Radiotherapy dose prediction method based on quantum classical hybrid convolutional neural network
Through the quantum classic hybrid convolutional neural network, the radiotherapy image data is integrated, and the quantum feature enhancement module and jump connection are used to solve the problems of complex and inefficient dose prediction in the radiotherapy plan design, achieving efficient dose prediction.
Patent Information
- Application Number
- CN202510288915.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-07-29
AI Technical Summary
Dose prediction in radiotherapy planning design is complex, traditional methods are inefficient and doctors have heavy workloads, and it is difficult for existing technologies to effectively integrate artificial intelligence to improve efficiency and accuracy.
Using a method based on quantum classical hybrid convolutional neural network, we use the method to obtain planned target areas, organ-threatening and CT images, generate distance images and integrate them into four-channel images, and use quantum feature enhancement modules and jump connection network structures for dose prediction, and combine quantum computing to improve feature extraction and processing capabilities.
Improve the computational efficiency of radiotherapy dose prediction and the ability to handle complex tasks, and promote the realization of clinical real-time dose prediction.
Smart Images
Figure CN120388181A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image processing, and relates to a radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network. Background Art
[0002] Radiation therapy is an important means for treating malignant tumors such as head and neck cancer. With the progress of technologies such as intensity-modulated radiotherapy and volumetric modulated arc therapy, its accuracy and effectiveness have been significantly improved. However, radiotherapy treatment planning still faces challenges such as complex dose prediction, low efficiency of traditional methods, and heavy workload of doctors. How to integrate advanced artificial intelligence technologies to improve efficiency and accuracy remains an urgent issue to be solved. Quantum Convolutional Neural Network (QCNN) combines the advantages of quantum computing and deep learning. By replacing traditional convolutional operations with quantum gates, it shows the potential to process high-dimensional data and complex tasks. Summary of the Invention
[0003] Object of the Invention. Aiming at the above problems, a radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network is provided to solve the problems raised in the above background art.
[0004] Technical Solution. To achieve the above object, the present invention proposes a radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network, and the dose prediction method includes:
[0005] (1) Obtain a Planning Target Volume (PTV) image, an Organ at Risk (OAR) image, and a CT image;
[0006] (2) Generate a distance image based on the PTV image;
[0007] (3) Integrate the PTV image, the OAR image, the CT image, and the distance image into a four-channel image;
[0008] (4) Input the four-channel image into the encoder of the quantum-classical hybrid convolutional neural network for convolutional processing, and the output of each layer of the encoder is a preliminary feature map;
[0009] (5) Input each preliminary feature map into a quantum feature enhancement module for processing to obtain a spatial attention feature map
[0010] (6) Jump-connect each spatial attention feature map to the decoder of the quantum-classical hybrid convolutional neural network for processing to obtain a dose prediction result.
[0011] Further, in the PTV image of the planned target volume, the pixel values outside the contour range of the planned target volume are set to 0, and the pixel values within the contour range of the planned target volume are set to the prescribed dose; in the OAR image of the organ at risk, the pixel values outside the contour range of the organ at risk are set to 0, and the pixel values within the contour range of the organ at risk are set to different integers.
[0012] Further, when generating the distance image, the value of each voxel is calculated from its minimum distance to the nearest PTV surface, and the calculation formula is as follows:
[0013]
[0014] where (i, j, k) are the coordinates of the voxel, and Ω s represents a point on the surface of the PTV image, and S c is a point in Ω s ;
[0015] In step (3), in order to distinguish the voxels inside or outside the PTV, the distance values of the voxels inside the PTV are multiplied by -1, the distance values are normalized to 50 mm, and the PTV image, OAR image, CT image, and distance image are converted to a common voxel spacing of 2.0×2.0×3.0 mm, and integrated into a four-channel image of 2D axial slices with the same size as the network input.
[0016] Further, the quantum-classical hybrid convolutional neural network includes:
[0017] An encoder that performs convolutional processing on the input image of the quantum-classical hybrid convolutional neural network to obtain a multi-scale feature map;
[0018] A quantum feature enhancement module that enhances the model's ability to extract key features;
[0019] A decoder that processes the feature map enhanced by the quantum features and outputs the dose prediction result.
[0020] Further, the encoder and decoder are composed of classical convolutional layers and pooling layers, and information is transmitted in a skip connection manner.
[0021] Further, the insertion points of the quantum feature enhancement module include the middle layers of the encoder region, decoder region, and skip connection part.
[0022] Further, the quantum feature enhancement module is composed of parameterized quantum circuits.
[0023] Furthermore, the quantum circuit includes:
[0024] A mapping sub-unit that is configured to map the input image data to a quantum state;
[0025] A proposed subunit, the proposed unit is configured to process the input parameters of the quantum state and extract spatial attention features;
[0026] A measurement subunit, the measurement unit is configured to measure the result obtained by processing the proposed subunit to convert the result into classical data to obtain spatial attention features.
[0027] Furthermore, the parameter optimization process of the quantum feature enhancement module is to continuously adjust the model parameter variables so that the output result of the model and the actual data reach the best preset fitting condition.
[0028] Furthermore, the quantum-classical hybrid convolutional neural network is trained through the following steps:
[0029] Input a four-channel image into the encoder of the quantum-classical hybrid convolutional neural network for convolution processing to extract features;
[0030] Jump-connect the feature map to the decoder of the quantum-classical hybrid convolutional neural network for processing to obtain a dose prediction result;
[0031] Calculate the loss function value according to the prediction result and the true value of the dose distribution;
[0032] Adjust and optimize the quantum-classical hybrid convolutional neural network according to the loss function value,
[0033] The loss function value is calculated from the voxel-level mean absolute error.
[0034] Beneficial effects. Compared with the prior art, the technical solution of the present invention has the following beneficial effects:
[0035] The technical solution of the present invention combines quantum computing with the image feature extraction function of convolutional neural networks, utilizes the advantages of quantum computing in processing high-dimensional data and complex tasks, enables the hybrid network to have higher computing efficiency and stronger processing capabilities; improves the model efficiency through quantum parameter compression and promotes clinical real-time dose prediction. Description of the Drawings
[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments are briefly introduced below.
[0037] Figure 1 It is a flowchart of the dose prediction method of the embodiment of the present application;
[0038] Figure 2 It is a schematic diagram of the quantum-classical hybrid convolutional neural network of the embodiment of the present application;
[0039] Figure 3Schematic diagram of the quantum spatial attention sub-module of the embodiment of the present application;
[0040] Figure 4 Schematic diagram of the quantum circuit of the embodiment of the present application;
[0041] Figure 5 Schematic diagram of the working process of the quantum convolution unit of the embodiment of the present application;
[0042] Figure 6 Schematic diagram of the training process of the quantum-classical hybrid convolutional neural network of the embodiment of the present application. Specific implementation mode
[0043] The following details the implementation modes of the present application. The embodiments of the implementation modes are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The implementation modes described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and should not be construed as a limitation to the present application.
[0044] Radiation therapy is an important means for treating malignant tumors such as head and neck cancer. With the progress of technologies such as intensity-modulated radiotherapy and volumetric modulated arc therapy, its accuracy and effectiveness have been significantly improved. However, radiotherapy treatment planning still faces challenges such as complex dose prediction, low efficiency of traditional methods, and heavy workload of doctors.
[0045] Based on the above possible problems, please refer to Figure 1 and Figure 2 , the embodiment of the present application provides a dose prediction method based on a quantum-classical hybrid convolutional neural network. In some embodiments, the dose prediction method includes:
[0046] 011: Obtain the planning target volume, organs at risk, and CT images;
[0047] 012: Generate a distance image based on the images;
[0048] 013: Integrate all the images into a four-channel image;
[0049] 014: Input the four-channel image into the encoder of the quantum-classical hybrid convolutional neural network for convolution processing, and the output of each layer of the encoder is a preliminary feature map;
[0050] 015: Input each preliminary feature map into a quantum spatial attention module 20 for processing to obtain a spatial attention feature map;
[0051] 016: Skip-connect each spatial attention feature map to the decoder of the quantum-classical hybrid convolutional neural network for processing to further obtain a dose prediction result.
[0052] Specifically, in the planned target volume (PTV) image, the pixel values outside the contour range of the PTV are set to 0, and the pixel values within the contour range of the PTV are set to the prescribed dose.
[0053] Specifically, in the organs at risk (OAR), the pixel values outside the contour range of the OAR are set to 0, and the pixel values within the contour range of the OAR are set to different integers. For example, the body is set to 1, the left lung is set to 2, the right lung is set to 3, the heart is set to 4, and the spinal cord is set to 5.
[0054] Furthermore, a distance image is generated, and the value of each voxel is calculated from the minimum distance to the nearest surface of the PTV. The calculation formula is as follows:
[0055]
[0056] where (i, j, k) are the coordinates of the voxel, Ω s represents a point on the surface of the PTV, and S c is a point in Ω s .
[0057] To distinguish the voxels inside or outside the PTV, the distance values of the voxels inside the PTV are multiplied by -1. Then the distance values are normalized to 50 mm to obtain a distance image that is easy for the neural network to process.
[0058] The PTV, OAR, CT image, and distance image are converted to a common voxel spacing of 2.0×2.0×3.0 mm and integrated into a four-channel image of 2D axial slices with the same size as the network input.
[0059] Please refer to Figure 2 , the quantum-classical hybrid convolutional neural network 1000 of the embodiment of the present application is designed based on the conventional Unet model. Among them, the Unet model includes a multi-layer encoder 100 and a multi-layer decoder 200. The sizes of the preliminary feature maps obtained by convolution of each layer of the encoder 100 can be different. For example, they can gradually decrease from top to bottom to form a U-shaped structure. Each convolution module 10 of each layer of the encoder 100 includes a double-depth separable convolution, a normalization layer, and a ReLU activation function. The quantum spatial attention module 20 is placed after each double-depth separable convolution. It should be noted that Figure 2 the actual number of layers and actual sizes of the encoder 100 and decoder 200 in
[0060] The quantum feature enhancement module 20 is a convolutional attention module designed based on quantum circuits, which learns the spatial features of the preliminary feature maps to obtain spatial attention feature maps.
[0061] Please refer to Figure 3, in some embodiments, the quantum communication spatial attention module is configured as a pooling unit 201 and a quantum convolution unit 202. The pooling unit 201 is configured to perform pooling processing on the intermediate feature map in the channel dimension; the quantum convolution unit 202 is configured to perform quantum convolution on the intermediate feature map after pooling processing through a quantum circuit to obtain a spatial attention feature map.
[0062] Specifically, the pooling unit 201 performs global max pooling and global average pooling on the input feature map in the channel dimension to obtain two single-channel feature maps, which is convenient for learning spatial features later; the results of global max pooling and global average pooling are concatenated according to the channels to obtain a two-channel intermediate feature map.
[0063] Specifically, the quantum convolution unit 202 performs convolution on the intermediate feature map after pooling processing through a quantum circuit as a convolution filter, and then passes through a Sigmoid activation function to obtain a spatial attention weight matrix 203. Perform pixel-level multiplication processing on the spatial attention weight matrix 203 and the preliminary feature map to extract spatial attention features and generate a spatial attention feature map.
[0064] In some embodiments, the quantum circuit includes: a mapping subunit, a hypothesis subunit, and a measurement subunit; the mapping subunit is configured to map the input image data to a quantum state; the hypothesis subunit is configured to process the input quantum state to obtain a processing result; the measurement subunit is configured to convert the processing result to classical information and connect to a classical neural network.
[0065] Specifically, the mapping subunit is used to map the input parameters of the quantum circuit from a classical state to a quantum state to realize the conversion from a classical state to a quantum state. The hypothesis subunit processes the input parameters of the quantum state and extracts spatial attention features. The hypothesis subunit includes trainable weights, and during network training, the trainable weights of the hypothesis are also continuously adjusted so that the quantum convolution unit 202 continuously adjusts and learns. The measurement subunit is used to measure the result processed by the hypothesis subunit to convert the result to a classical state to obtain spatial attention features.
[0066] In this way, spatial attention features can be extracted by using quantum computing through a quantum circuit, and the conversion between classical states and quantum states can be realized. In one embodiment, when the input feature is four-dimensional, the quantum circuit is as Figure 4 shown, and at this time the quantum convolution unit 202 is as Figure 5 shown.
[0067] In some embodiments, the quantum-classical hybrid convolutional neural network is trained through the following steps:
[0068] 021. Input the four-channel image into the encoder of the quantum-classical hybrid convolutional neural network for convolutional processing to extract features and obtain a preliminary feature map.
[0069] 022. Input the preliminary feature map into the quantum feature enhancement module 20 for processing to obtain a spatial attention feature map. The quantum feature enhancement module 20 is configured to extract features based on quantum computing.
[0070] 023. Skip-connect the spatial attention feature map to the decoder of the quantum-classical hybrid convolutional neural network for processing to obtain a dose prediction result.
[0071] 024. Calculate the loss function value according to the prediction result and the true value of the dose distribution.
[0072] 025. Adjust and optimize the quantum-classical hybrid convolutional neural network according to the loss function value.
[0073] Specifically, in some embodiments, the parameter optimization process uses the PennyLane framework. By continuously adjusting the model parameter variables, the output result of the model and the actual data reach the best fitting degree, and the hybrid quantum-classical computer is used to automatically derive the gradient of the hybrid quantum-classical loss function. Specifically, in some embodiments, the loss function value is calculated from the voxel-level mean absolute error.
[0074] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various deformations or modifications within the scope of the claims, which do not affect the essence of the present invention.
Claims
1. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network, characterized in that The described dose prediction method includes: (1) Obtain the planned target volume (PTV) image, the organ at risk (OAR) image, and the CT image; (2) Generate a distance image based on the PTV image; (3) Integrate the PTV image, the OAR image, the CT image, and the distance image into a four-channel image; (4) Input the four-channel image into the encoder of the quantum-classical hybrid convolutional neural network for convolutional processing, and the output of each layer of the encoder is a preliminary feature map; (5) Input each preliminary feature map into a quantum feature enhancement module for processing to obtain a spatial attention feature map; (6) Jump-connect each spatial attention feature map to the decoder of the quantum-classical hybrid convolutional neural network for processing to obtain the dose prediction result.
2. The radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 1, wherein In the described PTV image, the pixel values outside the contour range of the planned target volume are set to 0, and the pixel values within the contour range of the planned target volume are set to the prescribed dose; in the OAR image, the pixel values outside the contour range of the organ at risk are set to 0, and the pixel values within the contour range of the organ at risk are set to different integers.
3. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 1, characterized in that, In step (2), when generating the distance image, the value of each voxel is calculated from its minimum distance to the nearest PTV surface, and the calculation formula is as follows: where (i, j, k) are the coordinates of the voxel, and Ω s represents a point on the surface of the PTV image, and S c is a point in Ω s ; In step (3), in order to distinguish the voxels inside or outside the PTV, the distance values of the voxels inside the PTV are multiplied by -1, the distance values are normalized to 50 mm, the PTV image, the OAR image, the CT image, and the distance image are converted to a common voxel spacing of 2.0×2.0×3.0 mm, and integrated into a four-channel image with 2D axial slices of the same size as the network input.
4. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 1, characterized in that, The described quantum-classical hybrid convolutional neural network includes: An encoder that performs convolutional processing on the input image of the quantum-classical hybrid convolutional neural network to obtain multi-scale feature maps; A quantum feature enhancement module that enhances the model's ability to extract key features; A decoder that processes the feature maps enhanced by the quantum feature, and outputs the dose prediction result.
5. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 4, characterized in that, The encoder and the decoder are composed of classical convolutional layers and pooling layers, and information is transmitted in a skip connection manner.
6. The radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 4, wherein The insertion points of the quantum feature enhancement module include the encoder region, the decoder region, and the intermediate layer of the skip connection part.
7. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 4, characterized in that The quantum feature enhancement module is composed of parameterized quantum circuits.
8. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 7, characterized in that The quantum circuit includes: A mapping sub-unit configured to map the input image data to a quantum state; A hypothesizing sub-unit configured to process the input parameters of the quantum state and extract spatial attention features; A measurement sub-unit configured to measure the result processed by the hypothesizing sub-unit to convert the result into classical data to obtain the spatial attention feature.
9. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to claim 4, characterized in that, The parameter optimization process of the quantum feature enhancement module is to continuously adjust the model parameter variables so that the output result of the model and the actual data reach the best preset fitting condition.
10. A radiotherapy dose prediction method based on a quantum-classical hybrid convolutional neural network according to any one of claims 1 to 9, characterized in that, The quantum-classical hybrid convolutional neural network is trained through the following steps: Input the four-channel image into the encoder of the quantum-classical hybrid convolutional neural network for convolutional processing to extract feature maps; Jump-connect the feature maps to the decoder of the quantum-classical hybrid convolutional neural network for processing to obtain the dose prediction result; Calculate the loss function value according to the prediction result and the true value of the dose distribution; Adjust and optimize the quantum-classical hybrid convolutional neural network according to the loss function value; The loss function value is calculated from the voxel-level mean absolute error.