Computer-implemented method and system for detecting and diagnosing faults in radiotherapy machines

Through a predictive maintenance model based on deep learning, the complexity and applicability of DLG fault detection and diagnosis in the prior art are solved, and the accuracy of fault detection and maintenance efficiency are achieved, which is suitable for different radiation therapy systems.

CN118903712APending Publication Date: 2024-11-08ELEKTA SHANGHAI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411221387.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2019-11-14
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

In existing radiation therapy systems, the fault detection and diagnosis of dynamic blade guides (DLGs) relies on complex physical models, requires a lot of domain knowledge and expertise, and the model is limited in applicability, making it difficult to adapt to different machines and manufacturers. Differences, resulting in inaccuracy of fault prediction and inefficient maintenance.

Method used

Using a predictive maintenance model based on deep learning, the machine data of the radiation therapy machine is received and processed using trained deep learning models such as convolutional neural network (CNN), recurrent neural network (RNN), or long short-term memory (LSTM) Networking, detecting and diagnosing DLG failures, including classification failure types and predicting failure severity levels.

Benefits of technology

Improves the accuracy and efficiency of fault detection and diagnosis, reduces unnecessary machine repair and testing, reduces maintenance costs, and has high versatility in the model suitable for different radiation therapy machines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118903712A_ABST
    Figure CN118903712A_ABST
Patent Text Reader

Abstract

The present disclosure relates to computer-implemented methods and systems for detecting and diagnosing faults in radiotherapy machines. The method includes receiving machine data indicative of configuration and operation of target radiotherapy machine components; applying a trained deep learning model to the received machine data of the target radiotherapy machine components, the trained deep learning model trained to establish a relationship between: (i) machine data collected from normal components and faulty components having different fault severity levels, and the machine data is indicative of the configuration and operation of the normal component and the faulty component having different fault severity levels, and (ii) fault information of the normal component and the faulty component; and detecting a fault of the target radiotherapy machine component based on applying the trained deep learning model to the received machine data, and classifying the detected fault as one of a plurality of fault severity levels.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] This application is a divisional application of the application with national application number 201980103429.4, international application date November 14, 2019, national entry date July 7, 2022, and invention name “Predictive maintenance of dynamic blade guides based on deep learning”. Technical Field

[0002] This document relates generally to fault detection and diagnostics (FDD) of radiation therapy treatment systems, and more particularly, to systems and methods for detecting and diagnosing faults associated with dynamic blade guides (DLGs) in radiation therapy machines. Background Art

[0003] Radiation therapy (or "radiotherapy") can be used to treat cancer or other diseases in mammalian (e.g., human and animal) tissue. A radiation therapy technique is provided that uses a linear accelerator (also referred to as a "linac") to irradiate a tumor with high-energy particles (e.g., electrons, protons, ions, high-energy photons, etc.). The placement and dosage of the radiation beam must be precisely controlled to ensure that the tumor receives the prescribed radiation, and the placement of the beam should be such that damage to surrounding healthy tissue, which is often referred to as organs at risk (OARs), is minimized. Similar to the prescription of a drug, a physician prescribes a predetermined amount of radiation dose to the tumor and surrounding organs. Generally, ionizing radiation in the form of a collimated beam is directed toward the patient from an external radiation source.

[0004] Specified or selectable beam energies may be used, such as for delivering a diagnostic energy level range or a therapeutic energy level range. Modulation of the radiation beam may be provided by one or more attenuators or collimators, such as a multi-leaf collimator (MLC). The intensity and shape of the radiation beam may be adjusted by collimation to avoid damaging healthy tissue adjacent to the target tissue (e.g., OARs) by conforming the projected beam to the contours of the target tissue.

[0005] A radiotherapy system such as a linac system may include many components. Failure or malfunction of one or more components may result in operating errors, unexpected failures, or even system damage. In some cases, such component failure or malfunction may affect treatment effectiveness or patient safety. Preventive inspection and maintenance of equipment and components may help reduce or eliminate equipment failures and unintentional interruptions, and plan regular activities. Alternatively, predictive maintenance, which includes periodic or continuous monitoring and evaluation of the health and operating status of equipment in service, can be used to predict the likelihood of future component failure or malfunction. Accurate prediction, detection, and diagnosis of component failure or malfunction can reduce the costs associated with maintenance and service of radiotherapy systems. Summary of the invention

[0006] MR-linac is a radiotherapy system that combines linac radiotherapy with diagnostic grade magnetic resonance imaging (MRI). MR-linac enables in-room MRI of anatomical structures and physiological treatment adaptability and response monitoring, and has the potential to reduce treatment margins using real-time visualization and target tracking. Tumors and surrounding tissues can be precisely located, their movement tracked, and treatment adjusted in real time in response to changes in tumor position, shape, biology, and spatial relationship to critical organs during treatment.

[0007] The MR-linac system can include a multi-leaf collimator (MLC) that is used to shape, direct, or modulate the intensity of a radiotherapy beam to a specified target site within a patient. The MLC consists of collimating elements called leaves that can be independently moved in and out of the path of the radiotherapy beam to shape it and change its intensity. Conformal radiation therapy and intensity modulated radiation therapy (IMRT) can be delivered using the MLC. For example, in conformal radiation therapy, the MLC allows conformal shaping of the beam to match the boundaries of the target tumor. For intensity modulated therapy, the leaves of the MLC can be moved between fields to create an IMRT distribution.

[0008] The collimating elements of the MLC can move at high speeds during operation. TM The MLC (Elekta Ltd. of Sweden) has 160 interdigitating blades with a width of 5 mm at the isocenter. These blades are arranged in two groups of 80 blades, where each group of blades is contained in a dynamic blade guide (DLG) that moves with the MLC blades. The MLC blades and DLG can be digitally controlled to provide accurate blade positioning. The maximum speed of a single MLC blade can be up to 35 mm per second (mm / s), and the DLG can move at speeds of up to 30 mm / s. Therefore, when both the DLG and the MLC move in the same direction, the MLC blades can move at speeds of up to 65 mm / s.

[0009] MLC blades and DLGs may fail or malfunction during normal use of a radiotherapy machine. For example, a failed or malfunctioning DLG may be associated with a failure of one or more subcomponents (e.g., a brake, circuit board, drive motor, linear slide, or coupling of the DLG). Correctly and timely predicting or detecting DLG failures, and accurately diagnosing the root cause of the failure (e.g., classifying the detected failure into one or more fault types, such as a brake failure, circuit board failure, or drive motor failure) can be an important part of predictive maintenance of a radiotherapy system.

[0010] Conventional predictive maintenance methods face some challenges in the context of machine fault detection and diagnosis (FDD) of radiotherapy treatment systems, such as FDD of DLG or MLC in linac systems. For example, many conventional predictive maintenance are based on complex physical models. Such models usually involve complex mathematical formulas and a large number of parameters of machine characteristics (such as friction coefficient, vibration speed, pressure, temperature, current and voltage properties). The FDD process includes fitting machine or component data to a physical model. However, to establish a complex physical model, a large amount of domain knowledge, skills and expertise (such as physics, medicine and engineering) of human designers are usually required. This may be time-consuming and may increase design complexity and overall development costs. For example, feature extraction and feature engineering (such as feature selection, feature dimensionality reduction and feature optimization) required to establish a physical model may be a time-consuming and resource-intensive task. In addition, physical models are usually constructed based on some assumptions about the normal or abnormal operating characteristics of components in radiotherapy systems. However, the operating characteristics of components may be different from one model to another, and may also be different from one manufacturer to another. Therefore, the physical models developed under these assumptions may not be suitable for different machines or systems. When certain assumptions do not hold, FDD performance may suffer.The present inventors have recognized an unmet need for advanced techniques such as self-learning of various types of machine failures to improve predictive maintenance of radiation therapy systems.

[0011] This document describes a deep learning-based predictive maintenance model, and uses such a model to detect and diagnose faults associated with a portion of a linac system (e.g., a DLG). An exemplary predictive maintenance system includes a processor configured to: receive machine data indicating the configuration and operation of a DLG in a targeted radiation therapy machine; apply a trained deep learning model to the received machine data; and detect and diagnose DLG faults. The predictive maintenance system may train the deep learning model using a plurality of data sequences generated from received machine data of one or more normal DLGs and one or more faulty DLGs. Diagnosis of DLG faults in a targeted radiation therapy machine includes classifying the DLG faults into one or more fault types associated with various components of the DLG drive system.

[0012] In this document, terms such as "fault detection", "fault diagnosis", and "fault detection and diagnosis (FDD)" are used throughout. "Fault detection" includes detecting mature faults and / or impending faults. A mature fault can be a fault that has already caused a detectable fault or faulty operation of at least a portion of the radiotherapy system. An impending fault can be a fault that is expected to occur in the near future (e.g., according to a prediction algorithm) from the time the prediction is made. Therefore, "fault detection" as used in this document may refer to detecting a mature fault and / or predicting an impending fault. "Fault diagnosis" may refer to the process of identifying the root cause of a fault, classifying a detected fault (either a mature fault or an impending fault) into one of multiple fault types, classifying a detected fault into one of multiple fault severity levels, e.g., based on a DLG metric trend, and / or generating a fault analysis.

[0013] Example 1 is a computer-implemented method for detecting and diagnosing faults in a radiotherapy machine. The method includes the following steps: receiving machine data indicating the configuration and operation of a component in a target radiotherapy machine; applying a trained deep learning model to the received machine data of the component in the target radiotherapy machine, the trained deep learning model being trained to establish a relationship between: (1) machine data collected from normal components and faulty components in the corresponding radiotherapy machine, and (2) fault information of the normal components and the faulty components, the normal components and the faulty components being of the same type as the components in the target radiotherapy machine; and detecting and diagnosing faults associated with the components in the target radiotherapy machine.

[0014] In Example 2, the subject matter of Example 1 optionally includes the following steps: receiving machine data collected from normal components and faulty components having corresponding identified faults, the machine data indicating the configuration and operation of the corresponding components; constructing a training data set, the training data set comprising a plurality of data sequences generated from the received machine data of the normal components and the faulty components; and training a deep learning model using the constructed training data set to establish a trained deep learning model.

[0015] In Example 3, the subject matter of Example 2 optionally includes: a component in a targeted radiation therapy machine that may include a dynamic blade guide (DLG), a normal component that may include a normal DLG, and a faulty component that may include a faulty DLG with a corresponding identified DLG fault. The step of detecting and diagnosing a fault may include detecting and diagnosing a DLG fault in the targeted radiation therapy machine.

[0016] In Example 4, the subject matter of Example 3 optionally includes training a deep learning model, which may include: applying corresponding penalty weights to one or more of a plurality of data sequences in a training data set; and training the deep learning model using the constructed training data set, the constructed training data set including weighted data sequences.

[0017] In Example 5, the subject matter of any one or more of Examples 3 to 4 optionally includes a trained deep learning model, which may include one or more of the following: a convolutional neural network (CNN); a recurrent neural network (RNN); a long short-term memory (LSTM) network; a deep belief network (DBN); or a transfer learning module.

[0018] In Example 6, the subject matter of any one or more of Examples 3 to 5 optionally includes generating a plurality of data series including trends of DLG current measurements over time, the DLG currents respectively resulting from one or more DLG measurements at corresponding axes.

[0019] In Example 7, the subject matter of Example 6 optionally includes a DLG current trend, which may include one or more of the following: a trend of daily average current; a trend of daily varying current; a trend of daily maximum current; a trend of a multi-day moving average of current.

[0020] In Example 8, the subject matter of any one or more of Examples 3 to 7 optionally includes generating a plurality of data series including trends over time of DLG position metrics respectively calculated for one or more DLGs at respective axes.

[0021] In Example 9, the subject matter of Example 8 optionally includes a DLG location metric, which may include a count of DLG deviation-from-location events occurring within a particular time period, and the DLG location trend may include one or more of: a trend in the daily count of deviation-from-location events; or a trend in the cumulative count of deviation-from-location events over a specified number of days.

[0022] In Example 10, the subject matter of any one or more of Examples 3 to 9 optionally includes generating multiple data series that may include trends of alarms triggered by one or more alarm events, the alarm trend may include one or more of the following: a trend of daily counts of alarms; or a trend of cumulative counts of alarms over a specified number of days.

[0023] In Example 11, the subject matter of any one or more of Examples 3 to 10 optionally includes constructing a training data set, wherein constructing the training data set may include assigning a fault type to each of a plurality of data sequences, and wherein diagnosing a DLG fault in a target radiotherapy machine includes classifying the DLG fault into one or more fault types, the fault types comprising: a DLG brake fault; a DLG drive circuit board fault; a DLG drive motor fault; a DLG slide rail fault; or a DLG coupling unit fault.

[0024] In Example 12, the subject matter of any one or more of Examples 3 to 11 optionally includes constructing a training data set, wherein constructing the training data set may include assigning a corresponding fault severity level to each of a plurality of data sequences, and wherein diagnosing a DLG fault in a targeted radiotherapy machine includes classifying the DLG fault into one of a plurality of fault severity levels.

[0025] In Example 13, the subject matter of any one or more of Examples 3 to 12 optionally includes training a deep learning model, wherein the training deep learning model may include determining a corresponding remaining useful life (RUL) for each of a plurality of data sequences and establishing a relationship between the plurality of data sequences and the corresponding determined RUL. The method may also include using the trained deep learning model to predict the RUL of the DLG in the targeted radiation therapy machine.

[0026] In Example 14, the subject matter of any one or more of Examples 3 to 13 optionally includes training a deep learning model, wherein the training the deep learning model may include adjusting one or more model parameters to minimize a cost function, the cost function including a penalty term based on a Matthews correlation coefficient (MCC).

[0027] Example 15 is a system for detecting and diagnosing faults in a radiotherapy machine configured to provide radiotherapy to a patient. The system includes a processor configured to: receive machine data indicating the configuration and operation of a component in a target radiotherapy machine; apply a trained deep learning model to the received machine data of the component in the target radiotherapy machine, the trained deep learning model being trained to establish a relationship between: (1) machine data collected from normal components and faulty components in the corresponding radiotherapy machine, and (2) fault information of the normal components and the faulty components, the normal components and the faulty components being of the same type as the components in the target radiotherapy machine; and detect and diagnose faults associated with the components in the target radiotherapy machine.

[0028] In Example 16, the subject matter of Example 15 may optionally include a processor, which may include a training module configured to perform the following steps: receiving machine data collected from normal components and faulty components having corresponding identified faults, the machine data indicating the configuration and operation of the corresponding components; constructing a training data set, the training data set including a plurality of data sequences generated from the received machine data of the normal components and the faulty components; and training a deep learning model by using the constructed training data set to establish a trained deep learning model.

[0029] In Example 17, the subject matter of Example 16 optionally includes: a component in a targeted radiation therapy machine that may include a dynamic blade guide (DLG), a normal component that may include a normal DLG, and a faulty component that may include a faulty DLG with a corresponding identified DLG fault. The processor may be configured to detect and diagnose a DLG fault in the targeted radiation therapy machine.

[0030] In Example 18, the subject matter of Example 17 optionally includes a processor that can be configured to construct a training data set using fault information for each of a plurality of data sequences, the fault information including an indication of the presence or absence of a fault, a fault type, or a fault severity level.

[0031] In Example 19, the subject matter of any one or more of Examples 17 to 18 may optionally include a training module that can be configured to generate multiple data sequences including one or more of: a trend of DLG current measurements over time; a trend of a DLG position metric over time, the DLG position metric including a count of DLG deviation position events occurring within a specific time period; or a trend of a count of alarms triggered by one or more alarm events.

[0032] Example 20 is a non-transitory machine-readable storage medium, which includes the following instructions, which, when executed by one or more processors of the machine, cause the machine to perform operations including the following steps: receiving machine data indicating the configuration and operation of components in a target radiotherapy machine; applying a trained deep learning model to the received machine data of components in the target radiotherapy machine, the trained deep learning model being trained to establish a relationship between: (1) machine data collected from normal components and faulty components in the corresponding radiotherapy machine, and (2) fault information of normal components and faulty components, the normal components and the faulty components being of the same type as the components in the target radiotherapy machine; and detecting and diagnosing faults associated with components in the target radiotherapy machine.

[0033] In Example 21, the subject matter of Example 20 optionally includes operations further including the following steps: receiving machine data collected from normal components and faulty components with identified faults, the machine data indicating the configuration and operation of the corresponding components; constructing a training data set, the training data set comprising a plurality of data sequences generated from the received machine data of the normal components and the faulty components; and using the constructed training data set to train a deep learning model to establish a trained deep learning model.

[0034] In Example 22, the subject matter of Example 21 optionally includes: a component in a targeted radiation therapy machine that may include a dynamic blade guide (DLG), a normal component that may include a normal DLG, and a faulty component that may include a faulty DLG with a corresponding identified DLG fault. The option of detecting and diagnosing faults may include detecting and diagnosing DLG faults in the targeted radiation therapy machine.

[0035] In Example 23, the subject matter of Example 22 optionally includes also including an operation of diagnosing a DLG failure in a targeted radiotherapy machine, wherein diagnosing a DLG failure in a targeted radiotherapy machine includes classifying the DLG failure into one or more of the following: a DLG brake failure; a DLG drive circuit board failure; a DLG drive motor failure; a DLG slide rail failure; or a DLG coupling unit failure.

[0036] The present disclosure discusses systems and methods for detecting and diagnosing faults in a radiotherapy system, such as faults associated with dynamic blade guides (DLGs). An exemplary predictive maintenance system includes a processor configured to: receive machine data indicating a configuration and operation of a DLG in a targeted radiotherapy machine; apply a trained deep learning model to the received machine data; and detect and diagnose DLG faults. The predictive maintenance system may train the deep learning model using a data sequence constructed from received machine data of one or more normal DLGs and one or more faulty DLGs. Diagnosis of DLG faults in a targeted radiotherapy machine includes classifying DLG faults into different fault types or different fault severities.

[0037] A computer-implemented method for detecting and diagnosing faults in a radiotherapy machine, the method comprising: receiving machine data indicating the configuration and operation of a target radiotherapy machine component; applying a trained deep learning model to the received machine data of the target radiotherapy machine component, the trained deep learning model being trained to establish a relationship between: (i) machine data collected from normal components and faulty components with different fault severity levels, and the machine data indicating the configuration and operation of the normal components and the faulty components with different fault severity levels, and (ii) fault information of the normal components and the faulty components; and based on applying the trained deep learning model to the received machine data, detecting a fault of the target radiotherapy machine component, and classifying the detected fault into one of a plurality of fault severity levels.

[0038] A system for detecting and diagnosing faults in a radiotherapy machine, the system comprising: a memory for storing a trained deep learning model, the trained deep learning model being trained to establish a relationship between: (i) machine data collected from normal components and faulty components with different fault severity levels, and the machine data indicating the configuration and operation of the normal components and the faulty components with different fault severity levels, and (ii) fault information of the normal components and the faulty components; and a processor, the processor being configured to: receive machine data indicating the configuration and operation of a target radiotherapy machine component; apply the trained deep learning model to the received machine data; and based on applying the trained deep learning model to the received machine data, detect a fault of the target radiotherapy machine component and classify the detected fault into one of a plurality of fault severity levels.

[0039] The deep learning-based predictive maintenance discussed in this document improves FDD accuracy and maintenance efficiency. Compared with conventional FDD based on physical models, the deep learning model discussed in this document advantageously learns the characteristics of different fault types from a sequence of measurement results, and the measurement results come from components in a radiotherapy system (such as DLG in a linac system). The deep learning-based predictive maintenance system, device and method discussed in this document can also be applied to the maintenance of guns, vacuum devices, magnetrons and other key lianc parts and features. This document also discusses various techniques for improving the performance of deep learning, including training data balance based on penalty weights, fusion of different deep learning models and transfer learning. The generated model can effectively learn different fault features independently. The number of features learned by the deep learning model discussed in this article can be much higher than the number of features that can be provided by the manually designed feature extractor of the conventional FDD model. In addition, the deep learning model discussed in this article can be adapted to different radiotherapy machines with higher versatility than the platform-related conventional FDD model.

[0040] Conventional FDD models usually have a pipeline architecture, in which multiple intermediate modules (such as feature extraction and fault classification) are designed, trained and optimized separately. Such modular training and optimization require a lot of domain knowledge and longer development time and higher development cost. In contrast, the deep learning model discussed in this article provides an "end-to-end" (E2E) solution to FDD. According to various embodiments, a convolutional neural network (or other types of neural networks) can be regarded as a sequence (such as a time series) DLG parameter of the input measurement results, and directly generate fault detection and diagnosis as output. Compared with the pipeline architecture, all parameters and network structures can be trained at the same time. With the improvement of the accuracy and efficiency of fault prediction and diagnosis, the number of unnecessary machine repairs, tests and possible shutdowns and the associated maintenance costs can be greatly reduced. In some cases, expensive machine damage can be reduced or even eliminated due to early detection of faults before they can cause serious damage.

[0041] The above is intended to provide an overview of the subject matter of this patent application. It is not intended to provide an exclusive or exhaustive explanation of the present invention. This application includes a specific embodiment to provide further information about this patent application. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are not necessarily drawn to scale, and in the drawings, like reference numerals describe substantially similar components throughout the several views. Like reference numerals with different letter suffixes represent different instances of substantially similar components. The accompanying drawings generally illustrate various embodiments discussed in this document by way of example and not by way of limitation.

[0043] Figure 1 An exemplary radiation therapy system is shown.

[0044] Figure 2A An exemplary radiation therapy system that can provide a therapy beam is shown.

[0045] Figure 2B An exemplary combined system including a computed tomography (CT) imaging system and a radiation therapy system is shown.

[0046] Figure 3 A partial cross-sectional view of an exemplary combined system including a magnetic resonance (MR) imaging system and a radiation therapy system is shown.

[0047] Figure 4 is a diagram showing an exemplary collimator that may be used in a radiation therapy system such as a MR-linac machine.

[0048] Figure 5 is a diagram illustrating an exemplary MLC and a portion of an environment in which the exemplary MLC may be used.

[0049] FIG. 6A to FIG. 6C are graphs showing trends of machine data collected from the DLG of various MR-linac systems before, during, and after correction of a DLG failure.

[0050] Fig. 7A is a block diagram illustrating an exemplary process for training an FDD deep learning model and using the trained model to detect and diagnose DLG failures in a targeted radiation therapy system.

[0051] Figure 7B It shows that according to Fig. 7A A block diagram of an exemplary convolutional neural network (CNN) is shown in the process for constructing and training.

[0052] Figure 7C It shows that according to Fig. 7A A block diagram of the process shown in FIG. 1 to construct and train an exemplary recurrent neural network (RNN), specifically a long short-term memory (LSTM) network.

[0053] Figure 8 is a diagram showing an exemplary architecture of an FDD deep learning model that combines two models, also known as model fusion.

[0054] Fig. 9 is a diagram showing an example of generating training data for training an FDD deep learning model for fault detection and diagnosis.

[0055] Fig.10is a flow chart illustrating an exemplary method of training a deep learning model and using the deep learning model to detect and diagnose faults associated with DLG in a radiation therapy system.

[0056] Fig.11 An exemplary block diagram of a machine that may implement one or more of the methods as discussed herein is shown. DETAILED DESCRIPTION

[0057] In the following detailed description, reference is made to the accompanying drawings that form a part of the detailed description, and the detailed description is shown by way of a diagrammatic specific embodiment in which the present disclosure can be practiced. These embodiments, also referred to herein as "examples", are described in sufficient detail to enable those skilled in the art to practice the present disclosure, and it should be understood that the embodiments may be combined or other embodiments may be utilized and structural, logical, and electrical changes may be made without departing from the scope of the present disclosure. Therefore, the following detailed description is not restrictive, and the scope of the present disclosure is limited by the attached aspects and their equivalents.

[0058] Figure 1 An exemplary radiation therapy system 100 for providing radiation therapy to a patient is shown. The radiation therapy system 100 includes a data processing device 112. The data processing device 112 may be connected to a network 120. The network 120 may be connected to the Internet 122. The network 120 may connect the data processing device 112 with one or more of the following: a database 124, a hospital database 126, an oncology information system (OIS) 128, a radiation therapy device 130, an image acquisition device 132, a display device 134, and a user interface 136. The data processing device 112 may be configured to generate a radiation therapy treatment plan 142 to be used by the radiation therapy device 130.

[0059] The data processing device 112 may include a memory device 116, a processor 114, and a communication interface 118. The memory device 116 may store computer executable instructions such as an operating system 143, a radiation therapy treatment plan 142 (e.g., an original treatment plan, an adjusted treatment plan, etc.), a software program 144, and any other computer executable instructions to be executed by the processor 114. The memory device 116 may additionally store data including medical images 146, patient data 145, and other data required to implement the radiation therapy treatment plan 142.

[0060] The software program 144 may include radiation therapy treatment planning software that implements algorithms such as artificial intelligence, deep learning, and neural networks. In an example, the software program 144 may convert a medical image in one format (e.g., MRI) into another format (e.g., CT) by generating a synthetic image such as a pseudo CT image. For example, the software program 144 may include an image processing program that trains a predictive model for converting a medical image from a medical image 146 in one modality (e.g., an MRI image) to a synthetic image of a different modality (e.g., a pseudo CT image); alternatively, the trained predictive model may convert a CT image into an MRI image. In another example, the software program 144 may register a patient image (e.g., a CT image or an MR image) with a dose distribution (also represented as an image) of the patient so that corresponding image voxels and dose voxels are appropriately associated by the network. In yet another example, the software program 144 may replace a function of the patient image, such as a signed distance function or a processed version of the image that emphasizes some aspect of the image information. Such a function may emphasize edges or differences in voxel textures, or any other structural aspect useful for neural network learning. The software program 144 may replace a function of the dose distribution that emphasizes some aspect of the dose information. Such a function may emphasize steep gradients around the target or any other structural aspect useful for neural network learning.

[0061] In an example, the software program 144 can generate projection images for a set of two-dimensional (2D) and / or 3D CT or MR images depicting an anatomical structure (e.g., one or more targets and one or more OARs), the projection images representing different views of the anatomical structure at a first gantry angle of a radiotherapy device. For example, the software program 144 can process a set of CT or MR images and create a stack of projection images, the projection images depicting different views of the anatomical structure depicted in the CT or MR images at various viewing angles of the gantry of the radiotherapy device. In particular, one projection image can represent a view of the anatomical structure at 0 degrees of the gantry, a second projection image can represent a view of the anatomical structure at 45 degrees of the gantry, and a third projection image can represent a view of the anatomical structure at 90 degrees of the gantry. The degree can be the position of the MLC relative to a specific axis of the anatomical structure depicted in the CT or MR image. The axis can remain the same for each of the different degrees measured.

[0062] In an example, the software program 144 can generate a graphical aperture image representation of the MLC leaf position at various gantry angles. These graphical aperture images are also referred to as aperture images. In particular, the software program 144 can receive a set of control points that are used to control the radiotherapy device to produce a radiotherapy beam. The control points can represent the beam intensity, the gantry angle relative to the patient position, and the leaf position of the MLC and other machine parameters. Based on these control points, a graphical image can be generated to graphically represent the beam shape and intensity output by the MLC at each specific gantry angle. The software program 144 can align each graphical image of the aperture at a specific gantry angle with the corresponding projection image at that angle generated. The image is aligned with the projection and scaled so that each projection image pixel is aligned with the corresponding aperture image pixel.

[0063] In an example, the software program 144 stores treatment planning software. The treatment planning software may include a trained machine learning model that generates or estimates a graphical aperture image representation of MLC leaf positions at a given gantry angle for a projection image of an anatomical structure representing a view of the anatomical structure at a given gantry angle. The software program 144 may also include a beam model that calculates machine parameters or control points for a given type of machine to output a beam from the MLC that achieves the same or similar estimated graphical aperture image representation of the MLC leaf positions. That is, the treatment planning software may output an image that represents an estimated image of the beam shape and intensity for a given gantry angle and for a given projection image of the gantry at that angle, and the functionality may calculate control points for a given radiotherapy device to achieve that beam shape and intensity.

[0064] In some examples, the software program 144 may include a machine fault detection and diagnostic (FDD) software package 147. The FDD software package 147 may include a trained deep learning model, such as a convolutional neural network (CNN), a recurrent neural network (RNN), a deep belief network (DBN), or a hybrid neural network consisting of two or more neural network models of different types or different model configurations. A predictive maintenance system, which may be a subsystem of the radiation therapy system 100, may be configured to perform predictive machine maintenance using the FDD software package 147. In an example, the trained deep learning model may be used to detect and diagnose a fault in a portion of a radiation therapy machine, such as a DLG in a linac system. The following, such as with reference to FIGS. 7 to 8, shows a method for performing predictive machine maintenance on a radiotherapy system 100. Fig. 9 Examples of training deep learning models and using the models to detect and diagnose faults associated with DLGs are discussed.

[0065] In addition to the memory 116 storing the software program 144, the software program 144 may additionally or alternatively be stored on a removable computer medium, such as a hard drive, a computer disk, a CD-ROM, a DVD, an HD, a Blu-ray DVD, a USB flash drive, an SD card, a memory stick, or any other suitable medium; and the software program 144 may be executed by the data processor 114 when the software program 144 is downloaded to the data processing device 112.

[0066] The data processor 114 may be communicatively coupled to the memory 116, and the processor 114 may be configured to execute computer executable instructions stored in the memory 116. The processor 114 may send the medical image 146 to the memory 116 or receive the medical image 146 from the memory 116. For example, the processor 114 may receive the medical image 146 to be stored in the memory 116 from the image acquisition device 132 via the communication interface 118 and the network 120. The processor 114 may also send the medical image 146 stored in the memory 116 to the network 120 via the communication interface 118 to be stored in the database 124 or the hospital database 126.

[0067] The data processor 114 may create a radiation therapy treatment plan 142 using a software program 144 (e.g., treatment planning software) in conjunction with medical images 146 and patient data 145. The medical images 146 may include information such as imaging data associated with a patient's anatomical region, organ, or segmented data volume of interest. The patient data 145 may include information such as: (1) functional organ modeling data (e.g., serial versus parallel organs, appropriate dose response models, etc.); (2) radiation dose data (e.g., DVH information); or (3) other clinical information about the patient and treatment (e.g., other surgeries, chemotherapy, previous radiation therapy, etc.).

[0068] In some examples, the data processor 114 (or a separate processor) can be part of a predictive maintenance system that is configured to perform predictive machine maintenance such as detecting and diagnosing machine faults or malfunctions. The data processor 114 can execute the FDD software package 147 to generate detection and diagnosis of faults (such as faults associated with the DLG of the target radiotherapy machine). Machine data (also referred to as DLG data) indicating the configuration and operating status of the DLG can be sensed using one or more sensors 137 or sensors or measurement devices separate from the radiotherapy system 100. The DLG data can be stored in the database 124. In some examples, at least some of the DLG data can be provided to the radiotherapy system 100 via an input device such as in the user interface 136 and stored in the database 124. The data processor 114 may receive DLG data stored in the database 124 and execute the FDD software package 147 to detect a DLG fault, diagnose the DLG fault as being attributed to one or more of a brake fault, a circuit board fault, or a drive motor fault, thereby determining the severity of the DLG fault, or predicting the time to failure (or remaining useful life, or “RUL”).

[0069] In some examples, the data processor 114 as part of the predictive maintenance system can be configured to train a deep learning model using data collected from one or more normal (non-fault) DLGs of the corresponding linac machine and data collected from one or more faulty DLGs of the corresponding linac machine with known or expert-determined fault types. The DLG data together with the corresponding fault labels indicating the presence / absence of a fault or the type of fault are collectively referred to as training data and can be provided to the data processor 114 to train the deep learning model. The trained deep learning model can be stored in the memory 116 or the database 124 when the specified training convergence criteria are met.

[0070] In addition, the processor 114 may utilize a software program to generate intermediate data such as, for example, updated parameters to be used by a machine learning model such as a neural network model; or to generate an intermediate 2D or 3D image, which may then be subsequently stored in the memory 116. The processor 114 may then subsequently transmit an executable radiation therapy treatment plan 142 to the network 120 to the radiation therapy device 130 via the communication interface 118, wherein the radiation therapy plan may be used to treat the patient with radiation. In addition, the processor 114 may execute a software program 144 to implement functions such as image conversion, image segmentation, deep learning, neural networks, and artificial intelligence. For example, the processor 114 may execute a software program 144 that trains medical images or delineates medical image contours; such a software program 144, when executed, may train a boundary detector or utilize a shape dictionary.

[0071] The processor 114 may be a processing device including: one or more general-purpose processing devices, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), etc. More specifically, the processor 114 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor that implements other instruction sets, or a processor that implements a combination of instruction sets. The processor 114 may also be implemented by one or more special-purpose processing devices such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a system on a chip (SoC), etc. As will be appreciated by those skilled in the art, in some embodiments, the processor 114 may be a special-purpose processor rather than a general-purpose processor. The processor 114 may include one or more known processing devices, such as those from Intel. TM Pentium TM 、Core TM , Xeon TM or series of microprocessors from AMD TM Made by Turion TM , Athlon TM 、Sempron TM 、Opteron TM FX TM 、Phenom TM series of microprocessors, or any of a variety of processors manufactured by Sun Microsystems. Processor 114 may also include a graphics processing unit, such as a graphics processing unit from Nvidia TM Manufactured series, by Intel TM GMA, Iris TM series or by AMD TM Made by Radeon TM Processor 114 may also include an accelerated processing unit, such as an TM Xeon Phi TMThe disclosed embodiments are not limited to any type of processor that is otherwise configured to meet the computational requirements of identifying, analyzing, maintaining, generating and / or providing large amounts of data or manipulating such data to perform the methods disclosed herein. In addition, the term "processor" may include more than one processor (e.g., a multi-core design or multiple processors each having a multi-core design). The processor 114 may execute sequences of computer program instructions stored in the memory 116 to perform various operations, processes, methods as described in more detail below.

[0072] The memory device 116 can store medical images 146. In some embodiments, the medical images 146 can include one or more MRI images (e.g., 2D MRI, 3D MRI, 2D flow MRI, four-dimensional (4D) MRI, 4D volume MRI, 4D imaging MRI, etc.), functional MRI images (e.g., fMRI, DCE-MRI, diffusion MRI), CT images (e.g., 2D CT, cone beam CT, 3D CT, 4D CT), ultrasound images (e.g., 2D ultrasound, 3D ultrasound, 4D ultrasound), one or more projection images representing views of anatomical structures depicted in MRI, synthetic CT (pseudo CT) and / or CT images at different angles of the gantry relative to the patient axis, PET images, X-ray images, fluoroscopic images, radiotherapy portal images, SPECT images, computer-generated synthetic images (e.g., pseudo CT images), aperture images, graphical aperture image representations of MLC leaf positions at different gantry angles, etc. In addition, the medical images 146 can also include medical image data, such as training images and ground truth images, contour images, and dose images. In an embodiment, the medical image 146 may be received from the image acquisition device 132. Thus, the image acquisition device 132 may include an MRI imaging device, a CT imaging device, a PET imaging device, an ultrasound imaging device, a fluoroscopy device, a SPECT imaging device, an integrated linac and MRI imaging device, or other medical imaging device for obtaining medical images of a patient. The medical image 146 may be received and stored in any data type or any format type that the data processing device 112 may use to perform operations consistent with the disclosed embodiments.

[0073] The memory device 116 may be a non-transitory computer-readable medium such as a read-only memory (ROM), a phase-change random access memory (PRAM), a static random access memory (SRAM), a flash memory, a random access memory (RAM), a dynamic random access memory (DRAM) such as a synchronous DRAM (SDRAM), an electrically erasable programmable read-only memory (EEPROM), a static memory (e.g., a flash memory, a flash disk, a static random access memory), and other types of random access memory, a cache memory, a register, a CD-ROM, a DVD or other optical storage device, a cassette, other magnetic storage devices, or any other non-transitory medium that can be used to store information including images, data, or computer-executable instructions (e.g., stored in any format) that can be accessed by the processor 114 or any other type of computer device. The computer program instructions can be accessed by the processor 114, read from the ROM or any other suitable memory location, and loaded into the RAM for execution by the processor 114. For example, the memory 116 may store one or more software applications. The software applications stored in the memory 116 may include, for example, an operating system 143 for common computer systems and for software-controlled devices. In addition, the memory 116 may store an entire software application or only a portion of a software application that may be executed by the processor 114. For example, the memory device 116 may store one or more radiation therapy treatment plans 142.

[0074] The data processing device 112 can communicate with the network 120 via the communication interface 118, which can be communicatively coupled to the processor 114 and the memory 116. The communication interface 118 can provide a communication connection between the data processing device 112 and the components of the radiation therapy system 100 (e.g., allowing data to be exchanged with external devices). For example, in some embodiments, the communication interface 118 can have appropriate interface circuitry to connect to the user interface 136, which can be a hardware keyboard, keypad, or touch screen through which a user can enter information into the radiation therapy system 100.

[0075] The communication interface 118 may include, for example, a network adapter, a cable connector, a serial connector, a USB connector, a parallel connector, a high-speed data transmission adapter (e.g., such as optical fiber, USB 3.0, thunderbolt, etc.), a wireless network adapter (e.g., such as a WiFi adapter), a telecommunication adapter (e.g., 3G, 4G / LTE, etc.), etc. The communication interface 118 may include one or more digital and / or analog communication devices that allow the data processing device 112 to communicate with other machines and devices, such as remotely located components, via the network 120.

[0076] The network 120 may provide the functionality of a local area network (LAN), a wireless network, a cloud computing environment (e.g., software as a service, platform as a service, infrastructure as a service, etc.), a client-server, a wide area network (WAN), etc. For example, the network 120 may be a LAN or a WAN, which may include other systems S1 (138), S2 (140), and S3 (141). The systems S1, S2, and S3 may be the same as the data processing device 112, or may be different systems. In some embodiments, one or more of the systems in the network 120 may form a distributed computing / simulation environment that collaboratively performs the embodiments described herein. In some embodiments, one or more of the systems S1, S2, and S3 may include a CT scanner that obtains a CT image (e.g., a medical image 146). In addition, the network 120 may be connected to the Internet 122 to communicate with servers and clients remotely residing on the Internet.

[0077] Thus, the network 120 may allow data transfer between the data processing device 112 and a variety of other systems and devices such as the OIS 128, the radiation therapy device 130, and the image acquisition device 132. In addition, data generated by the OIS 128 and / or the image acquisition device 132 may be stored in the memory 116, the database 124, and / or the hospital database 126. Data may be transmitted / received via the network 120 through the communication interface 118 for access by the processor 114 as desired.

[0078] The data processing device 112 can communicate with the database 124 through the network 120 to send / receive a plurality of various types of data stored on the database 124. For example, the database 124 can store machine data associated with the radiotherapy device 130, the image acquisition device 132, or other machines related to radiotherapy. The machine data information may include control points such as radiation beam size, arc placement, duration of beam on and off, machine parameters, segmentation, MLC configuration, gantry speed, MRI pulse sequence, etc. In some examples, the database 124 may store machine data representing a part of the radiotherapy system 300, such as the configuration and operating state of the DLG of the linac machine. In addition, the database 124 may store training data that can be used to train a deep learning model for detecting and diagnosing component failures or malfunctions. The training data may include machine data obtained from a normal DLG and a faulty DLG from a radiotherapy machine. The database 124 may be a storage device and may be equipped with an appropriate database management software program. Those skilled in the art will appreciate that the database 124 may include multiple devices located in a central or distributed manner.

[0079] In some embodiments, the database 124 may include a processor-readable storage medium (not shown). Although the processor-readable storage medium in the embodiments may be a single medium, the term "processor-readable storage medium" should be deemed to include a single medium or multiple media (e.g., a centralized or distributed database and / or associated caches and servers) storing one or more sets of computer-executable instructions or data. The term "processor-readable storage medium" should also be deemed to include any medium capable of storing or encoding an instruction set that is executed by a processor and causes the processor to perform any one or more of the methods of the present disclosure. Therefore, the term "processor-readable storage medium" should be deemed to include, but is not limited to, solid-state memory, optical media, and magnetic media. For example, a processor-readable storage medium may be one or more volatile, non-transient, or non-volatile tangible computer-readable media.

[0080] The data processor 114 may communicate with the database 124 to read images into the memory 116 or store images from the memory 116 to the database 124. For example, the database 124 may be configured to store a plurality of images (e.g., 3D MRI, 4D MRI, 2D MRI slice images, CT images, 2D fluoroscopic images, X-ray images, raw data from MR scans or CT scans, Digital Imaging and Communications in Medicine (DICOM) data, projection images, graphic aperture images, etc.) received by the database 124 from the image acquisition device 132. The database 124 may store data to be used by the data processor 114 when executing the software program 144 or when creating the radiation therapy treatment plan 142. The database 124 may store the resulting predicted data and data generated by a trained machine learning model such as a neural network including network parameters constituting a model learned by the network. The data processing device 112 can receive image data such as medical images 146 (e.g., 2D MRI slice images, CT images, 2D fluoroscopic images, X-ray images, 3D MRI images, 4D MRI images, projection images, graphic aperture images, etc.) from a database 124, a radiotherapy device 130 (e.g., MR-Linac), and / or from an image acquisition device 132 to generate a treatment plan 142.

[0081] In an embodiment, the radiotherapy system 100 may include an image acquisition device 132, which can acquire medical images of the patient (e.g., MRI images, 3D MRI, 2D flow MRI, 4D volume MRI, CT images, cone beam CT, PET images, functional MRI images (e.g., fMRI, DCE-MRI, and diffusion MRI), X-ray images, fluoroscopic images, ultrasound images, radiotherapy portal images, SPECT images, etc.). The image acquisition device 132 may be, for example, an MRI imaging device, a CT imaging device, a PET imaging device, an ultrasound device, a fluoroscopic device, a SPECT imaging device, or any other suitable medical imaging device for obtaining one or more medical images of the patient. The images acquired by the image acquisition device 132 may be stored in the database 124 either as imaging data and / or as test data. By way of example, the images acquired by the image acquisition device 132 may also be stored in the memory 116 by the data processing device 112 as medical images 146.

[0082] In an embodiment, for example, the image acquisition device 132 can be integrated into a single device with the radiation therapy device 130. For example, an MR imaging device can be combined with a linear accelerator to form a system known as an "MR-Linac". Such an MR-Linac can be used, for example, to determine the location of a target organ or target tumor within a patient so as to accurately direct the radiation therapy to a predetermined target according to the radiation therapy treatment plan 142.

[0083] The image acquisition device 132 can be configured to acquire one or more images of the patient's anatomical structure for a region of interest (e.g., a target organ, a target tumor, or both). Each image, typically a 2D image or slice, can include one or more parameters (e.g., 2D slice thickness, orientation, and position, etc.). In an embodiment, the image acquisition device 132 can acquire 2D slices of any orientation. For example, the orientation of the 2D slice can include a sagittal orientation, a coronal orientation, or an axial orientation. The processor 114 can adjust one or more parameters such as the thickness and / or orientation of the 2D slice to include the target organ and / or target tumor. In an embodiment, the 2D slice can be determined based on information such as a 3D MRI volume. For example, when using a radiotherapy device 130, such a 2D slice can be acquired by the image acquisition device 132 in "real time" while the patient is receiving radiotherapy treatment, where "real time" means acquiring data at least in milliseconds or less.

[0084] The data processing device 112 may generate and store a radiation therapy treatment plan 142 for one or more patients. The radiation therapy treatment plan 142 may provide information about the specific radiation dose to be applied to each patient. The radiation therapy treatment plan 142 may also include other radiation therapy information such as control points including beam angles, gantry angles, beam intensities, dose histogram-volume information, the number of radiation beams to be used during treatment, the dose of each beam, etc.

[0085] The data processor 111 may be used to perform treatment planning using, for example, treatment planning software (e.g., manufactured by Elekta Ltd. of Sweden). ) to generate a radiation therapy treatment plan 142. To generate the radiation therapy treatment plan 142, the data processor 114 can communicate with an image acquisition device 132 (e.g., a CT device, an MRI device, a PET device, an X-ray device, an ultrasound device, etc.) to access images of the patient and depict a target such as a tumor. In some embodiments, a depiction of one or more OARs such as healthy tissue surrounding the tumor or close to the tumor may be required. Therefore, when the OAR is close to the target tumor, segmentation of the OAR can be performed. In addition, if the target tumor is close to the OAR (e.g., the prostate is close to the bladder and rectum), by segmenting the OAR from the tumor, the radiation therapy system 100 can study not only the dose distribution in the target, but also the dose distribution in the OAR.

[0086] In order to depict the target organ or target tumor from the OAR, medical images such as MRI images, CT images, PET images, fMRI images, X-ray images, ultrasound images, radiotherapy portal images, SPECT images, etc. of a patient undergoing radiotherapy can be non-invasively obtained through the image acquisition device 132 to reveal the internal structure of the body part. Based on the information from the medical image, the 3D structure of the relevant anatomical part can be obtained. In addition, during the treatment planning process, many parameters can be considered to achieve a balance between effective treatment of the target tumor (for example, so that the target tumor receives sufficient radiation dose for effective treatment) and low irradiation of the OAR (for example, the OAR receives the lowest possible radiation dose). Other parameters that can be considered include the location of the target organ and the target tumor, the location of the OAR, and the movement of the target relative to the OAR. For example, the 3D structure can be obtained by outlining the target in each 2D layer or slice of the MRI or CT image or outlining the OAR and combining the outlines of each 2D layer or slice. It can be manually (for example, by a physician, dosimetrist or medical staff using a computer such as that manufactured by Elekta Ltd. of Sweden). program) or automatically (for example, using an automatic segmentation program software such as ABAS based on Atlas manufactured by Elekta Ltd. of Sweden TMIn some embodiments, the 3D structure of the target tumor or OAR can be automatically generated by the treatment planning software.

[0087] After the target tumor and OAR have been located and delineated, the dosimetrist, physician or medical staff can determine the radiation dose to be applied to the target tumor and any maximum dose of the dose that can be received by the OARs close to the tumor (e.g., left and right parotid glands, optic nerves, eyes, lenses, inner ears, spinal cord, brainstem, etc.). After the radiation dose is determined for each anatomical structure (e.g., target tumor, OAR), a process known as inverse planning can be performed to determine one or more treatment planning parameters that will achieve the desired radiation dose distribution. Examples of treatment planning parameters include volume delineation parameters (e.g., it defines the target volume, contour sensitive structures, etc.), margins around the target tumor and OAR, beam angle selection, collimator settings, and beam on time. During the inverse planning process, the physician can define dose constraint parameters that set the limits of how much radiation the OAR can receive (e.g., define the full dose to the tumor target and zero dose to any OAR; define 95% of the dose to the target tumor; define the spinal cord, brainstem, and optic nerve structures to receive ≤45Gy, ≤55Gy, and <54Gy, respectively). The results of the inverse planning may constitute a radiation therapy treatment plan 142 that may be stored in the memory 116 or the database 124. Some of these treatment parameters may be interrelated. For example, adjusting one parameter in an attempt to change the treatment plan (e.g., weightings for different objectives, such as increasing the dose to the target tumor) may affect at least one other parameter, which in turn may result in the development of a different treatment plan. Thus, the data processing device 112 may generate a customized radiation therapy treatment plan 142 having these parameters so that the radiation therapy device 130 provides the radiation therapy treatment to the patient.

[0088] In addition, the radiation therapy system 100 may include a display device 134 and a user interface 136. The display device 134 may include one or more display screens that display medical images, interface information, treatment plan parameters (e.g., projection images, graphic aperture images, contours, doses, beam angles, etc.), treatment plans, targets, localization targets, and / or tracking targets, or any related information to a user. The user interface 136 may be a keyboard, keypad, touch screen, or any type of device through which a user can input information to the radiation therapy system 100. Alternatively, the display device 134 and the user interface 136 may be integrated into a computer such as a tablet computer (e.g., an Apple computer). Lenovo Samsung etc.) device.

[0089] In addition, any and all components of the radiation therapy system 100 can be implemented as virtual machines (e.g., VMWare, Hyper-V, etc.). For example, a virtual machine can be software used as hardware. Therefore, a virtual machine can include at least one or more virtual processors, one or more virtual memories, and one or more virtual communication interfaces that are used together as hardware. For example, the data processing device 112, OIS 128, and image acquisition device 132 can be implemented as virtual machines. Given the processing power, memory, and computing power available, the entire radiation therapy system 100 can be implemented as a virtual machine.

[0090] Figure 2A An exemplary radiotherapy device 202 is shown, which may include a radiation source (e.g., an x-ray source or linac), a bed 216, an imaging detector 214, and a radiotherapy output 204. The radiotherapy device 202 may be configured to emit a radiation beam 208 to provide treatment to a patient. The radiotherapy output 204 may include one or more attenuators or collimators such as an MLC. The MLC may be a leaf group, each leaf group consisting of a plurality of MLC leaves. The leaf groups may be mounted on corresponding DLGs that enable rapid movement of the MLC leaves so that the radiotherapy beam may be shaped and directed to a designated target site on the patient. Examples of MLCs and DLGs are described below such as with reference to Figures 3 to 5 discuss.

[0091] The patient can be positioned in the area 212 and supported by the bed 216 to receive a radiation therapy dose according to the radiation therapy treatment plan. The radiation therapy output 204 can be mounted or attached to the gantry 206 or other mechanical support. When the bed 216 is inserted into the treatment area, one or more chassis motors (not shown) can rotate the gantry 206 and the radiation therapy output 204 around the bed 216. In an embodiment, the gantry 206 can rotate continuously around the bed 216 when the bed 216 is inserted into the treatment area. In another embodiment, the gantry 206 can rotate to a predetermined position when the bed 216 is inserted into the treatment area. For example, the gantry 206 can be configured to rotate the treatment output 204 around an axis ("A"). Both the bed 216 and the radiation therapy output 204 can be independently movable to other positions around the patient, such as being movable in a transverse direction ("T"), being movable in a lateral direction ("L"), or as rotating around one or more other axes, such as rotating around a transverse axis (indicated as "R"). A controller communicatively connected to one or more actuators (not shown) can control the bed 216 to move or rotate so as to appropriately position the patient within or outside the radiation beam 208 according to the radiation therapy treatment plan. Both the bed 216 and the gantry 206 can move independently of each other in multiple degrees of freedom, which allows the patient to be positioned so that the radiation beam 208 can be precisely targeted to the tumor. The MLC can be integrated with the gantry 206 to deliver a specifically shaped radiation beam 208.

[0092] Figure 2A 206 , and may have an origin at an isocenter 210. The isocenter may be defined as the location where the central axis of the radiation beam 208 intersects the origin of the coordinate axes so as to deliver a prescribed radiation dose to a location on or in the patient. Alternatively, the isocenter 210 may be defined as the location where the central axis of the radiation beam 208 intersects the patient for various rotational positions of the radiation therapy output 204 positioned by the gantry 206 about the axis A. As discussed herein, the gantry angle corresponds to the position of the gantry 206 relative to the axis A, although any other axis or combination of axes may be referenced and used to determine the gantry angle.

[0093] The gantry 206 can have an attached imaging detector 214, which is preferably opposite the radiation therapy output 204. In an embodiment, the imaging detector 214 can be located within the field of the treatment beam 208. The imaging detector 214 can remain aligned with the treatment beam 208. The imaging detector 214 can rotate about the rotation axis as the gantry 206 rotates. In an embodiment, the imaging detector 214 can be a flat panel detector (e.g., a direct detector or a scintillator detector). In this way, the imaging detector 214 can be used to monitor the treatment beam 208, or the imaging detector 214 can be used to image the patient's anatomy, such as portal imaging. The control circuit system of the radiation therapy device 202 can be integrated into the system 100 or remote from the system 100.

[0094] In an illustrative embodiment, one or more of the couch 216, treatment output 204, or gantry 206 may be automatically positioned, and the treatment output 204 may establish a treatment beam 208 according to a specified dose for a particular treatment delivery instance. A treatment delivery sequence may be specified according to a radiation therapy treatment plan, such as using one or more different orientations or positions of the gantry 206, couch 216, or treatment output 204. Treatment deliveries may occur sequentially, but may intersect at a desired treatment site on or within the patient, such as at the isocenter 210. A prescribed dose of radiation therapy may thereby be delivered to the treatment site while reducing or avoiding damage to tissue near the treatment site.

[0095] Figure 2B An exemplary radiation therapy apparatus 202 is shown that combines a radiation system (e.g., a linac) with a CT imaging system. The radiation therapy apparatus 202 may include an MLC (not shown). The CT imaging system may include an imaging X-ray source 218 that provides X-ray energy, such as in the kiloelectronvolt (keV) energy range. The imaging X-ray source 218 may provide a fan and / or cone beam 208 directed toward an imaging detector 222, such as a flat panel detector. The radiation therapy apparatus 202 may be similar to that described with respect to Figure 2A The system depicted includes a radiation therapy output 204, a gantry 206, a couch 216, and another imaging detector 214, such as a flat panel detector. An X-ray source 218 may provide a relatively low energy diagnostic X-ray beam for imaging.

[0096] like Figure 2B As shown in , the radiotherapy output 204 and the x-ray source 218 can be mounted on the same rotating gantry 206, rotationally separated from each other by 90 degrees. In some examples, two or more x-ray sources can be mounted along the periphery of the gantry 206, so that each x-ray source has its own detector arrangement to provide diagnostic imaging at multiple angles simultaneously. Similarly, multiple radiotherapy outputs 204 can be provided.

[0097] Figure 3 An exemplary radiation therapy system 300 is shown that combines a radiation system (e.g., linac) with a nuclear MR imaging system, which is also referred to as an MR linac system. Such a combined system is also referred to as an MR linac system. The system 300 may include a couch 216, an image acquisition device 320, and a radiation delivery device 330. The system 300 may deliver radiation therapy to a patient according to a radiation therapy treatment plan (e.g., a treatment plan 142 created and stored in the memory 116). In some embodiments, the image acquisition device 320 may correspond to Figure 1 An image acquisition device 132 is provided in the apparatus and can acquire an image of a first modality (eg, an MIR image) or a target image of a second modality (eg, a CT image).

[0098] The bed 216 can support the patient during treatment. In some implementations, the bed 216 can move along a horizontal translation axis (labeled "I") so that the bed 216 can move a patient resting on the bed 216 into and / or out of the system 300. The bed 216 can also rotate around a central vertical rotation axis that is transverse to the translation axis. To allow such movement or rotation, the bed 216 can have motors (not shown) that enable the bed to move in various directions and rotate along various axes. A controller (not shown) can control these movements or rotations to appropriately position the patient according to the treatment plan.

[0099] In some embodiments, the image acquisition device 320 may include an MR imaging machine for acquiring 2D or 3D MRI images of the patient before, during, and / or after a treatment session. The image acquisition device 320 may include a magnet 321 for generating a main magnetic field for magnetic resonance imaging. The magnetic field lines generated by the operation of the magnet 321 may extend substantially parallel to the central translation axis "I". The magnet 321 may include one or more coils having axes extending parallel to the translation axis "I". In some embodiments, one or more coils in the magnet 321 may be spaced apart so that a central window 323 of the magnet 321 is free of coils. In other embodiments, the coils in the magnet 321 may be sufficiently thin or have a reduced density so that the coils are substantially transmissive to radiation of a wavelength generated by the radiotherapy device 330. In some embodiments, the image acquisition device 320 may also include one or more shielding coils that may generate magnetic fields of approximately equal amplitude and opposite polarity outside the magnet 321 so as to cancel or reduce any magnetic field outside the magnet 321. As described below, the radiation source 331 of the radiation therapy device 330 can be positioned in a region where the magnetic field is at least cancelled to a first order or reduced.

[0100] The image acquisition device 320 may also include two gradient coils 325 and 326, which may generate a gradient magnetic field superimposed on the main magnetic field. The coils 325 and 326 may generate a gradient in the magnetic field generated, which allows protons to be spatially encoded so that the position of the protons may be determined. The gradient coils 325 and 326 may be positioned around a common central axis with the magnet 321, and may be displaced along the central axis. The displacement may create a gap or window between the coils 325 and 326. In an embodiment where the magnet 321 includes a central window 323 between the coils, the two windows may be aligned with each other.

[0101] In some embodiments, the image acquisition device 320 may be an imaging device other than MRI, such as X-ray, CT, CBCT, spiral CT, PET, SPECT, optical tomography, fluorescence imaging, ultrasound imaging, radiation therapy portal imaging device, etc. As will be appreciated by one of ordinary skill in the art, the above description of the image acquisition device 320 relates to certain embodiments and is not intended to be limiting.

[0102] The radiotherapy device 330 may include a radiation source 331 (e.g., an x-ray source or linac) and a collimator such as an MLC 332. A collimator is a beam limiting device that can help shape the radiation beam coming out of the machine and can limit the maximum field size of the beam. The MLC 332 can be used to shape, direct, or modulate the intensity of the radiotherapy beam to a specified target site in the patient. The MLC 332 may include a metal collimator plate (also referred to as an MLC blade) that slides into a position that forms a desired field shape. The blade set may be mounted on a corresponding DLG 334 that enables the MLC blades to move quickly. The radiotherapy device 330 may be mounted on a chassis 335. When the bed 216 is inserted into the treatment area, one or more chassis motors (not shown) may rotate the chassis 335 around the bed 216. In an embodiment, the chassis 335 may rotate continuously around the bed 216 when the bed 216 is inserted into the treatment area. The chassis 335 may also have an attached radiation detector (not shown) which is preferably located opposite the radiation source 331 with the rotation axis of the chassis 335 positioned between the radiation source 331 and the detector. In addition, the device 330 may include control circuitry (not shown) for controlling, for example, one or more of the bed 216, the image acquisition device 320, and the radiation therapy device 330. The control circuitry of the radiation therapy device 330 may be integrated within the system 300 or remote from the system 300.

[0103] During radiation therapy treatment, a patient may be positioned on the couch 216. The system 300 may then move the couch 216 into a treatment area defined by the magnet 321 and coils 325, 326, and chassis 335. The control circuitry may then control the radiation source 331, MLC 332, and chassis motors to deliver radiation to the patient through the window between coils 325 and 326 according to the radiation therapy treatment plan.

[0104] FIG. 2A to FIG. 2B and Figure 3 The radiation therapy output configurations shown in , such as those in which the radiation therapy output can rotate about a central axis (e.g., axis "A"), are for purposes of illustration and not limitation. Other radiation therapy output configurations may be used. For example, the radiation therapy output may be mounted to a manipulator or robotic arm having multiple degrees of freedom. In yet another embodiment, the treatment output may be fixed (e.g., located in an area laterally separated from the patient), and a platform supporting the patient may be used to align the radiation therapy isocenter with a designated target site within the patient.

[0105] Figure 4 is a diagram showing an exemplary collimator 400 that may be used in a radiation therapy system such as an MR-linac machine. An example of a collimator is the Agility TM Collimator (Elekta Ltd. of Sweden). The collimator 400 may include a beam shaping component, which includes MLC blades 410, a diaphram 420 and a DLG 430. The MLC blades 410, which are an embodiment of the MLC blades 332, may be composed of an array of metal collimator plates called MLC blades. The MLC blades may have curved blade ends. In an example, the MLC blades may be made of tungsten alloy. The MLC blades may be arranged into two separate MLC blade groups. In an example, the MLC may include 160 blades, and the 160 blades are arranged into blade groups of 80 blades each. Each MLC blade group is contained in a corresponding DLG 430 and mounted thereon. The DLG 430, which is an embodiment of the DLG 334, may be arranged in a direction parallel to the arrangement of the MLC blades (Y direction, along Figure 4 The MLC blade 410 moves along the Y1 or Y2 axis shown in FIG. Figure 4A first blade group included in a first DLG is shown, and a second blade group located opposite to the first blade group and included in a second DLG. The MLC blades in the first group can move along the Y1 axis, and the MLC blades in the second group can move along the Y2 axis opposite to the Y1 axis. The first DLG is referred to as DLG-Y1, and the second DLG is referred to as DLG-Y2. Rapid MLC movement and effective beam shaping can be achieved by simultaneous blade and DLG movement. In a non-limiting example, the maximum speed of the MLC blade 410 is approximately 35 millimeters per second (mm / s), and the maximum speed of the DLG 420 is approximately 30 mm / s. Therefore, when the DLG 420 and the MLC blade 410 move in the same direction, the maximum possible blade speed can be approximately 65 mm / s. In a non-limiting example, the maximum stroke of the MLC blade within the DLG 420 can be approximately 200 mm, and the DLG has a stroke range of approximately 150 mm (measured at the isocenter). Therefore, the maximum overtravel of the blade can be approximately 150 mm.

[0106] A pair of contoured diaphragms 420 (also referred to as clips) may be mounted orthogonally to the MLC blades 410. In an example, the diaphragms 420 may overtravel a certain distance about the central axis, such as up to 150 mm in a non-limiting example. The diaphragms 420 may move in a direction (X direction) perpendicular to the direction of travel of the MLC blades (Y direction). Figure 4 , the first diaphragm can move along the X1 axis and the second diaphragm can move along the X2 axis opposite the X1 axis. In an example, the diaphragm 420 can have a maximum speed of about 90 mm / s. The diaphragm can have rounded radially defined edges. Both the MLC blades 410 and the diaphragm 420 have rounded ends. In some examples, the diaphragm 420 can have a "thick" area that provides full attenuation and a "thin" area where attenuation is provided by both the blades and the diaphragm.

[0107] The DLG 420 can be coupled to a DLG drive system, which includes a motor, a power transmission, a brake, and other components that drive the DLG 420 at a desired speed and direction. A controller can be coupled to the drive system to control the movement of the DLG 420. As previously discussed, the DLG, like other high-speed moving parts in a machine, may be subject to faults and malfunctions. These faults may be associated with one or more components of the DLG drive system or control system (such as a motor, a transmission, one or more brakes, a linear slide supporting the moving DLG, or a coupling unit connecting different components). Timely and accurate detection (including early detection or prediction) and diagnosis of these faults may be important to ensure proper system operation and accurate delivery of radiation therapy to patients. Various embodiments discussed in this document, including reference to FIG. 6A to FIG. 6C Figures 7 to Figure 8 The deep learning-based FDD discussed can improve the predictive maintenance of radiotherapy equipment.

[0108] Figure 5 An exemplary MLC 530 is shown, as well as a portion of an environment in which the exemplary MLC may be used. MLC 530 may be an embodiment of MLC 430 or MLC 232. By way of example and not limitation, MLC 530 may include blades 532A to 532J that may be automatically positioned to define an aperture that approximates a cross-section or projection of a tumor 540. Blades 532A to 532J allow for modulation of a radiotherapy beam. Blades 532A to 532J may be made of a material designated to attenuate or block the radiation beam in areas other than the aperture, depending on the radiotherapy plan. For example, blades 532A to 532J may include a metal plate (e.g., comprising tungsten) with the long axis of the plate oriented parallel to the beam direction and having ends oriented orthogonal to the beam direction (e.g., Figure 2A 540 or other target site. This is compared to using a static collimator configuration, or compared to using an MLC 332 configuration that is determined specifically using "offline" treatment planning techniques. Radiation therapy techniques that use MLC 332 to produce a specified radiation dose distribution to a tumor or a specific area within a tumor may be referred to as IMRT. The resulting beam shape output using MLC 332 is represented as a graphic aperture image. That is, a given graphic aperture image is generated to represent how the beam looks (beam shape) and its intensity after passing through and being output by MLC 332.

[0109] FIG. 6A to FIG. 6C Graphs showing exemplary trends of machine data collected from the DLGs of various MR-linac systems before and during a DLG failure and after the DLG failure has been corrected. Different types of DLG failures are considered in these examples. The machine data may include a sequence of measurements of a physical parameter over a specified time period. Measurements may be made from one or more parts of a DLG drive system or control system, such as a motor, a power transmission, a brake, a linear slide, or a coupling unit connecting different components.

[0110] In an example, the measured DLG data may include DLG current measurements when the collimator is in active operation. Figure 4 The DLG-Y1 or DLG-Y2 shown in FIG. 1 is measured from one or more DLGs, respectively.Y1 and I Y2 Indicated current. DLG current I Y1 and I Y2 may have floating point values. In another example, the DLG data may include the position of one or more of DLG-Y1 or DLG-Y2. The position data may be represented by numerical coordinates or a classification flag of a DLG deviation position event, which indicates that the DLG (e.g., DLG-Y1 or DLG-Y2) failed to reach a specified speed when it moved to a preset position. In an example, the DLG data may include the position of one or more of DLG-Y1 or DLG-Y2, respectively, which has occurred within a specified time period. Y1 and NP Y2 The count of out-of-position events associated with DLG-Y1 or associated with DLG-Y2 indicated. Out-of-position event count NP Y1 and NP Y2 Can have an integer value.

[0111] In yet another example, the DLG data may include alarms triggered by one or more alarm events, or a count of alarms during a specified time period indicated by N_Alarm. The alarm count N_Alarm may have an integer value. By way of example and not limitation, alarm events may include: (1) an "MLC terminated error alarm," which is a general MLC termination alarm that may be triggered by many components of the collimator including the DLG; (2) an "MLC terminated Not OK alarm," which corresponds to an unexpected termination of treatment, such as due to incorrect position, oscillation, or other problems with the DLG; (3) an "MLC Lost Leaf reflector alarm," which may occur when the DLG is traveling too slowly or too fast, so that the leaf reflector of a single leaf may not be in the expected position, and the leaf reflector is lost and found multiple times later in a short period of time; (4) an "reflector uncalibrated alarm," which may occur when the position of one or more reference reflectors has been lost, so that at least calibration of the reflector is required; (5) an "DLG-Y1 position alarm," which indicates termination of DLG-Y1; or (6) an "DLG-Y2 position alarm," which indicates termination of DLG-Y2.

[0112] DLG data as discussed above may be acquired continuously, periodically, or according to a specified data acquisition schedule. In some examples, a statistical metric (e.g., mean, median, maximum (max), minimum (min), range, variance, standard deviation (SD), or other high-order statistics) or a cumulative metric (e.g., sum, moving sum, or moving average) may be generated based on the DLG current measurements, DLG deviation event counts, or alarm counts. By way of example and not limitation, one or more metrics, hereinafter referred to as DLG current metrics, may be generated based on the DLG current measurements. Examples of DLG current metrics may include one or more of the following: (1) an hourly current metric, such as an hourly average I Y1 Or hourly average I Y2 , or maximum I per hour Y1 or maximum I per hour Y2 ; or (2) daily current measurement, such as I Y1 or I Y2 In an example, at least some daily current metrics (e.g., I Y1 or I Y2 In another example, at least some daily current metrics (e.g., I Y1 or I Y2 In some examples, the DLG metric may include a moving average of a daily current metric calculated over a moving time window of a specified number of days (e.g., a 5-day moving window). An example of a moving average current metric may include a moving average of a daily average current, where the daily average current may be determined based on an hourly average current or based on an hourly maximum current over a specified number of hours up to 24 hours. Y1 ) or DLG-Y2 (e.g. based on hourly Y2 ) calculates each of an hourly current metric, a daily current metric, or a moving average of the daily current metric.

[0113] In addition to the statistical DLG current metric, a metric of DLG out-of-position event counts can also be calculated. Examples of DLG out-of-position event metrics can include one or more of the following: (1) daily counts NP for a number of hours up to 24 hours Y1 or daily count NP Y2 ; (2) NP within a moving time window of a specified number of days (e.g., a 5-day moving window) Y1Moving sum or NP Y2 or (3) the moving sum of NPs within a moving time window of a specified number of days (e.g., a 5-day moving window). Y1 Moving average or NP Y2 Additionally or alternatively, a metric of the DLG alarm count may be calculated, which may include, by way of example and not limitation, one or more of the following: (1) a daily count of alarms (N_Alarm) over a number of hours up to 24 hours; (2) a moving sum of N_Alarms over a moving time window of a specified number of days (e.g., a 5-day moving window); or (3) a moving average of N_Alarms over a moving time window of a specified number of days (e.g., a 5-day moving window). A daily count, moving sum, or moving average of alarm counts may be calculated separately for one or more types of alarm events as discussed above. Alternatively or additionally, a daily count, moving sum, or moving average of alarm counts may be calculated across all alarm events, regardless of alarm type.

[0114] FIG. 6A to FIG. 6C Each shows exemplary trends for a number of DLG metrics. By way of example and not limitation, these data trends may include: daily average current I Y1 (Y1_mean_mean); daily maximum current I Y1 (Y1_max_mean, the maximum current per hour in a day I Y1 Average value of daily current I Y2 (Y2_mean_mean); daily maximum current I Y2 (Y2_max_mean, the maximum current per hour in a day I Y2 The average value of Y1 deviation event count NP within the 5-day moving window Y1 Moving sum of Y1_event_MS; Y2 deviation event NP within a 5-day moving window Y2 =The moving sum of (Y2_event_MS); and the moving sum of the alarm count N_Alarm within a 5 day moving window (Alarm_MS). In each of these figures, T1 represents the time when a predictive maintenance phase for predicting or detecting early signs of DLG failure is initiated. T2 represents the onset time of identification (e.g., by a human expert) of a fault type, and T3 represents the time when the fault type is corrected or resolved. Thus, a fault prediction window can be defined between T1 and T2, and a fault occurs in the window between T2 and T3.

[0115] Fig. 6ATrends of DLG metrics corresponding to a fault associated with a printed circuit board (PCB) of a DLG ("PCB fault") are shown. In this example, the PCB fault is accompanied by an increase in the four DLG current metrics (Y1_mean_mean, Y1_max_mean, Y2_mean_mean, and Y2_max_mean) between T2 and T3 over time. Correction of the PCB fault is accompanied by a regression of the four DLG current metrics after T3. Among the other metrics, the daily average currents of Y1 and Y2 (Y1_mean_mean and Y2_mean_mean) both increase before the onset of the PCB fault, demonstrating the ability of these metrics in predicting this type of PCB fault.

[0116] Figure 6B The trends of DLG metrics corresponding to a fault associated with a DLG brake ("Brake Fault") are shown. In this example, between T2 and T3, a brake fault is accompanied by a more significant rise over time in the metrics associated with Y2 (Y2_mean_mean, Y2_max_mean, and Y2_event_MS) than the metrics associated with Y1. Correction of the brake fault is accompanied by a regression of the metrics associated with Y2 after T3. The metrics associated with Y2 also increase before the onset of a brake fault, demonstrating the ability of these metrics in predicting this type of brake fault.

[0117] Figure 6C The trend of the DLG metrics corresponding to a hybrid fault involving both the PCB and the DLG brake is shown. In this example, the hybrid fault is accompanied by an increase in the four DLG current metrics (Y1_mean_mean, Y1_max_mean, Y2_mean_mean and Y2_max_mean) between T2 and T3 over time. The correction of the hybrid fault is accompanied by a regression of the four DLG current metrics after T3. Among the other metrics, the four DLG current metrics all increase before the onset of the hybrid fault, thereby demonstrating the ability of these metrics in predicting this type of hybrid fault.

[0118] Fig. 7A700A is a diagram illustrating an exemplary process 700A for training an FDD deep learning (DL) model (or machine learning model) and detecting and diagnosing faults associated with one or more components in a radiation therapy system (e.g., a DLG in a linac system) using the trained FDD model. The process 700A may be implemented as computer-readable and executable instructions stored in a memory 116 of the radiation therapy system 100, such as part of an FDD software package 147. The input 704 may include a specific deep learning model 710 with an initial set of values ​​and training data 720. Examples of the deep learning model 710 may include a convolutional neural network (CNN), a recurrent neural network (RNN), a deep belief network (DBN), or a hybrid neural network including two or more neural network models of different types or different model configurations.

[0119] Now refer to Figure 7B , CNN 700B can be constructed and used in process 700A. The CNN network can automatically learn the characteristics of the data from the sample, thereby eliminating the complex feature extraction in the conventional machine learning model that usually requires a lot of expertise. An additional advantage of CNN is that the scale of CNN parameters can be greatly reduced by weight sharing. As a result, the complexity of the training process can be reduced, the convergence speed can be increased, and the model generalization ability can be enhanced.

[0120] By way of example and not limitation, Figure 7B The CNN 700B shown in FIG. 700B may include an input layer 751, an intermediate learning layer 752, and a dense layer 753 forming a fully connected output layer. In the example shown, the intermediate learning layer 752 includes four convolutional layers (e.g., four one-dimensional convolutional layers or "CONV1D" as shown) and a global average pooling layer. The global average pooling layer can reduce the weight parameters of the network and reduce the training complexity. A batch normalization (BN) operation can be performed after each convolutional layer. In some examples, a nonlinear activation layer can be used in each intermediate layer. An example of an activation function is a rectified linear unit (ReLU). The BN layer can improve the gradient flowing through the network, solve the problem of gradient disappearance, allow a larger learning rate, improve training speed, reduce the strong dependence on initialization, improve regularization strategies, reduce the need for dropout, and effectively improve the generalization ability of the CNN network.

[0121] Now refer to Figure 7C, RNN can be constructed and used in process 700A. RNN includes connections between nodes to form a directed graph along a time series. The directed graph can use an internal state (memory) to store past information, and the network decision is affected by what it has learned from the past. This makes RNN suitable for processing sequences of inputs. In the example, a long short-term memory (LSTM) network can be used. LSTM is a type of RNN architecture characterized by feedback connections. A common LSTM unit can be composed of a cell, an input gate determines how much new information to add to the cell, a forget gate determines which information to discard or retain in the cell, and an output gate determines the value to be output. LSTM cells can remember values ​​within arbitrary time intervals, and three gates can adjust the flow of information in and out of the cell, which makes LSTM networks particularly suitable for sequential learning and pattern recognition (e.g., processing, classification, and making predictions based on an entire sequence of data such as time series data).

[0122] By way of example and not limitation, Figure 7C LSTM network 700C is shown, which may include an input layer 761, an intermediate learning layer 762, and a dense layer 763 that may serve as a fully connected output layer. The intermediate learning layer 762 may include two to four LSTM layers (e.g., Figure 7C ). To prevent overfitting of the model, a dropout operation may be performed after each LSTM layer. Dense layer 753 may be a fully connected output layer. In the example, L2 regularization may be implemented in the fully connected layer. Regularization is the process of controlling model complexity by adding a penalty term to an objective function (or cost function or loss function). For L2 regularization, the penalty term is defined as the L2 norm equal to the square of the magnitude of the feature weights in the neural network model. The resulting cost function may include an estimated error term (e.g., the sum of squared error terms) and a penalty term. The use of L2 regularization can help implement a training model in which a preferred balance between model complexity and predictive performance can be achieved.

[0123] Now return to reference Fig. 7A , the training data 720 may include DLG data 722 and expected results 724. The DLG data 722 may be collected from DLGs in a plurality of radiation therapy systems (e.g., linac machines), the DLGs including, for example, one or more normal (i.e., non-faulty) DLGs and one or more faulty DLGs with identified faults. For each DLG, the corresponding DLG data 722 may include information about the configuration and operating status of the DLG. In some examples, the DLG data 722 may include information as described above with reference to FIG. 6A to FIG. 6COne or more DLG metrics discussed, such as one or more statistical or cumulative DLG current metrics, one or more of the statistical or cumulative DLG deviation position event counts, or one or more of the statistical or cumulative DLG alarm counts. In some examples, the values ​​of the DLG metrics can present trends over a specified time period (e.g., 20 to 30 days) to form a metric sequence (e.g., a time series). Such DLG metric sequences corresponding to multiple DLG metrics can be generated for normal and faulty DLGs, respectively, to form training data 720. In an example, a specified number (N) of DLG metrics (including one or more of the statistical or cumulative DLG current indicators, statistical or cumulative DLG deviation position event counts, or statistical or cumulative DLG alarm counts as discussed above) can each present trends over a specified number of days (M). The training data 720 can be represented by a two-dimensional (M×N) data matrix. Such as referring to Fig. 9 , an example of constructing a training data set 720 using the DLG metric trend is discussed below.

[0124] The expected result 724 may include the designation of a normal DLG and a faulty DLG, such as a fault label indicating the presence / absence of a fault and the type of fault. The fault label may be provided by a human expert. In the example, a normal DLG is labeled "0", and the faulty DLGs each have a digital fault label indicating the type of fault, such as "1" representing a DLG brake fault, "2" representing a DLG circuit board fault, "3" representing a mixed fault involving both the brake and the circuit board, "4" representing a drive motor fault, "5" representing a DLG linear slide fault (e.g., a lubricated slide), or "6" representing a fault type such as a DLG coupling fault (e.g., a loose coupling).

[0125] The deep learning (DL) model training process 708 includes a training data preprocessing and selection module 731. Training data preprocessing can, in particular, alleviate certain defects of the training data. One of such defects is the data imbalance between the normal (fault-free) DLG data set (also referred to as "negative" samples) and the faulty DLG data set (also referred to as "positive" samples) in the training set 720. In an example, the ratio of negative samples to positive samples can be 6:1 to 10:1. Data imbalance may affect the performance of the trained deep learning model in predicting, detecting and diagnosing faults. Data preprocessing 731 can use one or more techniques (such as undersampling, oversampling, changing the penalty weights of positive and negative samples, enhancing the penalty term in the cost function of the deep learning model, and model integration) to alleviate data imbalance. According to various embodiments discussed in this document, oversampling, penalty weight adjustment of positive and negative samples, and model integration are used to preprocess the training data. These techniques can retain the information in the original data as much as possible. In an example, different penalty weights are assigned to positive and negative samples based at least on the sample size. In an example, the penalty weight can be inversely proportional to the sample size. For example, larger weights may be assigned to positive samples (corresponding to DLG failures) in the training set that typically have smaller sample sizes, while smaller weights may be assigned to negative samples (corresponding to normal DLGs) in the training set that typically have larger sample sizes.

[0126] A data enhancement process may be used additionally or alternatively to mitigate data imbalance to increase the number of positive samples associated with one or more DLG faults. For example, the data enhancement process may include exchanging DLG-Y1 data associated with identifying a brake fault with DLG-Y2 data. Through data exchange, the positive samples associated with the brake fault can be doubled. Additionally or alternatively, a synthetic minority oversampling technique (SMOTE) may be used to synthesize additional minority samples (positive samples or DLG data associated with identifying a fault) based on existing actual measured minority samples, and thus reduce the degree of data imbalance. SMOTE thinks from the perspective of existing minority samples and synthesizes new samples at a certain distance (in feature space) from them toward one of their neighbors. Clustering methods can be used to generate new samples to prevent model overfitting, which may be caused by simple oversampling of existing samples.

[0127] Another defect of training data is the inconsistent sample distribution between the training set and other sets (such as test sets or actual prediction sets). Inconsistent data distribution may lead to a trained model that performs well in the test set but cannot provide satisfactory prediction accuracy in the actual prediction set. In order to overcome the inconsistency of sample distribution between different data sets, a transfer learning module can be added to the deep learning model. Transfer learning is a method in which a model developed for a task is reused as the starting point of a model for different but related problems. Through model reuse, transfer learning can greatly speed up the model training process and improve performance when modeling the second problem. According to various embodiments, transfer learning can be applied to a trained deep learning model, such as an N-layer CNN or LSTM model trained on training data. The model parameters of the first N-1 layers can then be frozen, and a new data set (e.g., a prediction set different from the training set used to train the CNN or LSTM model) can be used to adjust the parameters of only the last layer (the Nth layer), and the last layer can be a fully connected layer. In an example, the prediction set used to train the last fully connected layer may include DLG data collected at a different time (e.g., a different date) from the training data.

[0128] like Fig. 7A As shown in , the pre-processed DLG data can be fed into a deep learning model to generate an estimated result 732, such as a decision whether a fault is present or not, or a fault type that determines whether a fault is present or not. The fault detection and diagnosis estimated at 732 can then be compared to the expected result 724 (e.g., "0" represents no fault, and non-zero numbers represent different identified fault types). The difference or estimated error can be calculated at 733. At 734, the estimated error can be compared to a model convergence or training stop criterion (such as a minimum value that persists for a specified number of training iterations).

[0129] If it is determined at 734 that convergence or training stopping criteria have not been met, the estimated error can be used to update the parameters of the deep learning model (e.g., layer node weights and biases), such as by backpropagation, to reduce or minimize the estimated error or error of the machine parameters during subsequent training trials. Another batch of training data can then be selected from the training data 720 and the expected results for another iteration of deep learning model training.

[0130] In an embodiment, the model parameter update using the estimation error may be performed to minimize or reduce a cost function (or objective function, or loss function), such as the cost function given in equation (1):

[0131] J(θ * ) = argmin Θ ||YY* || 2 (1)

[0132] where Y can represent machine parameters or fault detection or classification determined by a deep learning model, where Y * can represent known machine parameters or true value fault identification, and where Θ * It can be expressed as corresponding to Y and Y * The minimum squared error between the model parameters of the deep learning model (such as the layer node weights and biases as described above).

[0133] After updating the parameters of the deep learning model, the iteration index may be incremented by one. The iteration index may correspond to the number of times the parameters of the deep learning model have been updated. A convergence or training stop criterion may be checked at 734. In an embodiment, the convergence or stop criterion may include the value of the iteration index (e.g., the convergence or stop criterion may include whether the iteration index is greater than or equal to the maximum number of iterations determined). In an embodiment, the convergence or stop criterion may include the accuracy of the output set of machine parameters or the accumulated estimated error (e.g., the convergence or stop criterion may include whether the accumulated difference between the estimated DLG fault and the true value of the identified DLG fault across the batch of training data is less than a threshold). In an example, the threshold may correspond to the asymptotic minimum of all errors determined.

[0134] If it is determined at 734 that the convergence or stopping criteria have been met, the trained DL model 740 can be saved in a memory such as the memory device 116 of the data processing device 112, and the training can be stopped. The trained model 740 can be used during the testing or inference phase 712 to predict and diagnose DLG failures using test DLG data 742 collected from the target radiation therapy system and generate prediction results 744.

[0135] Deep learning models such as CNN models or RNN models (e.g., LSTM models) as discussed above may involve many model parameters, including, for example, sliding window length, the number of network layers, the number of hidden neurons, and the discard rate, etc. In order to improve fault diagnosis, a grid search method can be used to optimize one or more model parameters. Grid search, also known as parameter scanning, is a process that scans through a subset of the hyperparameter space of a manually specified learning algorithm and builds a model on each parameter combination. The grid search iterates over each parameter combination and stores the model for each combination, evaluates the performance metrics of each parameter combination, and determines the best parameter combination for a given model.

[0136] Compared with other hyperparameter optimization methods (such as genetic algorithm, random search algorithm, particle swarm optimization, Bayesian algorithm, etc.), the grid search method is easy to implement and is configured for parallel computing, and can meet the needs of fault detection and classification based on sequential DLG data.

[0137] In some examples, a two-step approach can be used to train a deep learning model. In the first step, a fault detection model is trained to detect the presence or absence of a fault. In the subsequent second step, a fault diagnosis model is trained to classify each of the faults detected in the first step into one of the fault categories. Because the first step essentially groups all positive samples of different fault types together, the two-step approach can advantageously help reduce the impact of sample size imbalance and improve FDD accuracy.

[0138] In some examples, the trained model 740 can be evaluated before it is used to predict and diagnose faults using the test DLG data 742. The trained model 740 can be applied to a validation set, which can include a set of DLG data samples, such as those shown in the attached Fig. 9 . The validation data may be different from the training data 720. Each sample in the validation set may be identified by the model 740 as one of a true positive (TP), a false positive (FP), a true negative (TN), and a false negative (FN). TP represents a DLG data sample in which the true value is positive (e.g., the DLG is faulty or has a specific fault type Fx) and the DL model predicts the same fault state or fault type Fx. FN represents a DLG data sample in which the true value is positive and the DL model predicts the opposite (e.g., the DLG is actually faulty but the model predicts no fault, or the DLG has a specific fault type Fx but the model predicts a different fault type Fy). FP represents a DLG data sample in which the true value is negative but the model predicts the opposite (e.g., the DLG is normal but the model predicts a fault, or the DLG has no fault Fx but the model predicts Fx). TN represents a DLG data sample in which the true value is negative (the DLG is normal, or has no fault Fx) and the DL model predicts the same.

[0139] Using the counts of TP, FN, FP, and TN samples identified from the validation set, one or more performance metrics may be generated. In an example, a performance metric may include a precision (also referred to as a positive prediction value) defined as TP / (TP+FP). Precision represents the fraction (e.g., percentage) of faulty DLGs detected by all models of all types of faults, or the fraction of faulty DLGs detected by all models of a specific type of fault type that actually have the specific fault type. Another example of a performance metric may include a recall (also referred to as sensitivity) defined as TP / (TP+FN). Recall represents the fraction of all faulty DLGs of all types identified as faults by the model, or the fraction of all faulty DLGs of a specific type identified by the model as having the specific fault type. Additionally or alternatively, a performance metric may include an F1 score defined as a balance of the harmonic mean of accuracy and recall as given in the following formula. The F1 score ranges between 0 and 1, with F1 reaching its best value (perfect precision and recall) at 1 and being worst at 0.

[0140]

[0141] Another performance measure is the Matthews correlation coefficient (MCC) defined as follows: The MCC is the correlation coefficient between the observed binary classification and the predicted binary classification and takes values ​​between -1 and 1.

[0142] The closer the value is to 1, the more accurate the model predictions are.

[0143]

[0144] In various examples, one or more performance metrics may be compared to a preset performance criterion (such as a performance threshold). If the performance metric determined based on the validation set meets the performance criterion, the rained DL model 740 may be used to predict and diagnose faults using the test DLG data 742; otherwise, one or more parameters of the model 740 may be adjusted, or the model 740 may be retrained using additional training DLG data.

[0145] In some examples, MCC can be used to modify the cost function used to train a DL model, such as the cost function given in equation (1) above. Adding a penalty for the imbalance of positive and negative samples in the cost function can not only improve the accuracy of the trained model, but also reduce the impact of data imbalance. The penalty term based on MCC (scaled by the penalty weight λ) is introduced into the cost function to determine the loss. Equation (4) is an example of a loss function that includes a base loss component based on binary cross entropy and an additional penalty term based on MCC:

[0146] Loss = Binary Cross Drop + λ*(1-mcc) (4) In some examples, process 700A can be used to train a DL model to classify a DLG fault into one of a plurality of fault severity levels. The fault severity can be based on the trend of the DLG metric. For example, if the DLG metric value exceeds a certain threshold, the fault is identified as a "serious" fault, or if the DLG metric value is below the certain threshold, the fault is identified as a "minor" fault. To train a deep learning model, such as according to process 700A, a training data set can be constructed such that each training sample (e.g., as shown in the attached Fig. 9 The DLG data sequence shown in FIG. 742 is assigned a severity level. The severity level, which may be part of the expected result 724, may take a categorical value or a numerical value. The trained model 740 may be used to classify the faults detected from the test DLG data 742 as minor faults or major faults. Fig. 9 An example of constructing a training data set 720 and assigning corresponding severity levels to training samples is discussed.

[0147] In some examples, the fault diagnosis may include a predicted time to failure, also referred to as remaining useful life (RUL). The RUL represents the time interval from the moment of failure of the machine component being predicted to when the failure actually occurs and requires maintenance or replacement. Information about the RUL can help maintenance personnel plan maintenance, optimize operational efficiency, and avoid unplanned downtime. During model training, such as according to process 700A, the training data set can be constructed so that a RUL value can be determined for each training sample (e.g., as shown in the attached figure). Fig. 9 ). In an example, the RUL value may be represented by the number of days between the time of the training sample and the time when the DLG metric value reaches a specified threshold. The RUL value, which may be part of the expected result 724, may be used to train a deep learning model that predicts the RUL of the test DLG 742. Fig. 9 An example of constructing a training data set 720 and determining corresponding RUL values ​​for training samples is discussed.

[0148] Although this document focuses on deep learning-based DLG fault detection and diagnosis, this is by way of example and not limitation. As will be appreciated by those skilled in the art, the systems, devices, and processes according to the various examples discussed in this document can be used to detect and diagnose faults associated with other parts or components of a radiotherapy system (e.g., MLC).

[0149] Figure 8800 is a diagram showing an exemplary architecture of an FDD deep learning model that combines two deep learning models, also referred to as model fusion (of sub-models). In an example, the fused sub-models may be models of the same type with at least one different model parameter (e.g., two CNN models with different model structures). In another example, the fused sub-models may be models of different types (e.g., a CNN combined with an LSTM model). Because different deep learning models have different expressive capabilities after training, the integrated model may combine the advantages of the sub-models, thereby improving the overall fault prediction and diagnosis performance.

[0150] like Figure 8 The example shown in includes a multi-layer deep learning model, which includes an input layer 810, a learning layer 820 and an output layer 830. The learning layer 820 includes two different DL models, or DL ​​models of the same type but different parameters. As an example, the model fusion indicated by "first model + second model" may include LSTM+LSTM, LSTM+CNN, or CNN+CNN. In the example shown, the first model 821 includes three CNN layers (e.g., three one-dimensional convolutional layers or "CONV1D" layers). Each CNN layer includes a rectified linear unit (ReLU) activation function and batch normalization (BN). ReLU activation can help improve the feature representation of the CNN model. The BU operation can alleviate the problem of internal covariate shift. The CONV-BN-ReLU sequence can improve the performance of the CNN network. The second model 822 includes two LSTM layers, with a dropout layer between each layer to prevent the LSTM layer from overfitting. It should be noted that Figure 8 The fusion of the two models shown in is for illustration and not limitation. In some examples, three or more models of the same or different types may be included in the fusion model, which has been contemplated by the inventors and is within the scope of this document.

[0151] The output layer 830 may include a cascade layer 831 and a softmax activation layer 832. The cascade layer 831 is a utility layer that cascades multiple outputs from the first model 821 and the second model 822 into a single output. The softmax activation layer 832 applies a softmax function to the cascade output to generate a classification output representing the presence or absence of a DLG fault or a DLG fault type. In an example, the classification output can be represented by an integer, such as "0" for no fault, "1" for a DLG brake fault, "2" for a DLG circuit board fault, "3" for a mixed fault involving both the brake and the circuit board, "4" for a drive motor fault, "5" for a DLG linear slide fault (e.g., a lubrication slide), or "6" for a DLG coupling fault, etc.

[0152] Fig. 9 is a diagram showing an example of generating training data for training an FDD deep learning model that can detect and diagnose faults associated with one or more components in a radiotherapy machine. The training data can be constructed using measurements obtained from a component such as a DLG in a linac system. Physical parameters such as the DLG current I Y1 or I Y2 ) or DLG metric measurements can be trended over time to form a data sequence 901 (e.g., a time series). The data sequence 901 can be generated by trending the physical parameter or DLG metric over time, such as FIG. 6A to FIG. 6C The time series shown in . A moving window 910 with a window size (L) of a specified number of days (e.g., 20 to 30 days) can slide along the time axis with a specified step size ΔL (e.g., 1 day). The DLG data within each data window forms a training sample. The "expected results" of the training samples, such as fault labels indicating the presence / absence of a fault, the type of fault, or the severity of a fault, can be determined based on the time position of the data window relative to one or more reference times T1, T2, Ts, or T3. T1 represents the time to start DLG fault prediction, when the window 910 begins to form a data series. T2 represents the time when the data sequence 901 crosses and exceeds the first threshold TH1, the first threshold TH1 indicating the beginning of an identified DLG fault (e.g., identified by a human expert). Ts represents the time when the data sequence 901 crosses and exceeds the second threshold TH2. T3 represents the time when the identified DLG fault is corrected or resolved. The first threshold TH1 and the second threshold TH2 can be determined by maintenance personnel based on engineering experience. In the example Fig. 9 In the example shown in , TH2 may be greater than TH1. These thresholds are set to prevent early or late alarms and may be dynamically adjusted according to the performance of the fault prediction and diagnosis.

[0153] When the window (e.g., window 910) completely drops before time T2 (i.e., the right edge of the window does not reach or exceed T2), the training samples of the window are assigned a fault label of "0" indicating that there is no DLG fault. When the moving window slides and begins to cross time T2 (e.g., window 920), the training samples of the window are assigned a non-zero fault label (e.g., "1") indicating that there is a DLG fault. When the moving window slides and begins to cross time T3 (e.g., window 930), the training samples of the window are assigned a label of "0" indicating that the DLG fault has been corrected. The window sliding process then continues until the left edge of the window reaches time T3 (e.g., window 940). At this point, the window sliding process stops. The data sequence thus generated (including windows 910 to 940 and intermediate windows) together with the fault labels associated with the training samples form a training data set. The training data set can be used to train as described above with reference to Fig. 7A The discussed deep learning model for detecting and diagnosing DLG faults.

[0154] Different fault labels may be assigned to corresponding training samples that distinguish different fault types or different fault severities. In an example, the fault severity may be based on the signal amplitude of the data sequence 901. As discussed above, the reference time Ts is the time when the data sequence 901 crosses and exceeds the second threshold TH2. Based on the time position of the data window relative to Ts, the data window may be labeled as a minor fault or a major fault. Fig. 9 As shown in , the minor fault zone can be defined as the time interval between T2 and Ts, and the major fault zone can be defined as the time interval between Ts and T3. The data window whose right edge falls between T2 and Ts (e.g., window 920A) is assigned a minor fault label (e.g., a digital fault label "1"). The data samples within the minor fault window are below the threshold TH2. The data window whose right edge falls between Ts and T3 (e.g., window 920B) is assigned a major fault label (e.g., a digital fault label "2"). At least a portion of the data samples within the major fault window exceeds the threshold TH2. The training samples in the data window and their associated fault labels can be used to train a deep learning model to distinguish minor faults from major faults.

[0155] Fig. 9 Also shown is the determination of the time to failure or remaining useful life (RUL) for the training samples in the data window. For a data window, the RUL can be determined as the interval between the right edge of the data window and the reference time Ts. Therefore, the RUL can be determined for each data window with fault labels "0" (no fault) and "1" (minor fault). The training samples in the data window and their associated RUL values ​​can be used to train a deep learning model to predict the RUL of a given DLG.

[0156] In some examples, having a value less than a RUL threshold (RUL TH ) may be assigned a different fault label (e.g., "11") to distinguish from data windows with a RUL greater than 100% that may be assigned a different fault label (e.g., "12"). TH In the example, RUL TH is about 10 days. Compared to the fault labeled "12", the fault labeled "11" has a short RUL and therefore may pose a higher risk of future failures and require increased awareness and preventive maintenance by service personnel. The training samples in the data window and their associated fault labels (e.g., "11" and "12") can be used to train a deep learning model to distinguish minor faults with different RULs.

[0157] Fig.10 1 is a flow chart illustrating an exemplary method 1000 for detecting and diagnosing a fault in a radiotherapy machine configured to provide radiotherapy to a patient, such as a fault associated with a dynamic blade guide (DLG) in a radiotherapy system such as system 202 or system 300. Method 1000 may be embodied in computer-readable instructions for execution by one or more processors, such that the operations of method 1000 may be performed in part or in whole by functional components of data processing device 112; therefore, method 1000 is described below by way of example with reference thereto. However, in other embodiments, at least some of the operations of method 1000 may be deployed on various other hardware configurations. Therefore, method 1000 is not intended to be limited to data processing device 112, but may be implemented in whole or in part by any other components. Some or all of the operations of method 1000 may be performed in parallel, out of order, or omitted entirely.

[0158] At 1010, machine data collected from one or more normal DLGs and one or more faulty DLGs having an identified fault may be received, such as by the data processing device 112. The received machine data may indicate the configuration and operation of the corresponding DLGs, and thus is also referred to as DLG data. In some examples, the DLG data may include the information described above with reference to FIG. 6A to FIG. 6COne or more DLG metrics discussed, such as one or more statistical or cumulative DLG current metrics, one or more of the statistical or cumulative DLG deviation position event counts, or one or more of the statistical or cumulative DLG alarm counts. Together with the DLG data received at 1010 are fault labels for normal DLGs and faulty DLGs, which can be provided by a human expert. In the example, the normal DLG has a fault label "0", and the faulty DLGs each have a fault label representing a specific fault type, such as "1" representing a DLG brake fault, "2" representing a DLG circuit board fault, "3" representing a mixed fault involving both the brake and the circuit board, "4" representing a drive motor fault, "5" representing a DLG linear slide fault (e.g., a lubricated slide), or "6" representing a fault type such as a DLG coupling fault (e.g., a loose coupling).

[0159] At 1020, a training data set may be constructed using machine data received from both a normal DLG and a faulty DLG. In some examples, the values ​​of the DLG metrics may present trends over a specified time period to form a metric sequence (e.g., a time series), and may be generated, such as according to Fig. 9 The process shown in FIG. 1 uses a moving data window to extract training samples. For each training sample, a fault label can be provided, such as by a human expert via a user interface, and assigned to the training sample. The fault label can take a numerical or categorical value and represent the presence / absence of a fault, the type of fault, or the severity of a fault. In some examples, the fault label can be determined based on the temporal position of the data window relative to one or more reference times, such as Fig. 9 As shown in . The training samples and the corresponding fault labels form input-output training pairs.

[0160] The training data set can be processed to mitigate certain defects, such as data imbalance between a normal (fault-free) DLG data set (also referred to as "negative" samples) and a faulty DLG data set (also referred to as "positive" samples). Various techniques can be used to process the training data set, including oversampling, penalty weight adjustment of positive and negative samples, and model integration for preprocessing the training data. In an example, different penalty weights are assigned to positive and negative samples based at least on the sample size. The penalty weight can be inversely proportional to the sample size. Data imbalance can be additionally or alternatively mitigated using a data augmentation process or a synthetic minority oversampling technique (SMOTE) to synthesize additional minority samples. Additionally or alternatively, in order to overcome the inconsistency of sample distribution between different data sets, a transfer learning module can be added to the deep learning model.

[0161] At 1030, the constructed training data set may be used, such as using Fig. 7AThe deep learning model is trained by the process shown in . The trained deep learning model may include a convolutional neural network (CNN), a recurrent neural network (RNN), a deep belief network (DBN), or a hybrid neural network including two or more neural network models of different types or different model configurations. Figure 7B and Figure 7C Examples of a CNN model and a Long Short-Term Memory (LSTM) network (a type of RNN model) are shown.

[0162] During model training, model parameters may be updated to minimize or reduce a cost function. The cost function may include an estimated error term (e.g., the sum of squared error terms) such as given in equation (1). The cost function may additionally include a penalty term. In an example, the penalty term is an L2 norm equal to the square of the magnitude of the feature weights in the neural network model and is used to control model complexity. In another example, the penalty term may be based on the Matthews correlation coefficient (MCC) such as given by equation (4). Convergence or training stop criteria may be checked during model training. If it is determined that the convergence or stop criteria have been met, training may be stopped. The trained model may be stored in a memory. In some examples, the trained model may be evaluated and one or more performance metrics may be generated, which may include precision, recall, balanced F1 score, or MCC, etc.

[0163] At 1040, machine data indicative of a configuration and operation of a DLG in a target radiation therapy machine may be received. DLG metrics may be generated from the machine data. At 1050, the trained deep learning model may be applied to the DLG metric data corresponding to the DLG of the target radiation therapy machine. Depending on the fault labels assigned to the training samples in the training data set, the trained model may be used to detect (including early detection or prediction) the presence or absence of a fault, classify the fault into one of a plurality of fault types, or predict a fault severity level.

[0164] At 1060, the presence or absence of a DLG fault may be detected using the trained model. If a DLG is detected, the model may output a fault diagnosis. Additionally or alternatively, a fault diagnosis may be performed, for example, by classifying the fault as one or more of a DLG brake fault, a DLG drive circuit board fault, a DLG drive motor fault, a DLG slide rail fault, or a DLG coupling unit fault. The fault detection and diagnosis results may be presented to a user. The user may review the fault diagnosis and take action (such as performing other tests), or perform necessary repairs or other preventive or corrective measures.

[0165] In some examples, for each training sample, a time to failure or remaining useful life (RUL) may be determined for the training sample, for example as described above with reference to Fig. 9 As discussed. A deep learning model can be trained to predict the RUL of DLG in a targeted radiotherapy machine.

[0166] Fig.11 A block diagram of an embodiment of a machine 1100 is shown, on which one or more of the methods discussed herein can be implemented. In one or more embodiments, one or more items of a data processing device 112 can be implemented by the machine 1100. In an alternative embodiment, the machine 1100 operates as an independent device, or can be connected (e.g., networked) to other machines. In one or more embodiments, the data processing device 112 may include one or more items of the items of the machine 1100. In a networked deployment, the machine 1100 can operate in the capacity of a server or client machine in a server-client network environment, or operate as a peer machine in a peer-to-peer (or distributed) network environment. The machine may be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a cellular phone, a network device, a network router, a switch or a bridge, or any machine capable of (sequentially or otherwise) executing instructions that specify actions to be taken by the machine. Further, while a single machine is illustrated, the term "machine" shall also be taken to include any collection of machines that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methodologies discussed herein.

[0167] The example machine 1100 includes a processing circuit system 1102 (e.g., a CPU, a graphics processing unit (GPU), an ASIC, a circuit system (e.g., one or more transistors, resistors, capacitors, inductors, diodes, logic gates, multiplexers, buffers, modulators, demodulators, radios (e.g., transmitting or receiving radios or transceivers)), a sensor 1121 (e.g., a transducer that converts one form of energy (e.g., light, heat, electricity, mechanical or other energy) into another form of energy), the like, or a combination thereof), a main memory 1104, and a static memory 1106, which communicate with each other via a bus 1108. The machine 1100 (e.g., a computer system) may also include a video display unit 1110 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)). The machine 1100 also includes an alphanumeric input device 1112 (e.g., a keyboard), a user interface (UI) navigation device 1114 (e.g., a mouse), a disk drive or mass storage unit 1116, a signal generating device 1118 (e.g., a speaker), and a network interface device 1120.

[0168] The disk drive unit 1116 includes a machine-readable medium 1122 on which is stored one or more sets of instructions and data structures (e.g., software) 1124 that embody or are utilized by any one or more of the methods or functions described herein. During execution of the instructions 1124 by the machine 1100, the instructions 1124 may also reside, in whole or in part, within the main memory 1104 and / or within the processor 1102, the main memory 1104 and the processor 1102 also constituting machine-readable media.

[0169] The machine 1100 as shown includes an output controller 1128. The output controller 1128 manages the flow of data to and from the machine 1100. The output controller 1128 is sometimes referred to as a device controller, where software that interacts directly with the output controller 1128 is referred to as a device driver.

[0170] Although in one embodiment the machine-readable medium 1122 is shown as a single medium, the term "machine-readable medium" may include a single medium or multiple media (e.g., centralized or distributed databases, and / or associated caches and servers) storing one or more instructions or data structures. The term "machine-readable medium" should also be considered to include any tangible medium capable of storing, encoding, or carrying instructions or capable of storing, encoding, or carrying data structures that are executed by a machine and cause the machine to perform any one or more of the methods of the present disclosure, which data structures are utilized by or associated with such instructions. Therefore, the term "machine-readable medium" should be considered to include, but are not limited to, solid-state memory and optical and magnetic media. Specific examples of machine-readable media include non-volatile memory, which includes, by way of example: semiconductor memory devices, such as erasable programmable read-only memory (EPROM), EEPROM, and flash memory devices; magnetic disks, such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0171] The instructions 1124 may also be transmitted or received over a communication network 1126 using transmission media. The instructions 1124 may be transmitted using a network interface device 1120 and any of a number of well-known transmission protocols (e.g., HTTP). Examples of communication networks include LANs, WANs, the Internet, mobile phone networks, plain old telephone (POTS) networks, and wireless data networks (e.g., WiFi and WiMax networks). The term "transmission medium" shall be deemed to include any intangible medium that is capable of storing, encoding, or carrying instructions for execution by a machine and includes digital or analog communications signals, or other intangible media that facilitates the communication of such software.

[0172] As used herein, "communicatively coupled between" means that entities on either side of the coupling must communicate through an item therebetween, and the entities cannot communicate with each other without communicating through the item.

[0173] Additional Notes

[0174] The above detailed description includes reference to the accompanying drawings, which form a part of the detailed description. The accompanying drawings show specific embodiments by way of illustration rather than by way of limitation, in which the present disclosure can be put into practice. These embodiments are also referred to as "examples" in this article. These examples may include elements other than those shown or described. However, the inventors have also contemplated providing examples of those elements that are only shown or described. In addition, with respect to a particular example (or one or more aspects thereof) or with respect to other examples (or one or more aspects thereof) shown or described in this article, the inventors have also contemplated using examples of any combination or permutation (or one or more aspects thereof) of those elements shown or described.

[0175] All publications, patents, and patent documents referenced in this document are incorporated herein by reference in their entirety, as if individually incorporated by reference. If there are inconsistent usages between this document and those documents incorporated by reference, the usage in the incorporated reference(s) should be considered supplementary to that of this document; for conflicting inconsistencies, the usage in this document controls.

[0176] In this document, when introducing elements of various aspects of the present disclosure or embodiments thereof, the terms "a", "an", "the", and "said" are used, as is common in patent documents, to include one or more than one or more of the elements, independent of any other instances or uses of "at least one" or "one or more". In this document, the term "or" is used to indicate a non-exclusive, or, such that "A or B" includes "A but not B", "B but not A", and "A and B" unless otherwise specified.

[0177] In the appended aspects, the terms "including" and "in which" are used as the plain-English equivalents of the respective terms "comprising" and "wherein." Furthermore, in the appended aspects, the terms "comprising," "including," and "having" are intended to be open-ended, meaning that there may be additional elements in addition to the listed elements, such that anything following such a term (e.g., comprising, including, having) in an aspect is still considered to fall within the scope of that aspect. Furthermore, in the appended aspects, the terms "first," "second," and "third," etc. are used merely as labels, and are not intended to impose numerical requirements on their objects.

[0178] The embodiments of the present disclosure can be implemented with computer executable instructions.Computer executable instructions (such as software code) can be organized into one or more computer executable components or modules.Any number of such components or modules and any organization of such components or modules can be used to implement the various aspects of the present disclosure.For example, the various aspects of the present disclosure are not limited to the specific computer executable instructions or specific components or modules shown in the drawings and described in this article.Other embodiments of the present disclosure may include different computer executable instructions or components with more or less functions than those shown and described herein.

[0179] The method examples (e.g., operations and functions) described herein may be at least partially machine or computer implemented (e.g., implemented as software code or instructions). Some examples may include a computer-readable medium or machine-readable medium encoded with instructions that can be operated to configure an electronic device to perform the method as described in the above examples. The implementation of such a method may include software code such as microcode, assembly language code, high-level language code, etc. (e.g., "source code"). Such software code may include computer-readable instructions (e.g., "objects" or "executable code") for performing various methods. The software code may form part of a computer program product. The software implementation of the embodiments described herein may be provided via an article having codes or instructions stored thereon, or via a method of operating a communication interface to send data via a communication interface (e.g., wirelessly, via the Internet, via satellite communications, etc.).

[0180] In addition, the software code may be tangibly stored on one or more volatile or non-volatile computer-readable storage media during execution or at other times. These computer-readable storage media may include any mechanism for storing information in a form that can be accessed by a machine (e.g., a computing device, an electronic system, etc.), such as, but not limited to, a floppy disk, a hard disk, a removable disk, any form of disk storage media, a CD-ROM, a magneto-optical disk, a removable optical disk (e.g., a compact disc and a digital video disk), a flash memory device, a cassette, a memory card or stick (e.g., a secure digital card), a RAM (e.g., a CMOS RAM, etc.), a recordable / non-recordable medium (e.g., a read-only memory (ROM)), an EPROM, an EEPROM, or any type of medium suitable for storing electronic instructions, etc. Such a computer-readable storage medium is coupled to a computer system bus so that it can be accessed by the processor and other parts of the OIS.

[0181] In an embodiment, the computer readable storage medium may have an encoded data structure for treatment planning, wherein the treatment plan may be adaptive. The data structure for the computer readable storage medium may be at least one of a Digital Imaging and Communications in Medicine (DICOM) format, an extended DICOM format, an XML format, and the like. DICOM is an international communication standard that defines a format for transmitting data related to medical images between various types of medical devices. DICOMRT refers to a communication standard specific to radiation therapy.

[0182] In various embodiments of the present disclosure, the methods of creating components or modules can be implemented in software, hardware, or a combination thereof. For example, the methods provided by various embodiments of the present disclosure can be implemented in software using standard programming languages ​​such as, for example, C, C++, Java, Python, etc., and combinations thereof. As used herein, the terms "software" and "firmware" are interchangeable and include any computer program stored in a memory for execution by a computer.

[0183] A communication interface includes any mechanism that interfaces with any of a hardwired medium, a wireless medium, an optical medium, etc. to communicate with another device, such as a memory bus interface, a processor bus interface, an Internet connection, a disk controller, etc. The communication interface may be configured by providing configuration parameters and / or sending signals to prepare the communication interface to provide data signals describing the software content. The communication interface may be accessed via one or more commands or signals sent to the communication interface.

[0184] The present disclosure also relates to a system for performing the operations herein. The system can be constructed specifically for the desired purpose, or can include a general-purpose computer selectively activated or reconfigured by a computer program stored in a computer. Unless otherwise specified, the operation or execution order of the operations in the embodiments of the present disclosure shown and described herein is not necessary. That is, unless otherwise specified, operations can be performed in any order, and compared with those operations disclosed herein, the embodiments of the present disclosure can include other operations or fewer operations. For example, it is contemplated that running or performing a specific operation before another operation, simultaneously with another operation, or after another operation falls within the scope of the various aspects of the present disclosure.

[0185] In view of the above, several purposes of the present disclosure can be achieved and other favorable results are obtained. Having described the various aspects of the present disclosure in detail, it will be apparent that modifications and variations are feasible without departing from the scope of the various aspects of the present disclosure as defined in the attached various aspects. Since various changes can be made in the above constructions, products and methods without departing from the scope of the various aspects of the present disclosure, it is intended that all contents contained in the above description and shown in the accompanying drawings should be interpreted as illustrative, rather than in a restrictive sense.

[0186] The above description is intended to be illustrative, rather than restrictive. For example, the above examples (or one or more aspects thereof) can be used in combination with each other. In addition, without departing from the scope of the present disclosure, many modifications can be made to adapt specific situations or materials to the teachings of the present disclosure. Although the size and type of the material and coating described herein are intended to limit the parameters of the present disclosure, they are by no means restrictive, but exemplary embodiments. After reviewing the above description, many other embodiments are obvious to those skilled in the art. Therefore, the scope of the present disclosure should be determined with reference to the full range of equivalents to which the attached various aspects and such various aspects are given.

[0187] In addition, in the above specific embodiments, various features can be combined together to simplify the present disclosure. This should not be interpreted as an unexpected disclosure feature being necessary for any aspect. Instead, the subject matter of the invention may be less than all the features of a specific disclosed embodiment. Therefore, the attached various aspects are thus incorporated into the specific embodiments, with each aspect acting alone as a separate embodiment. The scope of the present disclosure should be determined with reference to the attached various aspects and the full range of equivalents to which such aspects are assigned. In addition, the limitations of the attached various aspects are not written in a device plus function format, and are not intended to be interpreted based on the sixth paragraph of 35 U.S.C. §112, unless such limitations of various aspects are explicitly used with the phrase "device for..." followed by a functional statement without further structure.

[0188] The Abstract is provided to comply with 37 CFR §1.72(b), to allow the reader to quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the aspects.

[0189] According to the above description, the embodiments of the present invention also disclose the following technical solutions, including but not limited to:

[0190] Solution 1. A computer-implemented method for detecting and diagnosing faults in a radiation therapy machine, the method comprising:

[0191] receiving machine data indicating configuration and operation of components in a targeted radiation therapy machine;

[0192] Applying a trained deep learning model to the received machine data of a component in a target radiotherapy machine, the trained deep learning model being trained to establish a relationship between: (1) machine data collected from normal components and faulty components in the corresponding radiotherapy machine, and (2) fault information of the normal components and the faulty components, the normal components and the faulty components being of the same type as the components in the target radiotherapy machine; and

[0193] Faults associated with components in the targeted radiation therapy machine are detected and diagnosed.

[0194] Solution 2. The method according to Solution 1 further includes:

[0195] receiving machine data collected from the normal components and the faulty components having corresponding identified faults, the machine data indicating configuration and operation of the corresponding components;

[0196] constructing a training data set comprising a plurality of data sequences generated from the received machine data of normal components and faulty components; and

[0197] The deep learning model is trained using the constructed training data set to establish the trained deep learning model.

[0198] Option 3. The method of Option 2, wherein the component in the targeted radiation therapy machine comprises a dynamic blade guide (DLG), the normal component comprises a normal DLG, and the faulty component comprises a faulty DLG having a corresponding identified DLG fault, and

[0199] Wherein, detecting and diagnosing the fault includes detecting and diagnosing a DLG fault in the targeted radiation therapy machine.

[0200] Solution 4. The method according to Solution 3, wherein training the deep learning model comprises:

[0201] applying corresponding penalty weights to one or more of the plurality of data sequences in the training data set; and

[0202] The deep learning model is trained using the constructed training data set, where the constructed training data set includes weighted data sequences.

[0203] Solution 5. The method according to any one of Solution 3 to Solution 4, wherein the trained deep learning model comprises one or more of the following:

[0204] Convolutional Neural Networks (CNN);

[0205] Recurrent Neural Networks (RNNs);

[0206] Long short-term memory (LSTM) networks;

[0207] Deep Belief Network (DBN); or

[0208] Transfer learning module.

[0209] Embodiment 6. The method according to any one of embodiments 3 to 5, comprising: generating the plurality of data series comprising trends of DLG current measurements over time, the DLG currents being measured from one or more DLGs at corresponding axes, respectively.

[0210] Embodiment 7. The method according to embodiment 6, wherein the DLG current trend comprises one or more of the following:

[0211] Trends in daily average current;

[0212] Trends in daily changing current;

[0213] Trends in daily maximum current;

[0214] The multi-day moving average of the current trend.

[0215] Option 8. The method according to any one of Options 3 to 7 comprises: generating the plurality of data series including trends of DLG position metrics over time, and calculating the DLG position metrics for one or more DLGs at corresponding axes respectively.

[0216] Solution 9. The method according to Solution 8, wherein the DLG position metric comprises a count of DLG deviation-from-position events occurring within a specific time period, and the DLG position trend comprises one or more of the following:

[0217] Trends in daily counts of deviations from location events; or

[0218] The trend of the cumulative count of out-of-position events over the specified number of days.

[0219] Solution 10. The method according to any one of solutions 3 to 9, comprising generating the plurality of data series including trends of alarms triggered by one or more alarm events, wherein the trends of the alarms include one or more of the following:

[0220] Trends in daily counts of alerts; or

[0221] Trend of the cumulative count of alerts over the specified number of days.

[0222] Solution 11. The method according to any one of solutions 3 to 10, wherein constructing the training data set includes assigning a fault type to each of the plurality of data sequences, and wherein diagnosing a DLG fault in the targeted radiation therapy machine includes classifying the DLG fault into one or more fault types, the fault types comprising:

[0223] DLG brake failure;

[0224] DLG driver circuit board failure;

[0225] DLG drive motor failure;

[0226] DLG rail failure; or

[0227] DLG coupling unit fault.

[0228] Option 12. A method according to any one of Options 3 to 11, wherein constructing the training data set includes assigning a corresponding fault severity level to each of the multiple data sequences, and wherein diagnosing a DLG fault in the target radiotherapy machine includes classifying the DLG fault into one of multiple fault severity levels.

[0229] Solution 13. A method according to any one of Solutions 3 to 12, wherein training the deep learning model comprises: determining a corresponding remaining useful life (RUL) for each of the multiple data sequences; and establishing a relationship between the multiple data sequences and the corresponding determined RUL; and

[0230] The method includes using a trained deep learning model to predict the RUL of a DLG in the targeted radiation therapy machine.

[0231] Option 14. A method according to any one of Options 3 to 13, wherein training the deep learning model includes adjusting one or more model parameters to minimize a cost function, wherein the cost function includes a penalty term based on the Matthews correlation coefficient (MCC).

[0232] Embodiment 15. A system for detecting and diagnosing a fault in a radiation therapy machine configured to provide radiation therapy to a patient, the system comprising a processor configured to:

[0233] receiving machine data indicating configuration and operation of components in a targeted radiation therapy machine;

[0234] Applying a trained deep learning model to the received machine data of a component in a target radiotherapy machine, the trained deep learning model being trained to establish a relationship between: (1) machine data collected from normal components and faulty components in the corresponding radiotherapy machine, and (2) fault information of the normal components and the faulty components, the normal components and the faulty components being of the same type as the components in the target radiotherapy machine; and

[0235] Faults associated with components in the targeted radiation therapy machine are detected and diagnosed.

[0236] Embodiment 16. The system according to embodiment 15, wherein the processor comprises a training module configured to perform the following steps:

[0237] receiving machine data collected from the normal components and the faulty components having corresponding identified faults, the machine data indicating configuration and operation of the corresponding components;

[0238] constructing a training data set comprising a plurality of data sequences generated from the received machine data of normal components and faulty components; and

[0239] The deep learning model is trained by using the constructed training data set to establish the trained deep learning model.

[0240] Embodiment 17. The system of embodiment 16, wherein the component in the targeted radiation therapy machine comprises a dynamic blade guide (DLG), the normal component comprises a normal DLG, and the faulty component comprises a faulty DLG having a corresponding identified DLG fault, and

[0241] Wherein the processor is configured to detect and diagnose DLG failure in the targeted radiation therapy machine.

[0242] Option 18. A system according to Option 17, wherein the processor is configured to construct the training data set using fault information of each data sequence in the multiple data sequences, the fault information including an indication of the presence or absence of a fault, a fault type, or a fault severity level.

[0243] Embodiment 19. The system according to any one of embodiments 17 to 18, wherein the training module is configured to generate the plurality of data sequences including one or more of the following:

[0244] Trend of DLG current measurements over time;

[0245] a trend over time of a DLG position metric comprising a count of DLG out-of-position events occurring within a particular time period; or

[0246] A trend of the count of alarms triggered by one or more alarm events.

[0247] Embodiment 20. A non-transitory machine-readable storage medium, the non-transitory machine-readable storage medium comprising the following instructions, which, when executed by one or more processors of a machine, cause the machine to perform operations comprising the following steps:

[0248] receiving machine data indicating configuration and operation of components in a targeted radiation therapy machine;

[0249] Applying a trained deep learning model to the received machine data of a component in a target radiotherapy machine, the trained deep learning model being trained to establish a relationship between: (1) machine data collected from normal components and faulty components in the corresponding radiotherapy machine, and (2) fault information of the normal components and the faulty components, the normal components and the faulty components being of the same type as the components in the target radiotherapy machine; and

[0250] Faults associated with components in the targeted radiation therapy machine are detected and diagnosed.

[0251] Solution 21. The non-transitory machine-readable storage medium according to Solution 20, wherein the operation further comprises:

[0252] receiving machine data collected from the normal components and the faulty components having the identified faults, the machine data indicating configuration and operation of the respective components;

[0253] constructing a training data set comprising a plurality of data sequences generated from the received machine data of normal components and faulty components; and

[0254] The deep learning model is trained using the constructed training data set to establish the trained deep learning model.

[0255] Embodiment 22. The non-transitory machine-readable storage medium of embodiment 21, wherein the component in the targeted radiation therapy machine comprises a dynamic blade guide (DLG), the normal component comprises a normal DLG, and the faulty component comprises a faulty DLG having a corresponding identified DLG fault, and

[0256] Wherein, the option of detecting and diagnosing the fault includes detecting and diagnosing a DLG fault in the targeted radiation therapy machine.

[0257] Embodiment 23. The non-transitory machine-readable storage medium of embodiment 22, wherein the operations further comprise diagnosing a DLG fault in the targeted radiation therapy machine, wherein diagnosing the DLG fault in the targeted radiation therapy machine comprises classifying the DLG fault as one or more of the following:

[0258] DLG brake failure;

[0259] DLG driver circuit board failure;

[0260] DLG drive motor failure;

[0261] DLG rail failure; or

[0262] DLG coupling unit fault.

Claims

1. A computer-implemented method for detecting and diagnosing faults in a radiation therapy machine, the method comprising: receiving machine data indicating configuration and operation of targeted radiation therapy machine components; Applying a trained deep learning model to the received machine data of a target radiotherapy machine component, the trained deep learning model being trained to establish a relationship between: (i) machine data collected from a normal component and a faulty component having different fault severity levels, and the machine data indicating configuration and operation of the normal component and the faulty component having different fault severity levels, and (ii) fault information of the normal component and the faulty component; as well as Based on applying the trained deep learning model to the received machine data, a fault of the targeted radiation therapy machine component is detected and the detected fault is classified into one of a plurality of fault severity levels.

2. The method according to claim 1, further comprising: constructing a training data set using the machine data collected from the normal component and the faulty component having different fault severity levels; as well as The trained deep learning model is generated by training a deep learning model using the training data set.

3. The method according to claim 2, wherein: The training data set includes a plurality of data segments generated from the machine data collected from the normal component and the faulty component having different fault severity levels, each of the plurality of data segments being assigned a fault indicator indicating an absence of a fault or a fault severity level.

4. The method according to claim 3, wherein: generating the plurality of data segments by applying a moving window to the time series of machine data, Each of the plurality of data segments corresponds to a time window and is assigned the fault indicator based on a time position of the time window relative to one or more reference times.

5. The method according to claim 4, wherein: The time windows of any two adjacent data segments among the plurality of data segments at least partially overlap in time.

6. The method according to claim 4, wherein: Any one of the plurality of data segments is assigned one of the following: If the corresponding time window is outside the range defined between the first reference time and the second reference time, a first fault indicator indicating that no fault exists; or A second fault indicator of a fault severity level is indicated if the corresponding time window is at least partially within a range defined between the first reference time and the second reference time.

7. The method according to claim 6, in, The first reference time corresponds to a time series of the machine data exceeding a first threshold, The second reference time corresponds to a time series of the machine data being lower than a second threshold.

8. The method according to claim 7, wherein: a range defined between the first reference time and the second reference time includes a first sub-range and a second sub-range, Any one of the plurality of data segments is assigned one of the following: a first fault severity level if the corresponding time window is at least partially within the first sub-range but outside the second sub-range; or If the corresponding time window is at least partially within the second sub-range but does not exceed the second reference time, it is a second fault severity level.

9. The method according to claim 3, wherein: The fault indicator has a numerical or categorical value to indicate the absence of a fault or the severity level of the fault.

10. The method according to claim 2, wherein: The target radiation therapy machine component comprises a dynamic blade guide DLG, The training data set includes machine data collected from a normal DLG and a faulty DLG with different fault severity levels, and the machine data indicates configurations and operations of the normal DLG and the faulty DLG with different fault severity levels.

11. The method according to claim 10, wherein: The training data set includes a plurality of data segments generated from a series of measurements of DLG parameters from each of the normal DLG and the faulty DLG over time.

12. The method according to claim 11, wherein: The DLG parameter includes at least one of a DLG current or a DLG deviation position event count within a specific time period.

13. The method according to claim 1, wherein: The trained deep learning model is further trained to establish a relationship between: (i) machine data collected from normal components and faulty components with different fault severity levels, and the machine data indicates configuration and operation of the normal components and the faulty components with different fault severity levels, and (ii) remaining useful life (RUL) information of the normal components and the faulty components; and The method also includes predicting a RUL of the targeted radiation therapy machine component based on applying the trained deep learning model to the received machine data.

14. A system for detecting and diagnosing faults in a radiotherapy machine, the system comprising: A memory for storing a trained deep learning model, wherein the trained deep learning model is trained to establish a relationship between: (i) machine data collected from a normal component and a faulty component with different fault severity levels, and the machine data indicates configurations and operations of the normal component and the faulty component with different fault severity levels, and (ii) fault information of the normal component and the faulty component; as well as A processor configured to: receiving machine data indicating configuration and operation of targeted radiation therapy machine components; applying the trained deep learning model to the received machine data; as well as Based on applying the trained deep learning model to the received machine data, a fault of the targeted radiation therapy machine component is detected and the detected fault is classified into one of a plurality of fault severity levels.

15. The system of claim 14, wherein: The processor is configured to: constructing a training data set using the machine data collected from the normal component and the faulty component having different fault severity levels; as well as The trained deep learning model is generated by training a deep learning model using the training data set.

16. The system of claim 15, wherein: The training data set includes a plurality of data segments generated from the machine data collected from the normal component and the faulty component having different fault severity levels, each of the plurality of data segments being assigned a fault indicator indicating an absence of a fault or a fault severity level.

17. The system of claim 16, wherein: generating the plurality of data segments by applying a moving window to the time series of machine data, Each of the plurality of data segments corresponds to a time window and is assigned the fault indicator based on a time position of the time window relative to one or more reference times.

18. The system of claim 17, wherein: Any one of the plurality of data segments is assigned one of the following: A fault indicator indicating that there is no fault if the corresponding time window is outside the range defined between the first reference time and the second reference time; or A fault indicator of a fault severity level is indicated if the corresponding time window is at least partially within a range defined between the first reference time and the second reference time.

19. The system of claim 15, wherein: The target radiation therapy machine component comprises a dynamic blade guide DLG, The training data set includes a plurality of data segments generated from a series of measurements of DLG parameters from each of a normal DLG and a faulty DLG over time.

20. The system of claim 14, wherein: The trained deep learning model is further trained to establish a relationship between: (i) machine data collected from normal components and faulty components with different fault severity levels, and the machine data indicates configuration and operation of the normal components and the faulty components with different fault severity levels, and (ii) remaining useful life (RUL) information of the normal components and the faulty components; Wherein the processor is configured to predict the RUL of the targeted radiation therapy machine component based on applying the trained deep learning model to the received machine data.