Radiation dose prediction by artificial intelligence model using organ dose tradeoff

Abstract

Provided herein are methods and systems for training and executing a model that uses artificial intelligence approaches to learn and predict doses administered to different structures during radiation therapy treatment. The method includes receiving a value indicative of a priority between a first organ at risk of a patient receiving a radiation dose and a second organ at risk of the patient; executing an artificial intelligence model that uses the values (1024) to predict radiation doses for the first organ at risk, the second organ at risk, and the target structure, an artificial intelligence model is trained according to a training dataset comprising a set of participants, a dose applied to a target structure of the participants, a first organ at risk, and a second organ at risk; and outputting a predicted radiation dose for at least one of the first organ at risk, the second organ at risk, and the target structure (1026).

Description

Technical Field

[0001] This application generally relates to training an artificial intelligence model using machine learning techniques, the artificial intelligence model being configured to predict dose distribution properties of a radiation therapy treatment. Background Art

[0002] In the treatment planning of radiation therapy, predicting the expected radiation dose distribution to be delivered is an important task for improving treatment planning efficiency. Specifically, dose prediction can be used to effectively drive an optimizer computer model (also referred to herein as a plan optimizer). For example, the optimizer can ingest the dose prediction and generate a plan accordingly. In addition, the dose prediction can be used as a clinical decision support tool before running the optimizer.

[0003] Currently, many software solutions use algorithmic methods to calculate the predicted dose distribution of patient structures, such as a planning target volume (PTV) or an organ at risk (OAR). For example, many software solutions use computer models leveraging artificial intelligence (AI) to predict the dose that can or will be delivered to a certain structure (e.g., an anatomical structure or a patient organ). However, these traditional software solutions face technical challenges. For example, most current deep learning methods can only predict one dose distribution instance for a given set of planned structures, corresponding to a specific treatment intent (e.g., a specific trade-off pattern). Thus, current software solutions cannot simultaneously generate multiple dose distributions considering alternative trade-off patterns between planned structures. Summary of the Invention

[0004] For the reasons above, there is a desire for a system that can train one or more AI models that can predict radiation dose distributions using various trade-off patterns according to the requirements of an end user (e.g., a clinician or a treating physician). Using the methods and systems discussed herein, a clinician can select a desired trade-off level between different structures, which will then be used to generate a three-dimensional (3D) radiation dose distribution reflecting the selected trade-off pattern.

[0005] The methods and systems described herein utilize specially designed AI models trained using data-driven machine learning techniques, while other solutions typically rely on real-time calculations of multi-criteria optimization. Using the methods and systems described herein, dose prediction can be performed faster than other methods. For example, the dose can be predicted quickly (e.g., within a few seconds), while other data-driven techniques may take several minutes for the calculation. Thus, using the methods and systems discussed herein is more efficient than other methods. Specifically, the AI models disclosed herein are used as an input to the plan optimizer, while other solutions are used after a plan is generated by the plan optimizer.

[0006] Other solutions utilize known algorithms (such as the mixture density network algorithm), but there are some technical drawbacks. For example, these solutions can only predict a limited number of, sometimes predefined, radiation dose distributions for each patient, while the AI models disclosed herein can predict a large number of doses, including trade-off patterns not illustrated in the training set.

[0007] Other solutions use U-Net (a convolutional neural network) to attempt to achieve the same result. However, these solutions require a large amount of training data and a huge training data set. In contrast, the AI models disclosed herein may use as few as three plans for each patient in the training set to generate acceptable results, because the AI models disclosed herein can use a generative model dedicated to this task. In some embodiments, the generative model can use a conditional variational encoder to train the model. However, the methods and systems discussed herein are not limited to generative models and / or variational encoders (or conditional variational encoders).

[0008] In an embodiment, the method includes receiving, by a processor, a value indicating a priority between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient; executing, by the processor, an artificial intelligence model that uses the value to predict radiation doses for the first critical organ, the second critical organ, and a target structure, wherein the artificial intelligence model is trained according to a training data set including a group of participants, doses applied to the target structures of the participants, the first critical organ, and the second critical organ; and outputting, by the processor, a predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

[0009] Outputting the predicted dose may include displaying a dose volume histogram that depicts the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

[0010] The value indicating the priority can be received through a sliding scale input element.

[0011] The processor can receive multiple values and output multiple predicted radiation doses, the multiple values indicating multiple priorities between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient.

[0012] The method may further include transmitting, by the processor, the predicted radiation dose to a planning optimizer software solution.

[0013] The method may further include adjusting, by the processor, at least one attribute of a radiation therapy machine according to the predicted radiation dose.

[0014] A set of weighted tensors corresponding to the training data set can be used to train the artificial intelligence model.

[0015] A generative artificial intelligence model corresponding to a variational autoencoder or a conditional variational autoencoder can be used to train the artificial intelligence model.

[0016] In another embodiment, a system (e.g., a computer system) includes a server that includes a processor and a non-transitory computer-readable medium that contains instructions that, when executed by the processor, cause the processor to receive a value that indicates a priority between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient; execute an artificial intelligence model that uses the value to predict radiation doses for the first critical organ, the second critical organ, and a target structure, wherein the artificial intelligence model is trained according to a training data set that includes a group of participants, the doses applied to the target structures of the participants, the first critical organ, and the second critical organ; and output a predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

[0017] Outputting the predicted dose can include displaying a dose volume histogram that depicts the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

[0018] A value indicating a priority between a first critical organ of a patient and a second critical organ of the patient can be received through a sliding scale input element.

[0019] The processor can receive multiple values and output multiple predicted radiation doses, where the multiple values indicate multiple priorities between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient.

[0020] The instructions can also cause the processor to transmit the predicted radiation dose to a treatment planning optimizer software solution.

[0021] The instructions can also cause the processor to adjust at least one property of a radiation therapy machine according to the predicted radiation dose.

[0022] The artificial intelligence model can be trained using a set of weighted tensors corresponding to the training data set.

[0023] A generative artificial intelligence model corresponding to a variational autoencoder or a conditional variational autoencoder can be used to train the artificial intelligence model.

[0024] In another embodiment, the system includes a computer in communication with a server and configured to display a graphical user interface; a radiation therapy machine in communication with the server; and the server is configured to receive a value indicative of a priority between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient; execute an artificial intelligence model that uses the value to predict radiation doses for the first critical organ, the second critical organ, and a target structure, wherein the artificial intelligence model is trained based on a training data set including a group of participants, doses applied to the target structures of the participants, the first critical organ, and the second critical organ; and output a predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

[0025] Outputting the predicted dose may include displaying a dose volume histogram depicting the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

[0026] A value indicative of the priority may be received via a sliding scale input element.

[0027] The server may receive a plurality of values and output a plurality of predicted radiation doses, the plurality of values indicative of a plurality of priorities between a first critical organ of the patient receiving a radiation dose and a second critical organ of the patient. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] 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 indicated as representing the background art, the drawings represent aspects of the present disclosure.

[0029] Figure 1 Shows components of an artificial intelligence (AI) dose modeling system according to an embodiment.

[0030] Figure 2 Shows a process flow diagram of an AI dose modeling system according to an embodiment.

[0031] Figure 3 Shows a process flow diagram of an AI dose modeling system according to an embodiment.

[0032] Figure 4 Shows a process flow diagram of an AI dose modeling system according to an embodiment.

[0033] Figure 5 Shows a process flow diagram for generating a training data set according to an embodiment.

[0034] Figure 6 Shows a process flow diagram for generating a training data set (separating medical images into component parts) according to an embodiment.

[0035] Figure 7 A process flow diagram showing the process for generating a training dataset (how to generate a weighted anatomical tensor) according to an embodiment.

[0036] Figure 8 A process flow diagram showing the process for training an AI model according to an embodiment.

[0037] Figure 9 A process flow diagram showing the process for executing an AI model according to an embodiment.

[0038] Figures 10 to 11 Illustrates different process flow diagrams and step examples for training and executing an AI dose modeling system according to an embodiment.

[0039] Figure 12 Illustrates different doses predicted in an AI dose modeling system according to an embodiment. Detailed Description

[0040] Now, reference will be made to the illustrative embodiments shown in the 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 relevant art and having the present disclosure, changes and further modifications to the features of the invention described herein and additional applications of the principles of the subject matter described herein should be considered 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.

[0041] Figure 1 Illustrates the components of system 100 for an AI dose modeling system according to an embodiment. System 100 may include an analysis server 110a, a system database 110b, an AI model 111, electronic data sources 120a-d (collectively referred to as electronic data sources 120), end-user devices 140a-c (collectively referred to as 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 may belong to a radiation therapy clinic where patients may receive radiation therapy treatment in certain situations through one or more radiation therapy machines (e.g., medical device 160) located within the clinic.

[0042] System 100 is not limited to the components described herein and may include additional or other components (not shown for simplicity) that should be considered within the scope of the embodiments described herein.

[0043] The above components can be interconnected via network 130. Examples of network 130 can include, but are not limited to, private or public local area networks (LANs), wireless local area networks (WLANs), metropolitan area networks (MANs), wide area networks (WANs), and the Internet. Network 130 can include wired and / or wireless communication according to one or more standards and / or via one or more transport media. Communication on network 130 can be carried out according to various communication protocols, such as Transmission Control Protocol and Internet Protocol (TCP / IP), User Datagram Protocol (UDP), and Institute of Electrical and Electronics Engineers (IEEE) communication protocols. In one example, network 130 can include wireless communication according to the Bluetooth specification set or another standard or proprietary wireless communication protocol. In another example, network 130 can also include communication via a cellular network, including, for example, GSM (Global System for Mobile Communications), CDMA (Code Division Multiple Access), EDGE (Enhanced Data rates for GSM Evolution) network, 3G (UMTS), 4G (LTE), and / or 5G.

[0044] Analysis server 110a can generate and display an electronic platform configured to receive patient information and output the execution results of AI model 111. The electronic platform can include a graphical user interface (GUI) displayed on electronic data source 120, end-user device 140, medical device 160, and / or administrator computing device 150. Examples of the electronic platform generated and hosted by analysis server 110a can be a web-based application or a website configured to be displayed on various electronic devices (such as mobile devices, tablets, personal computers, etc.).

[0045] Analysis server 110a can be any computing device that includes a processor and a non-transitory machine-readable memory capable of performing the various tasks and processes described herein. Analysis server 110a can employ various processors, such as a central processing unit (CPU) and a graphics processing unit (GPU), etc. Non-limiting examples of such computing devices can include workstation computers, laptop computers, server computers, etc. Although system 100 includes a single analysis server 110a, analysis server 110a can include any number of computing devices operating in a distributed computing environment, such as a cloud environment.

[0046] The electronic data source 120 can represent various sources that contain, retrieve, and / or access data related to the medical device 160, such as operational information related to previously performed radiation therapy treatments (e.g., electronic log files or electronic profiles), data related to previously monitored patients (e.g., tumor location and calculated dose distributions), or participants in studies used to train the AI models discussed herein. For example, the analysis server 110a can use the clinic computer 120a, the medical professional device 120b, the server 120c (associated with the treating physician and / or clinic), and the database 120d (associated with the physician and / or clinic) to retrieve and receive data. The analysis server 110a can also retrieve data from the electronic data source 120, generate a training data set, and correspondingly train the AI model 111. The analysis server 110a can execute various algorithms to transform the raw data received or retrieved from the electronic data source 120 into a machine-readable object that can be stored and processed by other analysis processes described herein.

[0047] The end-user device 140 can be any computing device, including 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, and a server computer. During operation, various users can use the end-user device 140 to access the GUI operated and managed by the analysis server 110a. Specifically, the end-user device 140 can include the clinic computer 140a, the clinic server 140b, and the medical professional device 140c. Although referred to herein as an "end-user" device, these devices are not always operated by an end user. For example, the clinic server 140b cannot be directly used by an end user. However, the results stored on the clinic server 140b can be used to populate the various GUIs accessed by the end user through the medical professional device 140c.

[0048] The administrator computing device 150 can represent a computing device operated by a system administrator. The administrator computing device 150 can be configured to display the radiation therapy treatment attributes generated by the analysis server 110a (e.g., various analysis metrics determined during the training of one or more machine learning models and / or the system); monitor the various models 111 used by the analysis server 110a, the electronic data source 120, and / or the end-user device 140; review feedback; and / or facilitate the training or retraining (calibration) of the AI model 111 maintained by the analysis server 110a.

[0049] The medical device 160 can be a radiation therapy machine configured to perform radiation therapy treatment of a patient. The medical device 160 can also include an imaging device capable of emitting radiation such that the medical device 160 can perform imaging according to various methods to accurately image the internal anatomy of the patient. For example, the medical device 160 can include a rotation system (e.g., a static or rotating multi-view system). Non-limiting examples of multi-view systems can include stereoscopic systems (e.g., two systems arranged orthogonally). The medical device 160 can also communicate with a medical device computer 162 configured to display the various GUIs discussed herein. For example, the analysis server 110a can display the results predicted by the AI model 111 on the computing devices described herein.

[0050] The AI model 111 can be stored in the system database 110b. Data received or retrieved from the electronic data source 120 can be used to train the AI model 111 and data received from the end-user device, the medical device 160, and / or other sensors can be used to execute the AI model. In some embodiments, the AI model 111 can reside within a clinic-local or clinic-specific data repository. In various embodiments, the AI model 111 uses one or more engines (e.g., a deep learning engine) to generate a predicted dose for one or more anatomical structures of a patient. The predicted dose can then be used in downstream software applications or computer models, such as a plan optimizer.

[0051] It should be understood that any alternative and / or additional machine learning models can be used to implement a similar learning engine. The deep learning engine can include processing paths trained during a training phase. Once training is complete, the deep learning engine can be executed (e.g., by the analysis server 110a) to generate predicted results.

[0052] As described herein, the analysis server 110a can store the AI model 111 (e.g., a neural network, a random forest, a support vector machine, a regression model, and / or a recurrent model, etc.) in an accessible data repository. The analysis server 110a can retrieve the AI model 111 and train the AI model to predict deformities associated with one or more structures and / or organs of a patient. Various machine learning techniques may be involved in training a machine learning model for predicting the dose of one or more structures, including supervised learning techniques, unsupervised learning techniques, and / or semi-supervised learning techniques, etc.

[0053] One type of deep learning engine is a deep neural network (DNN). A DNN is a branch of neural networks and consists of a stack of layers, each performing a specific operation (such as convolution, pooling, and / or loss calculation, etc.). Each intermediate layer receives the output of the previous layer as its input. The starting layer is the input layer, which is directly connected to or receives an input data structure including data items in one or more machine-readable objects, and may have the same number of neurons as the data items in one or more machine-readable objects provided as input. For example, a machine-readable object can be a data structure, such as a list or a vector, which includes multiple data fields that contain data received for predicting a dose. Each neuron in the input layer can accept the content of one data field as input. The analysis server 110a can preprocess the machine-readable object (e.g., through an encoding process) so that the data fields can be accepted as input to the AI model 111 described herein.

[0054] The next layer group can include any type of layer that may exist in a DNN, such as a convolutional layer, a fully connected layer, a pooling layer, and / or an activation layer, etc. Some layers (such as a convolutional neural network layer) may contain one or more filters. These filters, commonly referred to as kernels, have an arbitrary size defined by the designer. Each neuron can only respond to a specific region of the previous layer, called the receptive field. The output of each convolutional layer can be regarded as an activation map, which highlights the effect of applying a specific filter at the input. A convolutional layer can be followed by an activation layer to apply a non-linearity to the output of each layer. The next layer can be a pooling layer, which helps to reduce the dimensionality of the convolutional output. In various implementations, high-level abstractions are extracted through fully connected layers. The weights of the neural connections and kernels can be continuously optimized during the training phase.

[0055] The AI model may also be (but not necessarily) a generative AI model. As used herein, a generative AI model can refer to a deep neural network model that attempts to learn the underlying probability distribution of its high-dimensional training data by learning a probability distribution on a low-dimensional (latent space) representation. In a later stage, such as during the "generative" mode of the model, it can sample the learned probability distribution on the low-dimensional (latent space) representation of the data and generate high-dimensional data similar to the data in the training dataset.

[0056] In a non-limiting example, a conditional generative neural network model can learn the probability distribution of handwritten digits, where it conditions on the "labels" ("one", "two", etc.) of the digits and then generates seemingly realistic handwritten digits according to any such label.

[0057] The AI model 111 can use a generative NN (neural network) model to learn the probability distribution of radiation therapy doses based on organ and tumor segmentation and the dose sparing ratio of such organs. Although the injection aspect of the AI model 111 is described as using a conditional variational autoencoder (CVAE) as the selected NN model, the methods and systems discussed herein are equally applicable to other generative models (e.g., generative adversarial networks (GANs)) when provided with the same input as the CVAE.

[0058] In practice, training data can be generated by the user through observation and experience and used to facilitate learning. For example, training data can be received and monitored during the provision of radiation therapy treatment to previous patients. In another example, the training data can be a dataset that includes the PTV dose distribution of a patient during treatment and the OAR dose distribution of the same patient. The training data can be preprocessed by any suitable data augmentation method (e.g., normalization, encoding, combinations thereof, etc.) to produce a new dataset with modified attributes to improve model training. The methods and systems described herein are not limited to training the AI model based on previously treated patients. For example, instead of previously treated patients, the training dataset can include data related to any group of participants (non-patients) willing to be monitored to generate the training dataset. Thus, participants who are not treated in a study can be connected to one or more electronic sensors, where the analysis server 110a includes the data collected by the sensors within the training dataset.

[0059] As Figure 2 shown, dose prediction and treatment planning have traditionally been done in a way that does not consider different trade-offs (e.g., the dose may only consider the average trade-off inherent in the dataset used to train the dose prediction algorithm, but not consider endless other potential trade-offs). For example, method 200 illustrates the different steps involved in generating a treatment plan. The clinician can first determine the clinical objectives, including the planned structure and dose volume histogram (DVH) objectives (step 210). After receiving the constraints and parameters from the clinician, a computer model is typically used to generate a dose prediction (step 220). The predicted dose is usually individualized and does not consider any trade-offs (e.g., may only consider the average trade-off). For example, the predicted dose may correspond to three organs considered separately. The predicted dose will drive the optimization and plan generation process. For example, a 3D dose can be generated based on the predicted dose (230), dose simulation can be achieved (240), and a treatment plan can be generated (250). The treatment plan can then be submitted to the clinician, who can approve or reject the treatment plan.

[0060] The methods and systems discussed herein use different and / or additional paradigms in combination with different computer modeling techniques than previously used. For example, now refer to Figure 3, Method 300 can use a computer model to predict multiple doses considering the trade - offs between different structures (step 310), such that a clinician can select a dose combination for planning purposes (step 320).

[0061] Using the methods and systems discussed herein, a software application can ingest data generated by the disclosed AI models (where the calculations are customized according to the needs of a specific patient) to generate a customized treatment plan. For example, a doctor treating a prostate cancer patient can customize the doses received by the prostate and its corresponding OARs (the rectum and the bladder). Traditional systems can predict the 3D dose distribution of the prostate, which is a standard treatment plan considering the average trade - offs (between different organs and structures) in the training data, and these trade - offs are not unique for each patient and result in a general outcome. However, a doctor may determine that a certain patient may need to transfer the dose for the rectum to the bladder and vice versa. Thus, traditional general treatment plans are not ideal.

[0062] In contrast, using the input received from a clinician (the treating physician), the AI models discussed herein can generate customized dose distribution predictions. Thus, downstream applications (such as a planning optimizer) can ingest the data predicted by the computer models discussed herein to generate a treatment plan that is customized for a specific patient and based on the priorities desired by the clinician.

[0063] In some embodiments, a clinician may receive multiple predictions for a series of organs and / or structures, where each prediction corresponds to a different trade - off. Thus, the clinician can determine which prediction is most suitable for the patient. Thus, using the methods and systems discussed herein, a clinician can use the AI model to make more informed decisions customized for each patient. For example, depending on the patient's comorbidities, a clinician may select different trade - offs for different patients.

[0064] Using the methods and systems discussed herein, different options can be provided to a clinician to generate a treatment plan. For example, assuming that the PTV (the prostate in this example) is adequately covered (receives a dose within the clinical target) and the OAR constraints are met, there can be a high degree of freedom to balance the bladder and rectum doses (the OAR constraints in clinical trial protocols tend to be relatively loose). Thus, the AI model may still make trade - off decisions within the constraints set by the clinician. When predicting doses, clinical dose limits may restrict the number of possible permutations and trade - offs. But this may not eliminate the possibility of generating multiple dose predictions. Thus, the dose predictions discussed herein may not deviate from the clinician's goals. Thus, using the methods and systems discussed herein, the doses of OARs can be customized for a patient while still remaining within the defined limits and clinical targets.

[0065] ReferenceFigure 4 shows an example of a data flow diagram 400 according to an embodiment, which shows how to train and execute an AI model to predict a dose. Method 400 may include steps 410-430. However, other embodiments may include additional or alternative steps or may omit one or more steps entirely. Method 400 is described as being executed by a server (such as the analysis server described in Figure 1 ). However, one or more steps of method 400 may be executed by any number of computing devices running in a distributed computing system described in Figure 1 . For example, one or more computing devices may execute some or all of the steps described in Figure 4 locally.

[0066] Using method 400, an AI model can be trained based on a training data set to predict and / or generate a 3D radiation dose distribution conditioned on a trade-off pattern between a patient's anatomy and two (or more) planning structures (such as a PTV and one or more OARs). In implementation, using method 400, an AI model can be executed to predict a dose distribution based on a priority value requested by a clinician.

[0067] Training of the AI model

[0068] Before executing the AI model, the analysis server may first train the AI model and ensure its accuracy. The analysis server can use a training data set containing data related to a group of patients (or participants in a clinical trial) to train the AI model.

[0069] The data required for model training may include 3D radiation dose volumes (three or more for each patient case) and weighted anatomical tensors (three or more for each patient case). Although each aspect of the example used to describe training the AI model uses three iterations for each patient case, it can be clearly understood that some other embodiments may use more or fewer iterations. Therefore, describing three different dose volumes for each patient is not intended to limit the methods or systems discussed herein. Additionally, the various examples discussed herein describe structures A and B as two examples of OARs. However, it can be clearly understood that the number of OARs can be more or fewer.

[0070] Now referring to Figure 5 , method 500 shows a process flow diagram for generating a training data set according to an embodiment. Specifically, method 500 illustrates how an analysis server generates a training data set that can be used to train an AI model such that the model is configured to generate a prioritized dose distribution.

[0071] The analysis server can query and retrieve a set of previously executed dose predictions. For example, the analysis server can retrieve a set of previously executed treatments and their corresponding treatment plans. In a non-limiting example, the analysis server can retrieve Digital Imaging and Communications in Medicine (DICOM) radiation therapy CT files and / or DICOM RT structure files related to the generation of the previously executed treatments or plans. In some embodiments, participants in the observed cases can be used to create the information required to generate the DICOM files (e.g., the participants did not receive treatment, but their information was recorded for training purposes). DICOM as used herein can refer to a standard method for storing and transmitting digital medical image information, where the files specifically handle the transmission of radiation therapy image data and / or auxiliary data.

[0072] The DICOM RT structure file and the DICOM CT file 510 can include all planned structures, including the calculation data regarding the PTV and two or more OARs (e.g., Structure A and Structure B, two structures of interest in the trade-off decision). The DICOM RT structure file and the DICOM CT file 510 can also include one or more medical images related to each patient / participant.

[0073] The analysis server may generate three iterations for each patient case (e.g., each patient case can be represented by a DICOM RT structure file and a DICOM CT file). For example, the analysis server can use a planning optimizer software solution (also known as a radiation planning optimizer) to generate different iterations. The planning optimizer may be a computer model that ingests DICOM files and generates a treatment plan (e.g., a dose distribution).

[0074] Using a radiation planning optimizer (e.g., a photon optimizer), the analysis server can generate a first radiation plan (Plan Equivalent Spare 522), where Structures A and B should be equally spared. That is, Structures A and B are equally considered OARs when exposed to accidental radiation. To achieve this, the analysis server can assign the value p as the priority value for the terms corresponding to Structures A and B in the objective function.

[0075] Using a radiation plan optimizer, the analysis server can generate a second radiation plan (referred to as "Plan Alternative A520"), where the radiation dose to Structure A should be lower than the radiation dose to Structure B (Structure A is prioritized over Structure B). Specifically, the dose received by Structure A may be less than the dose it received in Plan Equivalent Alternative 522, while the dose received by Structure B may be higher than the dose received by Structure B in Plan Equivalent Alternative 522. To achieve this, the analysis server can use the same set of goals and priority values as those used when generating Plan Alternative 522. The analysis server can modify the priority values of the terms corresponding to Structures A and B in the objective function to indicate that Structure A is prioritized over Structure B (e.g., 2p and p - 20 respectively). When calculating Plan Alternative A 520, the higher the priority value of Structure A relative to Structure B, the more likely it is that the radiation plan optimizer will reduce the dose allocated to Structure A.

[0076] The analysis server can also use the radiation plan optimizer to generate a third radiation plan (Plan Alternative B 524), where the radiation dose to Structure B should be lower than the radiation dose to Structure A (Structure B is prioritized over Structure A). Specifically, the dose received by Structure B may be less than the dose it received in Plan Equivalent Alternative 522, while the dose received by Structure A may be higher than the dose received by Structure A in Plan Equivalent Alternative 522. To achieve this, the analysis server can use the same set of goals and priority values as those used in other plans (520 and 522). The analysis server can modify the priority values of the items corresponding to Structures A and B in the objective function to indicate that Structure A is prioritized over Structure B (e.g., p - 20 and 2p respectively). When calculating Plan Alternative B 524, the higher priority value of Structure B relative to Structure A may cause the radiation plan optimizer to reduce the dose allocated to Structure B.

[0077] The analysis server may generate multiple iterations for each patient case such that each patient case has the following three plans:

[0078] A first plan that does not prioritize one OAR over another OAR;

[0079] A second plan that prioritizes a first OAR over a second OAR; and

[0080] A third plan that prioritizes the second OAR over the first OAR.

[0081] After generating the plans, the analysis server can generate a weighted anatomical tensor corresponding to the generated plans. The tensors used in this article can refer to any multi - dimensional data array (e.g., vectors and / or matrices). Before generating the tensor, the analysis server may first separate the medical images (corresponding to different plans) into component images.

[0082] Reference Figure 6 illustrates a non - limiting example of separating a medical image into component parts. In the depicted example, the medical image 600 can be separated into three medical images, where the medical image 610 shows the patient's body, the medical image 620 shows the PTV, and the medical image 630 shows two different OARs near the PTV.

[0083] After separating the medical image, the analysis server may generate a weighted anatomical tensor. The weighted anatomical tensor can be a four - dimensional tensor with three channels, where each channel is a 3D anatomical mask (e.g., the separated medical image).

[0084] Now refer to Figure 7 Example 700 illustrates how to generate a weighted anatomical tensor. As shown, channel 1712 (corresponding to the medical image 610) can represent the patient's body contour. As shown in 710, voxels that fall within the body can be assigned a value of 1 and voxels that fall outside the body can be assigned a value of 0. This pattern may be correct regardless of the type of corresponding plan (Plan Alternative A, Plan Equal Alternative, or Plan Alternative B).

[0085] Channel 2722 (corresponding to the medical image 620) can represent the PTV. As shown in 720, voxels that fall within the PTV can be assigned a value of 1 and voxels that fall outside the PTV can be assigned a value of 0. This pattern may be correct regardless of the type of corresponding plan (Plan Alternative A, Plan Equal Alternative, or Plan Alternative B).

[0086] Channel 3732 (corresponding to the medical image 630) can represent the OAR, but the method of assigning values may be different from that of channels 1 and 2 because the method may depend on the type of plan. As shown in 730, for Plan Alternative A, voxels within structure A can be assigned a value of 0.5, voxels within structure B can be assigned a value of 1.0, and voxels in the background can be assigned a value of 0. For Plan Equal Alternative, these values may be 0.75, 0.75, and 0 respectively. For Plan Alternative B, the values may be 1.0, 0.5, and 0 respectively. For other embodiments, the selection of these values can vary as long as the values of structure A and B add up to the same value and the lower value of the structure represents the lower dose of that structure. The analysis server can add different channels to generate the weighted anatomical tensor 740. That is, the weighted anatomical tensor 740 can be formed by concatenating the three depicted channels.

[0087] Then, the analysis server can use the weighted tensors and corresponding data (e.g., different plans) to train the AI model. In some embodiments, the analysis server can use a conditional variational autoencoder (conditional VAE) to train the AI model. Thus, the analysis server can use VAE or conditional VAE to train the model. However, the analysis server may also use other machine learning methods.

[0088] The analysis server can encode the training dataset. For example, the analysis server can use VAE or conditional VAE to analyze the training data and compress various data within the training dataset into a latent space (e.g., regions in the latent space representing similar data). In some examples, this can be referred to as the "deconstruction" of the training dataset. In some examples, the model can use a dimensionality reduction protocol where the content (within the training data) is transformed into the latent space. The latent space may refer to a mathematical representation of the weighted anatomical tensors (e.g., the mathematical space that maps back to the weighted anatomical tensors ingested by the AI model). By representing the data in the latent space, the AI model can analyze the training dataset rather than comparing the training dataset in its original form (e.g., medical images). Thus, training the AI model using VAE may create efficiencies not present in generically trained AI models.

[0089] Now refer Figure 8 , a non - restrictive flowchart depicting AI model training according to an embodiment. To train the AI model, the weighted anatomical tensor 800 (e.g., weighted anatomical tensor 740) and the corresponding dose volume 810 are ingested by the AI model. The weighted anatomical tensor 800 and the dose volume 810 can initially be concatenated into a single tensor, and then this single tensor is converted by the encoder into the means (μ) and variances (σ) of N Gaussian distributions. Subsequently, a value can be sampled from each of the N Gaussian distributions to form an N - element latent vector. This vector, along with three downsampled weighted anatomical tensors, can be ingested by the decoder to output the reconstructed dose volume 820. During training, the AI model can encode the data using 3D convolution (encoder 830) and then decode the data using upsampling and concatenation (840).

[0090] The loss function for training may be the sum of the reconstruction loss (between the original dose volume and the reconstructed dose volume) and the Kullback - Leibler divergence loss (between the latent distribution and a Gaussian distribution centered at 0 with a variance of 1). Mathematically, the loss function can be expressed as:

[0091] z = latent vector

[0092] x = input data

[0093] encoder = q θ (z|x)

[0094]

[0095] p(z) = Normal(0, 1)

[0096] Loss = Reconstruction Loss + Regularization Term

[0097]

[0098] Execution and implementation of the AI model

[0099] During training, the analysis server may iteratively generate new prediction results (suggestions) based on the training dataset. If the prediction results do not match the known or true results, the analysis server will continue training unless and until the computer-generated suggestions meet one or more accuracy thresholds and are within an acceptable range. For example, the analysis server can divide the training dataset into three groups (i.e., training, validation, and testing). The analysis server can train the AI model based on the first group (training). Then, the analysis server can execute (at least partially) the trained AI model to predict the results of the second group of data (validation). Then, the analysis server verifies whether the prediction is correct. Using the above method, the analysis server can evaluate whether the AI model is properly trained. The analysis server can continuously train and improve the AI model using this method. Then, the analysis server can use the remaining data points within the training dataset to measure the accuracy of the AI model (e.g., area under the curve, precision, and recall) (testing).

[0100] In implementation, the AI model may generate multiple preferred dose distributions. Now refer to Figure 9 , a non-limiting flowchart depicting the training of the AI model according to an embodiment. As shown, method 900 is partially similar to Figure 8 the training process shown in

[0101] The generation of different degrees of trade-off patterns of dose utilization can be achieved by manipulating the values in the weighted anatomical tensor, as shown in Table 1:

[0102] Background Structure A Structure B Voxel value 0 w 1.5-w

[0103] Table 1

[0104] As shown, voxels located outside any structure, inside structure A, or inside structure B can be assigned values of 0, w, and 1.5 - w, respectively. The value w can be any floating-point number between 0.5 and 1.0. In some embodiments, the closer w is to 0.5, the smaller the dose delivered to structure A and the larger the dose delivered to structure B.

[0105] To execute the trained AI model for dose generation (calculation), the inputs may include a weighted anatomical tensor 910 and a latent vector consisting of N elements, each element sampled from a Gaussian distribution. As a result, the trained AI model predicts a dose volume 920.

[0106] Referring again to Figure 2 , at 210, the analysis server can receive a value indicating the priority between a first OAR and a second OAR of a patient who has received a radiation dose. The analysis server can present an electronic platform that includes one or more input elements configured to receive a value indicating whether (and to what extent) a clinician prefers one OAR over another. For example, the analysis server may show an input element that resembles a range of priorities between the OARs from no priority to full priority. For example, a spectrum (a sliding scale with which an end user can interact) may include a range of 0 - 1 or 0 - 100, where 0 indicates no priority (or the corresponding OAR) and 100 indicates that the OAR is fully prioritized over the second OAR. In some embodiments, a different scale (e.g., 0 - 1) may be used. In another example, the analysis server may show an input element (e.g., radio buttons or a drop-down menu) that a clinician can use to enter a numerical representation of how one OAR is prioritized over another.

[0107] At 204, the analysis server can use the value to execute an artificial intelligence model to predict the radiation doses of the first OAR, the second OAR, and the target structure, where the artificial intelligence model is trained based on a training dataset that includes a group of participants, the doses applied to the target structures of the participants, the first OAR, and the second OAR.

[0108] The analysis server can use the data received in step 210 to execute the AI model discussed herein. Additionally, the analysis server can receive patient - related data, such as treatment data, clinical goals (e.g., dose limits for PTV and / or OARs), or one or more medical images related to the patient. Using the data received in step 210 and the additional data retrieved, the analysis server can execute the trained AI model. Thus, the AI model can predict different dose distribution values for the PTV and different OARs. In some embodiments, the trained AI model can generate a DVH that predicts the dose distributions of the PTV and the OARs.

[0109] In step 230, the analysis server can output the predicted radiation doses for at least one of the first OAR, the second OAR, and the target structure.

[0110] The analysis server can output the prediction results of the executed AI model in various ways. In one embodiment, the analysis server can output the prediction results on the GUI. For example, the GUI accessed by medical professionals may be populated with different dose distributions and their corresponding priority values. For example, the GUI can show multiple DVHs and a graphical indicator indicating which (if any) OARs are prioritized and the degree to which the OARs are prioritized (alternate).

[0111] In another example, the analysis server can transmit the data predicted by the AI model to downstream applications. For example, the dose prediction can be transmitted to a treatment planning optimizer, where the predicted data (combined with other treatment requirements and clinical goals) can be used to generate a detailed treatment plan for the patient. Alternatively, the AI models discussed herein can be incorporated into the treatment planning optimizer itself and become part of it.

[0112] In another example, the analysis server can use the data predicted by the AI model to modify one or more attributes of a patient's radiation therapy treatment. For example, the analysis server can modify the attributes of a multi-leaf collimator (MLC), move the treatment couch, pause the beam, or any combination of these examples. Specifically, in combination with one or more other software solutions, the analysis server can modify the opening of the MLC such that the radiation propagation is directed according to the prioritized dose distribution of the AI model.

[0113] Non-limiting example

[0114] In a non-limiting example, the analysis server can train an AI model (using the methods discussed herein) to generate a predicted dose distribution based on organ prioritization and trade-offs. Now refer to Figure 10 , Method 1000 illustrates the operational steps of training and executing an AI model to predict the dose distribution of a patient receiving prostate cancer treatment. In the described embodiment, the patient's prostate is considered the PTV, while their bladder and rectum are considered OARs. Using the methods and systems discussed herein, the analysis server can generate a dose distribution that prioritizes (alternate) one OAR to support another OAR while all organs remain within the clinical dose limits and / or other clinical goals required by the clinician.

[0115] As shown in the figure, during the training phase 1010, the analysis server can generate three plans for the medical image 1012, including three different dose distributions. In the depicted embodiment, the analysis server can generate a first plan that prioritizes or spares the bladder (bladder sparing plan 1014), a second plan that treats the two OARs equally (equal sparing plan 1016), and a third plan that prioritizes or spares the patient's rectum (rectum sparing plan 1018). Each plan can also include a numerical representation that indicates how the OARs are prioritized. For example, each plan may contain a numerical value (ranging from 0 - 1) to indicate its priority. For example, a value of -1 can be assigned to the bladder sparing plan 1014, which indicates that this plan prioritizes and spares the bladder rather than the rectum. In contrast, the rectum sparing plan is assigned a value of +1. Additionally, a value of 0 can be assigned to the equal sparing plan to indicate its neutrality.

[0116] After training, the AI model can be configured to receive a value indicating the priority between two OARs. Using the methods and systems discussed herein, the analysis server can train the AI model such that during inference, the trained AI model can receive a value that indicates how the user wishes to prioritize the OARs and predict various doses. For example, during the inference phase 1020, the analysis server can retrieve the medical image 1022 (showing the patient's bladder, prostate, and rectum). Then, the user can input a value indicating their desired priority (e.g., using the slider scale 1024). Thus, the AI model can output a dose distribution based on the input (including the medical image 1026).

[0117] Even if the AI model is trained using three iterations of the same plan (plan 1014 with value -1, plan 1015 with value 0, plan 1018 with value +1), the AI model may generate predictions for any input received within the same range. Therefore, the value received via the slider scale 1024 is not limited to -1, 0, or +1. For example, if a clinician inputs a value of 0.35 (using the slider scale 1024 or otherwise), the AI model can interpolate a prediction of the prioritized dose distribution accordingly. In Figure 12 depicts different DVHs corresponding to different priority values.

[0118] In Figure 11In another non - limiting example as shown, a clinician accesses a hosted (or associated with an analysis server) software application. The clinician can use a sliding scale to indicate values related to how the clinician wishes to prioritize the OARs receiving doses during treatment. Based on this input, the analysis server may predict different doses. For example, if the clinician selects - 1 via sliding scale 1110, the analysis server can generate a bladder spare plan 1112, if the clinician selects 0 via sliding scale 1120, the analysis server can generate a plan equal spare 1122, and if the clinician selects + 1 via sliding scale 1130, the analysis server can generate a rectum spare plan 1132.

[0119] Each plan may include medical images (medical images 1114 - 1124 and 1134). Each plan may also include a DVH corresponding to the dose values predicted by the AI model. For example, the bladder spare plan 1112 can include a DVH 1116, which includes line a (dose value corresponding to the PTV), line b (dose value corresponding to the rectum), and line c (dose value corresponding to the bladder). As shown, the dose value of the bladder (line c) is lower than the dose value of the rectum (line b).

[0120] The plan equal spare 1122 can include a DVH 1126, which includes line a (dose value corresponding to the PTV), b (dose value corresponding to the rectum), and c (dose value corresponding to the bladder). As shown, the dose value of the bladder is similar to the dose value of the rectum (within a tolerable threshold). In some embodiments, the dose values of the bladder and rectum are not similar to each other. However, compared to the bladder spare plan 1112 or the bladder spare plan 1132, the dose values of each organ are not prioritized or spared. Similarly, the plan equal spare 1122 generates a DVH 1126, which depicts substantially the same dose prediction for the rectum (line b) and the bladder (line c).

[0121] For example, the rectum spare plan 1132 can include a DVH 1136, which includes line a (dose value corresponding to the PTV), line b (dose value corresponding to the rectum), and line c (dose value corresponding to the bladder). As shown, the dose value of the rectum (line b) is lower than the dose value of the bladder (line c).

[0122] The clinician can modify the predicted dose by interacting with the sliding scale, thus modifying the predicted value multiple times as needed to determine their desired plan (including a dose that meets one or more criteria). Once the correct plan is identified, the clinician can accept the plan, where the plan is transmitted to a downstream software application and implemented for the treatment of the patient.

[0123] The various illustrative logical blocks, modules, circuits, and algorithmic steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. 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 such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure or the claims.

[0124] Embodiments implemented in computer software can be implemented using software, firmware, middleware, microcode, hardware description language, or any combination thereof. Code segments or machine-executable instructions can represent a procedure, function, subprogram, program, routine, subroutine, module, software package, class, or any combination of instructions, data structures, or program statements. 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 transmitted via any suitable means including memory sharing, message passing, token passing, network conveyance, etc.

[0125] 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. Accordingly, 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.

[0126] When implemented in software, these functions can be stored as one or more instructions or code on a non-transitory computer-readable or processor-readable storage medium. The steps of the methods or algorithms disclosed herein can be embodied in a processor-executable software module, which can exist 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 for transferring a computer program from one place to another. The non-transitory processor-readable storage medium can be any available medium accessible by a computer. By way of example, and not limitation, such non-transitory processor-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other tangible storage medium that can be used to store the desired program code in the form of instructions or data structures and that is accessible by a computer or a processor. As used herein, disk and disc include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk, and Blu-ray disc, where disks typically reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media. Additionally, the operations of a method or algorithm can exist as code and / or instructions in one or any combination or group on a non-transitory processor-readable medium and / or a computer-readable medium that can be included into a computer program product.

[0127] The foregoing description of the disclosed embodiments is provided to enable a person skilled in the art to make or use the embodiments described herein and their variations. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the principles defined herein can be applied to other embodiments without departing from the spirit or scope of the subject matter disclosed herein. Thus, this disclosure is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the appended claims and the principles and novel features disclosed herein.

[0128] Although various aspects and embodiments have been disclosed, other aspects and embodiments are contemplated. The various aspects and embodiments disclosed are for illustrative purposes only and are not intended to be limiting, and the true scope and spirit are indicated by the appended claims.

Claims

1. A method, comprising: receiving, by a processor, a value indicative of a priority between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient; executing, by the processor, an artificial intelligence model that uses the value to predict radiation doses for the first critical organ, the second critical organ, and a target structure, wherein the artificial intelligence model is trained based on a training data set comprising a group of participants, doses applied to the target structures of the participants, first critical organs, and second critical organs; and outputting, by the processor, the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

2. The method according to claim 1, wherein outputting the predicted dose comprises displaying a dose volume histogram depicting the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

3. The method according to claim 1 or 2, wherein the value indicative of the priority between the first critical organ of the patient and the second critical organ of the patient is received via a sliding scale input element.

4. The method according to claim 1, 2, or 3, wherein the processor receives a plurality of values indicative of a plurality of priorities between the first critical organ of the patient receiving a radiation dose and the second critical organ of the patient and outputs a plurality of predicted radiation doses.

5. The method according to any one of claims 1 to 4, further comprising: transmitting, by the processor, the predicted radiation dose to a treatment planning optimizer software solution.

6. The method according to any one of claims 1 to 5, further comprising: adjusting, by the processor, at least one attribute of a radiation therapy machine based on the predicted radiation dose.

7. The method according to any one of claims 1 to 6, wherein the artificial intelligence model is trained using a set of weighted tensors corresponding to the training data set.

8. The method according to any one of claims 1 to 7, wherein the artificial intelligence model is trained using a generative artificial intelligence model corresponding to a variational autoencoder or a conditional variational autoencoder.

9. A computer system, comprising: a server including a processor and a non-transitory computer-readable medium including instructions that, when executed by the processor, cause the processor to: receive a value indicative of a priority between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient; execute an artificial intelligence model that uses the value to predict radiation doses for the first critical organ, the second critical organ, and a target structure, wherein the artificial intelligence model is trained based on a training data set comprising a group of participants, doses applied to the target structures of the participants, first critical organs, and second critical organs; and Output the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

10. The system according to claim 9, wherein outputting the predicted dose includes displaying a dose volume histogram that depicts the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

11. The system according to claim 9 or 10, wherein the value indicating the priority between the first critical organ of the patient and the second critical organ of the patient is received via a sliding scale input element.

12. The system according to claim 9, 10, or 11, wherein the processor receives a plurality of values indicating a plurality of priorities between the first critical organ of the patient receiving the radiation dose and the second critical organ of the patient and outputs a plurality of predicted radiation doses.

13. The system according to any one of claims 9 to 12, wherein the instructions further cause the processor to transmit the predicted radiation dose to a treatment planning optimizer software solution.

14. The system according to any one of claims 9 to 13, wherein the instructions further cause the processor to adjust at least one attribute of a radiation therapy machine based on the predicted radiation dose.

15. The system according to any one of claims 9 to 14, wherein the artificial intelligence model is trained using a set of weighted tensors corresponding to the training data set.

16. The system according to any one of claims 9 to 15, wherein the artificial intelligence model is trained using a generative artificial intelligence model corresponding to a variational autoencoder or a conditional variational autoencoder.

17. A system, comprising: a computer, communicatively coupled to a server and configured to display a graphical user interface; a radiation therapy machine, communicatively coupled to the server; and the server, configured to: receive a value indicating a priority between a first critical organ of a patient receiving a radiation dose and a second critical organ of the patient; execute an artificial intelligence model that uses the value to predict a radiation dose for the first critical organ, the second critical organ, and a target structure, wherein the artificial intelligence model is trained based on a training data set comprising a group of participants, the dose applied to the target structure of the participants, the first critical organ, and the second critical organ; and output the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

18. The system according to claim 17, wherein outputting the predicted dose includes displaying a dose volume histogram that depicts the predicted radiation dose for at least one of the first critical organ, the second critical organ, or the target structure.

19. The system according to claim 17 or 18, wherein the value indicating the priority is received via a sliding scale input element.

20. The system according to claim 17, 18 or 19, wherein the server receives a plurality of values indicating a plurality of priorities between the first critical organ of the patient receiving the radiation dose and the second critical organ of the patient and outputs a plurality of predicted radiation doses.

Images