaf sorting, reduces the complexity of adaptation to different treatment sites, enhances the uniqueness and effectiveness of treatment plans, and reduces the cost of training and adaptation.
Smart Images

Figure CN119013081A_ABST
Abstract
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 disease comprises a known area of the prior art. For example, radiation therapy comprises an important component of many treatment plans for reducing or eliminating unwanted tumors. Unfortunately, the energy applied cannot inherently distinguish between unwanted material and adjacent tissues, organs, etc., which are desirable or even critical to the continued survival of the patient. Therefore, energy such as radiation is usually applied in a carefully administered manner to at least attempt to limit the energy to a given target volume. So-called radiation therapy plans often function in the above-mentioned aspects.
[0003] A radiation therapy plan typically includes specified values for each of the various treatment plateau parameters during each of a plurality of consecutive fields. Treatment plans for a radiation therapy session are often automatically generated through 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 optimized result is in fact the only optimal solution. Such optimization often includes automatically adjusting one or more physical treatment parameters (often while observing one or more corresponding constraints in these aspects) and mathematically calculating the possible corresponding treatment outcomes (such as drug administration levels) to identify a given set of treatment parameters that represents a good compromise between the desired therapy outcome and avoiding undesirable side effects.
[0004] Determining a reasonable order of control points for the utilized multi-leaf collimator comprises one of the more complex aspects of such optimization. Part of this complexity is due to the fact that the leaves of such collimators may be subject to various limitations and speed constraints that should preferably be taken into account. Additional complexity may result when a new multi-leaf collimator is used, completely new leaf ordering algorithms are required, different treatment facilities may require different tuning parameters, and / or typical leaf ordering methods focus on reproducing the currently required fluence map without considering other complexities. Summary of the invention
[0005] According to a first aspect of the present invention there is provided an apparatus for facilitating the administration of therapeutic radiation to a patient as defined in claim 1 .
[0006] According to a second aspect of the invention, there is provided a method for optimizing a radiation therapy plan as defined in claim 11.
[0007] Optional features are defined in the dependent claims.
[0008] A method for facilitating administration of therapeutic radiation to a patient is provided, the method comprising:
[0009] Accessing the memory in which the following is stored:
[0010] - Fluence map corresponding to the patient;
[0011] - a deep learning model trained to infer a leaf sequence of a multi-leaf collimator from a fluence map, wherein the deep learning model comprises a neural network model trained at least in part via a reinforcement learning method;
[0012] By a control circuit operatively coupled to the memory:
[0013] A radiation therapy plan is iteratively optimized to administer therapeutic radiation to a patient at least in part by employing multiple agents to each individually use the deep learning model to each generate a leaf sequence only for a single leaf pair of a multi-leaf collimator to generate a leaf sequence based on a deep learning model and a fluence map corresponding to the patient. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] The above needs are at least partially met by providing administration of therapeutic radiation using a deep learning model to generate a leaf sequence as described in the following detailed description, particularly when studied in conjunction with the accompanying drawings, in which:
[0015] Figure 1 Includes block diagrams of various embodiments configured according to these teachings; and
[0016] Figure 2 Included are flow charts of various embodiment configurations according to these teachings.
[0017] The elements in the figure are shown for simplicity and clarity, and are not necessarily drawn to scale. For example, the size and / or relative positioning of some of the elements in the figure may be exaggerated relative to other elements to help improve the understanding of the various embodiments of this teaching. In addition, the common but easy-to-understand elements that are useful or necessary in commercially feasible embodiments are usually not depicted, so that it is easier to understand 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 sequence is not actually required. Unless different specific meanings are otherwise set forth in this article, the terms and expressions used herein have the common technical meanings given to such terms and expressions by those skilled in the art of the above-mentioned technical field. Unless otherwise specifically indicated, the word "or" used in this article should be interpreted as having a disjunctive structure rather than a conjunction structure. DETAILED DESCRIPTION
[0018] In general, according to these various embodiments, a memory has stored therein a fluence map corresponding to a particular patient and a deep learning model. The deep learning model is trained to infer a leaf sequence of a multi-leaf collimator from the fluence map. The deep learning model includes a neural network model trained at least in part via a reinforcement learning method. The control circuitry accesses the memory and is configured to iteratively optimize a radiation therapy plan to administer therapeutic radiation to the patient at least in part by employing a plurality of agents to each individually use the deep learning model to each generate a leaf sequence for only a single leaf pair of the multi-leaf collimator to generate a leaf sequence according to the deep learning model and the fluence map.
[0019] By one approach, the neural network model comprises a convolutional neural network model. By one approach, the neural network model is trained at least in part via a supervised learning method. The neural network model may be trained using a training corpus that includes a fluence map for each of a plurality of corresponding fields / control points.
[0020] By one approach, the reinforcement learning method described above includes a deep learning method. By one approach, the reinforcement learning method provides a reward to the agent during training. Such a reward can be calculated, for example and at least in part, based on the degree to which the created leaf sequence produces a target fluence.
[0021] By one approach, the plurality of agents are each identical to one another. By another approach, the first agent and the second agent may be different from one another, such as when the multi-leaf collimator includes a first type of leaf and a second type of leaf, wherein the first and second types of leaves are different from one another. In this case, the first agent may generate a leaf sequence for a leaf pair including the first type of leaf, and the second agent may generate a leaf sequence for a leaf pair including the second type of leaf.
[0022] If desired, a radiation therapy platform including the above-described multi-leaf collimator can provide therapeutic radiation to a patient according to the above-described radiation therapy plan.
[0023] So configured, the above-described agents can be trained based on existing cases rather than by tuning one or more corresponding heuristic algorithms. Although such training may need to be repeated for each new multi-leaf collimator, the process is simple and therefore less burdensome than typical prior art methods. These teachings will also make it easy to adapt the trained agents to different treatment sites to better meet local needs and / or requirements. Under supervision, these teachings will also adapt the training based at least in part on the final plan quality, rather than just on the current fluence information. In many application settings, it is also expected that these teachings will provide improved leaf sorting results compared to typical prior art methods.
[0024] These and other advantages may become more apparent upon a thorough examination and study of the following detailed description. Figure 1 , an illustrative device 100 compatible with many of these teachings will first be presented.
[0025] In this particular example, enabling device 100 includes control circuit 101. As a "circuit," 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 carry electricity in an orderly manner, which paths also typically include corresponding electrical components (both passive (such as resistors and capacitors) and active (such as any of a variety of semiconductor-based devices)) to allow the circuit to implement the control aspects of these teachings.
[0026] Such control circuit 101 may include a fixed-purpose hardwired hardware platform (including but not limited to an application-specific integrated circuit (ASIC) (which is an integrated circuit customized by design for a specific purpose, rather than 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 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 that will be well understood by those skilled in the art).
[0027] The control circuit 101 is operably coupled to a memory 102. The memory 102 may be integrated into the control circuit 101, or may be physically separate (in whole or in part) from the 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 completely remote from the control circuit 101 (where, for example, the memory 102 is physically located in another facility, metropolitan area, or even country than the control circuit 101).
[0028] In addition to information such as optimization information for a specific patient and information about a specific radiation therapy platform as described herein, the memory 102 may also be used to, for example, non-transitory store computer instructions that, when executed by the control circuit 101, cause the control circuit 101 to operate as described herein. (As used herein, this reference to "non-transitory" will be understood to refer to the non-transitory state of the stored content (and thus except when the stored content merely constitutes a signal or wave), rather than the volatility of the storage medium itself, and thus includes both non-volatile memory (such as read-only memory (ROM)) and volatile memory (such as dynamic random access memory (DRAM)).
[0029] By an optional approach, the 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.
[0030] If desired, the control circuit 101 may also be operably coupled to a network interface (not shown). The control circuit 101 so configured may communicate with other elements (both internal and external to the device 100) via the network interface. Network interfaces, including both wireless and non-wireless platforms, are well understood in the art and need not be particularly detailed here.
[0031] By one approach, a computed tomography device 106 and / or other imaging device 107 known in the art may acquire some or all of any desired patient-related imaging information.
[0032] In this illustrative example, the control circuit 101 is configured to ultimately output an optimized energy-based treatment plan (such as, for example, an optimized radiation therapy plan 113). The energy-based treatment plan typically includes specified values for each of various treatment platform parameters during each exposure field in a plurality of consecutive exposure fields. In this case, the energy-based treatment plan is generated by an optimization process, examples of which are further provided herein.
[0033] By one approach, the control circuit 101 can be operably coupled to an energy-based treatment platform 114 that is configured to deliver therapeutic energy 112 to a corresponding patient 104 according to an optimized energy-based treatment plan 113. These teachings are generally applicable to 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.
[0034] By one approach, the radiation source 115 can be selectively moved along an arcuate path via a gantry (wherein the path includes the patient itself at least to some extent during treatment administration). Desirably, the arcuate path can include a complete or nearly complete circle. By one approach, the control circuit 101 controls the movement of the radiation source 115 along the arcuate path and can accordingly control when the radiation source 115 starts moving, stops moving, accelerates, decelerates, and / or the speed at which the radiation source 115 travels along the arcuate path.
[0035] As an illustrative example, radiation source 115 may include, for example, an x-ray source based on a radio frequency (RF) linear particle accelerator (linac-based). 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 generate ionizing radiation (e.g., x-rays) 116 and high-energy electrons.
[0036] 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 a 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 (e.g., beam shaping devices 117, such as clamps, multi-leaf collimators, etc.) to provide selective energy shaping and / or energy modulation as desired.
[0037] In a typical application setting, it is assumed herein that the patient support device 110 can be selectively controlled by the control circuit 101 to move in any direction (i.e., any X, Y, or Z direction) during an energy-based treatment session. Because the above-described elements and systems are well understood in the art, further details of these aspects are not provided herein unless otherwise relevant to the description.
[0038] Reference now Figure 2, a process 200 will be described that may be performed, for example, in conjunction with the above-described application arrangement (and more specifically, via the above-described control circuit 101). Generally speaking, the process 200 is used to facilitate generating an optimized radiation therapy plan 113, thereby facilitating treating a particular patient with therapeutic radiation according to the optimized radiation therapy plan using a particular radiation therapy platform.
[0039] At block 201, the process 200 provides access to the aforementioned memory 102 in order to access both at least one fluence map and a deep learning model corresponding to the patient 104. Those skilled in the art will appreciate that fluence represents the radiation fluence integrated over time and comprises a fundamental metric in dosimetry (i.e., measuring and calculating the absorbed dose of ionizing radiation in matter and tissue). The fluence map, in turn, comprises a map of fluence values for various parts of the patient's body.
[0040] The above-mentioned deep learning model is trained to infer the leaf sequence of a multi-leaf collimator from a fluence map. It is understood by those skilled in the art that machine learning includes branches of artificial intelligence. Machine learning typically employs learning algorithms such as Bayesian networks, decision trees, nearest neighbor methods, and the process can operate in a supervised or unsupervised manner as desired. Deep learning (sometimes also referred to as hierarchical learning, deep neural learning, or deep structured learning) is a subset of machine learning that employs a network capable of learning (usually unsupervised) from unstructured or unlabeled data. Deep learning architectures include deep neural networks, deep belief networks, recursive neural networks, and convolutional neural networks. Many machine learning algorithms build so-called "models" based on sample data (referred to as training data or training corpus) in order to make predictions or decisions without explicit programming.
[0041] For the sake of illustrative examples, it is assumed in this description that the above-mentioned deep learning model includes a neural network model trained at least in part via a reinforcement learning method. It will also be assumed, and again for the sake of illustrative examples, that the neural network model includes a convolutional neural network model, and that the neural network is trained at least in part via supervised learning.
[0042] By one approach, the neural network model may be trained using a training corpus that includes a fluence map for each of a plurality of corresponding fields / control points associated with a multi-leaf collimator and / or other characteristics of a given radiotherapy platform.
[0043] As described above, the neural network model is trained at least in part via a reinforcement learning method. In this illustrative example, the reinforcement learning method includes a deep learning method. As known to those skilled in the art, reinforcement learning includes the field of machine learning that uses intelligent agents and determines how these agents should act in a specific environment to maximize some reward. Therefore, this illustrative example provides for rewarding at least one agent during training. The latter can include calculating a reward based at least in part on the degree to which a given created leaf sequence reproduces a target fluence.
[0044] At block 202, the control circuit 101 iteratively optimizes the radiation therapy plan to administer therapeutic radiation 112 to the patient 104 at least in part by generating a leaf sequence of a multi-leaf collimator including the beam shaper 117 according to the deep learning model and the fluence map corresponding to the patient 104. In particular, the control circuit 101 employs a plurality of agents to each individually use the deep learning model to each generate a leaf sequence only for a single leaf pair of the multi-leaf collimator.
[0045] A multi-leaf collimator comprises a plurality of individual components (called "leaves") formed of a high atomic number material (such as tungsten) that can be independently moved in and out of the path of a radiotherapy beam in order to selectively block (and thus shape) the beam. Typically, the leaves of a multi-leaf collimator are arranged in pairs, aligned co-linearly with one another, and can be selectively moved toward and away from one another. A typical multi-leaf collimator has many such leaf pairs, often exceeding twenty, fifty, or even a hundred pairs.
[0046] By one approach, each of the plurality of agents is identical to one another. By another approach (e.g., when the multi-leaf collimator includes at least a first leaf and a second leaf, where the first leaf and the second leaf are different from one another (e.g., in width, thickness, material composition, etc.)), the plurality of agents may include some agents that are different from one another. For example, a first agent may generate a leaf sequence for a leaf pair including a first leaf, and a second agent may generate a leaf sequence for a leaf pair including a second, different leaf.
[0047] So configured, these teachings provide for the use of reinforcement learning in conjunction with machine learning models by using a reinforcement learning agent to observe and experiment with leaf orderings and assess relative success based on a reward mechanism that reflects how successful a given leaf sequence is in achieving a particular fluence outcome.
[0048] For illustration, some more detailed examples will now be given. It should be understood that the details of these examples are only for illustration and are not intended to be any specific limitation to these teachings.
[0049] In this example, the basic reinforcement learning infrastructure separates the task into an agent and an environment. The agent iteratively interacts with the environment by inferring appropriate actions based on observations. During training, the agent also receives a reward at each iteration. The reinforcement agent's policy of inferring actions from observations is modified by maximizing the cumulative reward. Once the agent is trained enough, the policy no longer changes and the agent no longer requires rewards. Assuming deep reinforcement learning is used, one approach is to infer actions by training a convolutional neural network.
[0050] In this example, this reinforcement learning-based leaf sorting employs a training environment in which a plan creation optimization algorithm can be automatically executed for a set of representative cases (e.g., various head and neck patient data). Once one or more agents are trained and validated, it can be used to guide the optimization process for new patient cases. (While the agents can use deep Q reinforcement learning, other reinforcement learning methods (such as policy gradient and actor-criticism) can also be employed as desired.)
[0051] By one approach, a leaf sorting agent converts sector fluences into corresponding leaf sequences, where the number of control points varies depending on the current multiresolution level from 16 to 2. If desired, each multiresolution level can have a separate training agent, and it is also possible that only certain multiresolution levels have agent-based leaf sorting.
[0052] By one approach, a single agent can be trained to infer the monitoring unit (MU) count of a single control point. These teachings will also be adaptable to use current MU count optimization algorithms.
[0053] These teachings will adapt to process each blade motion individually, allowing the same agent to make multiple observations of each blade, or to coordinate actions based on larger observations.
[0054] Since a given pair of leaves will typically affect adjacent fluence rows (via the tongue-and-groove effect), these teachings will accommodate allowing each single-leaf agent to interact with some or all of the adjacent leaves. In this case, a collaborative multi-agent approach can be implemented so that the agents controlling the adjacent leaves become part of the active agent's environment.
[0055] By one approach, reinforcement learning observations can only include the target fluence row associated with the current leaf pair. In a stacked (i.e., multi-layer) multi-leaf collimator design, each pair of leaves contributes two fluence rows. These teachings will also accommodate the addition of fluence maps from the previous or next sequence, or from a leaf sequence of adjacent control points.
[0056] The data in the observation can also be augmented by providing a full fluence row and / or adding the current MU weights for different control points to the observation.
[0057] By one approach, the observations are the target fluence row and the importance of each fluence pixel calculated by the optimizer using the second derivative of the cost function.
[0058] A simple reward calculation can be based on how well the created leaf sequence actually reproduces the target fluence. However, if desired, these teachings will adapt to optimize cumulative rewards rather than per-iteration rewards. So configured, the RL agent learns to benefit from future rewards as well, and part of the reward can be the final plan quality (e.g., optimized utility value).
[0059] By one approach, the reward may also derive a (weighted or unweighted) contribution from how well the leaf ordering satisfies the constraints and / or speed limits of the multi-leaf collimator.
[0060] Instead of the above, or in combination with it, the reward can also have a part where the agent is penalized for extending the optimization time. During the optimization process, the penalty may increase because the constraint violations are less severe while the optimization process is still ongoing and the solution may still change. One way to calculate the maximum allocation penalty size is to evaluate how much the fluence space changes when making the necessary changes to the leaf positions.
[0061] By one approach, the reward may be defined at least in part as a weighted mean squared error between a target fluence and a fluence generated by a proposed leaf sequence, and at least in part by penalizing leaf sequences that do not satisfy machine parameters and / or constraints.
[0062] If desired, and as indicated at optional block 203 , these teachings will accommodate use of a radiation therapy platform 114 including a multi-leaf collimator as described above to provide therapeutic radiation 112 to a patient 104 according to an optimized radiation therapy plan 113 .
[0063] 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 considered to be within the scope of the inventive concept.
Claims
1. An apparatus for facilitating the administration of therapeutic radiation to a patient, the apparatus comprising: A memory, wherein the memory stores the following items: - a fluence map corresponding to said patient; - a deep learning model trained to infer a leaf sequence of a multi-leaf collimator from a fluence map, wherein the deep learning model comprises a neural network model trained at least in part via a reinforcement learning method; control circuitry operably coupled to the memory and configured to iteratively optimize a radiation therapy plan to administer the therapeutic radiation to the patient at least in part by employing a plurality of agents to each individually use the deep learning model to each generate a leaf sequence only for a single leaf pair of the multi-leaf collimator to generate a leaf sequence based on the deep learning model and the fluence map corresponding to the patient.
2. The apparatus of claim 1, wherein the neural network model is trained at least in part via a supervised learning method.
3. An apparatus according to claim 1 or claim 2, wherein the neural network model is trained using a training corpus, the training corpus comprising a fluence map for each field / control point in a plurality of corresponding fields / control points.
4. An apparatus according to any preceding claim, wherein the neural network model comprises a convolutional neural network model.
5. An apparatus according to any preceding claim, wherein the reinforcement learning method comprises a deep learning method.
6. An apparatus according to any preceding claim, wherein the plurality of agents are each identical to one another.
7. An apparatus according to any preceding claim, wherein the multi-leaf collimator comprises a first type of leaf and a second type of leaf, wherein the first type of leaf and the second type of leaf are different from each other, and wherein the plurality of agents comprises a first agent and a second agent, the first agent generating a leaf sequence for a leaf pair comprising the first type of leaf, the second agent generating a leaf sequence for a leaf pair comprising the second type of leaf, wherein the first agent and the second agent are different from each other.
8. An apparatus according to any preceding claim, wherein the reinforcement learning method provides for rewarding the agent during training.
9. An apparatus according to any preceding claim, wherein the reinforcement learning method provides for calculating a reward based at least in part on the extent to which a created leaf sequence reproduces a target fluence.
10. The apparatus according to any preceding claim, further comprising: A radiation therapy platform includes the multi-leaf collimator and is configured to provide the therapeutic radiation to the patient according to the radiation therapy plan.
11. A method for optimizing a radiation therapy plan, the method comprising: Accessing the memory in which the following items are stored: - a fluence map corresponding to said patient; - a deep learning model trained to infer a leaf sequence of a multi-leaf collimator from a fluence map, wherein the deep learning model comprises a neural network model trained at least in part via a reinforcement learning method; By a control circuit operatively coupled to the memory: The radiation therapy plan is iteratively optimized to administer the therapeutic radiation to the patient at least in part by employing a plurality of agents to each individually use the deep learning model to each generate a leaf sequence only for a single leaf pair of the multi-leaf collimator to generate a leaf sequence based on the deep learning model and the fluence map corresponding to the patient.
12. The method of claim 11, wherein the neural network model is trained at least in part via a supervised learning method.
13. A method according to claim 11 or claim 12, wherein the neural network model is trained using a training corpus, the training corpus comprising a fluence map for each field / control point in a plurality of corresponding fields / control points.
14. A method according to any one of claims 11 to 13, wherein the neural network model comprises a convolutional neural network model.
15. The method according to any one of claims 11 to 14, wherein the reinforcement learning method comprises a deep learning method.
16. The method according to any one of claims 11 to 15, wherein the plurality of agents are each identical to one another.
17. A method according to any one of claims 11 to 16, wherein the multi-leaf collimator includes a first type of leaf and a second type of leaf, wherein the first type of leaf and the second type of leaf are different from each other, and wherein the multiple agents include a first agent and a second agent, the first agent generates a leaf sequence for a leaf pair including the first type of leaf, and the second agent generates a leaf sequence for a leaf pair including the second type of leaf, wherein the first agent and the second agent are different from each other.
18. A method according to any one of claims 11 to 17, wherein the reinforcement learning method provides for rewarding the agent during training.
19. A method according to any one of claims 11 to 18, wherein the reinforcement learning method provides for calculating a reward based at least in part on the extent to which a created leaf sequence reproduces a target fluence.
20. The method according to any one of claims 11 to 19, further comprising: By means of a radiation therapy platform including the multi-leaf collimator: The therapeutic radiation is provided to the patient according to the radiation therapy plan.
