CN-122025010-A - Double-flow neural network model applied to radiotherapy and dose reconstruction method

CN122025010ACN 122025010 ACN122025010 ACN 122025010ACN-122025010-A

Abstract

The specification discloses a dual-flow neural network model and a dose reconstruction method applied to radiotherapy, and relates to the technical field of radiotherapy. The model comprises a two-dimensional processing branch, a three-dimensional processing branch and a feature fusion and dose reconstruction branch. The two-dimensional processing branch is used for processing the input two-dimensional projection image to form a two-dimensional beam characteristic diagram. The three-dimensional processing branch is used for processing the input three-dimensional CT image to form a determined three-dimensional anatomical feature map. The dose reconstruction branch comprises a geometric position embedding module, a manifold mapping dimension increasing module and a cross-modal cross-attention fusion module. The cross-modal cross-attention fusion module is used for generating fusion feature tensors according to the three-dimensional anatomical feature map and the three-dimensional beam feature body, and outputting a three-dimensional dose distribution matrix through a three-dimensional dose regression network. Therefore, the model has low calculation complexity, can determine the three-dimensional dose distribution matrix in real time according to the data acquired in real time, has high reconstruction speed, and can meet the requirement of real-time accurate verification.

Inventors

JIN HAIJING
LIU XIN
SHI ZHONGYAN
LIANG RUNCHENG
ZHAO RI
LIU LIYE
LI HUA
CHEN FAGUO
ZHANG JING
GUO RONG
LIU ZHAOXING

Assignees

中国辐射防护研究院

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. The double-flow neural network model applied to radiotherapy is characterized by comprising a two-dimensional processing branch, a three-dimensional processing branch and a feature fusion and dose reconstruction branch; The two-dimensional processing branch comprises a two-dimensional encoder, and the two-dimensional encoder is used for processing an input two-dimensional projection image and determining a two-dimensional beam characteristic diagram; The three-dimensional processing branch comprises a three-dimensional encoder, and the three-dimensional encoder is used for processing the input three-dimensional CT image and determining a three-dimensional anatomical feature map; The dose reconstruction branch comprises a geometric position embedding module, a manifold mapping dimension-increasing module and a cross-modal cross-attention fusion module, wherein the geometric position embedding module is used for determining coordinate index mapping from a three-dimensional voxel space to a two-dimensional feature plane based on a portal geometric parameter, the manifold mapping dimension-increasing module is used for sampling and differentiating the two-dimensional beam feature map along the radial direction under the constraint of the coordinate index mapping to generate a three-dimensional beam feature body, the cross-modal cross-attention fusion module is used for carrying out weighted fusion on the beam feature according to the three-dimensional anatomical feature map and the three-dimensional beam feature body through a voxel-level cross-modal attention mechanism to generate a fusion feature tensor, and outputting a corresponding three-dimensional dose distribution matrix through a three-dimensional dose regression network based on the fusion feature tensor.
2. The dual flow neural network model for radiation therapy according to claim 1, wherein the dual flow neural network model for radiation therapy comprises a preprocessing module; The preprocessing module is used for carrying out filtering processing and/or image/dose calibration processing on the input two-dimensional projection image.
3. The dual flow neural network model for radiation therapy according to claim 1, wherein the two-dimensional encoder is configured to extract a two-dimensional beam profile containing beam intensity information and multi-leaf collimator shapes layer by layer from the two-dimensional projection image.
4. A dual flow neural network model for use in radiotherapy according to any of claims 1-3, characterized in that the dual flow neural network model is provided with a mixing loss function provided with both influencing factors of voxel level errors and of dose distribution structure errors.
5. A dose reconstruction method for radiation therapy, characterized by using the dual flow neural network model for radiation therapy as set forth in any one of claims 1-4, the method comprising: Acquiring a three-dimensional CT image, and acquiring a two-dimensional projection image of a patient in the treatment process in real time; inputting the three-dimensional CT image and the two-dimensional projection image into the double-flow neural network model, and determining a three-dimensional dose distribution matrix output in the double-flow neural network model; registering and comparing the three-dimensional dose distribution matrix with the planned dose in the treatment plan, and determining a comparison result in real time.
6. The method of dose reconstruction for radiation therapy according to claim 5, wherein said registering the three-dimensional dose distribution matrix with the planned dose in the treatment plan, determining the comparison in real time, comprises: determining a real-time three-dimensional pass rate and a dose volume histogram of the real-time organ at risk based on the three-dimensional dose distribution matrix; Determining a planned three-dimensional pass rate and a dose volume histogram of a planned organ at risk based on the planned dose; and comparing the real-time three-dimensional passing rate with the planned three-dimensional passing rate, comparing the dose volume histogram of the real-time organs at risk with the dose volume histogram of the planned organs at risk, and determining a comparison result in real time.
7. The method of claim 6, wherein the method further comprises: determining a treatment accident according to the comparison result; An alarm is raised to alert the physician to pause the treatment or to adjust a subsequent treatment plan.
8. A computer readable storage medium, characterized in that the storage medium stores a computer program which, when executed by a processor, implements the method of any of the preceding claims 5-7.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method of any of the preceding claims 5-7 when executing the program.
10. A computer program product, characterized in that instructions in the computer program product, when executed by a processor of an electronic device, cause the electronic device to perform the method of any of claims 5-7.

Description

Double-flow neural network model applied to radiotherapy and dose reconstruction method Technical Field The present disclosure relates to the field of radiotherapy, and in particular, to a dual-flow neural network model and a dose reconstruction method for radiotherapy. Background Radiation therapy is an important tool in clinical oncology therapy, and its therapeutic effect depends on precise dose control. During radiotherapy, how to accurately deposit the designed dose in the treatment plan to the tumor target area while avoiding excessive irradiation to surrounding normal tissues or Organs At Risk (OAR) is a key to improving the efficacy and reducing the side effects. The treatment planning system calculates a dose distribution that is ideally real-time, and the dose actually delivered to the patient may deviate due to factors such as machine performance fluctuations, patient placement errors, anatomical changes, etc. Therefore, verifying the actual delivered dose is a key element in ensuring that the treatment is safe and effective. The X-ray flat panel detector can measure the information of the outgoing X-ray beam passing through the phantom/patient before or during treatment, providing the possibility for dose verification. However, the existing treatment verification method has high requirement on computational resources, the reconstruction speed is slow, and the calculation accuracy near the heterogeneous tissue interface is obviously reduced due to the fact that the reconstruction speed is dependent on some simplifying assumptions, so that the requirement of real-time accurate verification cannot be met. Disclosure of Invention The present disclosure provides a dual-flow neural network model and a dose reconstruction method for radiation therapy to at least partially solve the above-mentioned problems of the prior art. The technical scheme adopted in the specification is as follows: the specification provides a dual-flow neural network model applied to radiotherapy, which comprises a two-dimensional processing branch, a three-dimensional processing branch and a feature fusion and dose reconstruction branch; The two-dimensional processing branch comprises a two-dimensional encoder, and the two-dimensional encoder is used for processing an input two-dimensional projection image and determining a two-dimensional beam characteristic diagram; The three-dimensional processing branch comprises a three-dimensional encoder, and the three-dimensional encoder is used for processing the input three-dimensional CT image and determining a three-dimensional anatomical feature map; The dose reconstruction branch comprises a geometric position embedding module, a manifold mapping dimension-increasing module and a cross-modal cross-attention fusion module, wherein the geometric position embedding module is used for determining coordinate index mapping from a three-dimensional voxel space to a two-dimensional feature plane based on a portal geometric parameter, the manifold mapping dimension-increasing module is used for sampling and differentiating the two-dimensional beam feature map along the radial direction under the constraint of the coordinate index mapping to generate a three-dimensional beam feature body, the cross-modal cross-attention fusion module is used for carrying out weighted fusion on the beam feature according to the three-dimensional anatomical feature map and the three-dimensional beam feature body through a voxel-level cross-modal attention mechanism to generate a fusion feature tensor, and outputting a corresponding three-dimensional dose distribution matrix through a three-dimensional dose regression network based on the fusion feature tensor. Preferably, the dual-flow neural network model applied to radiotherapy comprises a preprocessing module; The preprocessing module is used for carrying out filtering processing and/or image/dose calibration processing on the input two-dimensional projection image. Preferably, the two-dimensional encoder is configured to extract a two-dimensional beam profile containing beam intensity information and multi-leaf collimator shape layer by layer from the two-dimensional projection image. Preferably, the dual-flow neural network model is provided with a mixed loss function, and the mixed loss function is provided with two influencing factors of voxel level errors and dose distribution structure errors. In another aspect, the present disclosure provides a dose reconstruction method for radiotherapy, using the dual-flow neural network model for radiotherapy provided in the above aspect, the method comprising: Acquiring a three-dimensional CT image, and acquiring a two-dimensional projection image of a patient in the treatment process in real time; inputting the three-dimensional CT image and the two-dimensional projection image into the double-flow neural network model, and determining a three-dimensional dose distribution matrix output in the double-flow neura