Fast generation of layered multi-resolution matching multileaf collimator (MLC) openings

Abstract

An apparatus (30) for optimizing a radiation therapy plan for delivering therapeutic radiation to a patient using a therapeutic radiation source (16) that is concurrently modulated by a multi-leaf collimator (MLC) (14), the apparatus comprising at least one electronic processor (25) connected to a radiation therapy device (12). A non-transitory computer readable medium (26) stores instructions readable and executable by the at least one electronic processor to perform a radiation therapy plan optimization method (102) comprising optimizing MLC settings of the MLC with respect to an objective function, wherein the MLC settings define MLC leaf tip positions at a plurality of control points (CP) for a plurality of pairs of MLC leaves. The optimization is performed in two or more iterations, and a resolution of the MLC settings increases in successive iterations.

Description

TECHNICAL FIELD

[0001] The following relates generally to radiation therapy techniques, radiation therapy plan optimization techniques, conversion of radiation therapy plans defined by beam parameters to physically implementable parameters such as multi-leaf collimator settings, and related techniques. BACKGROUND

[0002] For external photon beam radiation therapy (RT), the creation of openings defined by the MLC is a central component in the treatment planning system (TPS). Different MLC opening configurations are applied to some variants for different treatment modalities and optimization types, e.g., for volumetric modulated arc therapy (VMAT); direct machine parameter optimization (DMPO); leaf sequencing; etc. The MLC settings are optimized for each individual patient to best deliver radiation to the target (typically a malignant tumor) while limiting unwanted radiation from being exposed to any neighboring critical organs (also referred to as organs at risk or OARs). In conventional RT, the radiation source is moved around the patient in a step-wise fashion and the MLC settings are optimized at each position so that the time-integrated radiation exposure meets various objectives designed by medical experts (e.g., a collaboration of a dosimetrist and an oncologist). In continuous arc RT methods such as VMAT, the radiation source is moved continuously over an arc of up to 360°. To perform continuous arc RT plan optimization, the continuous arc is discretized into a number of control points (CPs) which can in turn be similarly processed to the step positions of a conventional RT plan. For example, some radiation therapy planning systems produce plans in the Digital Imaging and Communications in Medicine (DICOM) format. The DICOM standard only allows the positions of the leaves of the MLC to be defined in a discrete manner in a sequence of beam positions and is also used by the planning system itself. It can be appreciated that in order to achieve sufficient resolution, the number of CPs can be quite large. A higher number of CPs includes a finer granularity of angular sampling rate which increases the number of parameters that can be adjusted to improve the RT plan characteristics. During dose calculation, finer granularity of angular sampling reduces the error from approximating continuous delivery to discrete gantry positions.

[0003] A typical MLC comprises multiple rows of leaf pairs (and optionally some larger blocking elements). The positions of the leaf tips can be controlled individually. The positions of the leaf tips are managed by some range constraints that are defined data-ly or dynamically (e.g., based on the characteristics of the MLC, the gantry mechanics, and the treatment modality). When performing MLC optimization, the positions of the leaf tips of each MLC at each CP are parameters, so that the total number of parameters to be optimized is large. In some two-step optimization methods, the beamlet shapes at each CP are defined by a series of optimized virtual beams such that the time-integrated radiation dose distribution delivered to the patient meets different clinical objectives (e.g., an objective regarding the minimum dose delivered to the target and an objective regarding the maximum dose delivered to a critical organ); in turn, the MLC settings are optimized to achieve the desired beamlets. In a one-pass method such as DMPO, the MLC settings are directly optimized to the respective clinical objectives. The RT plan with a large number of variables can be optimized to the clinical objectives separately. The RT plan with a large number of variables can be optimized using a column generation method that iterates between impact map optimization (FMO) and impact-based segment opening (re)generation.

[0004] For VMAT, the accuracy of the opening generation and the computational speed are essential in the treatment planning process, as a very large number of control points are to be considered. In VMAT, the motion of the MLC leaves is modeled along each treatment arc as a result of discrete control points. Each control point is associated with a particular MLC opening. The sequence of control points along each arc is at a pre-defined angular resolution, typically at 4° or 2° for a given setting. The number of openings to be optimized in a scenario involving multiple arcs is on the order of a few hundred.

[0005] Generally, the task of creating control points for the opening of the MLC is typically formulated as a problem of computing an approximate fit of the fluence to the ideal target given some boundary conditions. This approximate fit (e.g., in the least mean square error) can be implemented in several methods, which depend on the type of modeling used for the leaf tips and for the corresponding open field in between. A practical method based on the requirement of the system to define the leaf positions with a certain precision, including the modeling of the MLC leaves, is a discrete spatial grid, which defines the feasible leaf tip positions. For this grid model, the solution space of the optimal MLC fit problem is a discrete sequence of feasible MLC openings (e.g., depending on the grid size). The optimal MLC opening is found by pattern matching between the photon fluence formed by the leaf pairs opening assumed and the ideal target fluence. In a two-step optimization, the ideal target fluence is typically obtained by the FMO prior to the MLC opening optimization step, and the FMO can take into account the fluences of other control points in the current state of the treatment plan. In the DMPO method, the MLC opening is directly optimized to achieve the target fluence.

[0006] For the opening creation in a typical MLC system, an optimal fit of the multirow opening to a 2-D target fluence map is to be obtained; this involves computing the matching score at a single row for all feasible leaf pair openings according to the grid model. To obtain a final global fit criterion, the individual matching scores have to be appropriately combined in each MLC row. In addition, a more number, e.g., involving adjacent leaf rows (to control and adjust the opening profile composed of the rows of opened leaf pairs), can contribute to the overall fit criterion. According to this criterion, global optimization methods, e.g., based on shortest path graph search algorithms, are an efficient method to determine the best fit opening in the large set of feasible MLC openings.

[0007] In today's high-throughput semi-automatic RT treatment planning in medical therapy planning (still typically requiring some manual interaction of the treatment planning staff), a fast evaluation of the final plan is desirable. Therefore, the availability of very fast computational methods for treatment plan optimization (typically including multi-step MLC opening creation) is essential for the practical use of a TPS.

[0008] New and improved systems and methods are disclosed below to overcome these problems.

[0009] US 2017 / 087384 A1 describes a multi-layered multi-leaf collimation system. SUMMARY

[0010] In one disclosed aspect, an apparatus for optimizing a radiation therapy plan includes at least one electronic processor connected to a radiation therapy device for delivering a therapeutic radiation source to a patient using the therapeutic radiation source while being modulated by an MLC. A non-transitory computer readable medium stores readable instructions and performs an optimization method including optimizing MLC settings of the MLC of the radiation therapy device with respect to an objective function, where the MLC settings define MLC leaf tip positions for multiple rows of MLC leaf pairs at multiple CPs. The optimization is performed in two or more iterations with an increase in resolution of the MLC settings in successive iterations. The optimization also includes reducing a range of travel of the leaf tip positions and increasing resolution of the tip positions within the reduced range of travel in successive iterations.

[0011] In another aspect of the disclosure, a non-transitory computer readable medium stores instructions executable by at least one electronic processor to perform a radiation therapy plan optimization method. The method includes optimizing MLC settings of an MLC of a radiation therapy device with respect to an objective function, where the MLC settings define MLC leaf tip positions for multiple rows of MLC leaf pairs at multiple CPs. The optimization is performed in one or more iterations and resolution of the MLC settings in successive iterations. The optimization includes increasing resolution of the leaf tip positions in successive iterations and increasing a number of rows of MLC leaf pairs in successive iterations. The optimization also includes reducing a range of travel of the leaf tip positions and increasing resolution of the tip positions within the reduced range of travel in successive iterations.

[0012] In another aspect of the disclosure, a non-transitory computer readable medium stores instructions executable by at least one electronic processor to perform a radiation therapy plan optimization method. The method includes optimizing MLC settings of an MLC of a radiation therapy device with respect to an objective function, where the MLC settings define MLC leaf tip positions for multiple rows of MLC leaf pairs at multiple CPs. The optimization is performed in one or more iterations and resolution of the MLC settings in successive iterations. The optimization includes increasing resolution of the leaf tip positions in successive iterations and increasing a number of rows of MLC leaf pairs in successive iterations. The optimization also includes reducing a range of travel of the leaf tip positions and increasing resolution of the tip positions within the reduced range of travel in successive iterations.

[0013] An advantage is that a RT plan is generated with accurate dose delivery.

[0014] Another advantage is that a good MLC grid is utilized to optimize the positions of the MLC leaves.

[0015] Another advantage is that an optimized RT plan is quickly produced with optimized positions of the MLC leaves.

[0016] Another advantage is that precise positions of the MLC leaves are generated by utilizing a high resolution MLC grid (and angle grid for VMAT) while still maintaining fast computation speed in the multi-row opening creation process.

[0017] A given embodiment can provide zero, one, two, more or all of the above-mentioned advantages, and / or can provide other advantages as will readily occur to the skilled person when reading and understanding the disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0018] The disclosure can take form in various components and arrangements of components, and in various steps and arrangements of steps. The drawings are only for the purpose of illustrating preferred embodiments and should not be construed as limiting the disclosure.

[0019] Figure 1 A radiation delivery plan optimization system is shown in accordance with one aspect.

[0020] Figures 2 to 4 A different embodiment of a system that optimizes the position of leaves of an MLC is shown. Figure 1 DETAILED DESCRIPTION

[0021] The following relates to radiation therapy planning. This includes two steps. First, FMO is performed. This optimizes the fluence profile at each control point (CP), thus generating a fluence map for each control point. However, the fluence map is not physically realizable. Thus, in the second step, the fluence map at each CP is converted to a setting of MLC leaves. In various optimization processes - DMPO, the MLC setting is optimized directly against a clinical objective function, based on a dose objective based gradient for leaf sequencing. There are various possible combinations / sequences, e.g., a raw coarse MLC setting optimization based on clinical objectives following FMO, and back-and-forth iterations of both.

[0022] Each MLC includes multiple rows of pairs of leaves. In other words, each row of the MLC has two MLC leaves, and an opening is defined between facing ends of the two leaves (unless the ends are directed to touch to completely close the opening). By sliding the leaves along the row, this opening can be set larger or smaller, and can be positioned along the row or under a jaw of the MLC.

[0023] ​For planning purposes, the continuous arc employed in techniques such as VMAT is discretized into a number of CPs, e.g., in some examples with 2° to 4° increments. By simultaneously optimizing the MLC settings at all CPs, the imperfect match of the fluence map at the first CP (from the first step) can be compensated by the imperfect match of the expected fluence map at the second CP in a way that compensates for the imperfect fluence map match at the first CP. The MLC setting optimization is typically iterative and uses the overall fluence map as the objective function in a two-step approach. In a DMPO-like approach, the MLC setting optimization can use clinical objectives as the objective function.

[0024] Accurate dose delivery is a fundamental goal of high quality treatment planning. This makes accurate and high resolution placement of the MLC leaf tips mandatory. In cases where a discrete MLC grid model is used for leaf positioning, such requirements necessitate a high resolution grid spacing. However, a more fine-grained MLC grid during aperture creation results in a larger number of different hypothetical leaf pairs apertures involved in the matching process. This increase in the solution space can make pattern matching a time consuming step in the auto-optimization during treatment planning.

[0025] It is disclosed below to improve the efficiency of MLC optimization by using a sequence of resolutions from coarse to fine. In one aspect, the MLC leaf tips of each row are initially optimized over their full dynamic range using a coarse grid. In the next round, a finer grid is used to optimize over a smaller subset of that range; and so on until the finest physically achievable grid resolution is reached. The optimization sequence has found that in most cases this approach provides results similar to what would be achieved if the finest resolution was used from the start.

[0026] In other embodiments disclosed herein, the coarse-to-fine approach can be used in the "vertical" direction (that is, the direction transverse to the length of the rows) by coarsening the resolution in the vertical direction by combining several adjacent rows of MLC leaves.

[0027] In other embodiments disclosed herein, the coarse-to-fine approach can be used in the angular direction by coarsening the resolution along the arc by combining several CPs adjacent along the arc.

[0028] The coarse-to-fine sequence approach can be used independently in any one, two, or all three dimensions. Coarser positioning not only makes the leaves sequence faster (i.e., with fewer exploration possibilities), but potentially makes the dose calculation faster as well.

[0029] In embodiments further disclosed herein, the following selection is made: using MLC opening characteristics or qualities as targets in the MLC optimization. Typically, it is not desirable to have an optimized MLC setting that produces ragged edges to the opening due to the large difference in tip positions of the neighboring MLC leaf rows in the vertical direction. Similarly, it is not desirable to have a jagged edge to the opening. In principle, these opening characteristics can be (softly) enforced by adding suitable MLC opening targets in the optimization. Such targets can also be included in connection with the coarse-to-fine optimization sequence disclosed herein. However, the coarse-to-fine optimization sequence can naturally favor smoother opening edges, since, for example, when the optimization transitions from combining (e.g.) 12 rows together in the vertical direction to combining (e.g.) 4 rows together in the vertical direction, the rows in the original 12-row combination have the same tip position at the start of the 4-row resolution optimization.

[0030] Reference is made to Figure 1 An illustrative embodiment of a radiation delivery plan optimization system is shown. The radiation therapy can be any type of radiation therapy employing therapeutic radiation beams, e.g., electron beams, photon beams, high-energy X-ray beams, etc. The radiation treatment can employ a discrete "step-and-shoot" approach, in which a radiation beam source steps between fixed positions along a partial or full encircling trajectory of the patient. Alternatively, the radiation treatment can employ a continuous arc radiation therapy, such as VMAT, intensity modulated arc therapy (IMAT), step and short RT delivery, or other, in which the radiation beam source continuously irradiates the patient as the beam rotates around the patient along a partial or full encircling trajectory. For example, traversing of the trajectory includes moving the therapeutic radiation source along a continuous arc. The number of beams can be one, two, three, or more. In the case of continuous arc radiation therapy, the number of arcs performed in a treatment session can typically be one, two, three, four, or more. For planning purposes, the continuous arc is discretized into discrete control points, e.g., at 2° or 4° intervals (larger or smaller intervals are also contemplated depending on the desired resolution). In other examples, the trajectory can be circular, non-circular, or other shape, and can be traversed in a step-and-shoot approach or continuously. In the latter case, the continuous trajectory is typically discretized into control points along the trajectory in order for the radiation therapy plan to be more tractable.

[0031] In one example, the radiation delivery planning optimization system can be a linear accelerator (LINAC) 12 having a multi-leaf collimator (MLC) 14 configured to shape and deliver a high-energy electron beam that impinges a target (e.g., an x-ray or gamma-ray generator as a source of radiation 16 and associated hardware) that emits x-rays (i.e., photons) in response, causing a treatment beam to be delivered to a patient (not shown). The MLC 14 includes a beam aperture BA formed from a plurality of MLC leaves (details of which are shown in Figures 2-4 FIG. 3) that are adjustably disposed to shape the radiation beam during radiation treatment. The LINAC 12 includes a fixed base 13 that supports the LINAC, and a gantry portion 15 mounted to the base 13 having an arm 17 that rotates about a horizontal axis H (not shown) around a patient lying on a table or treatment couch 19. In some embodiments, the treatment couch 19 provides three translational degrees of freedom of motion and optionally three rotational degrees of freedom of motion, with the treatment couch position settings preferably being parameters of a parameterized radiation treatment plan to be executed by the radiation delivery planning optimization system 10. The treatment radiation beam is emitted from the LINAC 12 and passes through (and is shaped by) the MLC 14 in a direction toward the table 19 (e.g., downward as shown in Figure 1 FIG. 1). For a trajectory for delivering RT that includes a continuous arc, the arm 17 is rotated along the axis H over an angular range that constitutes the arc to deliver the treatment radiation to the patient, and as the arm 17 is rotated, the leaves of the MLC 14 are opened and closed in accordance with a predetermined radiation treatment plan to modulate the shape of the radiation beam at different angles of the arc (modeled in the plan as control points CP).

[0032] The radiation treatment delivery device 12 also includes (or is controlled by) a computing device 18 (e.g., a typical workstation computer, or more generally a computer, although other form factors such as a tablet, a smart phone, etc. are also contemplated). The workstation 18 includes a computer or other electronic data processing device having typical components, such as at least one electronic processor 20, at least one user output device 22 (e.g., a mouse, keyboard, trackball, and / or the like), and a display device 24. In some embodiments, the display device 24 can be separate from the computing device 18. The RT plan is typically run on a computer 25 having high computing capacity, such as a server computer accessed via a hospital network. While a single server 25 is shown, it will be appreciated that the desired computing capacity can be obtained by multiple cooperating server computers (e.g., a computing cluster, cloud computing resources, etc.), and it will be appreciated that the computer 25 includes such multiple computer embodiments.

[0033] One or more non-transitory storage media 26 are also provided to store data and instructions (e.g., software) that are readable and executable by the computer 25 to perform the RT planning as disclosed herein, and / or by the workstation or other controller 18 to control the RT delivery device 12 to deliver therapeutic radiation to a patient in accordance with a plan. The non-transitory storage media 26 can include, by way of non-limiting examples, one or more magnetic disks, RAID or other magnetic storage; solid state drives, flash drives, electronically erasable read-only memory (EEROM) or other electronic memory; optical disks or other optical storage; various combinations thereof; and the like. The storage media 26 can include multiple different media, optionally of different types, and can be variously distributed (e.g., a hard drive installed on the workstation 18, a RAID accessed via a hospital network, and / or the like). The optimized RT plan 30 is generated by the server 25, stored appropriately on the non-transitory storage media 26, and then downloaded by the workstation 18 and LINAC 12 controlled in accordance with the RT plan 30 to deliver therapeutic radiation to a patient. In another embodiment, the optimized RT plan 30 is transmitted in DICOM format to and from a tumor information resource library (not shown), the radiation delivery plan optimization system 10, a PACS resource library (not shown), and others. As shown herein, the non-transitory storage media 26 also stores instructions that can be read and executed by the at least one electronic processor 20, 25. These instructions include instructions that can be read and executed by the server 25 to optimize a treatment of radiation therapy for delivery to a patient by the LINAC 12. The optimized treatment plan 30 of radiation therapy can be saved to the one or more non-transitory storage media 26, sent (e.g., wirelessly) to the controller computer 18 of the LINAC 12 to control the LINAC 12 to deliver radiation therapy to a patient in accordance with the optimized treatment plan 30 of radiation therapy.

[0034] The non-transitory storage media 26 stores instructions that are readable and executable by the at least one electronic processor (e.g., the server 25) to perform the disclosed operations, including performing the radiation therapy plan optimization method or process 102 to generate the RT plan 30.

[0035] With continued reference to Figure 1An illustrative embodiment of the radiation therapy plan optimization method 102 is shown in the form of a flowchart. In operation 104, a coarse-grained resolution for the MLC is set. This can be a coarse-grained resolution along the MLC leaf pair direction, and / or a coarse-grained resolution in the perpendicular direction (i.e., the direction transverse to the MLC leaf pair direction). By "coarse-grained resolution", it is meant that the resolution is coarser than the physical resolution of the MLC 14. For the MLC leaf pair direction, the resolution is coarse-grained because the resolution for positioning the leaf tip is not as fine as the resolution actually achieved by the mechanical actuators that move the leaves. For example, if the MLC 14 can actually position its leaf tips with a precision of 1.0 mm, then the coarse-grained resolution can only allow the leaf tips to be positioned with a precision of 5.0 mm (i.e., the coarse-grained resolution). For the transverse MLC leaf pair direction (i.e., the "perpendicular" direction), the resolution is coarse-grained because the MLC 14 is modeled as having fewer leaf pairs than actually exist in the MLC 14. For example, if the MLC 14 has 40 leaf pairs, then it can be represented with a coarse-grained resolution that has only 10 leaf pairs. This requires that groups of four leaf pairs be modeled as a single leaf pair. In general, operation 104 can set the coarse-grained resolution only in the leaf pair direction, or only in the direction transverse to the leaf pair direction, or in both. (If the coarse-grained resolution is not used in one of these directions, then the full resolution achievable by the MLC 14 is employed in that direction.)

[0036] In operation 106, MLC optimization is performed using the coarse-grained resolution defined at 104. The MLC optimization 106 can employ any known MLC optimization algorithm, and the MLC optimization 106 can be performed with respect to a fluence map, where the fluence map is generated by a previous FMO, in which case the MLC settings that are coarse-grained are optimized with respect to the goals defined by the fluence map; or, for example, in DMPO, the MLC optimization 106 can be optimized with respect to clinical goals.

[0037] After MLC optimization 106, in operation 108, a finer granularity of MLC resolution is set than the initial coarse granularity resolution of operation 104. For example, if the MLC leaf tip has a physically achievable resolution of 1.0 mm, and the initial coarser granularity resolution at 104 was 5.0 mm, then in operation 108 this can be reduced to a resolution of 2.0 mm (which is still coarse compared to the 1.0 mm physical resolution, but is finer than the starting 5.0 mm resolution). In addition to using a finer granularity resolution, it can be considered to reduce the optimization range: for example, if the initial optimization of leaf tip positions was performed over the full dynamic range of MLC 14, operation 108 can now optionally set a range of half the full range, centered on the leaf tip positions (for each leaf tip) output by MLC optimization 106. Similarly, if MLC 14 has 40 leaf pairs, and the initial coarse granularity resolution was 10 leaf pairs, then in operation 108 this can be adjusted to 20 leaf pairs. As indicated by return arrow 110, the flow then returns to the next iteration of MLC optimization 106, where MLC optimization 106 is now performed with the finer granularity resolution set at operation 108. This cycle 106, 108, 110 can optionally be repeated one or more times, until in the last pass of operation 108 the resolution is set equal to the physically achievable resolution of MLC 14. In Figure 1 In another considered variation, not shown in FIG. 1, in addition to the successive resolutions of MLC settings as per each operation 104, 108, it is further considered to include successive reduction of angular resolution of CPs along the arc.

[0038] At 102, MLC settings of MLC 14 are optimized for a respective objective function, e.g., as described above. The MLC settings define tip positions of MLC leaves 18 for a plurality of MLC leaf pairs at a plurality of CPs along the arc. Optimization 106 is performed in two or more iterations, and for each operation 108 in successive iterations, the resolution of the MLC settings is increased. In one example, the objective function is defined by an optimized fluence map using FMO. The optimization includes optimizing the fluence map by optimizing beams at the CPs along the arc for clinical objectives of radiation therapy plan 30, e.g., clinical objectives including: minimum number of pairs of leaves 18 in an opening, minimum opening width, smoothness of the opening, and shape of the opening, etc. In another example, the objective function includes clinical objectives of radiation therapy plan 30 (e.g., DPMO optimization loop).

[0039] In one embodiment, the optimization operation 102 includes increasing the resolution of the leaf tip positions of the MLC 14 in successive iterations. For example, in a first iteration, the leaf tip positions of the leaves 18 are optimized on a coarse grid. In each successive iteration, the leaf tip positions are optimized on a finer grid than the previous iteration, until the mechanical resolution of the MLC leaf tips is reached. Optionally, the optimization can further include reducing the range over which the leaf tip positions are optimized in successive iterations (e.g., the range can be the full mechanical range of the leaf tips in the first iteration, then the range can be + / - one half of the range from the position of the first iteration in the second iteration, etc.).

[0040] Embodiments of the optimization operation 102 perform a pattern matching operation with progressively adjusted MLC grid spatial resolution, repeated in multiple iterations. The first iteration described above is computationally efficient because the pattern matching process is performed for a small set of hypothesized positions of the leaves 18 defined in a coarse MLC grid. In each subsequent iteration, the resulting openings between the leaves 18 are used to limit the search space for the pattern matching process in the next iteration (i.e., the range over which the optimization is performed is reduced). Typically, such a restricted next-level search space is constructed as an interval of feasible openings ("similar" to the best matching opening at the current level). As used herein, "similarity" can refer to the spatial distance of the corresponding leaf tip positions along the row dimension (e.g., lateral length). The pattern matching in the next iteration searches over a new set of hypothesized openings defined in a finer MLC grid, but with additional search space limitations imposed from the results of the previous iteration. The optimization operation 102 is performed over multiple iterations 106, 108, 110 until a sufficient final (physically achievable) resolution of the MLC grid has been reached. With this design, the number of hypothesized openings between the leaves 18 that will be evaluated in each iteration is kept under control, which limits the overall run time.

[0041] The non-transitory storage medium 26 stores instructions for performing an embodiment of the optimization operation 102. The instructions may include a set of hierarchies of the blades 18. The instructions also include instructions for performing matching to generate a "best fit" multi-row pairing of openings between the blades 18 by, for example, performing matching against a given target flux. The instructions may impose local range constraints (or other segment opening requirements) on individual blade tips. A shortest path search algorithm may be implemented to determine a globally optimized opening within the range constraints. The instructions may further include instructions for generating local position range constraints for the blade tip positions in each row of the MLC 14. The blade tip positions for a given reference opening between the blades 18 are used to construct a range constraint in each row that defines a (symmetrical) spatial neighborhood around each given tip position. Typically, the spatial width of each range can be selected by the user. However, a suitable choice is to set the left and right ranges of the intervals equal to the MLC grid spacing of the iteration currently being considered.

[0042] Figure 2 An example of a first embodiment of the optimization operation 102 is shown. In this illustrative example, it is assumed that hierarchical opening creation is performed in, for example, three different levels (ie, three channels 106, 108, 110 in a cycle). The MLC 14 has Figure 2 The physically achievable resolution in is represented by PRR (e.g., 1 mm as an example). The first level utilizes an MLC grid of a first resolution RES 1 (e.g., 5 mm grid point spacing) and optimizes the blade positions only to the first resolution. Figure 2 , the subsequent second and third levels utilize successively increasing MLC grid resolutions, e.g., RES 2 for the second level (e.g., 2 mm grid spacing), and RES 3 for the third level (e.g., 1 mm grid spacing). More generally, the spatial resolution of the final level can be determined by considering both: (i) the mechanical specifications of the MLC and (ii) parameters explicitly specified by the user or by the TPS.

[0043] Figure 2 A single blade 18 of a blade pair is shown in FIG. (i.e., Figure 2The opposite leaf pairs are not shown. In each iteration, a travel range is defined around the leaf tip position for each row of openings, and the optimization is performed over this travel range, (e.g., a travel range TR1 equal to the full leaf travel range for the first iteration, a reduced travel range TR2 for the second iteration, and a further reduced travel range TR3 for the final iteration). Stated another way, the travel ranges TR1, TR2, and TR3 for each iteration are constraints on the optimization. In each iteration after the first iteration, the travel range TR2, TR3 is centered on the last leaf tip position obtained in the previous iteration (e.g., the travel range TR2 is centered on the leaf tip position obtained in the first iteration), and extends a certain amount to either side (e.g., left and right) of the last leaf tip position. In the illustrative Figure 2 In the middle, the travel range TR2 extends to the left at the previous coarse resolution RES 1 and also to the right at RES 1. Similarly, the travel range TR 3 for the final iteration extends to the left at the previous coarse resolution RES 2 and also to the right at RES 2, until it reaches the resolution RES 3 (which is equal to the physically achievable resolution PRR).

[0044] Referring back to Figure 1 In another embodiment (optionally combined with the first embodiment), the optimization operation 102 includes increasing the number of rows of pairs of leaves 18 of the MLC 14 in successive iterations. That is, the resolution of the MLC setting in the "vertical" direction is increased. For example, in the first iteration, the number of rows of pairs of leaves 18 of the MLC 14 is reduced by combining adjacent rows of leaves, compared to the physical number of rows of MLC leaf pairs in the MLC, so that the MLC setting is optimized on a coarse grid in the direction transverse to the rows. In each subsequent iteration, the rows of leaves 18 are combined using a finer grid, until the physical number of rows of MLC leaf pairs in the MLC is reached. This approach also helps to find optimal values (e.g., local optimal values) that comply with the requirements, such as one or more segment openings having a minimum number of leaf pairs. Smoother larger segment openings have less difference between the calculated dose and the actually delivered dose.

[0045] Embodiments of the optimization operation 102 involve fast computation of a match score of a single leaf pair against a row of target fluence maps. To obtain the best MLC opening consisting of multiple rows of leaf pair positions, this match operation (e.g., in a graph search based optimization framework) is performed row by row. This process can be further accelerated by performing the row match operation at a hierarchical multi-resolution on the MLC rows in multiple levels: in other instances, the possible segment openings are evaluated intensively and clustered (e.g., k-means) and converted to segment openings.

[0046] Figure 3 An example of a second embodiment of the optimization operation 102 is shown. In a first level, only some pre-selected rows (e.g., Figure 3 0 and 4 shown in FIG ). For all other rows (e.g., rows 1, 2, and 3, depicted as white bars), the blade tip positions are set in an approximate, computationally less expensive (relative to the blade pair matching operation) manner. This approximation of the blade positions can be, for example, based on interpolation of the blade tips in those pre-selected rows. This interpolation also guides the generation of the opening towards a smooth opening profile shape that is conducive to accurate calculation of the patient radiation dose. The blade pairs of all (matched or interpolated) rows generated by this first level can be used in combination with the hierarchical multi-level approach for blade tip positioning mentioned above (e.g., as Figure 2 Return to reference Figure 3 In the second level, a different set of rows (e.g., rows 0, 2, and 4) is preselected that covers more rows than in the previous level. Thus, in the second level, blade pairs for more rows are accurately optimized based on the matching operation, and blade pairs for fewer rows (e.g., rows 1 and 3) are set by interpolation. This process iterates over multiple levels until the blade tip positions for all rows in the MLC are optimized by the matching operation (e.g., rows 0-4 as shown in the final level).

[0047] Continue to refer Figure 3 In another approach, adjacent rows can be grouped together to form a single virtual row of greater width in a "vertical" direction transverse to the rows. Figure 3 This is shown diagrammatically by the dashed box in the middle example of , where the two virtual blades 18v are indicated by dashed lines. This approach eliminates the computational cost of interpolation but may be less effective in smoothing the opening profile.

[0048] In some examples, the optimization operation 102 may include performing the optimization operation in successive iterations (e.g., Figure 2 ) and in successive iterations (e.g., as shown in Figure 3 The number of rows of MLC blade pairs is increased in FIG.

[0049] In another aspect, the method 102 of gradually increasing the resolution of the MLC settings can be combined with gradually increasing (or otherwise controlling the resolution of) the angular positions of the CPs along the arc. This embodiment is particularly applicable to VMAT, which is based on an arc composed of a plurality of sequential control points. The openings are generated for the CPs positioned at different angular resolutions along the sequential (i.e., angular) dimension. In conventional (i.e., non-multiresolution) VMAT, the entire set of CPs along the arc are simultaneously optimized (e.g., by applying a matching operation). Typically, this creation of all control points along the arc is iterated multiple times to achieve a gradual improvement in the fidelity of the fit to the target fluence.

[0050] In this embodiment, only a small number of CPs on the VMAT arc (i.e., those defined at a coarse angular resolution) are created by the matching operation. All other CPs are set by using a computationally less expensive method. This computationally less expensive method can be based on copying or interpolating from the openings created at the coarse angular resolution. Thus, the computationally expensive matching operation is initially performed only on the small number of CPs defined by the coarse angular resolution. In subsequent iterations, the resolution of the angular spacing is gradually increased so that more CPs are subjected to the computationally expensive matching operation. A larger set of CPs (which can or can not be interpolated) can be used for dose calculation during this process to avoid inaccuracies.

[0051] The non-transitory storage medium 26 stores instructions to perform an embodiment of the optimization operation 102. These instructions can include the angular resolution of the CPs along the VMAT arc. The first iteration has a coarse angular spacing, which gradually becomes finer until the final angular CP spacing. The instructions also include instructions to generate the best fit leaf pair opening at a particular control point on the arc by performing a match against a given target fluence. For VMAT, these instructions are able to impose dynamic sequence constraints on individual leaf tips. The instructions further include instructions to create a synthesized leaf pair opening from other openings in the sequence. This is performed in a computationally efficient manner. This can be achieved, for example, by i) copying the most recent (i.e., in the angular context) existing opening from the sequence; ii) suitably interpolating from some existing openings in the sequence, etc.

[0052] Figure 4 An example of the iterative increase in angular resolution is shown. In a typical VMAT optimization, a continuous arc is discretized to a resolution of control points along the arc with an angular spacing of 4°. For the illustrative multiresolution method, there are three levels of angular resolution: 16° for the first level; 8° for the second level; and 4° for the third level. Figure 4Various angles of the leaflets 18 along the CP are shown, labeled A-H at degrees 0°, 4°, 8°, 12°, 16°, 20°, 24°, and 32°, respectively. Based on this example, in a first iteration step, openings (e.g., 0°, 16°, and 32°) are generated by applying a matching operation (as shown in dark shading), while openings at, e.g., 4°, 8°, 12°, and 20° are synthesized from other openings (as shown in no shading) at low cost. In a second iteration step, openings at 0°, 8°, 16°, 24°, and 32° are generated by applying a matching operation, while openings at 4°, 12°, 20°, and 24° are synthesized from other openings at low cost. In a third iteration step, all openings are generated at local angular intervals of 0°, 4°, 8°, 12°, 16° (and so on) by applying a matching operation.

[0053] Referring back to Figure 1 When the radiation therapy plan 30 is optimized, it is stored in the non-transitory memory 26. On the day of the radiation therapy session, the radiation therapy plan 30 is retrieved from the memory 26 to the workstation 18 and used to control the LINAC 12 to deliver radiation therapy according to the plan 30, including configuring the MLC 14 at each CP to use the LINAC 12 to deliver therapeutic radiation to the patient throughout an arc, while modulating the LINAC 12 by the MLC in accordance with the optimized radiation therapy plan 30.

[0054] The present disclosure has been described with reference to the preferred embodiments. Modifications and alterations can occur to others upon reading and understanding the preceding detailed description. It is intended to include all modifications and alternations insofar as they come within the scope of the appended claims.

Claims

1. An apparatus for optimizing a radiation therapy plan (30) for delivering therapeutic radiation to a patient using a therapeutic radiation source modulated by a multi-leaf collimator (MLC) (14), the apparatus comprising: at least one electronic processor (25) connected to a radiation therapy device (12); and a non-transitory computer readable medium storing instructions readable and executable by the at least one electronic processor for performing a radiation therapy plan optimization method (102), the method comprising: optimizing MLC settings of the MLC with respect to an objective function, wherein the MLC settings define MLC leaf tip positions for pairs of MLC leaves at a plurality of control points (CPs), wherein the optimization is performed in two or more iterations, and a resolution of the MLC settings increases in successive iterations, wherein the optimization includes reducing a range of travel of the leaf tip positions, and increasing a resolution of the leaf tip positions within the reduced range of travel in the successive iterations.

2. The apparatus of claim 1, wherein the optimization includes: in a first iteration, optimizing the MLC leaf tip positions on a coarse grid; and in subsequent iterations, optimizing the MLC leaf tip positions on a finer grid than the last iteration until a mechanical resolution of the MLC leaf tips is reached.

3. The apparatus of claim 1, wherein the optimization includes increasing a number of rows of pairs of MLC leaves in successive iterations.

4. The apparatus of claim 3, wherein the optimization includes: in a first iteration, reducing a number of rows of pairs of MLC leaves compared to a physical number of rows of pairs of MLC leaves in the MLC by combining adjacent rows of pairs of MLC leaves, such that the MLC settings are optimized on a group grid in a direction transverse to the rows; and in subsequent iterations, combining adjacent rows of pairs of MLC leaves using a finer grid until a physical number of rows of pairs of MLC leaves in the MLC is reached.

5. The apparatus of claim 4, wherein the optimization includes: in a first iteration, reducing a number of rows of pairs of MLC leaves compared to a physical number of rows of pairs of MLC leaves in the MLC by selecting a subset of rows of pairs of MLC leaves and interpolating rows of pairs of MLC leaves that are not selected between selected rows of pairs of MLC leaves of the subset, such that the MLC settings are optimized on a coarse grid in a direction transverse to the rows; and in subsequent iterations, increasing a number of rows of pairs of MLC leaves in the selected subset until the physical number of rows of pairs of MLC leaves in the MLC is selected.

6. The apparatus of any of claims 1-5, wherein the optimization includes controlling a plurality of angular positions of the CPs along a trajectory, the trajectory including an arc.

7. The device of any one of claims 1 to 5, wherein the radiation therapy plan optimization method (102) further comprises: optimizing a fluence map by optimizing a beam at the CPs along a trajectory comprising an arc line with respect to clinical objectives of the radiation therapy plan (30); wherein the objective function is defined by the optimized fluence map.

8. The device of any one of claims 1 to 5, wherein the objective function comprises clinical objectives of the radiation therapy plan (30), the clinical objectives comprising a minimum number of leaf (18) pairs in an aperture, a minimum aperture area, a smoothness of an aperture, and / or other properties of an aperture shape.

9. The device of any one of claims 1 to 5, further comprising a radiation therapy device (12) configured to deliver therapeutic radiation to the patient using the therapeutic radiation source (16) of the radiation therapy device throughout a trajectory comprising an arc line and simultaneously modulated by the MLC (14) according to a radiation therapy plan (30) optimized by the radiation therapy plan optimization method (102); wherein the traversal of the trajectory comprises moving the therapeutic radiation source (16) along the arc line.

10. A non-transitory computer readable medium (26) storing instructions executable by at least one electronic processor (25) to perform a radiation therapy plan optimization method (102), the method comprising: optimizing MLC settings of a multi-leaf collimator, MLC, (14) of a radiation therapy device (12) with respect to an objective function, wherein the MLC settings define MLC leaf tip positions for pairs of MLC leaves in a plurality of rows at a plurality of control points, CPs, wherein the optimization is performed in two or more iterations, and a resolution of the MLC settings is increased in successive iterations; and wherein the optimization comprises reducing a range of travel of the leaf tip positions, and increasing a resolution of the leaf tip positions within the reduced range of travel in the successive iterations.

11. A radiation therapy plan optimization method (102), the method comprising: optimizing MLC settings of a multi-leaf collimator, MLC, (14) of a radiation therapy device (12) with respect to an objective function, wherein the MLC settings define MLC leaf tip positions for pairs of MLC leaves in a plurality of rows at a plurality of control points, CPs, wherein the optimization is performed in two or more iterations, and a resolution of the MLC settings is increased in successive iterations; and wherein the optimization comprises reducing a range of travel of the leaf tip positions, and increasing a resolution of the leaf tip positions within the reduced range of travel in the successive iterations.

Images