Multi-leaf collimator prediction method, prediction system and computing device
Through the multi-leaf collimator prediction method based on volumetric intensity-modulated radiotherapy, the encoder and decoder are used to generate the motion sequence of the multi-leaf collimator, which solves the problem of low prediction efficiency of multi-leaf collimator sequence in the prior art, and realizes efficient treatment plan generation and adaptive radiotherapy.
Patent Information
- Application Number
- CN202510305762.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-14
- Publication Date
- 2025-06-13
AI Technical Summary
The lack of effective prediction of multi-leaf collimator sequences in existing radiotherapy technologies has led to low efficiency and time-consuming in traditional methods when serializing leaves, making it difficult to achieve efficient treatment plan generation.
Through a multi-leaf collimator prediction method based on volumetric intensity-modulated radiotherapy, the motion sequence of the multi-leaf collimator is encoded and decoded using an encoder, latent space rotation and control point decoder.
It effectively improves the prediction efficiency of multi-leaf collimator motion sequences, reduces the time for the generation of treatment plans, and achieves more efficient execution plan generation and adaptive radiotherapy.
Smart Images
Figure CN120148761A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of healthcare informatics, and particularly to a multi-leaf collimator prediction method, a prediction system, and a computing device. Background Art
[0002] Radiation therapy is one of the important treatment means for tumors. Volumetric Modulated Arc Therapy (VMAT) is a new radiotherapy technology developed on the basis of conventional intensity-modulated radiotherapy technology, which can rotate and irradiate tumors within any angular range set by 360° single arc or multi-arc.
[0003] The multi-leaf collimator MLC is a mechanical moving part used to generate a conformal radiation field, and the shielding of rays is controlled by the movement of the leaves. The lead door is generally made of high-density lead material because lead has a high density and radiation absorption ability, and can effectively shield radiation.
[0004] Current radiotherapy mainly focuses on predicting the dose distribution of market quality, preliminary Segment generation, or generating a fluence map of the plan; while lacking the prediction of the multi-leaf collimator sequence. Traditional methods take a lot of time in leaf serialization, and the serialization efficiency is low; it is difficult to effectively predict the sequence of related components of the multi-leaf collimator.
[0005] Therefore, a method is needed to convert the dose distribution of the treatment plan into an executable plan and predict the movement sequence of the multi-leaf collimator (MLC). Summary of the Invention
[0006] Aiming at the above technical problems existing in the prior art, the present invention provides a multi-leaf collimator prediction method, a prediction system, and a computing device, which improve the prediction efficiency of the multi-leaf collimator movement sequence.
[0007] The first aspect of the present invention discloses a multi-leaf collimator prediction method based on volumetric modulated arc radiotherapy, including the following steps: obtaining a radiotherapy plan based on volumetric modulated arc radiotherapy; encoding the radiotherapy plan and the multi-leaf collimator through an encoder to obtain a second vector; projecting the second vector based on latent space rotation to obtain a projection vector; decoding the projection vector through a decoder of control points to obtain a prediction vector; and obtaining a movement sequence according to the prediction vector.
[0008] Preferably, the multi-leaf collimator includes 240 control points, 160 leaves, and 2 lead doors.
[0009] Preferably, the network structure of the encoder includes a convolutional layer, a batch normalization layer, a pooling layer, a convolutional layer, a batch normalization layer, an activation function, a convolutional layer, and a batch normalization layer connected in sequence.
[0010] Preferably, the method for latent space rotation includes:
[0011] Project the second vector in the beam direction to obtain a third vector based on the beam direction view;
[0012] Multiply the dose of the third vector by the gradient field tensor to obtain a fourth vector;
[0013] After rotating the fourth vector along the longitudinal axis or the beam direction and passing through an activation function, a projection vector is obtained.
[0014] Preferably, the calculation method of the fourth vector is expressed as:
[0015] D′ = D ⋅ G
[0016] where D′ represents the fourth vector, D represents the dose of the voxel in the beam direction, and G represents the gradient value of the voxel in the gradient field.
[0017] Preferably, the control points are arranged with any one of the following decoders or their combinations: vane decoder, jaw decoder, and machine unit decoder.
[0018] Preferably, the network structure of the vane decoder or the jaw decoder includes a plurality of convolutional blocks with residual connections, a convolutional layer, a batch normalization layer, an activation function, a convolutional layer, and a hyperbolic tangent activation function connected in sequence.
[0019] Preferably, the network structure of the machine unit decoder includes a plurality of convolutional blocks with residual connections, an average pooling layer, and an output convolutional layer connected in sequence.
[0020] The second aspect of the present invention discloses a prediction system for implementing the above multi-leaf collimator prediction method, including an encoding module, a projection module, and a decoding module.
[0021] The encoding module is used to encode the radiotherapy plan and the multi-leaf collimator through an encoder to obtain a second vector;
[0022] The projection module is used to project the second vector based on latent space rotation to obtain a projection vector;
[0023] The decoding module is used to decode the projection vector through the decoder of the control points to obtain a prediction vector, and obtain a motion sequence through the prediction vector.
[0024] The third aspect of the present invention discloses a computing device, including a memory, and the memory stores a program, which when processed, executes the above multi-leaf collimator prediction method.
[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: converting the radiotherapy plan into a motion sequence and an executable plan effectively avoids the time waste and efficiency reduction caused by leaf serialization, and enables more efficient generation of executable plans and adaptive radiotherapy. Description of the Drawings
[0026] Figure 1 is a flowchart of the multi-leaf collimator prediction method based on volumetric modulated arc therapy of the present invention;
[0027] Figure 2 is the network structure for implementing the prediction method;
[0028] Figure 3 is a schematic diagram of the network of the encoder;
[0029] Figure 4 is a schematic diagram of the network of the latent space rotation module;
[0030] Figure 5 is a schematic diagram of the network of the leaf decoder;
[0031] Figure 6 is a schematic diagram of the network of the machine unit decoder;
[0032] Figure 7 is a schematic diagram of the network of the convolutional block;
[0033] Figure 8 is a logical block diagram of the prediction system. Detailed Embodiments
[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0035] In the present invention, it should be understood that terms such as "including" or "having" are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof disclosed in this specification, and do not exclude the possibility of the presence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0036] In addition, it should be noted that, without conflict, the embodiments in this disclosure and the features in the embodiments may be combined with each other. The present invention will be further described in detail below with reference to the accompanying drawings:
[0037] As Figure 1As shown in the figure, the first aspect of the present invention provides a multi-leaf collimator prediction method based on volumetric modulated arc therapy (VMAT), comprising the following steps:
[0038] Step 101: Obtain a radiotherapy plan based on volumetric modulated arc therapy.
[0039] Step 102: Encode the radiotherapy plan and the multi-leaf collimator through an encoder to obtain a second vector.
[0040] Step 103: Project the second vector based on latent space rotation to obtain a projection vector.
[0041] Step 104: Decode the projection vector through a decoder of control points to obtain a prediction vector, and obtain a motion sequence through the prediction vector.
[0042] The plan is divided into multiple control points. More specifically, it is divided into 240 control points (resolution of 1.5°), which is the maximum number of control points included in all plans. However, the number of control points and the resolution can be adjusted according to the actual situation. The components of each control point include: leaves (collimator), lead doors, and machine units (MUs).
[0043] Step 105: Correct the outliers of the motion sequence and generate an executable plan.
[0044] Step 106: Control the multi-leaf collimator according to the executable plan.
[0045] The present invention can convert the dose distribution of the radiotherapy plan into an executable plan, which can be recalculated and audited in an independent plan quality assurance (QA) software for dose verification, and sent to a medical linear accelerator (LINAC), effectively avoiding the time waste and efficiency reduction caused by leaf serialization, and can achieve more efficient execution plan generation and adaptive radiotherapy.
[0046] In a specific test, the executable plan can be generated within a few seconds without performing other time-consuming tasks in the treatment planning system (TPS).
[0047] In step 101, the voxel size of the original dose data in the radiotherapy plan is usually 4 mm × 4 mm × 4 mm; the slice thickness of the CT image is 2 mm, and the planar resolution is 0.81 mm. The dose data can be preprocessed: the dose data is cropped and resampled centered on the isocenter, and the final shape is S=(128, 128, 320), and the size is l=(400 mm, 400 mm, 400 mm). After linear interpolation, the voxel size of the resampled data becomes 3.125 mm × 3.125 mm × 1.250 mm.
[0048] The radiotherapy plan is processed to include 160 multi-leaf collimator (MLC) positions (2×80), two jaw positions, and the MU values at 240 control points. Linear interpolation is performed on the leaves, jaws, and MU values. For VMAT arcs that are not fully closed (less than 360°), the control points affected by them can be set to 0 MU.
[0049] Figure 2 The network structure for implementing the prediction method of the present invention is shown.
[0050] Figure 3 The network structure of the encoder is shown, including a 3D convolutional layer Conv 3D, a 3D batch normalization layer BatchNorm 3D, a 3D max pooling layer MaxPool 3D, a 3D convolutional layer Conv 3D, a 3D batch normalization layer BatchNorm 3D, a ReLU activation function, a 3D convolutional layer Conv 3D, and a 3D batch normalization layer BatchNorm 3D, which are connected in sequence. One encoder can be set for each control point. After being encoded by the encoder, the dose distribution at this control point can be obtained.
[0051] For example, the first vector input to the encoder is: [b, c, x, y], where b represents the number of patients or the patient ID, 320 represents the layer / channel of the three-dimensional vector, and x and y represent that there are x*y dose pixels in each layer. The specific first vector is [b, 320, 128, 128], which is a four-dimensional vector.
[0052] The second vector output by the encoder is: [b, f, z, x, y], where b represents the patient, f represents the multi-leaf collimator / control point, and x, y, and z represent the radiotherapy doses in the X, Y, and Z directions. The specific second vector is [b, 64, 80, 32, 32], which is a five-dimensional vector.
[0053] Figure 4 The method for latent space rotation is shown, including:
[0054] Step 301: Directionalize the second vector in the beam direction to obtain a third vector based on the beam's eye view (BEV), and the third vector represents the dose in the beam direction.
[0055] Step 302: Multiply the dose of the third vector by the gradient field tensor to obtain the simulated absorbed dose, that is, the fourth vector.
[0056] Step 303: After rotating the fourth vector along the longitudinal axis / beam direction and passing through the ReLU activation function, a projection vector is obtained. The specific projection vector is: [b, 64, 240, 80, 32], which is represented as: [number of patients, projection result in the X direction, incident angle, multi-leaf collimator, projection result in the Y direction].
[0057] The calculation method of the fourth vector is expressed as:
[0058] D′ = D˙G
[0059] where D′ represents the fourth vector, which is used to simulate the absorbed dose of the voxel, D represents the dose of the voxel in the beam direction, and G represents the gradient value of the voxel in the gradient field, that is, the weight along the beam path.
[0060] Voxels on the same side should be given greater weight than voxels on the opposite side. The generated dose projection images (one for each machine unit) are stacked to form the input array of the decoder.
[0061] In step 104, the decoder of the control point includes: a leaf decoder, a jaw decoder, and a machine unit decoder. More specifically, it includes a total of 160 leaf decoders, 2 jaw decoders, and multiple machine units. Corresponding components and their decoders can be arranged for each control point.
[0062] For the network structure of the leaf decoder, see Figure 5 , which includes multiple convolutional blocks with residual connections ProjectionResNetBlock, a 3D convolutional layer Conv 3D, a 3D batch normalization layer BatchNorm 3D, an activation function ReLU, a 3D convolutional layer Conv 3D, and a hyperbolic tangent activation function. A specific output vector of the leaf decoder is: [b, 2, 240, 80, 1]. Through eight convolutional blocks with residual connections, the data volume is gradually reduced to the shape of the expected output: MLC 240×80×2, jaw position 240×2×1; through a depthwise separable convolutional layer to reduce the model size and memory requirements
[0063] The network structure of the jaw decoder and the leaf decoder adopts the same design. A specific output vector of the jaw decoder is: [b, 2, 240, 1, 1].
[0064] As Figure 6 , the network structure of the machine unit (MU) decoder includes multiple convolutional blocks with residual connections (ProjectionResNetBlock), an adaptive average pooling layer AdaptiveAvgPool3D, and an output convolutional layer Conv3D. Specifically, 6 convolutional blocks are designed, as Figure 7 , the network structure of the convolutional block includes a depthwise separable convolution, a 3D batch normalization layer BatchNorm3D, an activation function LeakyReLU, a depthwise separable convolution, a 3D batch normalization layer BatchNorm 3D, and an activation function LeakyReLU. Among them, the batch size of batch normalization is 5.
[0065] The leaf decoder and the jaw decoder use a hyperbolic tangent output activation function; the MU decoder uses the output of the final convolutional layer.
[0066] Through an encoder, latent space rotation, and vane decoders, lead gate decoders, and machine unit decoders corresponding to control points, the present invention predicts a prediction vector of a multi-leaf collimator (MLC) and generates a motion sequence through the prediction vector. These plans are generated within seconds without other time-consuming tasks in a treatment planning system (TPS). They can be immediately recalculated and audited in a stand-alone plan quality assurance (QA) software for dose verification and sent to a medical linear accelerator (LINAC), effectively avoiding the time waste and efficiency reduction caused by the traditional method of vane serialization. Therefore, more efficient radiotherapy plan generation and adaptive radiotherapy can be achieved.
[0067] The above-mentioned motion sequence can be stored in a radiotherapy plan file. The generated radiotherapy plan meets clinical standards (in terms of PTV coverage, maximum dose, and organ at risk (OAR) dose) and has good consistency with the original dose data.
[0068] In the present invention, the radiotherapy plan can be implemented through a deep learning-based automatic planning module. The automatic planning module predicts a high-quality dose distribution based on CT images and DICOM structure sets.
[0069] The present invention also provides a prediction system for implementing the above prediction method, such as Figure 8 , including an encoding module 1, a projection module 2, and a decoding module 3.
[0070] The encoding module 1 is used to encode the radiotherapy plan and multi-leaf collimator through an encoder to obtain a second vector; the projection module 2 is used to project the second vector based on latent space rotation to obtain a projection vector; the decoding module 3 is used to decode the projection vector through a decoder of control points to obtain a prediction vector and obtain a motion sequence through the prediction vector.
[0071] Among them, the decoding module 3 includes a vane decoding sub-module 31, a lead gate decoding sub-module 32, and a machine unit decoding sub-module 33.
[0072] Among them, the vane decoding sub-module 31 is used to obtain a prediction vector of the vane through a vane decoder and a projection vector; the lead gate decoding sub-module 32 is used to obtain a prediction vector of the lead gate through a lead gate decoder and a projection vector; the machine unit decoding sub-module 33 is used to obtain a prediction vector of the machine unit through a machine unit decoder and a projection vector.
[0073] A third aspect of the present invention further provides a computing device, including a memory and a processor. The processor may be a multi-core processor or may include multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more special co-processors, such as a graphics processing unit (GPU), a digital signal processor (DSP), and so on. In some embodiments, the processor may be implemented using custom circuits, such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).
[0074] The memory may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. Among them, the ROM may store static data or instructions required by the processor or other modules of the computer. The permanent storage device may be a readable and writable storage device. The permanent storage device may be a non-volatile storage device that does not lose the stored instructions and data even when the computer is powered off. In some embodiments, the permanent storage device uses a mass storage device (such as a magnetic or optical disk, flash memory) as the permanent storage device. In some other embodiments, the permanent storage device may be a removable storage device (such as a floppy disk, optical drive). The system memory may be a readable and writable storage device or a volatile readable and writable storage device, such as dynamic random access memory. The system memory may store some or all of the instructions and data required by the processor during operation. In addition, the memory may include any combination of computer-readable storage media, including various types of semiconductor storage chips (DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and magnetic disks and / or optical disks may also be used. In some embodiments, the memory may include a removable storage device that is readable and / or writable, such as a compact disc (CD), a read-only digital versatile disc (such as DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, a high density optical disc, a flash memory card (such as SD card, min SD card, Micro-SD card, etc.), a magnetic floppy disk, and so on. The computer-readable storage medium does not include carrier waves and instantaneous electronic signals transmitted wirelessly or wiredly. An executable code / program is stored on the memory, and when the executable code / program is processed by the processor, the processor can execute the above multi-leaf collimator prediction method.
[0075] In some processes described in the specification, claims, and the above-mentioned drawings of this application, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any order of execution. Additionally, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent a sequence, and do not limit that "first" and "second" are of different types.
[0076] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A multi-leaf collimator prediction method based on volumetric intensity modulated radiotherapy, characterized in that: The following steps are involved: Obtain radiotherapy plans based on volumetric intensity-modulated radiotherapy; Encoding the radiotherapy plan and the multileaf grating through an encoder to obtain a second vector; Based on the latent space rotation, the second vector is projected to obtain a projection vector; Decoding the projection vector through a decoder of the control point to obtain a prediction vector; A motion sequence is obtained based on the prediction vector.
2. The multi-leaf collimator prediction method according to claim 1, characterized in that: The multi-leaf collimator includes 240 control points, 160 leaves and 2 jaws.
3. The multi-leaf collimator prediction method according to claim 1, characterized in that: The network structure of the encoder includes a convolutional layer, a batch normalization layer, a pooling layer, a convolutional layer, a batch normalization layer, an activation function, a convolutional layer, and a batch normalization layer connected in sequence.
4. The multi-leaf collimator prediction method according to claim 1, characterized in that: Methods for latent space rotation include: Directing the second vector to a field of view to obtain a third vector based on a field of view; Multiplying the dose of the third vector by the gradient field tensor to obtain a fourth vector; After rotating the fourth vector along the longitudinal axis or the beam direction, the projection vector is obtained by passing it through an activation function.
5. The multi-leaf collimator prediction method according to claim 4, characterized in that: The calculation method of the fourth vector is expressed as: D′=D˙G Wherein, D′ represents the fourth vector, D represents the dose of the voxel in the direction of the field, and G represents the gradient value of the voxel in the gradient field.
6. The multi-leaf collimator prediction method according to claim 1, characterized in that: The control points are arranged with any one or a combination of the following decoders: blade decoder, jaw decoder and machine unit decoder.
7. The multi-leaf collimator prediction method according to claim 6, characterized in that: The network structure of the leaf decoder or jaw decoder includes multiple convolutional blocks with residual connections, convolutional layers, batch normalization layers, activation functions, convolutional layers, and hyperbolic tangent activation functions connected in sequence.
8. The multi-leaf collimator prediction method according to claim 6, characterized in that: The network structure of the machine unit decoder consists of multiple convolutional blocks with residual connections, average pooling, and output convolutional layers connected sequentially.
9. A prediction system, characterized in that: Used to implement the multi-leaf collimator prediction method according to any one of claims 1 to 8, the prediction system comprises an encoding module, a projection module and a decoding module, The encoding module is used to encode the radiotherapy plan and the multi-leaf grating through an encoder to obtain a second vector; The projection module is used to project the second vector based on the latent space rotation to obtain a projection vector; The decoding module is used to decode the projection vector through a decoder of the control point to obtain a prediction vector, and obtain a motion sequence through the prediction vector.
10. A computing device, characterized in that The invention comprises a memory storing a program, and when the program is processed, the multi-leaf collimator prediction method according to any one of claims 1 to 8 is executed.