A radiotherapy quality control error analysis method based on interpretable neural network
Through the radiotherapy quality control error analysis method based on interpretable neural network, the problem that traditional methods cannot fully consider individual differences and dynamic changes is solved, and the accurate modeling and error detection of radiotherapy plan data is achieved, which improves the accuracy of treatment and security.
Patent Information
- Application Number
- CN202410656782.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-24
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-05-24
AI Technical Summary
Traditional radiotherapy quality control methods cannot fully consider individual differences in patients and dynamic changes in the treatment process, which makes it difficult to detect and correct the errors in radiotherapy planning in a timely manner, and the results are subjective and uncertain.
Using radiotherapy quality control error analysis method based on interpretable neural networks, the dose difference histogram and distance consistency graph are extracted by establishing and training error classification models integrated with interpretable mechanisms, including metrological error classification submodels and spatial error classification submodels. , used to determine and interpret dose error and spatial error categories.
Accurate modeling and error detection of radiotherapy planning data is achieved, real-time quality assurance and intervention means are provided, the accuracy and safety of radiotherapy are improved, and the dependence on professionals and the subjectivity of results is reduced.
Smart Images

Figure CN118471434B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of radiotherapy, and in particular to a radiotherapy quality control error analysis method based on an interpretable neural network. Background Art
[0002] With the widespread application of radiotherapy in the treatment of brain tumors, ensuring the accuracy and quality of radiotherapy plans has become an important challenge in the medical field, which directly affects the treatment effect and survival rate of patients. In intensity modulated radiotherapy (IMRT), the quality assurance (QA) of the patient's radiotherapy plan is crucial for the accurate implementation of the treatment plan. The comparison of the radiation plan dose distribution calculated by the treatment planning system (TPS) and the actual measured dose distribution is the basis for approving or rejecting the treatment plan. Dose distribution comparison is the core of IMRT quality control. Its effectiveness can determine our ability to identify errors in the detection plan and delivery system, and assist in error detection and quality control in IMRT. Traditional radiotherapy quality control methods have some limitations. For example, manual inspection cannot fully consider individual differences of patients and dynamic changes during treatment, which makes it difficult to detect and correct errors in radiotherapy plans in a timely manner; gamma analysis only provides quantitative comparison of dose distribution, which requires doctors and professionals to interpret the category of analysis errors, resulting in subjectivity and uncertainty in the results.
[0003] Therefore, there is a need to provide a radiotherapy quality control error analysis method based on an interpretable neural network to perform error analysis on radiotherapy planning data and provide real-time quality assurance and intervention measures for the medical team, thereby improving the accuracy and safety of radiotherapy. Summary of the invention
[0004] The present invention provides a radiotherapy quality control error analysis method based on an interpretable neural network, comprising: establishing and training an error classification model integrated with an interpretable mechanism, wherein the error classification model integrated with the interpretable mechanism includes a metrology error classification sub-model integrated with the interpretable mechanism and a spatial error classification sub-model integrated with the interpretable mechanism; obtaining dose distribution data pairs to be classified, wherein the dose distribution data pairs to be classified include reference dose distribution information and measured dose distribution information; extracting dose difference histograms and distance consistency diagrams corresponding to the dose distribution data pairs to be classified; inputting the dose difference histograms into the metrology error classification sub-model integrated with the interpretable mechanism to determine dose error categories and dose error classification explanations; inputting the distance consistency diagrams into the spatial error classification sub-model integrated with the interpretable mechanism to determine spatial error categories and spatial error classification explanations.
[0005] Furthermore, the metrology error classification sub-model incorporating an explainable mechanism includes a metrology error classification module and a metrology error classification explanation module, wherein the metrology error classification module is used to determine the dose error category based on the dose difference histogram, and the metrology error classification explanation module is used to generate a dose error classification explanation corresponding to the dose error category; the metrology error classification module includes an input layer, a self-attention layer, a residual connection layer, a normalization layer, a fully connected layer and an output layer.
[0006] Furthermore, the spatial error classification sub-model incorporating an explainable mechanism includes a spatial error classification module and a spatial error classification explanation module, wherein the spatial error classification module is used to determine the spatial error category based on the distance consistency map, and the spatial error classification explanation module is used to generate a spatial error classification explanation corresponding to the spatial error category: the spatial error classification module includes an input layer, 3 convolutional layers, a pooling layer, a self-attention layer, a fully connected layer and an output layer.
[0007] Furthermore, extracting a dose difference histogram corresponding to the dose distribution data pair to be classified includes: determining a pixel error corresponding to each pixel based on the reference dose distribution information and the measured dose distribution information included in the dose distribution data pair to be classified; determining an error pixel based on the pixel error corresponding to each pixel and a preset error range; determining the number of error pixels corresponding to each error segment of the preset error range based on the pixel error corresponding to the error pixel; and establishing a dose difference histogram corresponding to the dose distribution data pair to be classified based on the number of error pixels corresponding to each error segment of the preset error range.
[0008] Furthermore, extracting the distance consistency map corresponding to the dose distribution data pair to be classified includes: for each radiation point, determining the associated radiation point corresponding to the radiation point, wherein the associated radiation point is the radiation point with the same radiation dose as the radiation point and the closest distance; based on the relative position relationship between each of the radiation points and the associated radiation point of the radiation point, generating the distance consistency map corresponding to the dose distribution data pair to be classified.
[0009] Furthermore, the dose difference histogram is input into the metrology error classification sub-model integrated with the interpretable mechanism to determine the dose error category and the dose error classification explanation, including: sampling the dose difference histogram to obtain dosimetric feature sampling information; inputting the dosimetric feature sampling information into the metrology error classification sub-model integrated with the interpretable mechanism to determine the dose error category and the dose error classification explanation.
[0010] Furthermore, the distance consistency map is input into the spatial error classification sub-model integrated with the interpretable mechanism to determine the spatial error category and the spatial error classification explanation, including: performing image segmentation and sampling on the distance consistency map to obtain multiple pixel blocks; inputting the multiple pixel blocks into the spatial error classification sub-model integrated with the interpretable mechanism to determine the spatial error category and the spatial error classification explanation.
[0011] Furthermore, the distance consistency map is sampled, including: when the sampling number is less than a preset number threshold, the distance consistency map is sampled using a Gaussian distribution; when the sampling number is greater than or equal to the preset number threshold, the distance consistency map is sampled using a distance linear distribution.
[0012] Furthermore, the objective function of the measurement error classification interpretation module is:
[0013] l(f(z),g(z′))=(f(z)-g(z′)) 2
[0014]
[0015] Among them, l(f(z),g(z′)) is the error of the measurement error classification and interpretation module in the Z neighborhood, f(z) is the output of the measurement error classification module in the Z neighborhood, g(z′) is the output of the measurement error classification and interpretation module in the Z neighborhood, and w LIME is the neuron parameter of the spatial error classification explanation module, and R(g) is the regularization term of the measurement error classification explanation module.
[0016] Furthermore, the measurement error classification sub-model incorporating the interpretable mechanism is trained based on the following loss function:
[0017] Loss = -αCE(Y c′ ,Y c )+β|V * -V|
[0018] Among them, Loss is the loss function of the measurement error classification explanation module, α and β are preset parameters, and Y c′ is the dose error category label, Y c is the dosage error category output by the metering error classification explanation module, CE() represents the cross entropy loss function, and V * is the dose error classification explanation label, and V is the dose error classification explanation output by the metering error classification explanation module.
[0019] Compared with the prior art, the radiotherapy quality control error analysis method based on an interpretable neural network provided by the present invention has at least the following beneficial effects:
[0020] The quality control method for intensity modulated radiotherapy based on artificial neural networks can achieve efficient modeling and generation of complex data. Different from traditional quality control methods, it can not only effectively handle the noise and changes in medical imaging data, detect the errors it misses, but also classify the root causes of the errors. By integrating the explainability mechanism into the efficient modeling capability of neural networks, it aims to achieve accurate modeling and error detection and classification of radiotherapy planning data, provide real-time quality assurance and intervention means for medical teams, and thus improve the accuracy and safety of radiotherapy. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:
[0022] Figure 1 It is a flowchart of a radiotherapy quality control error analysis method based on an interpretable neural network according to some embodiments of this specification;
[0023] Figure 2 is a schematic diagram of the structure of an error classification model incorporating an explainable mechanism according to some embodiments of this specification;
[0024] Figure 3 is a schematic diagram of the structure of a measurement error classification module according to some embodiments of this specification;
[0025] Figure 4 is a schematic diagram of the structure of a spatial error classification module according to some embodiments of this specification;
[0026] Figure 5is a schematic diagram of a dose difference histogram according to some embodiments of the present specification;
[0027] Figure 6a is a schematic diagram of a distance consistency diagram in the X-axis direction according to some embodiments of this specification;
[0028] Figure 6b It is a schematic diagram of a distance consistency diagram along the Y-axis direction shown in some embodiments of the present specification. DETAILED DESCRIPTION
[0029] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.
[0030] Figure 1 is a flow chart of a radiotherapy quality control error analysis method based on an interpretable neural network according to some embodiments of this specification, such as Figure 1 As shown, a radiotherapy quality control error analysis method based on an interpretable neural network may include the following process.
[0031] Step 110, establish and train an error classification model that incorporates an explainable mechanism.
[0032] Figure 2 is a schematic diagram of the structure of an error classification model incorporating an interpretable mechanism according to some embodiments of this specification, such as Figure 2 As shown, the error classification model incorporating an interpretable mechanism includes a measurement error classification sub-model incorporating an interpretable mechanism and a spatial error classification sub-model incorporating an interpretable mechanism.
[0033] In some embodiments, the metrology error classification submodel incorporating an interpretable mechanism includes a metrology error classification module and a metrology error classification explanation module, wherein the metrology error classification module is used to determine the dose error category based on the dose difference histogram (DDH), for example, no error (1), incorrect multi-leaf collimator transmission (±1%) (2, 3), monitor (MU) scale error (±1% and ±2%) (4, 5, 6, 7). The metrology error classification explanation module is used to generate dose error classification explanations corresponding to the dose error categories.
[0034] The measurement error classification module can be an artificial neural network integrated with the attention mechanism. Figure 3is a schematic diagram of the structure of the measurement error classification module according to some embodiments of this specification, such as Figure 3 As shown, in some embodiments, the metering error classification module includes an input layer, a self-attention layer, a residual connection layer, a normalization layer, a fully connected layer and an output layer, wherein the number of neurons in the input layer is equal to the number of columns of the dose difference histogram. In the self-attention layer, the attention weight between each position is calculated to distinguish the impact of different dose differences on error classification. The residual connection layer and the normalization layer are added to improve the stability of model training. The fully connected layer consists of a hidden layer with a size of 32. The activation function selects the sigmoid function, and the loss function is the cross entropy function. The size of the output layer is determined based on the total number of dose error categories.
[0035] In some embodiments, a spatial error classification sub-model incorporating an interpretable mechanism includes a spatial error classification module and a spatial error classification explanation module, wherein the spatial error classification module is used to determine the spatial error category based on a distance consistency map, and the spatial error classification explanation module is used to generate a spatial error classification explanation corresponding to the spatial error category.
[0036] Figure 4 is a schematic diagram of the structure of a spatial error classification module according to some embodiments of this specification, such as Figure 4 As shown, in some embodiments, the spatial error classification module includes an input layer, 3 convolutional layers, a pooling layer, a self-attention layer, a fully connected layer, and an output layer. The size of the input layer is 201*201*2, and it contains a distance consistency map (DTA) in both the x and y axis directions. Construct a network structure residual structure having an input layer, a convolutional layer, a pooling layer, a self-attention layer, a fully connected layer, and an output layer. In the convolutional layer, three convolution kernels of 33×33, 17×17, and 3×3 are used respectively, and the ReLU activation function is applied; then the maximum pooling operation is used in the pooling layer to reduce the spatial dimension; the residual network adds a jump connection link to solve the problem of gradient disappearance in ordinary CNN. Then the feature map output by the pooling layer is converted into a one-dimensional vector, connected to the fully connected layer, and finally the softmax output layer is used for classification. The spatial error category consists of 8 categories: no error (1), incorrect source size (2), 1mm MLC leaf group overtravel or undertravel (3, 4), 2mm single MLC leaf overtravel or undertravel (5, 6), 1mm device misalignment error (in the x or y direction, respectively) (7, 8).
[0037] Compared with gamma analysis, the dual neural network error classification system of the metrological error classification submodel and the spatial error classification submodel can well determine the type of error. Professionals can adjust the system more conveniently according to the type of error to eliminate the error between the calculated two-dimensional dose distribution and the reference dose distribution.
[0038] The measurement error classification module and the spatial error classification module are composed of neural networks, which are typical black box models. In order to better understand the basis of classification and ensure the accuracy of classification, a classification explanation module is added to explain the classification model. Here, the improved LIME (Local Interpretable Model-agnostic Explanations) is used to perform interpretable analysis on the measurement error classification module. LIME locally fits the classification model through a simple model with good interpretability.
[0039] In some embodiments, since the sampling probability is considered to be greater when the point is closer to the Z point, the influence of the distance weight on the training of the model g can be discarded in the objective function, and the objective function of the measurement error classification explanation module is:
[0040] l(f(z),g(z′))=(f(z)-g(z′)) 2
[0041]
[0042] Among them, l(f(z),g(z′)) is the error of the measurement error classification and interpretation module in the Z neighborhood, f(z) is the output of the measurement error classification module in the Z neighborhood, g(z′) is the output of the measurement error classification and interpretation module in the Z neighborhood, and w LIME is the neuron parameter of the spatial error classification explanation module, and R(g) is the regularization term of the measurement error classification explanation module. l(f(z), g(z′)) means that by minimizing the expected domain error between the original model f and the fitted model g, the model g is made to fit the original model f better in the neighborhood of Z, where l(f(z), g(z′)) represents the error in the neighborhood of Z. The regularization term is on the right to ensure the simplicity of the new model g. The simpler it is, the more interpretable the model g is.
[0043] In some embodiments, the metrology error classification sub-model incorporating an interpretable mechanism is trained based on the following loss function:
[0044] Loss = -αCE(Y c′ ,Y c )+β|V * -V|
[0045] Among them, Loss is the loss function of the measurement error classification explanation module, α and β are preset parameters, and Y c′ is the dose error category label, Y c is the dosage error category output by the metering error classification explanation module, CE() represents the cross entropy loss function, and V *is the dose error classification explanation label, and V is the dose error classification explanation output by the metering error classification explanation module.
[0046] The objective function of the spatial error classification explanation module is consistent with the objective function of the error classification explanation module. The loss function of the spatial error classification sub-model that incorporates the interpretable mechanism is consistent with the loss function of the measurement error classification sub-model that incorporates the interpretable mechanism, and will not be repeated here.
[0047] For dose error, the dose error classification explanation may include the importance of each column in the dose difference histogram for making a dose error classification judgment. The output is in the form of a one-dimensional vector whose length is consistent with the number of columns in the dose difference histogram.
[0048] For spatial errors, the spatial error classification explanation can include the importance of each pixel block in making a spatial error category judgment. The output is also in the form of a one-dimensional vector, and the length is consistent with the length of the encoding result generated by encoding the distance consistency map, that is, consistent with the number of pixel blocks.
[0049] The sample data set can be divided into training set, validation set and test set in the ratio of 16:2:1, and the error classification model incorporating the explainable mechanism can be trained.
[0050] Step 120: Obtain dose distribution data pairs to be classified.
[0051] The dose distribution data to be classified includes reference dose distribution information and measured dose distribution information. The IMRT plan for a radiotherapy patient includes a reference dose distribution without error and noise calculated using an internal dose calculation algorithm.
[0052] Step 130 , extracting the dose difference histogram and distance consistency graph corresponding to the dose distribution data to be classified.
[0053] Figure 5 is a schematic diagram of a dose difference histogram according to some embodiments of the present specification, such as Figure 5 As shown, in some embodiments, extracting a dose difference histogram corresponding to a pair of dose distribution data to be classified includes:
[0054] Determine a pixel error corresponding to each pixel based on the dose distribution data to be classified, including reference dose distribution information and measured dose distribution information;
[0055] Determine an error pixel based on a pixel error corresponding to each pixel and a preset error range;
[0056] Determine the number of error pixels corresponding to each error segment of a preset error range based on the pixel errors corresponding to the error pixels;
[0057] Based on the number of error pixels corresponding to each error segment within a preset error range, a dose difference histogram corresponding to the dose distribution data to be classified is established.
[0058] For example, the pixel error can be calculated based on the following formula:
[0059]
[0060] in, express Reference dose distribution at the pixel point, express Measured dose distribution at a pixel.
[0061] The vertical axis of the dose difference histogram is the number of error pixels corresponding to the error segment. As an example only, the dose difference histogram is mainly used to analyze the low gradient part. The left and right boundaries of the horizontal axis are set to ±10%, that is, only the part with an error between ±10% is analyzed. Figure 1 The width of the bins is 0.5%, so the dose difference histogram Figure 1 There are 22 columns in total.
[0062] In some embodiments, extracting a distance consistency map corresponding to a pair of dose distribution data to be classified includes:
[0063] For each radiation point, determining an associated radiation point corresponding to the radiation point, wherein the associated radiation point is the radiation point with the same radiation dose as the radiation point and closest to the radiation point;
[0064] Based on the relative positional relationship between each radiation point and the radiation point associated with the radiation point, a distance consistency map corresponding to the dose distribution data pair to be classified is generated.
[0065] Figure 6a is a schematic diagram of a distance consistency diagram in the X-axis direction according to some embodiments of this specification, Figure 6b is a schematic diagram of a distance consistency diagram along the Y axis according to some embodiments of this specification, such as Figure 6a and Figure 6b As shown in the figure, the distance consistency map (DTA) mainly contains features used to represent spatial differences, which are divided into distance consistency maps in the X and Y axes. Figure 6a and Figure 6bThe rectangular points in the represent the detection points. The horizontal and vertical axes in each DTA represent two-dimensional coordinates, and the corresponding coordinate values are expressed as the distance to the nearest position with the same dose as the point. If the nearest point is on the left, the DTA is positive, otherwise it is negative. Since DTA is more sensitive to high-difference areas, all parts with difference gradients below 10% are set to 0. At the same time, in order to reduce the amount of calculation, only the 30×30 cm area in the middle of the DTA is taken, and the scale is set to 0.1 cm, so the size of the DTA here is 301×301×2.
[0066] Step 140, input the dose difference histogram into the measurement error classification sub-model incorporating the interpretable mechanism to determine the dose error category and the dose error classification explanation.
[0067] In some embodiments, the dose difference histogram is input into a dosing error classification sub-model incorporating an interpretable mechanism to determine a dose error category and a dose error classification explanation, including:
[0068] Encode the dose difference histogram to obtain dosimetric feature encoding information;
[0069] The dosimetric feature encoding information is input into the metrological error classification sub-model that incorporates an interpretable mechanism to determine the dose error category and dose error classification explanation.
[0070] As an example only, encoding the dose difference histogram specifically includes encoding the data of each column of the dose difference histogram. For example, the dose difference histogram includes two columns, and the number of error pixels included in the two columns is 2 and 2 respectively. Then the dose difference histogram after normalization encoding is x_ddh=(0.5, 0.5).
[0071] Step 150, input the distance consistency map into the spatial error classification sub-model incorporating the interpretable mechanism to determine the spatial error category and the spatial error classification explanation.
[0072] In some embodiments, the distance consistency map is input into a spatial error classification sub-model incorporating an interpretable mechanism to determine a spatial error category and a spatial error classification explanation, including:
[0073] Perform image segmentation and sampling on the distance consistency map to obtain multiple pixel blocks;
[0074] Multiple pixel blocks are input into a spatial error classification sub-model that incorporates an interpretable mechanism to determine the spatial error category and the spatial error classification explanation.
[0075] Specifically, after the distance consistency map is segmented by U-NET, the segmented pixel blocks are encoded, the selected pixel blocks are represented by 1 and the unselected pixel blocks are represented by 0. According to the selection of N pixel blocks in the distance consistency map, the entire distance consistency map is encoded. For example, the distance consistency map is segmented into 4 pixel blocks, and the first two pixel blocks are selected to work, then the encoding of the distance consistency map is 1100. The role of the encoding here is to convert the distance consistency map from two-dimensional data to one-dimensional data, while reducing the amount of data. The spatial error classification submodel that incorporates an interpretable mechanism can only process one-dimensional data. After converting it to one-dimensional data, it is easier to determine whether the pixel blocks corresponding to the one-dimensional data play a role in the final spatial error classification.
[0076] In some embodiments, sampling the distance consistency map includes:
[0077] When the number of samples is less than the preset number threshold, the distance consistency map is sampled using Gaussian distribution;
[0078] When the number of samples is greater than or equal to a preset number threshold, the distance consistency map is sampled using distance linear distribution.
[0079] By way of example only, the dose difference histogram may be sampled based on the following formula:
[0080]
[0081] Where S represents the number of samples. When the number of samples is less than 200, Gaussian distribution is used, and the sampling point Z is used as the expected sampling point in its vicinity. When the number of samples is greater than 200, the distance linear distribution is used, and π(z′, z) represents the distance between the sampling points z′ and Z. is a uniform probability distribution, which indicates the sampling probability of each point in the original LIME. It means integrating the product of uniform distribution probability and distance of all points around Z. Since the sampling probability of this method is linearly related to the distance, it does not consider the specific position of the sampling point, and only focuses on the distance. When the number of samples is low, if this idea is followed, the sampled objects may be concentrated in one area. Therefore, when the number of samples S is small, a two-dimensional Gaussian distribution is used. When there are more data, the Gaussian distribution calculation is more distributed, and then converted to a distance linear distribution.
[0082] Just as an example, based on the intensity modulated radiation therapy plan for brain tumor patients collected in the hospital database, the reference dose distribution information and measured dose distribution information related to each patient are extracted. There may be certain errors between the two. According to the above data, the causes of errors are divided into two categories: dose error and spatial error. The two categories of error causes contain 7 and 8 categories respectively. Then, the measurement error classification module and the spatial error classification module are used to complete the classification of the error causes of the two. The specific application steps are as follows: an intensity modulated radiotherapy plan for a brain tumor patient is imported, and the reference dose distribution information and measured dose distribution information corresponding to the patient's radiotherapy plan are used as inputs, and then the dose difference histogram and the distance consistency map are extracted for the errors between the reference dose distribution information and the measured dose distribution information, and the dose difference histogram is input into the measurement error classification module to determine the dose error category; the distance consistency map is input into the spatial error classification module to determine the spatial error category, and the measurement error classification module and the spatial error classification module are classified and interpreted by the measurement error classification interpretation module and the spatial error classification interpretation module at the same time. First, some preprocessing operations are performed on the dose difference histogram and the distance consistency map: the dose difference histogram is sampled, and the distance consistency map is segmented and sampled to obtain different pixel blocks and encode them. The relative weight of each pixel block is calculated to explain the influence of each pixel block on the classification judgment made, so as to correctly classify the patient's dose distribution error, and finally the types of dosimetric errors and spatial errors and the basis for making error classification judgments are output.
[0083] Since the error classification model with interpretable mechanism can better distinguish the errors of calculated dose and measurement record of quality control of intensity modulated radiotherapy, the output error type takes into account both dosimetric error and spatial error. By analyzing DDH and DTA respectively, the specific error type can be distinguished to help doctors determine the cause of the error and make targeted modifications, which reduces the time required for professional doctors to determine the source of the error. The output interpretable linear equation G can explain and analyze the error classification, which can be used in aspects such as medical AI assisted training and assisting doctors to make more perfect radiotherapy plans. In medical assisted training, since the amount of training set data is generally small during the training process of medical AI, this may cause the trained model to be biased. In some cases, although the output result is correct in the test, the underlying judgment of the model may be different from the real logic, which will result in high accuracy in the test set, but it is likely to be wrong in subsequent practical applications. At this time, it is important to understand how the model makes classifications, so the interpretability of the model is critical. The interpretable part of the error classification model with interpretable mechanism can be used alone, which can well explain all models. It can help understand the underlying logic of model classification and can be used to ensure the correctness of the final model. It can assist doctors to improve patients' intensity-modulated radiotherapy plans more efficiently, guide doctors in underdeveloped medical areas to conduct quantitative analysis on the causes of errors in the calculation and measurement of dose distribution in intensity-modulated radiotherapy plans, improve the quality of primary medical services, and can also be applied to the training and teaching of medical students in related majors. Visualization helps medical students understand the basis for making judgments more easily.
[0084] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.
Claims
1. A radiotherapy quality control error analysis method based on an interpretable neural network, characterized in that: include: Establishing and training an error classification model incorporating an interpretable mechanism, wherein the error classification model incorporating an interpretable mechanism includes a measurement error classification sub-model incorporating an interpretable mechanism and a spatial error classification sub-model incorporating an interpretable mechanism; Acquiring a dose distribution data pair to be classified, wherein the dose distribution data pair to be classified includes reference dose distribution information and measured dose distribution information; Extracting the dose difference histogram and distance consistency diagram corresponding to the dose distribution data to be classified; Inputting the dose difference histogram into the metering error classification sub-model incorporating an interpretable mechanism to determine the dose error category and the dose error classification explanation; Inputting the distance consistency map into the spatial error classification sub-model incorporating an interpretable mechanism to determine the spatial error category and the spatial error classification explanation; The spatial error classification sub-model incorporating an interpretable mechanism includes a spatial error classification module and a spatial error classification interpretation module, wherein the spatial error classification module is used to determine the spatial error category based on the distance consistency map, and the spatial error classification interpretation module is used to generate the spatial error classification interpretation corresponding to the spatial error category; The spatial error classification sub-model incorporating an interpretable mechanism comprises a spatial error classification module and a spatial error classification interpretation module, wherein the spatial error classification module is used to determine the spatial error category based on the distance consistency map, and the spatial error classification interpretation module is used to generate the spatial error classification interpretation corresponding to the spatial error category; Extracting a dose difference histogram corresponding to the dose distribution data to be classified, comprising: Determine a pixel error corresponding to each pixel based on the dose distribution data to be classified, including reference dose distribution information and measured dose distribution information; Determine an error pixel based on a pixel error corresponding to each pixel and a preset error range; Determining the number of error pixels corresponding to each error segment of the preset error range based on the pixel errors corresponding to the error pixels; Based on the number of error pixels corresponding to each error segment of the preset error range, establishing a dose difference histogram corresponding to the dose distribution data to be classified; Extracting a distance consistency map corresponding to the dose distribution data to be classified, including: For each radiation point, determining an associated radiation point corresponding to the radiation point, wherein the associated radiation point is the radiation point with the same radiation dose as the radiation point and is closest to the radiation point; generating a distance consistency map corresponding to the dose distribution data pair to be classified based on the relative position relationship between each of the radiation points and the radiation point associated with the radiation point; The objective function of the measurement error classification interpretation module is: l(f(z),g(z′))=(f(z)-g(z′)) 2 Among them, l(f(z),g(z′)) is the error of the measurement error classification and interpretation module in the Z neighborhood, f(z) is the output of the measurement error classification module in the Z neighborhood, g(z′) is the output of the measurement error classification and interpretation module in the Z neighborhood, and w LIME is the neuron parameter of the spatial error classification and interpretation module, R(g) is the regularization term of the measurement error classification and interpretation module; The measurement error classification sub-model incorporating the interpretable mechanism is trained based on the following loss function: Loss=-αCE(Y c′ ,AND c )+β|V * -V| Among them, Loss is the loss function of the measurement error classification explanation module, α and β are preset parameters, and Y c′ is the dose error category label, Y c is the dosage error category output by the metering error classification explanation module, CE() represents the cross entropy loss function, and V * is the dose error classification explanation label, and V is the dose error classification explanation output by the metering error classification explanation module.
2. The radiotherapy quality control error analysis method based on an interpretable neural network according to claim 1, characterized in that: The measurement error classification module includes an input layer, a self-attention layer, a residual connection layer, a normalization layer, a fully connected layer and an output layer.
3. The radiotherapy quality control error analysis method based on interpretable neural network according to claim 2, characterized in that: The spatial error classification module includes an input layer, 3 convolutional layers, a pooling layer, a self-attention layer, a fully connected layer and an output layer.
4. A radiotherapy quality control error analysis method based on an interpretable neural network according to any one of claims 1 to 3, characterized in that: The dose difference histogram is input into the metering error classification sub-model incorporating an interpretable mechanism to determine the dose error category and the dose error classification explanation, including: Encoding the dose difference histogram to obtain dosimetric feature encoding information; The dosimetric feature coding information is input into the metrology error classification sub-model incorporating an interpretable mechanism to determine the dose error category and dose error classification explanation.
5. A radiotherapy quality control error analysis method based on an interpretable neural network according to any one of claims 1 to 3, characterized in that: The distance consistency map is input into the spatial error classification sub-model incorporating an interpretable mechanism to determine the spatial error category and the spatial error classification explanation, including: Performing image segmentation and sampling on the distance consistency map to obtain a plurality of pixel blocks; The plurality of pixel blocks are input into the spatial error classification sub-model incorporating an interpretable mechanism to determine the spatial error category and the spatial error classification explanation.
6. The radiotherapy quality control error analysis method based on interpretable neural network according to claim 5, characterized in that: Sampling the distance consistency map includes: When the number of samples is less than a preset number threshold, the distance consistency map is sampled using Gaussian distribution; When the sampling number is greater than or equal to a preset number threshold, the distance consistency map is sampled using distance linear distribution.
Citation Information
Patent Citations
Predication method for three-dimensional dose distribution in intensity modulated radiation therapy plan and application of predication method
CN107441637A
Automatic design method of radiotherapy plan
CN110021399A
Methods and systems for radiation dose estimation
CN110022934A
Fine-grained image classification method based on attention transfer mechanism
CN110598029A
Tumor intensity modulated radiation therapy dose prediction method based on deep learning
CN114155934A
