System and method for generating radiotherapy treatment plan

Abstract

The invention relates to a system and a method for generating a radiotherapy treatment plan. In some embodiments, a system may include one or more processors for receiving radiation therapy data associated with a set of treatments administered to a set of patients; generating a beam direction view (BEV) projection for each patient in the set of patients; and for each patient in the set of patients, providing treatment data associated with the patient and data associated with the BEV projection corresponding to the patient to the model to train the model to generate an output. The output may represent a fluence map. Systems and methods for generating a sequence of leaves are also provided.

Description

Technical Field

[0001] The present application generally relates to systems and methods for generating radiation therapy treatment plans. Background Art

[0002] Radiation therapy, or RT for short, is a common technique used to treat certain forms of cancer. Radiation therapy involves careful planning to deliver radiation to a target within the patient's body. During treatment, a multi-leaf collimator (MLC), which comprises a plurality of leaves, can be moved relative to the patient. Between each control point of energy delivery, the leaves of the MLC can be positioned and repositioned during the delivery of radiation by a linear accelerator (LINAC) to form the shape of the beam at each control point. The goal of carefully planned LINAC operation is to deliver radiation to the tumor while minimizing the delivery of radiation to surrounding healthy tissue. It will be appreciated that programming the LINAC to position and reposition so as to deliver radiation to the tumor is a complex process that relies heavily on the expertise of trained clinicians.

[0003] Recently, advanced software-based systems have been developed to assist in developing complex RT plans that minimize normal tissue damage and target margins. For example, knowledge-based planning, which leverages prior knowledge to guide subsequent treatment planning, has recently garnered significant attention in the medical imaging and therapy fields. However, some aspects of the radiotherapy treatment planning process cannot be easily addressed using existing knowledge-based planning techniques. For example, techniques suitable for analyzing two-dimensional images are often difficult to transfer to the three-dimensional domain. Consequently, traditional inverse optimization problems remain standard for implementing certain aspects of the treatment planning process. These techniques are costly in terms of both the time required and the computational resources consumed. Summary of the Invention

[0004] For the reasons stated above, there is a need for systems and methods that can quickly and accurately analyze a patient's scans and automate one or more aspects of the radiation therapy treatment planning process.

[0005] In an embodiment, a system may include one or more processors programmed to: receive radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients, wherein each treatment in the set of treatments corresponds to a patient in the group of patients; generate a beam direction view (BEV) projection for each patient in the group of patients based on the treatments in the set of treatments administered to each patient in the group of patients; and for each patient in the group of patients, provide treatment data associated with the patient and data associated with the BEV projection corresponding to the patient to a model to train the model to generate an output, the output representing a fluence map.

[0006] The one or more processors that can be programmed to generate a BEV projection for each patient in the group of patients can be programmed to: generate a set of digitally reconstructed radiographs (DRRs) based on the treatments in the set of treatments for each patient in the group of patients; and concatenate the set of DRRs for each patient to form a BEV projection for the patient.

[0007] One or more processors that can be programmed to generate the set of DRRs can be programmed to: generate the set of DRRs based on the treatments in the set of treatments for each patient in the set of treatments and the configuration of the multi-leaf collimators (MLCs) participating in the treatments in the set of treatments.

[0008] The one or more processors that can be programmed to provide treatment data associated with the patient and data associated with the BEV projection of each patient to the model can be programmed to: provide treatment data associated with the patient and data associated with the BEV projection of each patient to the model, the treatment data associated with the patient including a representation of the patient's planning target volume (PTV) or a representation of organs at risk (OAR).

[0009] The one or more processors that can be programmed to provide treatment data associated with the patient and data associated with a BEV projection of each patient to the model can be programmed to: provide treatment data associated with the patient and data associated with a BEV projection of each patient to the model, the treatment data associated with the patient including one or more leaf configurations of a multi-leaf collimator (MLC) during energy delivery to the patient by a linear accelerator (LINAC).

[0010] The one or more processors that can be programmed to provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to train the model to generate an output can be programmed to: compare the output with a target annotation volume map to determine a difference between the output and the target annotation volume map; and update one or more weights of the model based on the difference between the output and the target annotation volume map.

[0011] The one or more processors that can be programmed to provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to the model to train the model to generate an output can be programmed to: provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to the model to train the model to generate a first output, the first output being associated with a first scale; and provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to the model to train the model to generate a second output, the second output being associated with a second scale.

[0012] In another embodiment, a method may include: receiving, by at least one processor, radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients, wherein each treatment in the set of treatments corresponds to a patient in the group of patients; generating, by at least one processor, a beam direction view (BEV) projection for each patient in the group of patients based on the treatments in the set of treatments administered to each patient in the group of patients; and providing, by at least one processor, for each patient in the group of patients, the treatment data associated with the patient and data associated with the BEV projection corresponding to the patient to a model to train the model to generate an output, the output representing a fluence map.

[0013] In yet another embodiment, a non-transitory computer-readable medium stores instructions thereon that, when executed by one or more processors, cause the one or more processors to: receive radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients, wherein each treatment in the set of treatments corresponds to a patient in the group of patients; generate a beam direction view (BEV) projection for each patient in the group of patients based on the treatments in the set of treatments administered to each patient in the group of patients; and, for each patient in the group of patients, provide the treatment data associated with the patient and data associated with the BEV projection corresponding to the patient to a model to train the model to generate an output, the output representing a fluence map.

[0014] By implementing the techniques described in connection with the above-described systems and methods, the time required for treatment planning in the context of IMRT and VMAT therapies can be significantly reduced by replacing the expensive inverse planning optimization process with a direct, single-path approach using a model (e.g., a machine learning-based model as described herein). With particular focus on VMAT therapy, the fluence maps predicted using the above-described techniques take into account the mechanical constraints of the linear accelerator and reduce or eliminate the need for traditional iterative processes when generating fluence maps. Furthermore, in IMRT therapy, while traditional approaches focus on predicting fluence maps for each angle separately, the techniques described herein can simultaneously and jointly predict fluence maps for each angle. This helps improve the overall optimization of the treatment plan, as each angle specified by the treatment plan must be determined relative to other angles to avoid overshooting or undershooting certain parts of the body. Similarly, for VMAT therapy, jointly predicting fluence maps is even more important, as the resulting leaf sequence must be continuous at control points to avoid mechanical failures due to sudden movements of the LINAC, collisions with the leaves of the MLC, and the like.

[0015] In an embodiment, a system may include one or more processors configured to: receive data associated with a set of fluence maps associated with operation of a multi-leaf collimator to focus energy generated by a linear accelerator (LINAC); and provide data associated with each fluence map to a model so that the model generates an output, the output representing a leaf sequence of the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences.

[0016] One or more processors programmable to provide data associated with each fluence map to the model to cause the model to generate an output may be programmed to train the model to map the input fluence map to a subset of blade sequences in a set of possible blade sequences.

[0017] One or more processors that may be programmed to train a model to map an input fluence map to a subset of blade sequences in a set of possible blade sequences may be programmed to: map the input fluence map to a subset of blade sequences in a set of possible blade sequences, wherein the subset of blade sequences is associated with one or more operating parameters of the LINAC.

[0018] The one or more processors may be programmed to: compare the output to a target leaf sequence corresponding to each fluence map; determine a difference between the leaf sequence and the target leaf sequence; and update one or more weights of the model based on the difference between the leaf sequence and the target leaf sequence.

[0019] The one or more processors programmable to receive the set of fluence maps may be programmed to receive the set of fluence maps based on generation of a treatment plan for a patient undergoing intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment.

[0020] The one or more processors programmable to receive the set of fluence maps may be programmed to receive the set of fluence maps based on generation of a training set of sample fluence maps comprising one or more randomly generated fluence maps.

[0021] One or more processors that may be programmed to receive the set of fluence maps may be programmed to receive the set of fluence maps based on generation of a training set of sample fluence maps and generation of a treatment plan for a patient undergoing intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment.

[0022] The one or more processors that provide data associated with each fluence map to the model to cause the model to generate the output may be programmed to provide data associated with each fluence map to the generative transformer model to cause the generative transformer model to generate the output.

[0023] One or more processors that provide data associated with each fluence map to the model so that the model generates an output can be programmed to: provide data associated with each fluence map to the model so that the model generates an output, the output representing a leaf sequence of the fluence map and at least one monitoring unit (MU) value, the leaf sequence including one or more sets of multi-leaf collimator opening sequences corresponding to MU values ​​in the at least one MU value.

[0024] In another embodiment, a method may include receiving, by at least one processor, data associated with a set of fluence maps associated with operation of a multi-leaf collimator to focus energy generated by a linear accelerator (LINAC); and providing, by at least one processor, the data associated with each fluence map to a model so that the model generates an output, the output representing a leaf sequence of the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences.

[0025] In yet another embodiment, a non-transitory computer-readable medium stores instructions thereon that, when executed by one or more processors, cause the one or more processors to: receive data associated with a set of fluence maps associated with operation of a multi-leaf collimator to focus energy generated by a linear accelerator (LINAC); provide data associated with each fluence map to a model so that the model generates an output, the output representing a leaf sequence of the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences.

[0026] By implementing the techniques associated with the above-described systems and methods, leaf sequences can be calculated more quickly than using conventional techniques that require repeatedly calculating leaf motion, simulating dose delivery, and updating a fluence map used to recalculate leaf motion. By converging leaf motion more quickly, computational resources can be conserved, and corresponding leaf sequences can be generated more quickly than using conventional techniques.

[0027] In an embodiment, one or more processors may be programmed to: receive data associated with treatment properties associated with a patient; generate data associated with a fluence map for the patient based on a first model trained to receive the data associated with the treatment properties of the patient as input; generate data associated with a leaf sequence for the fluence map based on a second model trained to receive the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences; and transmit the data associated with the leaf sequence to cause a linear accelerator (LINAC) to operate according to the leaf sequence during energy delivery to the patient.

[0028] In another embodiment, a method may include receiving, by one or more processors, data associated with treatment properties associated with a patient; generating, by the one or more processors and based on a first model trained to receive as input the data associated with the treatment properties of the patient, data associated with a fluence map for the patient; generating, by the one or more processors, data associated with a leaf sequence for the fluence map based on a second model trained to receive the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences; and sending, by the one or more processors, the data associated with the leaf sequence to cause a linear accelerator (LINAC) to operate according to the leaf sequence during energy delivery to the patient.

[0029] In yet another embodiment, a non-transitory computer-readable medium stores instructions that, when executed by one or more processors, cause the one or more processors to: receive data associated with treatment properties associated with a patient; generate data associated with a fluence map for the patient based on a first model trained to receive the data associated with the treatment properties of the patient as input; generate data associated with a leaf sequence for the fluence map based on a second model trained to receive the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences; and transmit the data associated with the leaf sequence to cause a linear accelerator (LINAC) to operate according to the leaf sequence during energy delivery to the patient.

[0030] By implementing a model to determine fluence maps and leaf sequences, the need for repeated calculations (e.g., regarding inverse optimization or repeated dose calculations and fluence map updates) and the overall time required to generate a treatment plan for a patient (e.g., from data associated with the patient's geometry, planning target volume, and organs at risk to a set of control points that can be used to operate the LINAC) can be reduced. This, in turn, can lead to greater throughput of patients and higher utilization of the LINAC, which in turn can lead to positive health outcomes, as patients may be able to be treated more quickly than using traditional techniques. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Non-limiting embodiments of the present disclosure are described by way of example with reference to the accompanying drawings, which are schematic and not drawn to scale. Unless otherwise indicated to represent background art, the drawings represent aspects of the present disclosure.

[0032] Figure 1 A diagram illustrating a system for generating a radiation therapy treatment plan according to an embodiment is illustrated.

[0033] Figure 2 A flow chart illustrating a process for generating a fluence map according to an embodiment is illustrated.

[0034] Figure 3A An example of geometric properties of multiple beams represented in a BEV image according to an embodiment is illustrated.

[0035] Figure 3B A model for predicting fluence maps according to an embodiment is described.

[0036] Figure 3C A diagram illustrating a multi-scale approach for training a model according to an embodiment.

[0037] Figure 4 A flow chart illustrating a process for generating a blade sequence according to an embodiment is shown.

[0038] Figure 5A An example flow chart illustrating a model that may be trained to receive data associated with a fluence map as input and provide data associated with a blade sequence as output is illustrated in accordance with an embodiment.

[0039] Figure 5B Examples of target fluence and possible fluences that can be predicted using a model are described according to embodiments.

[0040] Figure 5C A model is described that receives fluence map data and multi-leaf collimator data as input and outputs a leaf sequence as output, according to an embodiment.

[0041] Figure 5DIllustrated is a visual depiction of input-output pair mappings for supervised training of one or more models, according to an embodiment.

[0042] Figure 5E Illustrated are a set of linking models and an example visual depiction of the progression of the linking models through a search space, according to an embodiment.

[0043] Figure 6 A flow chart illustrating a process for generating fluence maps and blade sequences according to an embodiment is illustrated.

[0044] Figure 7 An example flow chart illustrating a sequential implementation of two models according to an embodiment. DETAILED DESCRIPTION

[0045] Reference will now be made to the illustrative embodiments shown in the accompanying drawings and specific language will be used to describe the illustrative embodiments herein. However, it should be understood that this does not limit the scope of the claims or the present disclosure. For those skilled in the art who are familiar with the relevant art and possess the present disclosure, changes and further modifications to the inventive features described herein and additional applications of the principles of the subject matter described herein should be considered to be within the scope of the subject matter disclosed herein. Other embodiments may be used and / or other changes may be made without departing from the spirit or scope of the present disclosure. The illustrative embodiments described in the detailed description are not intended to limit the subject matter presented.

[0046] Figure 1 Components of a system 100 for planning and delivering radiation therapy according to an embodiment are illustrated. The system 100 may include an analysis server 114a, a system database 114b, a model 111 (e.g., a machine learning-based model as described herein), electronic data sources 120a-d (collectively, electronic data sources 120), end-user devices 140a-c (collectively, end-user devices 140), an administrator computing device 150, a medical device 160, and a medical device computer 162. Figure 1 The various components shown in the figure may belong to a radiation therapy clinic where a patient may, in certain circumstances, receive radiation therapy treatments via one or more radiation therapy machines (e.g., medical device 160) located within the clinic. System 100 is not limited to the components described herein and may include additional or other components (not shown for the sake of brevity) that should be considered within the scope of the embodiments described herein.

[0047] The above-mentioned components can be interconnected through network 130. Examples of network 130 may include, but are not limited to, private or public local area networks (LANs), wireless LAN (WLAN) networks, metropolitan area networks (MANs), wide area networks (WANs), and the Internet. Network 130 may include wired and / or wireless communications according to one or more standards and / or via one or more transport media. Communications on network 130 may be carried out according to various communication protocols, such as Transmission Control Protocol and Internet Protocol (TCP / IP), User Datagram Protocol (UDP), and IEEE communication protocols. In one example, network 130 may include wireless communications according to the Bluetooth specification set or another standard or proprietary wireless communication protocol. In another example, network 130 may also include communications through a cellular network, which, for example, includes GSM (Global System for Mobile Communications), CDMA (Code Division Multiple Access), and EDGE (Enhanced Data for Global Evolution) networks.

[0048] The analysis server 114a can generate and display an electronic platform configured to use the model 111 (including an artificial intelligence and / or machine learning model) to receive patient information and output the results of executing the model 111. The electronic platform can include a graphical user interface (GUI) displayed on one or more electronic data sources 120, end-user devices 140, medical devices 160, and / or administrator computing devices 150. Examples of the electronic platform generated and hosted by the analysis server 114a can be a web-based application or a website configured to be displayed on different electronic devices (e.g., mobile devices, tablets, personal computers, etc.).

[0049] The information displayed by the electronic platform may include, for example, input elements for receiving data associated with the patient being treated, synchronizing one or more sensors, and displaying predicted results generated by the model 111. For example, the analysis server 114a may execute the model 111 (e.g., a machine learning model trained to generate fluence maps, blade sequences, etc., as described herein for patients being treated by the medical device 160). The analysis server 114a may then display the results to the clinician and / or directly modify one or more operational properties of the medical device 160.

[0050] The analysis server 114a can be any computing device that includes a processor and non-transitory machine-readable memory capable of performing the various tasks and processes described herein. The analysis server 114a can employ various processors, such as a central processing unit (CPU) and a graphics processing unit (GPU). Non-limiting examples of such computing devices may include workstation computers, laptop computers, server computers, and the like. Although the system 100 includes a single analysis server 114a, the analysis server 114a may include any number of computing devices operating in a distributed computing environment, such as a cloud environment.

[0051] Electronic data sources 120 can represent various electronic data sources that contain, retrieve, and / or access data associated with medical devices 160, such as operational information associated with previously performed radiation therapy treatments (e.g., electronic log files or electronic profiles), data associated with previously monitored patients (e.g., CT scans, tumor locations, deformation information, etc.), or participants in a study used to train the AI ​​models discussed herein. For example, analysis server 114a can use clinic computer 120a, medical professional device 120b, server 120c (associated with a clinician and / or clinic), and database 120d (associated with a clinician and / or clinic) to retrieve / receive data associated with medical devices 160. Analysis server 114a can retrieve data from end-user devices 120, generate a training dataset, and train AI model 111. Analysis server 114a can execute various algorithms to convert the raw data received / retrieved from electronic data sources 120 into machine-readable objects that can be stored and processed by other analysis processes described herein.

[0052] The end-user device 140 can be any computing device that includes a processor and a non-transitory machine-readable storage medium capable of performing the various tasks and processes described herein. Non-limiting examples of the end-user device 140 can be a workstation computer, a laptop computer, a tablet computer, or a server computer. In operation, various users can use the end-user device 140 to access the GUI operated and managed by the analysis server 114a or the execution results of the model 111. Specifically, the end-user device 140 can include a clinic computer 140a, a clinic server 140b, and a medical professional device 140c. Although referred to as "end-user" devices in this article, these devices are not always operated by end users. For example, the clinic server 140b cannot be used directly by the end user. However, the results stored on the clinic server 140b can be used to populate various GUIs accessed by the end user through the medical professional device 140c. In some embodiments, the end-user device 140 can be associated with one or more clinicians associated with one or more treatment plans of the patient (e.g., involving the preparation of one or more treatment plans).

[0053] Administrator computing device 150 may represent a computing device operated by a system administrator. Administrator computing device 150 may be configured to display radiation therapy treatment attributes generated by analysis server 114a (e.g., various analytical metrics determined during training of one or more machine learning models and / or systems); monitor various models 111 used by analysis server 114a, electronic data sources 120, and / or end-user devices 140; review feedback; and / or facilitate training or retraining (calibration) of models 111 maintained by analysis server 114a.

[0054] In some embodiments, the medical device 160 can be a diagnostic imaging device or a treatment delivery device. For example, the medical device 160 can include one or more computed tomography (CT) scanners, a linear accelerator (LINAC) with a multi-leaf collimator (referred to herein as a collimator for ease of description), or other similar devices configured to deliver energy to a target tissue associated with a patient (referred to as a planning target volume) and, in some cases, measure the energy delivered to the target tissue. The multi-leaf collimator is composed of a plurality of small lead leaves that can be moved individually to form a radiation beam and deliver a dose to the tumor while minimizing the dose to surrounding healthy tissue. The medical device 160 can also include one or more sensors configured to monitor the patient being treated. That is, the medical device 160 and / or the analysis server 114a can communicate with various sensors that can monitor the patient's external biosignals. Non-limiting examples of sensors can include 3D surface mechanisms and optical (or other) sensors configured to monitor the patient's movement (e.g., how the patient moves and / or breathes).

[0055] The model 111 can be stored in a system database 114b. The model 111 can be trained using data received / retrieved from the electronic data source 120 and can be executed using data received from end-user devices, medical devices 160, and / or sensors 163. In some embodiments, the model 111 can reside in a clinic-local or clinic-specific data repository. In various embodiments, the model 111 can use one or more deep learning engines to develop a treatment plan for a patient receiving radiation therapy. For example, the analysis server 114a can receive patient attributes from the sensor 163 and execute the model 111 accordingly. The analysis server 114a can then display the results on one or more end-user devices 140. In some embodiments, the analysis server 114a can change one or more configurations of the medical device 160 based on the results predicted by the model 111.

[0056] refer to Figure 2 Flowchart illustrating a process 200 for generating a fluence map according to an embodiment. Process 200 includes operations 202-206. However, other embodiments may include additional or alternative operations or may omit one or more operations entirely. Process 200 is described as being performed by an analysis server that may communicate with Figure 1 However, one or more steps of process 200 may be implemented by the analysis server 114a described in Figure 1 For example, one or more computing devices may execute locally Figure 2 Some or all of the steps described in .

[0057] At operation 202, an analysis server receives radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients. For example, the analysis server may receive radiation therapy treatment data associated with a group of patients, wherein each patient was treated using a LINAC configured to deliver energy according to an intensity modulated radiation therapy (IMRT) plan or a volumetric modulated arc therapy (VMAT) plan. In some embodiments, the radiation therapy treatment data associated with the group of patients may include (e.g., represent) a set of LINAC configurations (e.g., LINAC gantry positions, multi-leaf collimator (MLC) configurations, and corresponding monitoring unit (MU) values). For example, the radiation therapy treatment data associated with the group of patients may include a set of LINAC configurations at multiple time points associated with the patient treatments. In this example, the radiation therapy treatment data may represent one or more leaf sequences as described herein. In some embodiments, the radiation therapy treatment data may represent one or more PTVs or OARs as described herein.

[0058] In some embodiments, the radiation therapy treatment data associated with the group of patients may include (e.g., represent) a group of computed tomography (CT) scans. In an example, the radiation therapy treatment data associated with the group of patients may include (e.g., represent) a group of magnetic resonance imaging (MRI) scans. In some examples, the radiation therapy treatment data associated with the group of patients may include (e.g., represent) a group of ultrasound scans.

[0059] At operation 204, the analysis server generates a beam direction view (BEV) projection for each patient in the group of patients. For example, the analysis server may generate at least one BEV projection for each patient based on radiation therapy treatment data associated with the group of treatments. In an example, the analysis server may generate the BEV projection based on a subset of the radiation therapy treatment data corresponding to each patient in the group of patients. In some embodiments, the analysis server generates multiple BEV projections for each patient in the group of patients. It will be appreciated that the BEV projection may correspond to treatment of each patient during a time period associated with the radiation therapy treatment assigned to the patient.

[0060] In some embodiments, the analysis server generates a BEV projection for each patient based on a set of digitally reconstructed radiographs (DRRs) generated by the analysis server. For example, the analysis server can generate a set of DRRs based on radiation therapy treatment data and one or more configurations of a LINAC that performs the radiation therapy. In this example, the analysis server can generate the set of DRRs, where one or more DRRs correspond to a patient in one or more patients whose treatment is represented by the radiation therapy treatment data. The one or more configurations of the LINAC can represent a position of the gantry of the LINAC relative to the patient and / or a configuration of blades of an MLC supported by the LINAC. The analysis server can then combine each DRR corresponding to a treatment session for a particular patient to form a BEV projection.

[0061] In one example, the analysis server can combine each DRR for a patient by connecting each DRR associated with a treatment (e.g., a treatment session). The connected DRRs can represent the radiation delivered to the patient from a particular perspective. For example, the connected DRRs can be used to determine the BEV of a plane extending through a portion of the planning target volume (PTV). In this example, the delivery of radiation associated with multiple control points can be represented in a single image. The resulting BEV image can represent the delivery of radiation to the PTV and one or more organs near the PTV. These organs near the PTV are referred to as organs at risk (OARs).

[0062] At operation 206, for each patient in the set of patients, the analysis server provides treatment data associated with the patient and data associated with the BEV projection corresponding to the patient to the model to train the model to generate an output representing a fluence map. For example, the analysis server may provide the model with treatment data associated with the patient and data associated with the BEV projection corresponding to the patient, wherein the treatment data represents one or more of: a representation of a PTV associated with the patient and / or a representation of one or more OARs of the patient. In an example, the PTV and the one or more OARs may be represented as being associated with (e.g., included in) a three-dimensional representation of a part of the patient's body. In an example, the analysis server may determine the representation of the body part based on one or more scans (e.g., CT scans, MRI scans, etc.) of the patient.

[0063] In some embodiments, the analysis server may provide treatment data associated with a patient and data associated with BEV projections corresponding to the patient to a model to train the model, wherein the treatment data associated with the patient represents one or more blade configurations of an MLC supported by a LINAC. For example, the one or more blade configurations may be blade configurations corresponding to each of a plurality of beams generated by the LINAC during treatment of the patient. In an example, the one or more blade configurations may be further associated with one or more MU values, the MU values ​​representing an amount of radiation delivered when the LINAC supports the MLC in a given blade configuration.

[0064] In some embodiments, an analysis server may provide treatment data associated with a patient and data associated with a BEV projection corresponding to the patient to a model, where the model includes one or more encoders and one or more decoders. For example, if the model is associated with a U-net architecture, the model may include one or more encoders associated with a contracting path and one or more corresponding decoders associated with an expanding path, wherein the one or more encoders and one or more decoders are associated with one or more levels. In this example, the analysis server may provide the treatment data associated with the patient and data associated with the BEV projection corresponding to the patient to the model by providing a single BEV image and corresponding treatment data (e.g., a corresponding LINAC configuration and MU value) to a first encoder at a first level of the model. The first encoder may be configured to receive the single BEV image and provide two outputs: a first output provided by the first encoder along the contracting path to a second encoder, and a second output provided to a first decoder corresponding to the first encoder, the first decoder being associated with the expanding path. Depending on the number of layers in the model, this process may be repeated for the corresponding encoders and decoders at each successive layer of the model. In some embodiments, at the last layer (or nth layer) of the model, the output of the encoder at layer n is provided as input to the decoder at layer n. The decoder then provides output to the decoder associated with the layer above the nth layer, which continues to each successive layer until it reaches the first decoder.

[0065] In some embodiments, the output of each encoder may be provided to one or more convolutional layers associated with the encoder before being provided to the next encoder along the contracting path of the model, or in the case of the nth encoder, to the nth decoder along the expanding path of the model. Additionally or alternatively, the output of each encoder may be provided to a max pooling layer (e.g., after passing through one or more convolutional layers) before being provided to the next encoder along the contracting path of the model. In some embodiments, the output of each decoder along the expanding path of the model may be provided to an upper convolutional layer. Additionally or alternatively, the output of each decoder may be provided to one or more convolutional layers (e.g., after passing through an upper convolutional layer). In some embodiments, the last decoder of the model may provide one or more fluence maps as output. For example, the last decoder of the model may provide one or more fluence maps as output, wherein the one or more fluence maps correspond to one or more BEV images provided as input to the model. Reference Figure 3B The diagram depicts an example model that implements the U-net architecture.

[0066] Although the present disclosure is discussed with respect to models that may be associated with a U-net architecture in examples, those skilled in the art will understand that other models (e.g., a V-net model, a convolutional neural network (CNN), etc.) may be involved in one or more operations of process 200.

[0067] In some embodiments, the analysis server may update one or more weights of the model based on the output of the model. For example, when the model receives as input treatment data associated with a patient and data associated with a BEV projection corresponding to the patient, the model may provide data associated with a fluence map as output. In this example, the analysis server may iteratively provide a single instance of the BEV projection and corresponding treatment data associated with a given patient so that the model outputs a corresponding fluence map. In this example, the analysis server may compare the fluence map with a target annotated fluence map to determine the difference between the fluence map output by the model and the target annotated fluence map. In some embodiments, the target annotated fluence map is associated with the treatment data associated with the patient (e.g., included in the treatment data or stored in association therewith).

[0068] In some embodiments, the analysis server may update one or more weights of the model based on the analysis server comparing a fluence map output by the model at each iteration with a target-annotated fluence map, where the target-annotated fluence map represents one or more dose height variation (DHV) curves. For example, the analysis server may compare a DHV curve represented by the fluence map output by the model with a DHV curve of the target-annotated fluence map. The analysis server may then determine the difference between the DHV curve output by the model and the DHV curve of the target-annotated fluence map and update one or more weights of the model as described above.

[0069] In some embodiments, the analysis server may update one or more weights associated with the model based on the difference between the fluence map output by the model and the target-annotated fluence map. For example, when the difference meets a threshold amount, the analysis server may update one or more weights associated with the model so that the same data continuously input to the model more closely approximates the target-annotated fluence map. In this example, the threshold amount may represent a change in one or more corresponding pixels, a change in one or more intensity values ​​of corresponding pixels that reaches or exceeds a predetermined intensity value, etc. The analysis server may repeat this aspect of operation 206 until the model converges.

[0070] In some embodiments, the analysis server may provide treatment data associated with the patient and data associated with the BEV projection corresponding to the patient in one or more stages. For example, the analysis server may provide treatment data associated with the patient and data associated with the BEV projection corresponding to the patient in one or more stages. In this example, the one or more stages may correspond to one or more scales at which the model is trained. Figure 3C A description of multi-stage training is discussed. For example, at a first scale, an analysis server may provide treatment data associated with a patient and data associated with a BEV projection corresponding to the patient, wherein the treatment data is limited to cases where the MLC delivers radiation at a predetermined first set of beam angles (e.g., angles between beams formed by the MLC supported by the LINAC, measured relative to an axis extending along the height of the patient and relative to the isocenter). Additionally or alternatively, at the first scale, the analysis server may provide treatment data associated with the patient and data associated with a BEV projection corresponding to the patient, wherein the treatment data is limited to cases where the MLC delivers radiation and the blades of the MLC are confined to a second predetermined range. In this example, at a second scale, the analysis server may provide treatment data associated with the patient and data associated with a BEV projection corresponding to the patient, wherein the treatment data is limited to cases where the MLC delivers radiation at a predetermined second set of beam angles. Here, the predetermined second set of beam angles may be greater than the predetermined first set of beam angles (e.g., the beam angles may be more closely spaced within the predetermined second set of beam angles as compared to the predetermined first set of beam angles). Additionally or alternatively, in the second scale, the analysis server may provide treatment data associated with the patient and data associated with BEV projections corresponding to the patient, wherein the treatment data is limited to the case where the MLC is delivering radiation and the blades of the MLC are limited to within a second predetermined range. In some embodiments, one or more beam angles associated with the first scale may be included as a subset of the beam angles in the second scale.

[0071] When training at different scales, given a number of training iterations i, (0 < i < N), the input BEV projections and output fluence maps, as well as the arc size (representing the size K of the control points described herein) are gradually refined over multiple stages from the first stage to the second stage, the third stage, etc. At each successive stage, the scales can be scaled down (e.g., reduced) so as to include more and more details when training at subsequent stages, thereby gradually refining the network features into high-frequency components. Figure 4 An example illustrating how to train a network in multiple training stages, starting with low-resolution input and output pairs (e.g., a first predetermined scale) and gradually introducing high-frequency components (e.g., a second predetermined scale, a third predetermined scale, etc.).

[0072] By implementing the disclosed techniques, the need for initial fluence map iterations based on inverse planning techniques prior to leaf sequencing can be eliminated. Furthermore, the need to repeatedly confirm that a given fluence map corresponds to a leaf sequence that results in an acceptable dose calculation can also be reduced or eliminated entirely. This is because the fluence maps predicted by the model described herein can be trained on fluence maps represented by earlier RT plans optimized during the treatment planning process. Furthermore, fluence maps generated based on the techniques described herein are more likely to meet the operational constraints of the LINAC and the supported MLC and, by extension, are more likely to be executed by the LINAC.

[0073] Furthermore, in the case of VMAT, certain techniques for optimizing VMAT treatment plans can rely on a multiresolution approach, where a given arc is first divided into large sectors containing, for example, 16 control points, and then the sectors are split in two at each subsequent multiresolution stage, starting with sectors with, for example, 16 control points, to 8, to 4, and finally to 2. The goal of this multiresolution approach is to avoid convergence to a local minimum, which can occur if there is a difference between the optimal fluence map and the fluence map that the LINAC can achieve given its mechanical constraints. Since the model described herein is trained on fluence maps that are the result of an optimized VMAT treatment plan, the target fluence map already takes into account the machine constraints. Therefore, only few, if any, additional steps need to be performed to confirm that the fluence map described herein is optimal.

[0074] refer to Figure 3A An example of geometric properties of multiple beams represented in a BEV image according to an embodiment is described. More specifically, Figure 3A Multiple beams generated by an MLC supported by a LINAC are illustrated, directed toward an isocenter associated with a patient's PTV. In this example, the individual DRRs are concatenated to form the illustrated BEV image. In some embodiments, an analysis server can predict a fluence map based on the BEV image generated according to process 200 described above. In some embodiments, predictions can be made for each control point (or at least a portion of the control points) associated with different gantry angles of the LINAC described herein.

[0075] As described, the projection represented by the BEV image is obtained based on one or more DRRs. As described above, a DRR may refer to an image generated based on one or more images from a radiation simulation projection (referred to as a "projection") through the patient's anatomy. The projection may then be described along one or more planes extending through the patient's anatomy. In some embodiments, the analysis server may create the DRR by digitally simulating the radiation attenuation properties of the patient's tissue based on medical images of the patient (e.g., computed tomography (CT) and / or magnetic resonance imaging (MRI) data) generated during the delivery of radiation to the patient (e.g., toward the PTV). In some embodiments, the analysis server may simulate different MUs for different beam angles represented by the BEV image.

[0076] In some embodiments, the analysis server can use the location 300 of the radiation source and the distance from location 300 to the MLC (referred to as the source-to-detector distance, or SDD) to predict how the radiation will interact with (e.g., reach) the patient's tissue. In this projection, the analysis server can also consider the leaf position 302 (e.g., representing the shape of the MLC opening when forming a given beam) and the fluence map 304, as well as the source axis distance (SAD distance). The analysis server can then predict a BEV image 306 specific to location 300. In some embodiments, the BEV image 306 is specific to the beam angle associated with the location of the radiation source relative to the patient (e.g., location 300). In some embodiments, the analysis server can iteratively predict additional BEV images by iteratively moving location 300 (e.g., based on defined increments or based on different control points). In some embodiments, the analysis server can connect different DRRs (associated with different beams formed by the MLC) as multi-channel inputs (e.g., 2D) to a model configured to receive data associated with the BEV and provide a fluence map as described herein as output.

[0077] In the example above, where DRRs are used, the analysis server can simulate the interaction between the patient's anatomy and the radiation delivered to the patient and the analysis server can represent the simulated interaction via radiographic images without exposing the patient to actual radiation. Furthermore, the analysis server can generate DRRs at various angles based on available treatment data, thereby eliminating the need to retrieve the same or similar data from previously administered treatments. Thus, the analysis server can generate (e.g., simulate) treatment data for training a model as described herein without being constrained to a dataset that includes only previously generated data associated with a given patient.

[0078] refer to Figure 3B and 3C A non-limiting example of a model 308 for predicting the fluence map described with respect to process 200 is illustrated. Figure 3B A model based on a U-net architecture is depicted, but the techniques described herein are not limited to such models, and those skilled in the art will readily appreciate that the techniques described herein can be applied to any suitable model (e.g., any suitable neural network), such as a conditional variational autoencoder (CVAE), etc.

[0079] As in Figure 3B As described in , the BEV image 306 is generated by the analysis server (e.g., based on projections acquired in a DRR used to generate the BEV image) and can be provided as an input to the model 308. In one example, the analysis server can provide the BEV image 306 as an input to the model 308, where the BEV image 306 is based on the simulated projections. The analysis server can then cause the model 308 to determine a correspondence between the fluence map 312 and the BEV image 306. In some embodiments, the analysis server can train the model 308 to recognize a mapping between the BEV image 306 and the fluence map 312 corresponding to the projections. For example, the model 308 can include various convolutional layers 310 that the analysis server trains to generate the BEV image 306 based on the MU weights and the position of the MLC corresponding to each BEV image 306 (e.g., with respect to the MU weights and the position of the MLC corresponding to each BEV image 306). Figure 3A The BEV image 306 is encoded and subsequently decoded (the blade pair position or MLC opening in question).

[0080] In some embodiments, in addition to providing the BEV image 306, the analysis server may also provide data associated with different organs (e.g., organ outlines), outlines including the PTV and OARs, etc., via additional input channels associated with the inputs to the model 308. For example, during training, the analysis server may provide data associated with different organs, outlines including the PTV and OARs, etc., via additional input channels associated with the inputs to the model 308. In an example, during inference, the analysis server may provide data associated with different organs via additional input channels associated with the inputs to the model 308.

[0081] In some embodiments, after the analysis server trains the model 308, the analysis server may configure the model 308 to output a fluence map 312, which may represent the predicted values ​​for all K beams. For example, during training, the analysis server may update the weights of model 306 based on various loss functions to improve the accuracy of model 306. In one embodiment, the loss function for training is defined as:

[0082]

[0083] where y i is the reference fluence map of the i-th beam, Dist(·,·) is y i and In some embodiments, Dist(·,·) may represent a distance function that measures the distance between a predicted value and a ground truth value, the distance being determined based on calculation of one or more of an L1 norm (sometimes referred to as an L1 distance), an L2 norm (sometimes referred to as an L2 distance), a mean square error (MSE), a structural similarity index measure (SSIM), and / or a combination thereof. Additionally or alternatively, the loss function may be represented as a loss function applicable to a generative adversarial network (GAN) including a Wasserstein GAN (WGAN) or a loss function calculated in a specific space, such as a dose space.

[0084] For example, the loss function can be expressed as

[0085]

[0086] where dosec(.) represents a differentiable function that calculates the dose distribution from the output fluence map 312, and target_dose=dosec(y) is the target dose distribution.

[0087] When executing the model 308 based on VMAT therapy, the K beams may correspond to K control points (e.g., discretization of an arc into K continuous angles). In these embodiments, in addition to the predicted fluence map 312, an additional branch (sometimes referred to as a detection head) may be added to the model 308 to output a leaf ordering prediction (e.g., a set of right leaves, left leaves, and MU weights for each of the K angles). This leaf ordering may be used as an initialization for the leaf ordering module.

[0088] In some embodiments, during training or validation, the analysis server can implement various evaluation protocols (e.g., to evaluate the accuracy of the predictions). For example, in addition to validating the target annotated dose map (using appropriate thresholds as described herein), the dose domain can also be validated (e.g., by utilizing DVH curves).

[0089] refer to Figure 3C A diagram illustrating a multi-scale approach for training a model as described herein. More specifically, the analysis server can implement a multi-scale approach during training to improve robustness during training. Figure 3C A schematic diagram of the technique is depicted.

[0090] In some embodiments, during training iterations, both the BEV images provided as input to the model and the fluence maps (and arc sizes in some embodiments) provided as output of the model can be progressively refined from downscaled starting versions to include more and more detail, thereby progressively refining the network features to high-frequency components. Figure 3CIt illustrates how the model 308 is trained in multiple training phases, starting with low-resolution input and output pairs at an initial learning state 314 and gradually introducing high-frequency components (e.g., a second learning phase 316) until the model 308 is fully trained (e.g., a final learning phase 318).

[0091] Figure 4 Flowchart illustrating a process 400 for generating a leaf sequence according to an embodiment. The process 400 includes operations 402-406. However, other embodiments may include additional or alternative operations or may omit one or more operations entirely. The process 400 is described as being performed by an analysis server that may communicate with Figure 1 However, one or more steps of process 400 may be implemented by the analysis server 114a described in Figure 1 For example, one or more computing devices may execute locally Figure 4 Some or all of the steps described in .

[0092] At operation 402, an analysis server receives data associated with a set of fluence maps for use in operation of an MLC to focus energy generated by a LINAC. For example, the analysis server may receive data associated with the set of fluence maps for operation of the MLC based on treatment of one or more patients. In one example, the analysis server may receive the set of fluence maps based on a treatment plan generated for the patient. In this example, the patient may be receiving intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment. In some embodiments, the data associated with the set of fluence maps may be further associated with MUs. For example, for each fluence map represented by the set of fluence maps, a corresponding MU representing radiation delivered at one or more points associated with the fluence map may be included in (e.g., associated with) the data associated with the set of fluence maps.

[0093] In some embodiments, the analysis server receives data associated with the set of fluence maps based on the generation of the sample fluence map training set. For example, the analysis server may receive data associated with the set of fluence maps based on a computing device (e.g., an analysis server or another computing device) that generates the sample fluence map training set. In this example, the sample fluence training set may include one or more randomly generated fluence maps. In an example, the sample fluence training set may include one or more fluence maps associated with the treatment of one or more patients (as described above). In some embodiments, the analysis server may receive data associated with the fluence maps based on the treatment of one or more patients, and the analysis server may generate the sample fluence map training set. For example, the analysis server may receive data associated with the fluence maps based on the treatment of one or more patients, and the analysis server may add one or more randomly generated fluence maps. In this way, the analysis server can augment the sample fluence training set by including fluence maps generated by clinicians and randomly generated fluence maps.

[0094] At operation 404, the analysis server provides data associated with each fluence map to the model so that the model generates an output, the output representing a leaf sequence of the fluence map, the leaf sequence comprising one or more sets of MLC opening sequences. For example, the analysis server may provide data associated with each fluence map to the model so that the model generates an output during model training. In some embodiments, the analysis server provides data associated with each fluence map to the model so that the model generates an output, wherein the model comprises a generative transformer model. The generative transformer model may then provide an output. In each of these examples, the output comprises a leaf sequence. For example, the output may comprise a set of leaf sequences, each leaf sequence in the set of leaf sequences representing movement of one or more leaves of an MLC based on a fluence map intended to deliver radiation to patient tissue based on the fluence map. In some embodiments, the input to the model (e.g., a given fluence map) is associated with the output of the model (e.g., a given leaf sequence).

[0095] In some embodiments, the output includes data associated with one or more MUs. For example, the output can be associated with one or more MUs corresponding to each leaf sequence. In this example, when the MLC is in a given configuration and positioned relative to the patient by the LINAC, the MUs can be associated with delivering radiation. In this manner, the model can output data configured to cause the LINAC to control the operation of the MLC according to a provided fluence map.

[0096] At operation 406, the analysis server updates one or more weights of the model based on the difference between the leaf sequence and the target leaf sequence. For example, the analysis server may update one or more weights of the model when training the model to map a fluence map provided as a model input to the leaf sequence. In some embodiments, the analysis server trains the model based on the analysis server comparing an output of the model (e.g., an output leaf sequence) to a target leaf sequence. In this example, the target leaf sequence may correspond to a given fluence map provided by the analysis server as a model input.

[0097] In some embodiments, the analysis server can update one or more weights associated with the model based on the difference between the fluence map output by the model and the target annotated fluence map. For example, as described above, the analysis server can compare the leaf sequence output by the model with the target leaf sequence. In this example, the analysis can determine the difference based on the comparison of the output leaf sequence and the target leaf sequence. In some embodiments, when the analysis server determines that the difference meets a threshold amount, the analysis server can update one or more weights associated with the model so that the same data is continuously input to the model to get closer to the target annotated fluence map. In this example, the threshold amount can represent a change in one or more corresponding pixels, a change in one or more intensity values ​​of corresponding pixels that reaches or exceeds a predetermined intensity value, etc. The analysis server can repeat this aspect of operation 406 until the model converges.

[0098] In some embodiments, the analysis server can train a model to map a fluence map provided as a model input to a blade sequence, where the blade sequence is included in a subset of blade sequences in a set of blade sequences. For example, the analysis server trains a model to map a fluence map provided as a model input to a subset of possible blade sequences (e.g., associated with blade motion sequences and positions that a LINAC can perform). In this example, the subset of possible blade sequences can be associated with a set of all possible blade sequences.

[0099] refer to Figure 5A An example flow chart illustrating a model 500 that can be trained to receive data associated with a fluence map as input and provide data associated with a blade sequence as output. In some embodiments, the model 500 can be used with respect to Figure 4 As illustrated, the model 500 (e.g., a machine learning-based and / or deep learning-based model as described herein) learns to map the fluence to parameters describing various aspects of the blades of the MLC during operation (including the MU level (θ), the blade velocity sequence s indicating the position of a given blade l at a control point i), and the parameters describing various aspects of the blades of the MLC during operation. l，i , control points p1, p2, ..., p nThose skilled in the art will readily appreciate that the described techniques may be applied to any suitable model (e.g., any suitable neural network) in addition to the models described herein.

[0100] In some embodiments, the model 500 can learn the dynamics of mapping from fluence to MLC without extracting data associated with the treatment plan. It will be appreciated that in some embodiments, the model 500 can be provided with existing or historical data associated with the treatment plan to enhance its performance. By eliminating the reliance on existing or historical data associated with the treatment plan, the analysis server can configure the model 500 to predict the MLC sequence for a "new" plan (e.g., a plan that is not based on a patient's previous treatment) without having to retrieve large amounts of historical data, which can be very time-consuming and resource-intensive. Furthermore, because the model 500 does not rely on historical and previous radiotherapy treatment data associated with the treatment plan, the model 500 can predict new radiotherapy treatment plans that can better optimize inference performance compared to a model trained only using radiotherapy treatment data associated with the treatment plan involved in an actual treatment.

[0101] In some embodiments, model 500 may be a controllable parameter model of a LINAC, capturing the dynamics of the MLC blades, dose rate settings, etc. Using model 500 (which is trained based on one or more of the techniques described herein), a large number of fluence maps may be predicted, thereby establishing a correspondence between dynamic parameters associated with a treatment plan and fluence.

[0102] The model 500 can extract an MLC sequence 500a and a corresponding fluence map 500b that have been simulated and generated by the analysis server (or another server). The MLC sequence 500a can be simulated using machine-specific constraints. In some embodiments, the model 500 can map the MLC sequence to its corresponding fluence map, so that the model 500 can be trained to determine how to generate the MLC sequence from the fluence map, for example, using various parameters (e.g., using dynamic parameters such as MU values, control points, and blade speed sequences). For example, the model 500 can map the MLC sequence to the corresponding fluence map, where the model 500 includes a generative transformer. In some embodiments, the model 500 can optionally extract existing or historical treatment plans (e.g., existing or historical fluence maps) to improve the training of the model 500.

[0103] refer to Figure 5B and 5C 502 and an example of a target fluence that can be predicted using the model 500. In some embodiments, the model 500 can learn from a "one-to-many" mapping scheme. Thus, the model 500 can predict different MLC data for the same fluence map, such as Figure 5BFor example, model 500 can predict fluence maps 504 and 506 to achieve target dosing 502.

[0104] In some embodiments, the analysis server may first define an input / output (I / O) framework, where, for example, for VMAT, the input consists of all the fluence maps of the arc and the output contains the dynamic parameters of the entire arc. and MLC The MLC data can be estimated from the data rather than from the fluence map alone. Figure 5C As shown in FIG, a network (e.g., model 500) can extract a fluence map and MLC data and outputs a single data point (e.g., x). This approach can be conceptualized as a trained projection operator that maps optimization variables (fluence map, MLC) to a feasible set of machine capabilities.

[0105] In some embodiments, x may represent a leaf sequence (e.g., a leaf sequence containing leaf positions and MU weights for each control point). In some embodiments, y may represent a fluence map. Thus, in is the set of constraints on the MLC dynamics. Constraints, as used herein, may be machine-specific and may be described mathematically. In operation, the constraints may correspond to maximum / minimum blade speed (e.g., 2-4 cm / s), maximum / minimum dynamic blade clearance, maximum / minimum blade span, maximum / minimum blade position, etc.

[0106] Because the input leaf sequence is provided in addition to the target labeling map, the operation of the model 500 can be simplified. More specifically, instead of searching within a wide search space and optimizing for an optimal solution throughout the search space, the model 500 can be guided toward a nearby solution in the optimization space. As a result, the model 500 can generate output using less time and / or computational resources than conventional methods would otherwise require.

[0107] As discussed herein, the model 500 can identify a variety of different equivalent solutions to a single set of problems. Thus, in some embodiments, the model 500, when trained using a supervised learning method based on a general search space, can produce network predictions that reduce (e.g., attempt to minimize) the average error with respect to all solutions. Since the set of equivalent optimal solutions may not generally be convex (e.g., in Figure 5B 508 depicted in FIG), so the model 500 cannot predict an optimal solution. However, the solution can be modified based on various machine MLC constraints.

[0108] refer to Figure 5DA visual depiction illustrating input-output pair mappings for supervised training of one or more models. In some embodiments, to train the model 500 in a supervised manner (e.g., to predict a blade sequence based on a fluence map and an initial blade sequence), the analysis server can create a dataset of input-output pairs as described herein. In an example, the input-output pairs can represent a set of fluence maps (e.g., all possible fluence maps that can be generated by one or more models described herein) and a set of blade sequences (e.g., all feasible blade sequences that can be generated by one or more models described herein). As discussed herein, feasible blade sequences include blade sequences that satisfy the operational constraints of the LINAC.

[0109] In some embodiments, the set of blade sequences can be mapped to a subset of the fluence maps included in the set of fluence maps. For example, the analysis server can map the set of blade sequences to the subset of fluence maps included in the set of fluence maps. In this example, the analysis server can use a forward model to map the set of blade sequences to the subset of fluence maps. In this manner, the analysis server can identify a set of achievable fluence maps and, if the analysis server receives a fluence map that does not satisfy the mapping to the set of feasible blade sequences, the analysis server can abandon generating the blade sequence using the fluence map.

[0110] In some embodiments, another set of blade sequences can be mapped to the fluence map subset. For example, the analysis server can map a set of blade sequences associated with an optimal blade sequence (e.g., a blade sequence that optimizes LINAC operation) to a subset of fluence maps (e.g., one or more fluence maps) in the set of achievable fluence maps. In this example, the analysis server can again map the set of blade sequences to the fluence map subset using a forward model.

[0111] refer to Figure 5E An example visual depiction illustrating a set of chained models and the progression of chained models through a search space. As a result of training model 500 using the methods and systems discussed herein, even if the initial leaf sequence predicted by model 500 is far from the optimal leaf sequence, model 500 can be "chained" to refine the leaf sequence in multiple steps, as shown in FIG5F (graphs 514 and 516). Chaining leaf sequences can improve or stabilize the solution predicted by model 500.

[0112] In some embodiments, the linking of leaf sequences can be used to find In some embodiments, each cascade can predict the minimum value of k Nearby The optimal solution, for example The minimum value of .

[0113] As shown, instead of training model 500 to predict the final result (X3), model 500 can be trained to iteratively modify the results so that model 500 achieves the final result. For example, model 500 can receive MLC data and a fluence map and predict an initial result (X1), then modify that result so that it is closer to the optimized result (e.g., a local maximum or minimum within the search space). Specifically, model 500 identifies (X2) and then (X3). This iterative modification of the results is also visually depicted in graph 516, where the darker portion indicates the convergence of the results within the search space.

[0114] refer to Figure 6 Flowchart illustrating a process 600 for generating a fluence map and a blade sequence according to an embodiment. The process 600 includes operations 602-608. However, other embodiments may include additional or alternative operations or may omit one or more operations entirely. The process 600 is described as being performed by an analysis server that may communicate with Figure 1 However, one or more steps of process 600 may be implemented by the analysis server 114a described in Figure 1 is executed in any number of computing devices running in the distributed computing system described in the preceding text. For example, one or more computing devices may execute locally Figure 6 Some or all of the steps described in .

[0115] At operation 602, an analysis server receives radiation therapy treatment data associated with a patient treatment. For example, the analysis server may receive radiation therapy treatment data associated with a patient treatment, where the patient's treatment involves IMRT or VMAT therapy. In an example, the radiation therapy treatment data associated with the patient treatment may include a representation of one or more of the following: an anatomical structure of the patient (e.g., a CT scan, an MRI scan, etc.), a PTV of the patient, one or more OARs of the patient, etc.

[0116] At operation 604, the analysis server executes the first model. For example, the analysis server may execute the first model, wherein the first model is related to Figure 2The analysis server may train the first model based on the same or similar model as described in process 200 of . In some embodiments, the analysis server may execute the first model to predict a fluence map for the patient. For example, the analysis server may provide the model with radiation therapy treatment data so that the model generates an output. In this example, the output may include one or more fluence maps corresponding to a BEV image generated by the analysis server as described herein. As described above, the analysis server may train the first model based on BEV images associated with a set of patients that have previously been treated. In some embodiments, the analysis server may generate a BEV image based on the analysis server generating DRRs based on the radiation therapy treatment data and the analysis server concatenating the set of DRRs to form the BEV image.

[0117] At operation 606, the analysis server executes the second model. For example, the analysis server may execute the second model, wherein the second model is related to Figure 4 In some embodiments, the analysis server may execute a second model to predict a leaflet sequence based on the patient's fluence map. For example, the analysis server may execute a first model to predict a patient's fluence map, and the analysis server may provide the fluence map output by the first model as input to the second model. In this example, the output may include a leaflet sequence corresponding to the fluence map generated by the first model as described herein.

[0118] At operation 608, the analysis server sends the blade sequence to the plan optimizer computer model. For example, the analysis server sends the blade sequence to the plan optimizer computer model to cause the plan optimizer computer model to output a treatment plan. In this example, the treatment plan may include data representing one or more MLC positions when supported by a LINAC and one or more corresponding MU values, which, in combination, cause the LINAC to deliver energy according to the fluence map provided as input to the second model. In some embodiments, the analysis server also sends data associated with the PTV, OAR, and / or one or more DVHs.

[0119] Figure 7 An example flow chart illustrating an implementation 700 of two models in sequence according to an embodiment. The implementation 700 includes operations 710-720. However, other embodiments may include additional or alternative operations or may omit one or more operations entirely. The implementation 700 is described as being performed by an analysis server that may communicate with Figure 1 However, one or more steps of implementation 700 may be implemented by Figure 1 The distributed computing system described in the present invention may be executed by any number of computing devices running in the distributed computing system.

[0120] At operation 710, the analysis server 702 receives radiation therapy treatment data. For example, the analysis server 702 may receive radiation therapy treatment data associated with a patient treatment involving IMRT or VMAT therapy. In some embodiments, the radiation therapy treatment data represents one or more of the following: a PTV of the patient, one or more OARs of the patient, a scan of an anatomical structure of the patient, and the like.

[0121] At operation 712, the analysis server 702 provides data to the first model 704. For example, the analysis server 702 may generate one or more BEV images based on radiation therapy treatment data. In this example, the analysis server 702 may provide data associated with the one or more BEV images to the first model 704 to cause the first model 704 to generate an output. In this example, the first model 704 is associated with the Figure 2 The analysis server 702 may generate one or more BEV images based on the radiation therapy treatment data.

[0122] At operation 714, the analysis server 702 causes the first model 704 to generate a fluence map. For example, the analysis server 702 may cause the first model 704 to generate a fluence map based on the analysis server 702 providing data (eg, data associated with one or more BEV images) to the first model.

[0123] At operation 716, the analysis server 702 provides the fluence map to the second model 706. For example, the analysis server 702 may provide data associated with one or more fluence maps generated by the first model 704 to the second model 706 to cause the second model 706 to generate an output. In an example, the second model 706 is associated with the second model 706. Figure 4 The model described in process 400 is the same or similar.

[0124] At operation 718, analysis server 702 causes second model 706 to generate a blade sequence. For example, analysis server 702 may cause second model 706 to generate a blade sequence based on analysis server 702 causing first model 704 to generate a fluence map and analysis server 702 providing the fluence map to the second model.

[0125] At operation 720, the analysis server 702 sends the blade sequence. For example, the analysis server 702 may send the blade sequence to cause the LINAC to operate according to the blade sequence. For example, the analysis server 702 may send data associated with the blade sequence, the data being configured to cause the LINAC to operate according to the blade sequence.

[0126] The various illustrative logic blocks, modules, circuits, and algorithmic steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether this functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present disclosure or the claims.

[0127] Embodiments implemented in computer software can be implemented in software, firmware, middleware, microcode, hardware description language, or any combination thereof. A code segment or machine executable instruction can represent any combination of a procedure, function, subroutine, program, routine, subroutine, module, software package, class, or instruction, data structure, or program statement. A code segment can be coupled to another code segment or hardware circuit by passing and / or receiving information, data, arguments, parameters, or memory contents. Information, arguments, parameters, data, etc. can be passed, forwarded, or sent by any suitable means including memory sharing, message passing, token passing, network delivery, etc.

[0128] The actual software code or specialized control hardware used to implement these systems and methods does not limit the claimed features or the present disclosure. Therefore, the operation and behavior of the systems and / or methods are described without reference to specific software code, it being understood that software and control hardware can be designed to implement the systems and methods based on the description herein.

[0129] When implemented in software, these functions can be stored as one or more instructions or codes on a non-transitory computer-readable or processor-readable storage medium. The steps of the method or algorithm disclosed herein can be embodied in a processor-executable software module, which can reside on a computer-readable or processor-readable storage medium. Non-transitory computer-readable or processor-readable media include computer storage media and tangible storage media, which help transfer computer programs from one place to another. Non-transitory processor-readable storage media can be any available medium that can be accessed by a computer. For example (but not limited to), such non-transitory processor-readable media may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage device or any other tangible storage medium, which can be used to store the desired program code in the form of instructions or data structures and can be accessed by a computer or processor. As used herein, disks and optical disks include compact discs (CDs), laser discs, optical discs, digital versatile discs (DVDs), floppy disks and Blu-ray discs, where disks typically reproduce data magnetically, while optical discs use lasers to reproduce data optically. The above combinations should also be included within the scope of computer-readable media. Additionally, the operations of a method or algorithm may reside as one or any combination or set of codes and / or instructions on a non-transitory processor-readable medium and / or computer-readable medium, which may be incorporated into a computer program product.

[0130] The foregoing description of the disclosed embodiments is provided to enable those skilled in the art to make or use the embodiments described herein and variations thereof. Various modifications to these embodiments will be apparent to those skilled in the art, and the principles defined herein may be applied to other embodiments without departing from the spirit or scope of the subject matter disclosed herein. Therefore, the present disclosure is not intended to be limited to the embodiments shown herein, but should be accorded the widest scope consistent with the appended claims and the principles and novel features disclosed herein.

[0131] While various aspects and embodiments have been disclosed, other aspects and embodiments are contemplated. The disclosed aspects and embodiments are for purposes of illustration only and are not intended to be limiting, with the true scope and spirit being indicated by the following claims.

Claims

1. A system comprising: One or more processors programmed to: receiving radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients, wherein each treatment in the set of treatments corresponds to a patient in the group of patients; generating a beam direction view (BEV) projection for each patient in the group of patients based on a treatment in the group of treatments administered to each patient in the group of patients; as well as For each patient in the set of patients, treatment data associated with the patient and data associated with the BEV projection corresponding to the patient are provided to a model to train the model to generate an output, the output representing a fluence map.

2. The system according to claim 1, wherein: The one or more processors programmed to generate the BEV projection for each patient in the group of patients are programmed to: generating a set of digitally reconstructed radiographs (DRRs) based on the treatments in the set of treatments for each patient in the set of patients; as well as The set of DRRs for each patient is concatenated to form the BEV projection for the patient.

3. The system according to claim 2, wherein: The one or more processors programmed to generate the set of DRRs are programmed to: The set of DRRs is generated based on the treatments in the set of treatments for each patient in the set of patients and a configuration of a multi-leaf collimator (MLC) involved in the treatments in the set of treatments.

4. The system according to claim 1, wherein: The one or more processors programmed to provide the model with the treatment data associated with the patients and the data associated with the BEV projections of each patient are programmed to: The model is provided with the treatment data associated with the patients and the data associated with the BEV projections for each patient, the treatment data associated with the patients including a representation of a planning target volume (PTV) or a representation of organs at risk (OAR) for the patients.

5. The system according to claim 1, wherein The one or more processors programmed to provide the model with the treatment data associated with the patients and the data associated with the BEV projections of each patient are programmed to: The model is provided with the treatment data associated with the patients and the data associated with the BEV projections for each patient, the treatment data associated with the patients including one or more leaf configurations of a multi-leaf collimator (MLC) during energy delivery to the patients by a linear accelerator (LINAC).

6. The system according to claim 1, wherein: The one or more processors programmed to provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to a model to train the model to generate an output are programmed to: comparing the output to a target annotated quantity map to determine a difference between the output and the target annotated quantity map; and One or more weights of the model are updated based on the difference between the output and the target annotation map.

7. The system according to claim 1, wherein: The one or more processors programmed to provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to a model to train the model to generate an output are programmed to: providing the treatment data associated with the patient and the data associated with the BEV projection associated with the patient to the model to train the model to generate a first output, the first output being associated with a first scale; as well as The treatment data associated with the patient and the data associated with the BEV projection associated with the patient are provided to the model to train the model to generate the output, the output being associated with a second scale.

8. A method comprising: receiving, by at least one processor, radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients, wherein each treatment in the set of treatments corresponds to a patient in the group of patients; generating, by the at least one processor, for each patient in the group of patients a beam direction view (BEV) projection based on a treatment in the group of treatments administered to each patient in the group of patients; as well as For each patient in the set of patients, treatment data associated with the patient and data associated with the BEV projection corresponding to the patient are provided to the model by the at least one processor to train the model to generate an output, the output representing a fluence map.

9. The method according to claim 8, wherein Generating the BEV projection for each patient in the group of patients includes: generating, by the at least one processor, a set of digitally reconstructed radiographs (DRRs) based on the treatments in the set of treatments for each patient in the set of patients; and The set of DRRs for each patient are concatenated via the at least one processor to form the BEV projection for the patient.

10. The method of claim 9, generating the set of DRRs comprises: The set of DRRs is generated, by the at least one processor, based on the treatments in the set of treatments for each patient in the set of patients and a configuration of a multi-leaf collimator (MLC) participating in the treatments in the set of treatments.

11. The method according to claim 8, wherein Providing the model with the treatment data associated with the patients and the data associated with the BEV projections of each patient comprises: The treatment data associated with the patients and the data associated with the BEV projections of each patient are provided to the model by the at least one processor, the treatment data associated with the patients including a representation of a planning target volume (PTV) or a representation of organs at risk (OAR) of the patients.

12. The method according to claim 8, wherein Providing the model with the treatment data associated with the patients and the data associated with the BEV projections of each patient comprises: The model is provided with the treatment data associated with the patients and the data associated with the BEV projections for each patient, the treatment data associated with the patients including one or more leaf configurations of a multi-leaf collimator (MLC) during energy delivery to the patients by a linear accelerator (LINAC).

13. The method according to claim 8, wherein Providing treatment data associated with the patient and data associated with the BEV projection associated with the patient to a model to train the model to generate an output includes: comparing, by the at least one processor, the output to a target annotated quantity map to determine a difference between the output and the target annotated quantity map; and One or more weights of the model are updated, by the at least one processor, based on the difference between the output and the target annotation map.

14. The method according to claim 8, wherein Providing treatment data associated with the patient and data associated with the BEV projection associated with the patient to a model to train the model to generate an output includes: providing, by the at least one processor, the treatment data associated with the patient and the data associated with the BEV projection associated with the patient to the model to train the model to generate a first output, the first output being associated with a first scale; and The treatment data associated with the patient and the data associated with the BEV projection associated with the patient are provided to the model by the at least one processor to train the model to generate the output, the output being associated with a second scale.

15. A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to: receiving radiation therapy treatment data associated with a set of treatments administered to a group of previously treated patients, wherein each treatment in the set of treatments corresponds to a patient in the group of patients; generating a beam direction view (BEV) projection for each patient in the group of patients based on a treatment in the group of treatments administered to each patient in the group of patients; as well as For each patient in the set of patients, treatment data associated with the patient and data associated with the BEV projection corresponding to the patient are provided to a model to train the model to generate an output, the output representing a fluence map.

16. The non-transitory computer-readable medium of claim 15, wherein: The instructions causing the one or more processors to generate the BEV projection for each patient in the group of patients cause the one or more processors to: generating a set of digitally reconstructed radiographs (DRRs) based on the treatments in the set of treatments for each patient in the set of patients; as well as The set of DRRs for each patient is concatenated to form the BEV projection for the patient.

17. The non-transitory computer-readable medium of claim 16, wherein: The instructions that cause the one or more processors to generate the set of DRRs cause the one or more processors to: The set of DRRs is generated based on the treatments in the set of treatments for each patient in the set of patients and a configuration of a multi-leaf collimator (MLC) involved in the treatments in the set of treatments.

18. The non-transitory computer-readable medium of claim 15, wherein: The instructions causing the one or more processors to provide the treatment data associated with the patients and the data associated with the BEV projections of each patient to the model cause the one or more processors to: The model is provided with the treatment data associated with the patients and the data associated with the BEV projections for each patient, the treatment data associated with the patients including a representation of a planning target volume (PTV) or a representation of organs at risk (OAR) for the patients.

19. The non-transitory computer-readable medium of claim 15, wherein: The instructions causing the one or more processors to provide the treatment data associated with the patients and the data associated with the BEV projections of each patient to the model cause the one or more processors to: The model is provided with the treatment data associated with the patients and the data associated with the BEV projections for each patient, the treatment data associated with the patients including one or more leaf configurations of a multi-leaf collimator (MLC) during energy delivery to the patients by a linear accelerator (LINAC).

20. The non-transitory computer-readable medium of claim 15, wherein: The instructions causing the one or more processors to provide treatment data associated with the patient and data associated with the BEV projection associated with the patient to a model to train the model to generate an output cause the one or more processors to: comparing the output to a target annotated quantity map to determine a difference between the output and the target annotated quantity map; and One or more weights of the model are updated based on the difference between the output and the target annotation map.

21. A system comprising: One or more processors for: receiving data associated with a set of fluence maps associated with operation of a multi-leaf collimator to focus energy generated by a linear accelerator (LINAC); as well as Data associated with each fluence map is provided to a model to cause the model to generate an output representing a leaf sequence for the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences.

22. The system of claim 21, wherein: The one or more processors programmed to provide the data associated with each fluence map to the model so that the model generates the output are programmed to: The model is trained to map an input fluence map to a subset of blade sequences in a set of possible blade sequences.

23. The system of claim 22, wherein: The one or more processors programmed to train the model to map an input fluence map to a subset of blade sequences in a set of possible blade sequences are programmed to: The input fluence map is mapped to the subset of blade sequences in a set of possible blade sequences, wherein the subset of blade sequences is associated with one or more operating parameters of the LINAC.

24. The system of claim 21, wherein: The one or more processors are further programmed to: comparing the output to a target leaf sequence corresponding to each fluence map; determining differences between the leaf sequence and the target leaf sequence; as well as One or more weights of the model are updated based on the differences between the leaf sequence and the target leaf sequence.

25. The system of claim 21, wherein: The one or more processors programmed to receive the set of fluence maps are programmed to: The set of fluence maps is received based on generation of a treatment plan for a patient undergoing intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment.

26. The system of claim 21, wherein: The one or more processors programmed to receive the set of fluence maps are programmed to: The set of fluence maps is received based on generation of a training set of sample fluence maps, the training set of sample fluence maps comprising one or more randomly generated fluence maps.

27. The system of claim 21, wherein: The one or more processors programmed to receive the set of fluence maps are programmed to: The set of fluence maps is received based on generation of a training set of sample fluence maps and generation of a treatment plan for a patient undergoing intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment.

28. The system of claim 21, wherein: The one or more processors that provide the data associated with each fluence map to the model to cause the model to generate the output are programmed to: The data associated with each fluence map is provided to a generative transformer model to cause the generative transformer model to generate the output.

29. The system of claim 21, wherein: The one or more processors that provide the data associated with each fluence map to the model to cause the model to generate the output are programmed to: The data associated with each fluence map is provided to the model so that the model generates the output, the output representing a leaf sequence and at least one monitoring unit MU value of the fluence map, the leaf sequence including one or more sets of multi-leaf collimator opening sequences corresponding to MU values ​​in the at least one MU value.

30. A method comprising: receiving, by at least one processor, data associated with a set of fluence maps associated with operation of a multi-leaf collimator to focus energy generated by a linear accelerator (LINAC); as well as Data associated with each fluence map is provided to the model via the at least one processor to cause the model to generate an output representing a leaf sequence for the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences.

31. The method according to claim 30, wherein The one or more processors programmed to provide the data associated with each fluence map to the model so that the model generates the output are programmed to: The model is trained to map an input fluence map to a subset of blade sequences in a set of possible blade sequences.

32. The method according to claim 31, wherein The one or more processors programmed to train the model to map an input fluence map to a subset of blade sequences in a set of possible blade sequences are programmed to: The input fluence map is mapped to the subset of blade sequences in a set of possible blade sequences, wherein the subset of blade sequences is associated with one or more operating parameters of the LINAC.

33. The method according to claim 30, wherein The one or more processors are further programmed to: comparing the output to a target leaf sequence corresponding to each fluence map; determining differences between the leaf sequence and the target leaf sequence; as well as One or more weights of the model are updated based on the differences between the leaf sequence and the target leaf sequence.

34. The method of claim 30, wherein: The one or more processors programmed to receive the set of fluence maps are programmed to: The set of fluence maps is received based on generation of a treatment plan for a patient undergoing intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment.

35. The method of claim 30, wherein: The one or more processors programmed to receive the set of fluence maps are programmed to: The set of fluence maps is received based on generation of a training set of sample fluence maps, the training set of sample fluence maps comprising one or more randomly generated fluence maps.

36. The method of claim 30, wherein: The one or more processors programmed to receive the set of fluence maps are programmed to: The set of fluence maps is received based on generation of a training set of sample fluence maps and generation of a treatment plan for a patient undergoing intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT) treatment.

37. The method of claim 30, wherein: The one or more processors that provide the data associated with each fluence map to the model to cause the model to generate the output are programmed to: The data associated with each fluence map is provided to a generative transformer model to cause the generative transformer model to generate the output.

38. The method of claim 30, wherein: The one or more processors that provide the data associated with each fluence map to the model to cause the model to generate the output are programmed to: The data associated with each fluence map is provided to the model so that the model generates the output, the output representing a leaf sequence and at least one monitoring unit MU value of the fluence map, the leaf sequence including one or more sets of multi-leaf collimator opening sequences corresponding to MU values ​​in the at least one MU value.

39. A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to: receiving data associated with a set of fluence maps associated with operation of a multi-leaf collimator to focus energy generated by a linear accelerator (LINAC); and Data associated with each fluence map is provided to a model to cause the model to generate an output representing a leaf sequence for the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences.

40. The non-transitory computer readable medium of claim 39, wherein: The instructions causing the one or more processors to provide the data associated with each fluence map to the model to cause the model to generate the output cause the one or more processors to: The model is trained to map an input fluence map to a subset of blade sequences in a set of possible blade sequences.

41. A system comprising: One or more processors programmed to: receiving data associated with treatment attributes associated with a patient; generating data associated with a fluence map for a patient based on a first model trained to receive as input data associated with a treatment attribute of the patient; generating data associated with a leaf sequence of the fluence map based on a second model trained to receive the fluence map, the leaf sequence comprising one or more sets of multi-leaf collimator opening sequences; as well as Data associated with the blade sequence is transmitted to cause a linear accelerator (LINAC) to operate according to the blade sequence during energy delivery to the patient.

Images