CN114882296A

CN114882296A - Neural network-based in-vivo EPID dose verification error classification and grading algorithm

Info

Publication number: CN114882296A
Application number: CN202210737529.6A
Authority: CN
Inventors: 杨志勇; 张盛; 杨坤禹
Original assignee: Tongji Medical College of Huazhong University of Science and Technology
Current assignee: Tongji Medical College of Huazhong University of Science and Technology
Priority date: 2022-06-27
Filing date: 2022-06-27
Publication date: 2022-08-09

Abstract

The invention discloses an in-vivo EPID (extended peripheral identification) dose verification error classification and classification algorithm based on a neural network, which comprises the following specific algorithms: s1, adopting a neural network to carry out error classification on the EPID images containing various errors; s2, preparing a training set, training a classification neural network, needing a large number of matched data sets, inputting data, needing to give correct labels, and dividing errors in a daily treatment process into three types: mechanical errors, positioning errors and anatomical structure errors, wherein the simulated mechanical errors comprise multi-leaf grating positioning errors and machine output errors, and the positioning errors comprise translation errors and rotation errors; s3, structure and training of the model, the invention classifies and evaluates the EPID dose distribution deviation collected by the patient during each treatment through the tree hierarchy classification CNN model, can be used for judging the dose deviation and the deviation type actually accepted by the patient, and provides a judgment basis for the self-adaptive plan design of the patient.

Description

Neural network-based in-vivo EPID dose verification error classification and grading algorithm

Technical Field

The invention belongs to the technical field of medicine, and particularly relates to an in-vivo EPID dose verification error classification and classification algorithm based on a neural network.

Background

In clinical practice, radiotherapy is also one of the cancer treatments commonly used at present. In recent years, the applications of artificial intelligence and deep learning in radiotherapy have been increasing year by year, and convolutional neural networks have been used clinically for medical image segmentation, treatment prescription optimization, clinical result prediction, and the like. The CNN combines the shallow feature and the deep feature through the cascade in the up-sampling process, and is favorable for providing an object type identification basis in the medical image.

The invention with the authorization notice number of CN110237445B discloses an EPID-based in-vivo three-dimensional dose monitoring and verifying method, which collects images of each radiation field during radiotherapy by an electronic radiation field imaging device; extracting an original ray gray value in a field image by a ratio of corresponding scattered rays to original rays, converting the gray value of the original rays of the EPID plane into an EPID plane original ray intensity value, further carrying out back-stepping by combining a mold body CT value to obtain an original ray intensity value before entering a mold body, and obtaining a three-dimensional dose value in the mold body by convolution of the original ray intensity value before entering the mold body and an energy deposition kernel; the accuracy of calculation and execution of the radiotherapy planning system can be verified by comparing the calculated three-dimensional dose value with the calculated value of the radiotherapy planning system, but in clinical practice of radiotherapy, the EPID dose actually received by a patient is different due to the positioning deviation and structural change of the patient during each treatment, so that the subsequent patient is not favorable for self-adaptive adjustment of a treatment plan according to the EPID dose deviation, and therefore an in-vivo EPID dose verification error classification and grading algorithm based on a neural network is provided.

Disclosure of Invention

The invention aims to provide an on-body EPID dose verification error classification and classification algorithm based on a neural network, which classifies and evaluates EPID dose distribution deviation acquired by a patient during each treatment through a tree-shaped hierarchical classification CNN model, is used for judging the actually accepted dose deviation and deviation type of the patient and provides an evaluation basis for the self-adaptive planning design of the patient.

In order to achieve the purpose, the invention provides the following technical scheme: the neural network-based in-vivo EPID dose verification error classification and grading algorithm comprises the following specific algorithms:

s1, adopting a neural network to carry out error classification on the EPID images containing various errors;

s2, preparing a training set, training a classification neural network, needing a large number of matched data sets, inputting data, needing to give correct labels, and dividing errors in a daily treatment process into three types: mechanical errors, positioning errors and anatomical structure errors, wherein the simulated mechanical errors comprise multi-leaf grating positioning errors and machine output errors, and the positioning errors comprise translation errors and rotation errors;

and S3, structure and training of the model.

Preferably, said S2 is specifically,

s2.1, error simulation;

s2.2, extracting features, namely calculating the signal intensity difference between the EPID image containing errors and the EPID image not containing errors to obtain a dose difference map (DD map) as the input of a network;

s2.3, carrying out normalization processing on the data;

s2.4, cutting a pixel matrix acquired by the EPID image;

and S2.5, resampling the matrix, wherein the cut matrix is too large, so that the parameters of the model are increased, the training speed of the network is reduced, meanwhile, in order to ensure the universality of the network structure on various disease types, a linear interpolation method is adopted, resampling is carried out on the matrix, and finally a 256-pixel-by-256-pixel dose error map is unified as the input of the network.

Preferably, in S2.1, since the clinically generated EPID image includes a plurality of errors and it is necessary to spend a large amount of time analyzing the types and degrees of errors included therein to obtain a correct label, the EPID image including errors and its corresponding label are obtained as a training set by simulating the errors determined by the types and degrees by the in-vivo EPID image prediction model, and for the in-place error simulation of the multi-leaf grating, the simulation is calculated by modifying the leaf positions forming each control point and reintroducing into the planning system, wherein the introduction of the systematic errors is to translate the leaf forming the field in all the control points by 2mm in the same direction, and the introduction of the random errors performs random translation on the leaf forming the field in each control point, the degree of translation follows a gaussian distribution with an average value of 2mm and a variance of 1mm, and for the machine output errors, obtained by scaling the machine output of each control point and re-accumulating, similar to the in-place error of a multi-leaf grating, 3% of the systematic error and random error are simulated, for the in-place error and anatomical structure error, the CT image of the patient is modified, wherein the target volume is simulated to be retrograded by adding quantitative deformation vector field deformation CT to the target volume and the surrounding normal tissue, the tissue fluid filling is simulated by assigning the CT value of the nasal cavity to 0 (CT value of water), after introducing corresponding errors to the plan and CT, the EPID image containing various errors is calculated by the established EPID image calculation model, at the same time, the original CT and radiotherapy plans of the patient are used to calculate the EPID image containing no errors, since the EPID image actually generated in the clinical process contains various errors, when the error introduction of the simulation data is performed, a single error is introduced to form a data set 1, two error formation data sets 2 are introduced, so that the detection capability of an error detection model on clinical data is improved.

Preferably, in S2.3, to accelerate network convergence and only need to learn a relative difference distribution through a network, a dose difference map under various errors is processed to [0,1] by a max-min normalization method, where the specific formula is as follows:

in the above formula, Data is the dose in the dose difference map, and min (Data) and max (Data) are the minimum and maximum values of the respective dose error maps.

Preferably, in S2.4, the range of acquiring the EPID image is large (the pixel matrix size is 1280 × 1280, and the effective area is 30cm × 40cm), while the field of the patient tends to be concentrated near the target area, which will cause the input data to include too much blank information, in order to make the network learn more useful information, the EPID image without errors is divided into reserved areas and non-reserved areas by a threshold of 10% of the highest dose, and on the premise of ensuring that the reserved areas are included as much as possible, the cropping area is selected by using a rectangular frame of 1024 pixels × 1024 pixels, because the types of the diseases are different, and the size difference of the reserved areas is large, the size of the rectangular frame is adjusted according to the type of the disease.

Preferably, said S3 is specifically,

s3.1, establishing a tree-shaped hierarchical classification model;

s3.2, establishing a basic classifier;

s3.3, calculating a loss function;

and S3.4, calculating training parameters.

Preferably, in S3.1, specifically, in the hierarchical classification idea from coarse to fine, the sub-class result (fine sub-class) of the error classification should follow the parent-class result (coarse large sub-class), and the tree-shaped hierarchical classification model is used in the error detection task.

Preferably, S3.2 is specifically that, in the tree-like hierarchical model, an infrastructure of efficientnet-b7 is used as a feature extraction part of the model, the tree-like hierarchical model includes a shallow convolutional layer, seven moving and turning bottleneck convolutional blocks (MB conv blocks) and a deep convolutional layer, and is composed of five basic modules, where the MB conv block mainly combines a residual error structure and an attention mechanism, after image features are extracted, a feature matrix is flattened, and a classifier composed of a full connection layer and a sigmoid function is input, since a data set has multiple tags at the same time, that is, the model should be a multi-tag classifier, a sigmoid function is used as an activation function output by the full connection layer, after the activation of the sigmoid function, each output value is a probability of occurrence of a corresponding error, probabilities between the categories are independent and are between 0 and 1, and 0.5 is used as a threshold for classification, i.e. a probability greater than 0.5, the corresponding error is considered to be present.

Preferably, in S3.3, for each category, the classification task is two classifications, that is, a positive classification (existing) or a negative classification (nonexistence), and to ensure the accuracy of the category with high classification difficulty, the following formula is used for calculating a single category:

wherein

For a classifier with N classes, the following formula for the local loss calculation of all classes is given as follows:

preferably, S3.4 specifically includes that an Adam adaptive matrix optimizer is adopted, the initial momentum is 0.5, the batch size is 3, the initial learning rate is set to 0.0001, the learning rate is updated adaptively, the learning rate is gradually reduced as the number of training rounds increases, and the update formula is as follows:

Lr _epoch ＝Lr ₀ *10 ^-epoch/3 ，

because the training initialization is based on a model of pre-training, the training loss convergence is faster, and in order to prevent overfitting, the maximum number of training rounds is 50, Lr _epoch The value is the change value of the loss function along with the number of training rounds, and meanwhile, the training is stopped in advance when the loss of the verification set does not decrease in 5 rounds.

Compared with the prior art, the invention has the beneficial effects that:

the invention classifies and evaluates the EPID dose distribution deviation collected by the patient during each treatment through the tree hierarchy classification CNN model, can be used for judging the dose deviation and the deviation type actually accepted by the patient and provides a judgment basis for the self-adaptive plan design of the patient.

Drawings

FIG. 1 is a schematic diagram of a tree-level classification model according to the present invention.

FIG. 2 is a schematic diagram of a classifier structure based on the efficientnet-b7 in the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Referring to fig. 1-2, the present invention provides a technical solution: the neural network-based in-vivo EPID dose verification error classification and grading algorithm comprises the following specific algorithms:

s2.1, error simulation, because the EPID image generated in clinic contains various errors and needs to spend a great deal of time to analyze the types and the error degrees of the errors contained in the EPID image to obtain a correct label, simulating the errors determined by the types and the degrees through an in-vivo EPID image prediction model to obtain the EPID image containing the errors and the corresponding label thereof as a training set, and for the in-place error simulation of the multi-leaf raster, calculating by modifying the positions of the leaves forming each control point and reintroducing the leaves into a planning system, wherein the introduction of the system errors is to translate the leaves forming the field in the same direction by 2mm in all the control points, and the introduction of the random errors is to randomly translate the leaves forming the field in each control point, the degree of translation follows a Gaussian distribution with the average value of 2mm and the variance of 1mm, and for the machine output errors, the machine output of each control point is obtained by reducing and reintegrating the machine output of each control point, similar to the multi-leaf raster in-place error, 3% of both the systematic and random errors are modeled, and for the in-place error, as well as the anatomical error, the patient's CT image is modified, wherein the target volume is simulated to recede by adding quantitative deformation vector field deformation CT to the target volume and surrounding normal tissue, interstitial fluid filling is simulated by assigning the CT value of the nasal cavity to 0 (CT value of water), after introducing corresponding errors to the plan and the CT, calculating EPID images containing various errors through the established EPID image calculation model, at the same time, the original CT and radiotherapy plans of the patient are used to calculate EPID images containing no errors, since the EPID images actually generated during the clinical procedure contain various errors, when the error of the simulation data is introduced, a single error is introduced to form a data set 1, and two errors are introduced to form a data set 2, so that the detection capability of an error detection model on clinical data is improved;

s2.3, normalizing the data, in order to accelerate network convergence, and only needing network learning relative difference distribution, carrying out max-min normalization on the dose difference graph under various errors to [0,1], wherein the specific formula is as follows:

in the above formula, Data is the dose in the dose difference map, and min (Data) and max (Data) are the minimum and maximum values of the dose error maps respectively;

s2.4, a pixel matrix acquired by the EPID image is cut, the range of the acquired EPID image is large (the size of the pixel matrix is 1280 multiplied by 1280, the effective area is 30cm multiplied by 40cm), the field of a patient is often concentrated near a target area, the input data contains excessive blank information, in order to enable the network to learn more useful information, the EPID image without errors is divided into reserved areas and non-reserved areas through a threshold value of 10% of the highest dose, and on the premise that the reserved areas are contained as much as possible, a rectangular frame of 1024 pixels multiplied by 1024 pixels is adopted to select the cut areas, and the sizes of the reserved areas are large in difference due to different types of diseases, so the sizes of the rectangular frame are adjusted according to the types of the diseases;

s2.5, resampling the matrix, wherein the cut matrix is too large to increase the parameters of the model and slow down the training speed of the network, and meanwhile, in order to ensure the universality of the network structure on various disease types, the matrix is resampled by adopting a linear interpolation method, and finally a dose error map with 256 pixels multiplied by 256 pixels is unified as the input of the network, wherein the error detection network does not need to learn at the pixel level, so that the resolution ratio does not need to be completely unified;

s3, structure and training of the model;

s3.1, establishing a tree-shaped hierarchical classification model, wherein in the hierarchical classification idea from coarse to fine, the subclass result (fine subclass) of error classification should follow the parent result (coarse major class), and the tree-shaped hierarchical classification model is used in an error detection task, as shown in FIG. 1, a classifier 1 firstly roughly divides the major class, and when a dose difference diagram is judged to be corresponding error, such as mechanical error, the corresponding sub-classifier divides the minor class;

s3.2, establishing a basic classifier, using a convolutional neural network as a common classifier network, having certain advantages in the aspect of extracting the features of the image, using a basic structure of efficientnet-b7 as a feature extraction part of the model in a tree-shaped hierarchical model, comprising a shallow convolutional layer, seven moving and turning bottleneck convolutional blocks (MB conv block) and a deep convolutional layer, and consisting of five basic modules, wherein the MB conv block mainly combines a residual error structure and an attention mechanism, after the image features are extracted, flattening a feature matrix, and inputting a classifier consisting of a full connection layer and a sigmoid function, because a data set simultaneously has a plurality of labels, namely the model is a multi-label classifier, using the sigmoid function as an activation function output by the full connection layer, and after the activation of the sigmoid function, each output value is the probability of the occurrence of corresponding error, the probabilities among all the categories are independent and are between 0 and 1, 0.5 is used as a threshold value of the classification, namely, the probability is more than 0.5, and the corresponding error exists;

s3.3, loss function calculation, wherein for each category, classification tasks are classified into two categories, namely a positive category (existing) or a negative category (nonexistence), and meanwhile, for ensuring the accuracy of the category with high classification difficulty, the local loss is adopted as a main loss function, and a calculation formula for a single category is as follows:

wherein

s3.4, calculating training parameters, adopting an Adam self-adaptive matrix optimizer, setting the initial momentum to be 0.5, setting the batch size to be 3, setting the initial learning rate to be 0.0001, adopting self-adaptation to update the learning rate, gradually reducing the learning rate along with the increase of the number of training rounds, and adopting an update formula as follows:

Lr _epoch ＝Lr ₀ *10 ^-epoch/3 ，

because the training initialization is based on a model of pre-training, the training loss convergence is faster, and in order to prevent overfitting, the maximum number of training rounds is 50, Lr _epoch With value of loss function over number of training roundsThe values were varied, with the setting that training stopped prematurely when the loss of validation set did not drop within 5 rounds.

In conclusion, the algorithm is an error classification algorithm for in-vivo dose verification of clinical radiotherapy patients, can be used as an evaluation tool for radiotherapy quality assurance of patients, can judge the actually accepted dose deviation and deviation type of the patients, and provides a judgment basis for the adaptive planning design of the patients.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. The neural network-based in-vivo EPID dose verification error classification and grading algorithm comprises the following specific algorithms:

and S3, structure and training of the model.

2. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 1, wherein: specifically, the step S2 is,

s2.1, error simulation;

s2.3, carrying out normalization processing on the data;

s2.4, cutting a pixel matrix acquired by the EPID image;

3. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 2, wherein: s2.1 is specifically that, because EPID images generated in clinic contain various errors and a large amount of time is needed to analyze the types and the error degrees of the errors contained in the EPID images so as to obtain correct labels, the EPID images containing the errors and the corresponding labels thereof are obtained as training sets by simulating the errors determined by the types and the degrees through a body EPID image prediction model, and for the in-place error simulation of the multi-leaf raster, the in-place error simulation is obtained by modifying the positions of leaves forming all control points and reintroducing the leaves into a planning system for calculation, wherein the introduction of the systematic errors is to translate the leaves forming the field in all the control points in the same direction by 2mm, while the introduction of the random errors is to randomly translate the leaves forming the field in each control point, the degree of translation follows a Gaussian distribution with the average value of 2mm and the variance of 1mm, and for the machine output errors, the machine output of each control point is obtained by reducing and reintegrating the machine output of each control point, similar to the multi-leaf raster in-place error, 3% of both the systematic and random errors are modeled, and for the in-place error, as well as the anatomical error, the patient's CT image is modified, wherein the target volume is simulated to recede by adding quantitative deformation vector field deformation CT to the target volume and surrounding normal tissue, interstitial fluid filling is simulated by assigning the CT value of the nasal cavity to 0 (CT value of water), after introducing corresponding errors to the plan and the CT, calculating EPID images containing various errors through the established EPID image calculation model, at the same time, the original CT and radiotherapy plans of the patient are used to calculate EPID images containing no errors, since the EPID images actually generated during the clinical procedure contain various errors, when the error of the simulation data is introduced, a single error forming data set 1 is introduced, and two error forming data sets 2 are introduced, so that the detection capability of the error detection model on the clinical data is improved.

4. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 2, wherein: the S2.3 is specifically that, in order to accelerate network convergence and only need to learn relative difference distribution through a network, a dose difference map under various errors is processed to [0,1] by a max-min normalization method, and a specific formula is as follows:

5. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 2, wherein: the S2.4 is specifically that the scope of acquiring the EPID image is large (the pixel matrix size is 1280 × 1280, and the effective area is 30cm × 40cm), and the radiation field of the patient is often concentrated near the target area, which will cause the input data to include too much blank information, in order to enable the network to learn more useful information, the EPID image without errors is divided into reserved areas and non-reserved areas through a 10% threshold of the highest dose, and on the premise that the reserved areas are included as much as possible, a rectangular frame of 1024 pixels × 1024 pixels is used to select the clipping area, and since the types of the disease types are different, the size difference of the reserved areas is large, the size of the rectangular frame is adjusted according to the type of the disease.

6. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 1, wherein: specifically, the step S3 is,

s3.1, establishing a tree-shaped hierarchical classification model;

s3.2, establishing a basic classifier;

s3.3, calculating a loss function;

and S3.4, calculating training parameters.

7. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 6, wherein: the step S3.1 is specifically to, based on the hierarchical classification idea from coarse to fine, follow the sub-class result (fine sub-class) of the error classification with the parent-class result (coarse large sub-class), and use the tree-shaped hierarchical classification model in the error detection task.

8. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 6, wherein: s3.2 specifically is that an infrastructure of the efficientnet-b7 is adopted in the tree-shaped hierarchical model as a feature extraction part of the model, the tree-shaped hierarchical model comprises a shallow convolutional layer, seven mobile turning bottleneck convolutional blocks (MB conv blocks) and a deep convolutional layer, and the tree-shaped hierarchical model is composed of five basic modules, where MB conv block mainly combines residual structure and attention mechanism, after the image features are extracted, the feature matrix is flattened and is input into a classifier consisting of a full connection layer and a sigmoid function, and as a data set simultaneously has a plurality of labels, namely, the model is a multi-label classifier, a sigmoid function is adopted as an activation function output by a full connection layer, and after the sigmoid function is activated, each output value is the probability of the occurrence of the corresponding error, the probabilities between the classes are independent and between 0 and 1, and 0.5 is used as a threshold value of the classification, namely the probability is greater than 0.5 to consider that the corresponding error exists.

9. The neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 6, wherein: s3.3 specifically is that, for each category, the classification task is two classifications, i.e., positive classification (existing) or negative classification (nonexistence), and meanwhile, to ensure the accuracy of the category with high classification difficulty, Focalloss is used as a main loss function, and a calculation formula for a single category is as follows:

wherein

10. the neural network-based in-vivo EPID dose verification error classification and ranking algorithm of claim 6, wherein: s3.4 specifically is, adopt Adam adaptive matrix optimizer, initial momentum is 0.5, batch size is 3, sets up initial learning rate and is 0.0001 to adopt self-adaptation to update the learning rate, increase along with training round, reduce the learning rate gradually, the update formula as follows:

Lr _epoch ＝Lr ₀ *10 ^-epoch/3 ，