WO2020199836A1 - 食管放射治疗计划中风险器官剂量学评估方法、评估*** - Google Patents

食管放射治疗计划中风险器官剂量学评估方法、评估*** Download PDF

Info

Publication number
WO2020199836A1
WO2020199836A1 PCT/CN2020/077844 CN2020077844W WO2020199836A1 WO 2020199836 A1 WO2020199836 A1 WO 2020199836A1 CN 2020077844 W CN2020077844 W CN 2020077844W WO 2020199836 A1 WO2020199836 A1 WO 2020199836A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature vector
layer
dose
training
geometric feature
Prior art date
Application number
PCT/CN2020/077844
Other languages
English (en)
French (fr)
Inventor
李腾
江大山
刘剑飞
王妍
Original Assignee
安徽大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 安徽大学 filed Critical 安徽大学
Priority to US17/437,845 priority Critical patent/US11515025B2/en
Publication of WO2020199836A1 publication Critical patent/WO2020199836A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G16H20/10—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance relating to drugs or medications, e.g. for ensuring correct administration to patients
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G16H20/40—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance relating to mechanical, radiation or invasive therapies, e.g. surgery, laser therapy, dialysis or acupuncture
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]

Definitions

  • the invention relates to the technical field of esophageal radiotherapy planning, in particular to a dosimetry assessment method and an assessment system for risk organs in an esophageal radiotherapy plan.
  • Esophageal radiotherapy planning is a time-consuming process that requires tedious parameter adjustments to achieve the delivery of the maximum radiation dose value under the planned target volume (PTV) while minimizing the radiation damage to the organ at risk (OAR).
  • An important indicator to guide adjustment is the dose-volume histogram (DVH), which is used to measure the volume percentage of organs receiving different radiation doses.
  • DVH is highly correlated with the geometric relationship between PTV and OAR. Modeling this correlation can significantly reduce the adjustment time, and quickly realize a clinically accepted treatment plan by providing parameters close to the optimal plan.
  • Previous related studies have simulated the correlation between geometric relationships and DVH. In previous work, researchers proposed a three-parameter probability function of distance to describe the function of distance between PTV and OAR and polynomial by fitting the evolution of the skew normal parameters, and simulated their correlation with DVH. There are also workers who calculate the volume fraction of OAR corresponding to different distance PTVs, use the distance target histogram (DTH) to represent the geometric relationship between them, and use the correlation between DTH and DVH through component analysis and support vectors Regression modeling.
  • DTH distance target histogram
  • the purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for evaluating the dosimetry of risk organs in an esophageal radiotherapy plan based on a deep confidence network.
  • the deep learning method is beneficial to reduce the burden on the physicist and shorten the formulation of radiotherapy.
  • the planning time and effective specificity for clinical individuals improve the accuracy of radiotherapy planning.
  • the present invention solves the above-mentioned technical problems through the following technical means: a method for evaluating the dosimetry of risk organs in an esophageal radiotherapy plan, including the following steps:
  • Step 2 Calculate the data information of the esophageal cancer VMAT plan to obtain the distance target histogram
  • Step 3 Use the coding layer in the auto-encoder structure to reduce the dimensionality of the geometric feature vector of the distance target histogram in step 2, to obtain the reduced geometric feature vector;
  • Step 4 Establish a deep confidence network model, iterate until it converges, and complete the training of the deep confidence network model;
  • Step 5 The geometric feature vector after dimensionality reduction is nonlinearly fitted to the correlation between the dose feature vector and the geometric feature vector after dimensionality reduction through the trained depth belief network, and the dimension is the same as that of the geometric feature vector after dimensionality reduction.
  • Eigenvector of dose
  • the distance target histogram is established by calculating the volume percentage of the organ at risk
  • v i OAR represents the i-th voxel of the dangerous organ
  • v k PTV represents the k- th voxel of the planned target area
  • S PTV is the voxel set of the planned target area
  • r(v i OAR , PTV) is the body of the dangerous organ The Euclidean distance from the surface of the planned target area.
  • n coordinate points ((x i ,y i )i ⁇ 1,2,...,n) are selected equidistantly from the x-axis of the curve of the target histogram; the selected n discrete coordinate points are used to express This curve; each coordinate point contains the volume fraction value y i (i ⁇ 1,2,...,n) and the corresponding distance value x i (i ⁇ 1,2,...,n); y i represents the i-th
  • the volume fraction value of the coordinate, y i (i ⁇ 1,2,...,n) are vector elements used to construct the geometric feature vector of the n-dimensional distance target histogram; xi represents the distance value of the i-th coordinate.
  • the encoder is composed of several coding layers and several decoding layers.
  • V i represents the i-th layer coding layer as an input through i-1 times reduced feature vector dimensionality, H i expressed by the i-th reduction feature vector dimensionality, H i i-th layer coding layer as an output, v i ⁇ ⁇ 0,1 ⁇ , W i represents the weight of the i layer weight matrix layer encoding;
  • C i is the i-th layer encoded deviation layer; where h i is smaller than the dimension v i,
  • the decoding layer is a process opposite to the encoding layer, which reconstructs the original input feature vector by increasing the dimension of the input feature vector;
  • the deep belief network model in step 4 includes a stack of several Boltzmann machines; the training process is to pre-train each Boltzmann machine first, and then combine the pre-trained Boltzmann machines. Man machines are stacked as a whole of the deep belief network for network fine-tuning training.
  • the training dose-volume histogram (DVH)
  • the distance target histogram for training the dose feature vector of the training dose-volume histogram and the geometric feature vector of the training distance target histogram are calculated; and the dimensionality reduction training is obtained through the decoding layer of the autoencoder structure
  • each Boltzmann machine structure Pre-train the structure of each Boltzmann machine; each Boltzmann machine structure includes the corresponding visible layer v and hidden layer h;
  • the iterative steps of the Boltzmann machine in the training process are: first map the geometric feature vector of the reduced distance target histogram in the visible layer to the hidden layer, and then use the vector in the hidden layer to reconstruct the new distance
  • the geometric feature vector of the target histogram, and then the reconstructed geometric feature vector of the target histogram is mapped to the hidden layer.
  • This process is represented as a loop, and each Boltzmann machine repeats three loop updates during the pre-training process. Weight matrices in the visible and hidden layers of the Boltzmann machine;
  • the objective loss function is:
  • dt is the geometric feature vector of the training distance target histogram after dimensionality reduction, and is also the input vector of the visible layer of each Boltzmann machine structure.
  • ⁇ s is the constraint coefficient determined empirically in the experiment
  • the function Y represents the visible layer in the Boltzmann machine structure
  • the energy function of the process to the hidden layer is:
  • Wi j represents the weight between the visible vector and the hidden vector.
  • the visible layer has i units of the visible vector (the vector dt of the distance target histogram feature for training after dimensionality reduction), and the hidden layer has j hidden vector units, a i represents the offset of the i-th visible vector unit, and b j represents the offset of the j-th hidden vector unit;
  • Y′ is the reverse process of Y, which means that the vector in the hidden layer is used to reconstruct the new distance target histogram
  • 1 means that the sparsity expression in the formula is guaranteed by a non-zero penalty factor,
  • 1 means the L1 regularity of the weight
  • each Boltzmann machine goes through an iterative loop.
  • its objective loss function is less than 0.05, it tends to converge stably and ends the training process while saving the parameter values of the Boltzmann machine to obtain the pre-trained Boltzmann machine.
  • Man machine model a number of pre-trained Boltzmann machine models are superimposed into a pre-trained deep confidence network.
  • the loss function uses the mean square error loss function:
  • the function D represents a pre-trained deep belief network model stacked with three Boltzmann machine structures; dt is the geometric feature vector of the distance target histogram for training after dimensionality reduction; dv_i is the dimensionality reduction step 3
  • said step six layer reconstructed by decoding the n-dimensional feature vector dose, the reconstructed n-dimensional feature vector as a new dose dose - volume histogram ordinate coordinate points, i.e. m 'j (j ⁇ 1,2,...,n), take the dose fraction value n j (j ⁇ 1,2,...,n) as the abscissa of the coordinate points, and get new n coordinate points ((n j ,m' j )j ⁇ 1,2,...,n); The curve drawn by connecting these n coordinate points is the dose-volume histogram for predicting dangerous organs.
  • the invention also discloses a dosimetry assessment system for organs at risk in an esophageal radiotherapy plan, including:
  • the coding layer is used to reduce the dimension of the geometric feature vector of the distance target histogram to obtain the reduced geometric feature vector
  • the geometric feature vector after dimensionality reduction is nonlinearly fitted to the correlation between the dose feature vector and the geometric feature vector after dimensionality reduction through the trained deep confidence network, and the relationship between the geometric feature vector and the geometric feature vector after dimensionality reduction is obtained.
  • the decoding layer is used to reconstruct the dose feature vector to obtain the dose feature vector with the same dimension as the geometric feature vector before dimensionality reduction.
  • the deep belief network model includes a stack of several Boltzmann machines; the training process is to first pre-train each Boltzmann machine, and then stack the pre-trained Boltzmann machines as the depth The whole belief network is trained for network fine-tuning.
  • the present invention samples the dose values of 50 coordinate points of the DVH at the beginning of the corresponding experiment, and then obtains 50 coordinate points, and each coordinate point corresponds to the volume fraction value under the corresponding dose value.
  • the advantage of the present invention is that when predicting the DVH of the OAR for a new patient, the DTH of the OAR is first calculated, and then the 50-dimensional DTH feature vector can be established by sampling the DTH curve, and then the feature vector can be simplified to three-dimensional by the autoencoder Feature vector. The corresponding 3D DVH feature vector can be obtained by mapping the corresponding deep belief network model. Finally, the decoding layer in the autoencoder is used to reconstruct the DVH feature vector, and finally the predicted OAR DVH is obtained.
  • the present invention is an effective application of deep learning technology.
  • the OAR of the esophageal radiotherapy plan is performed according to the geometric relationship between OAR and PTV Automatic evaluation of the dose.
  • the model method can achieve accurate DVH prediction and can provide near-optimal parameters for esophageal treatment planning, which can significantly shorten the time for formulating esophageal cancer radiotherapy plans to reduce the burden on physical radiotherapists.
  • Figure 1 is a structural flowchart of Embodiment 1 of the present invention.
  • Figure 2 is a dose-volume histogram of Example 1 of the present invention.
  • Fig. 3 is a histogram of a distance target according to Embodiment 1 of the present invention.
  • Fig. 4 is a structural diagram of an automatic encoder according to Embodiment 1 of the present invention.
  • Fig. 5 is a structural diagram of a deep confidence network according to Embodiment 1 of the present invention.
  • FIG. 6 is a comparison diagram of the effect of the model predicted dose volume histogram and the clinical actual dose volume histogram of the spinal cord of Example 2 of the present invention.
  • the upper boundary is the curve with the upper position
  • the lower boundary is the curve with the lower position.
  • Fig. 7 is an effect comparison diagram of the predicted dose volume histogram of the heart model and the clinical actual dose volume histogram of Example 2 of the present invention.
  • the upper boundary is the curve with the upper position
  • the lower boundary is the curve with the lower position.
  • FIG. 8 is a comparison diagram of the effect of the model predicted dose volume histogram and the clinical actual dose volume histogram of the right lung in Example 2 of the present invention.
  • the upper boundary is the curve with the upper position
  • the lower boundary is the curve with the lower position.
  • Fig. 9 is an effect comparison diagram of the model predicted dose volume histogram and the clinical actual dose volume histogram of the left lung in Example 2 of the present invention.
  • the upper boundary is the curve with the upper position
  • the lower boundary is the curve with the lower position.
  • This embodiment discloses a method for evaluating the dosimetry of organs at risk in an esophageal radiotherapy plan, including the following steps:
  • Step 1 From collecting VMAT plan data information of patients with esophageal cancer, this embodiment uses the Matlab-based radiology research platform CERR package to extract CT images and structural contour images from the original radiotherapy data;
  • Step 2 Calculate the VMAT plan data for esophageal cancer to obtain the distance target histogram (DTH), and construct the geometric feature vector of the distance target histogram through the distance target histogram;
  • DTH distance target histogram
  • DTH is based on structural contour images and CT images (CT images are slices, cross-sectional views, each cross-sectional view has a thickness of 3mm, representing the z-axis height, and the cross-sectional view shows the dangerous organs and planned targets)
  • CT images are slices, cross-sectional views, each cross-sectional view has a thickness of 3mm, representing the z-axis height, and the cross-sectional view shows the dangerous organs and planned targets
  • the geometric structure information of the area, the spatial geometric distance between the dangerous organ and the planned target area and the corresponding overlap volume of the corresponding distance are calculated, and the DTH is drawn.
  • v i OAR represents the i-th voxel of the dangerous organ
  • v k PTV represents the k- th voxel of the planned target area
  • S PTV is the voxel set of the planned target area
  • r(v i OAR , PTV) is the body of the dangerous organ Euclidean distance from the surface of the target area
  • the distance target histogram uses the distance target histogram to express the contours of the planned target area through equidistant expansion or get different distances of the expansion or contraction of the planned target area contours, and calculate the volume percentage of the overlapping area between the dangerous organs and the contours of the expanded plan target area; use different The percentage of overlap volume in the distance indicates the geometric relationship between the planned target area and the dangerous organ. In particular, when the distance value is a negative value, it is used to indicate the invasion of the dangerous organ by the planned target area.
  • 50 coordinate points ((x i ,y i )i ⁇ 1,2,...,50) are selected equidistantly from the x-axis of the curve of the target histogram; the selected 50 discrete coordinate points are used to express This curve; each coordinate point contains the volume fraction value y i (i ⁇ 1,2,...,50) and the corresponding distance value x i (i ⁇ 1,2,...,50); y i means the i-th
  • the volume fraction value of the coordinate, y i (i ⁇ 1,2,...,50) is a vector element used to construct the geometric feature vector of the 50-dimensional distance target histogram; xi represents the distance value of the i-th coordinate;
  • Step 3 Use the encoding layer in the autoencoder to reduce the dimensionality of the geometric feature vector of the distance target histogram in step 2, and obtain the reduced geometric feature vector, that is, the dimensionality reduction feature vector of the distance target histogram;
  • the autoencoder of this embodiment includes 4 coding layers and 3 decoding layers.
  • the number in the figure refers to the number of neurons in a certain layer of the encoder or decoder.
  • Each coding layer is reduced by a fully connected layer to compress the dimensionality of the input.
  • the activation function of each neuron in this layer is:
  • V i represents the i-th layer coding layer as an input through i-1 times reduced feature vector dimensionality, H i expressed by the i-th reduction feature vector dimensionality, H i i-th layer coding layer as an output, v i ⁇ ⁇ 0,1 ⁇ , W i represents the weight of the i layer weight matrix layer encoding;
  • C i is the i-th layer encoded deviation layer; where h i is less than the dimension v i, so the encoding can be reduced input feature layer The dimension of the vector;
  • the decoding layer is a process opposite to the encoding layer, which reconstructs the original input feature vector by increasing the dimension of the input feature vector;
  • h j represents the eigenvector of the j-th layer of decoding layer as the input after j-1 dimension upgrade
  • v j represents the eigenvector of the j- th layer of decoding layer as the output after the j-th dimension ascending
  • W j represents the weight matrix of the j-th decoding layer
  • b j is the deviation value of the j-th decoding layer, where the dimension of v j is greater than the dimension of h j , so the decoding layer can Play the role of feature reconstruction.
  • Step 4 Establish a deep confidence network model, iterate until it converges, and complete the training of the deep confidence network model;
  • the DVH in this embodiment uses the Matlab-based radiology research platform CERR package, and directly uses the built-in function to extract. You can also extract the dose information from the dose map and the volume information at the corresponding position, and draw DVH.
  • the x-axis of DVH represents the percentage of the dose received by the dangerous organ, and the y-axis represents the volume percentage, indicating that the dose received by the volume is equal to or greater than x The dose indicated on the axis.
  • the present invention selects 50 coordinate points ((n j ,m j )j ⁇ 1,2,...,50) equidistantly from the x-axis of the curve x-axis of the dose-volume histogram (DVH) for training; use the selected 50 Discrete coordinate points to express this curve; each coordinate point contains the volume fraction value m j (j ⁇ 1,2,...,50) and the corresponding dose fraction value n j (j ⁇ 1,2,...,50) ; M j represents the volume fraction value of the j-th coordinate, m j (j ⁇ 1,2,...,50) is a vector element used to construct the dose feature vector of the 50-dimensional training dose-volume histogram; n j represents The dose fraction value of the j coordinate;
  • the deep belief network includes a stack of several Boltzmann machines, and this embodiment uses three Boltzmann machines.
  • the training process is to first pre-train each Boltzmann machine, and then stack the pre-trained Boltzmann machine as the whole of the deep confidence network for network fine-tuning training;
  • Each Boltzmann machine structure includes the corresponding visible layer v and hidden layer h.
  • the iterative steps of the Boltzmann machine in the training process are: first map the geometric feature vector of the reduced distance target histogram in the visible layer to the hidden layer, and then use the vector in the hidden layer to reconstruct the new distance
  • the geometric feature vector of the target histogram, and then the reconstructed geometric feature vector of the target histogram is mapped to the hidden layer.
  • This process is represented as a loop, and each Boltzmann machine repeats three loop updates during the pre-training process. Weight matrices in the visible and hidden layers of the Boltzmann machine;
  • the objective loss function is:
  • dt is the geometric feature vector of the training distance target histogram after dimensionality reduction, and is also the input vector of the visible layer of each Boltzmann machine structure.
  • ⁇ s is the constraint coefficient determined empirically in the experiment
  • the function Y represents the visible layer in the Boltzmann machine structure
  • the energy function of the process to the hidden layer is:
  • W ij represents the weight between the visible and the hidden vector weight vectors, a total visible visible layer vector (after training using dimensionality reduction target histogram feature vector distance dt) th cell i
  • V i represents the i-th visible layer visible vectors Unit
  • the hidden layer has a total of j hidden vector units
  • h j represents the j-th unit of the hidden vector of the hidden layer
  • a i represents the offset of the i-th visible vector unit
  • b j represents the offset of the j-th hidden vector unit
  • Y′ is the reverse process of Y, which means that the vector in the hidden layer is used to reconstruct the new distance target histogram
  • the number of visible vectors i in the visible layer of each Boltzmann machine structure is different. Because three Boltzmann machine structures are stacked together at the end, the number of vector units in the hidden layer of the first layer j is the number of vector units in the visible layer of the second layer, that is, the number of outputs of the first layer is the number of inputs of the second layer.
  • the number of visible vector units in the visible layer is 30, and the number of hidden vector units in the hidden layer is 25; in the same way, for the second Boltzmann machine Structure, the number of visible vector units of the visible layer is 25, the number of hidden vector units of the hidden layer is 10; and so on, the third Boltzmann machine structure, the number of visible vector units of the visible layer is 10 , The number of hidden vector units in the hidden layer is 5.
  • each Boltzmann machine goes through an iterative loop.
  • its objective loss function is less than 0.05, it tends to converge stably and ends the training process while saving the parameter values of the Boltzmann machine to obtain the pre-trained Boltzmann machine.
  • Man machine model a number of pre-trained Boltzmann machine models are superimposed into a pre-trained deep confidence network.
  • the loss value of the mean square error loss function is less than 0.05, it tends to stabilize and converge, save the parameter values, and complete the training of the deep confidence network model.
  • the function D represents a pre-trained deep belief network model stacked with three Boltzmann machine structures; dt is the geometric feature vector of the distance target histogram for training after dimensionality reduction; dv_i is the dimensionality reduction step 3
  • the ith component of the dose feature vector of the training dose-volume histogram; the dose feature vector of the training dose-volume histogram after dimensionality reduction in this embodiment is the feature vector of the three-dimensional dose-volume histogram, so there are three Components, the three components need to build three different deep belief networks, and the predicted output of the three deep belief networks corresponds to a set of reduced-dimensional dose feature vectors.
  • Step 5 The dimensionality-reduced geometric feature vector is used to non-linearly fit the correlation between the dose feature vector and the dimensionality-reduced geometric feature vector through the trained deep confidence network, and the dimensionality-reduced dose feature vector is predicted;
  • the nonlinear correlation between the reduced geometric feature vector and the reduced dose feature vector is composed of three nonlinear functions to form three deep confidence network models to fit the function; the reduced three-dimensional dose feature vector and the reduced
  • the nonlinear correlation between the geometric feature vectors of the three dimensions after dimension is:
  • the acquisition of dvp of the present invention exists in the processing of VMAT plan data information of patients with esophageal cancer for verification or clinical use.
  • the dose-volume histogram obtained by the method of the present invention is compared with the dose-volume histogram obtained by the prior art to further verify the accuracy of the method of the present invention. If the set requirements are not met, the deep confidence network model is further trained.
  • the difference between this embodiment and the foregoing embodiment is that a total of 80 VMAT plans for esophageal cancer were collected and divided into 8 evenly, each with 10 plans.
  • the first set of experiments select the first one as the test set data, the rest as the training set data, the second set of experiments select the second as the test set data, and the rest as the training set data, and so on, complete eight sets of experiments, Calculate the average of eight groups of experiments as the final experimental result.
  • the dose information corresponding to the patient's dangerous organs is calculated from the collected high-quality esophageal cancer VMAT plans.
  • the target dangerous organs are the left lung, right lung, heart, and spinal cord. Calculate the corresponding dose-volume histogram respectively.
  • Each dangerous organ gets 80 sets of DVH. Sampling 50 coordinate points in the DVH curve at equal dose intervals, each coordinate point contains a dose fraction value and a corresponding volume fraction value. Select the volume fraction value to construct a 50-dimensional DVH feature vector.
  • the distance target histograms of the patient's target dangerous organs are calculated separately to describe the geometric relationship between the radiation target area and the dangerous organ.
  • Each dangerous organ gets 80 groups of DTH. Select 50 coordinate points with equal distance intervals from the DTH curve. Each coordinate point contains the volume fraction value and the corresponding distance value. Select the volume fraction value to construct a 50-dimensional DTH feature vector.
  • the present invention When predicting the DVH of the OAR for a new patient, the present invention first calculates the DTH of the OAR, and secondly, a 50-dimensional DTH feature vector can be established by sampling the DTH curve, and then the feature vector is simplified into a three-dimensional feature vector by an auto encoder. The corresponding 3D DVH feature vector can be obtained by mapping the corresponding deep belief network model. Finally, the decoding layer in the autoencoder is used to reconstruct the DVH feature vector, and finally the predicted OAR DVH is obtained.
  • the present invention is an effective application of deep learning technology.
  • the esophagus is based on the geometric relationship between OAR and PTV. Automatic evaluation of OAR dose of radiotherapy plan.
  • the model method can achieve accurate DVH prediction and can provide near-optimal parameters for esophageal treatment planning, which can significantly shorten the time for formulating esophageal cancer radiotherapy plans to reduce the burden on physical radiotherapists.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • Theoretical Computer Science (AREA)
  • Public Health (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Urology & Nephrology (AREA)
  • Surgery (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Chemical & Material Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medicinal Chemistry (AREA)
  • Probability & Statistics with Applications (AREA)
  • Radiation-Therapy Devices (AREA)
  • Image Analysis (AREA)

Abstract

一种食管放射治疗计划中风险器官剂量学评估方法以及***，包括以下步骤：集食管癌病人的VMAT计划数据信息，得到DTH；对DTH的几何特征向量进行降维，得降维后的几何特征向量；建立深度置信网络模型，并完成深度置信网络模型的训练；非线性地拟合剂量特征向量和降维后几何特征向量之间的相关性，得到与降维后的几何特征向量的维度相同的剂量特征向量；通过自动编码器结构的解码层将从步骤五的剂量特征向量重构，得到与降维前的几何特征向量的维度相同的剂量特征向量，最终得到预测危险器官的DVH。该方法具有显著缩短制定食管癌放疗计划时间以达到减轻物理放疗师的负担的优点。

Description

食管放射治疗计划中风险器官剂量学评估方法、评估*** 技术领域
本发明涉及食管放射治疗计划技术领域，尤其是涉及食管放射治疗计划中风险器官剂量学评估方法、评估***。
背景技术
食管放射治疗计划是一个耗时的过程，需要繁琐的参数调整以在计划靶区(PTV)下实现最大放射剂量值的传递，同时最小化危险器官(OAR)的辐射损伤。指导调整的一个重要指标是剂量-体积直方图(DVH)，它用于测量接收不同辐射剂量下的器官的体积百分数。
DVH与PTV和OAR之间的几何关系高度相关。对这种相关性建模可以显著减少调整时间，通过提供接近最佳的计划的参数以快速实现临床接受的治疗计划。之前的相关研究已经模拟了几何关系和DVH之间的相关性。以往工作中研究者通过拟合偏斜法线参数的演化，提出了距离的三参数概率函数来描述PTV和OAR与多项式之间距离的函数，并模拟了其与DVH的相关性。也有工作者通过计算由不同距离PTV相应的OAR的体积分数，利用距离目标直方图(DTH)来表示他们之间的几何关系，并使用DTH和DVH之间的相关性通过成分分析法和支持向量回归建模。
然而，现有的研究主要集中在提取几何特征的线性模型和传统的机器学习方法来模拟他们之间的相关性，这些方法都有着模型的鲁棒性不高而且预测精度较低适用性不强等缺点。这些问题一方面增加了物理治疗师的工作量，另一方面根据相关的临床人群统计标准的不同对放射治疗方案的 质量审核标准也不同，从而导致计算的DVH的精度较低。
发明内容
本发明的目的在于克服现有技术的不足，提供了一种基于深度置信网络的食管放射治疗计划中风险器官剂量学评估方法，用深度学习的方法，有利于减轻物理师负担，缩短了制定放疗计划的时间以及有效的针对临床个体的特异性提高放射治疗计划的精度。
本发明通过以下技术手段实现解决上述技术问题的：一种食管放射治疗计划中风险器官剂量学评估方法，包括以下步骤：
步骤一、收集食管癌病人的VMAT计划数据信息，包括提取出的CT图像、结构轮廓图像；
步骤二、对食管癌VMAT计划数据信息计算，得到距离目标直方图；
步骤三、采用自动编码器结构中的编码层对步骤二中的距离目标直方图的几何特征向量进行降维，得到降维后的几何特征向量；
步骤四、建立深度置信网络模型，迭代直至其收敛，完成深度置信网络模型的训练；
步骤五、降维后的几何特征向量通过训练后的深度置信网络非线性地拟合剂量特征向量和降维后几何特征向量之间的相关性，得到与降维后的几何特征向量的维度相同的剂量特征向量；
步骤六、通过自动编码器结构的解码层将从步骤五的剂量特征向量重构，得到与降维前的几何特征向量的维度相同的剂量特征向量，最终得到预测危险器官的剂量-体积直方图。
优选地，所述步骤二中通过计算风险器官的体积百分数来建立距离目 标直方图，
Figure PCTCN2020077844-appb-000001
其中v i OAR表示危险器官的第i个体素；v k PTV表示计划靶区的第k个体素；S PTV为计划靶区的体素集合；r(v i OAR，PTV)为危险器官的体素到计划靶区表面的欧式距离。
优选地，从距离目标直方图的曲线x轴等距离选取n个坐标点((x i,y i)i∈1,2,…,n)；用选取的这n个离散坐标点来表达出这条曲线；每个坐标点包含体积分数值y i(i∈1,2,…,n)和相应的距离值x i(i∈1,2,…,n)；y i表示第i个坐标的体积分数值，y i(i∈1,2,…,n)为向量元素，用以构建n维距离目标直方图的几何特征向量；x i表示第i个坐标的距离值。
优选地，编码器由若干个编码层和若干个的解码层构成。
每个编码层通过一个完全连接层缩小来压缩输入的维数，此层中每个神经元的激活函数为：
P(h i＝1|v i)＝sigmoid(c i+W iv i)
其中v i表示第i层编码层中作为输入的经i-1次降维后的特征向量，h i表示第i层编码层中作为输出的经i次降维后的特征向量，h i,v i∈{0,1},W i表示第i层编码层的权重矩阵；c i是第i层编码层的偏差值；这里h i的维数小于v i，
解码层是一个与编码层相反的过程，它通过增加输入特征向量的维数来重建原始输入的特征向量；
P(v j＝1|h j)＝sigmoid(b j+W jh j)
其中h j表示第j层解码层中作为输入的经j-1次升维后的特征向量，v j 表示第i层解码层中作为输出的经j次升维后的特征向量，h j,v j∈{0,1},W j表示第j层解码层的权重矩阵，b j是第j层解码层的偏差值，这里v j的维数大于h j的维数。
优选地，所述步骤四中的深度置信网络模型包括若干个玻尔兹曼机堆叠形成的；训练过程为先对每一个玻尔兹曼机进行预训练，再把预训练好的玻尔兹曼机堆叠起来作为深度置信网络的整体进行网络微调训练。
优选地，首先搜集训练用食管癌病人的VMAT计划数据，作为训练集数据，并从训练集数据中提取出训练用CT图像、训练用结构轮廓图像，计算训练用剂量-体积直方图(DVH)、训练用距离目标直方图；并计算得到训练用剂量-体积直方图的剂量特征向量、训练用距离目标直方图的几何特征向量；并通过自动编码器结构的解码层得到降维后的训练用剂量-体积直方图的剂量特征向量、降维后的训练用距离目标直方图的几何特征向量；
对每一个玻尔兹曼机的结构进行预训练；每一个玻尔兹曼机结构包括相应的可见层v和隐藏层h；
玻尔兹曼机在训练过程中的迭代步骤为：首先将可见层中的降维后的距离目标直方图的几何特征向量映射到隐藏层中，然后用隐藏层中的向量重构新的距离目标直方图的几何特征向量，再将重构的距离目标直方图的几何特征向量映射到隐藏层中，这个过程表示为一个循环，每个玻尔兹曼机在预训练过程中重复三次循环更新玻尔兹曼机可见层与隐藏层中的权重矩阵；
在深度置信网络的每一层中，目标损失函数是：
Figure PCTCN2020077844-appb-000002
式中dt为降维后的训练用距离目标直方图的几何特征向量，也是每一个玻尔兹曼机结构的可见层的输入向量。式中
Figure PCTCN2020077844-appb-000003
表示降维后的训练用距离目标直方图的几何特征向量dt到隐藏层特征向量g的均方误差，α s为实验中经验确定的约束系数，函数Y表示玻尔兹曼机结构中可见层到隐藏层的过程，其能量函数为：
Figure PCTCN2020077844-appb-000004
其中θ＝{W ij,a i,b j}是玻尔兹曼机的参数，他们均为实数。其中Wi j表示可见向量与隐藏向量之间的权重，可见层共有可见向量(降维后的训练用距离目标直方图特征的向量dt,)单元i个，隐藏层共有隐藏向量单元j个，a i表示第i个可见向量单元的偏置，b j表示第j个隐藏向量单元的偏置；
Figure PCTCN2020077844-appb-000005
表示由隐藏向量g提供的输入与重构成新的距离目标直方图的几何特征向量之间的均方误差，Y′是Y的相反过程，表示用隐藏层中的向量重构新的距离目标直方图的几何特征向量；α r||W|| 1表示通过非零的惩罚因子来保证式中的稀疏性表达，||W|| 1表示对深度置信网络每一层的权重矩阵的L1正则化，α r为稀疏系数；
每一个玻尔兹曼机的训练过程通过迭代循环，当其目标损失函数小于0.05，则趋于稳定收敛而结束训练过程同时保存玻尔兹曼机的参数值，得到预训练完成的玻尔兹曼机模型，若干个预训练完成的玻尔兹曼机模型叠加成预训练完成的深度置信网络。
优选地，重新把降维后的训练用距离目标直方图的几何特征向量输入到所述预训练好的深度置信网络模型中进行微调训练，损失函数使用均方误差损失函数:
Figure PCTCN2020077844-appb-000006
当均方误差损失函数损失值小于0.05，则趋于稳定收敛，保存参数值，完成深度置信网络模型的训练；
其中函数D表示三个玻尔兹曼机结构堆叠成的经过预训练的深度置信网络模型；dt为降维后的训练用距离目标直方图的几何特征向量；dv_i为采用步骤三降维后的训练用剂量-体积直方图的剂量特征向量的第i个成分；i个成分需要建立i个不同的深度置信网络，i个深度置信网络的预测输出对应的一组降维后的剂量特征向量。
优选地，所述步骤六通过解码层重构的n维剂量特征向量，把重构的n维剂量特征向量作为新的剂量-体积直方图的坐标点的纵坐标，即m’ j(j∈1,2,…,n),把剂量分数值n j(j∈1,2,…,n)作为坐标点的横坐标，得到新的n个坐标点((n j,m’ j)j∈1,2,…,n)；连接这n个坐标点绘制的曲线即为预测危险器官的剂量-体积直方图。
本发明还公开一种食管放射治疗计划中风险器官剂量学评估***，包括，
编码层，用以对距离目标直方图的几何特征向量进行降维，得到降维后的几何特征向量；
深度置信网络模型，降维后的几何特征向量通过训练后的深度置信网络非线性地拟合剂量特征向量和降维后几何特征向量之间的相关性，得到与降维后的几何特征向量的维度相同的剂量特征向量；
解码层，用以将剂量特征向量重构，得到与降维前的几何特征向量的维度相同的剂量特征向量。
优选地，深度置信网络模型包括若干个玻尔兹曼机堆叠形成的；训练过程为先对每一个玻尔兹曼机进行预训练，再把预训练好的玻尔兹曼机堆叠起来作为深度置信网络的整体进行网络微调训练。
本发明通过对应实验开始采样DVH的50个坐标点的剂量值，进而得到50个坐标点，每个坐标点对应相应剂量值下的体积分数值。
本发明的优点在于：在对于新患者预测OAR的DVH时，首先计算OAR的DTH，其次可以通过对DTH曲线进行采样来建立50维DTH特征向量，然后通过自动编码器将该特征向量简化为三维特征向量。其相应的三维DVH特征向量可以用相应的深度置信网络模型映射得到。最后，使用自动编码器中的解码层来重建DVH特征向量，最终得到预测的OAR的DVH。
本发明是深度学习技术的一次有效应用，相对于提取几何特征的线性模型和传统的机器学习方法来预测食管癌放疗剂量学特征而言，根据OAR与PTV的几何关系进行食管放射治疗计划的OAR的剂量的自动评估。实验证明，该模型方法能够实现准确的DVH预测并且可以为食管治疗计划提供接近最优的参数，这可以显著缩短制定食管癌放疗计划时间以达到减轻物理放疗师的负担。
附图说明
图1为本发明实施例1的结构流程图。
图2为本发明实施例1的剂量-体积直方图。
图3为本发明实施例1的距离目标直方图。
图4为本发明实施例1的自动编码器的结构图。
图5为本发明实施例1的深度置信网络的结构图。
图6为本发明实施例2的脊髓的模型预测剂量体积直方图和临床实际剂量体积直方图效果对比图。图中，上边界是位置在上的曲线，下边界是位置在下的曲线。
图7为本发明实施例2的心脏的模型预测剂量体积直方图和临床实际剂量体积直方图效果对比图。图中，上边界是位置在上的曲线，下边界是位置在下的曲线。
图8为本发明实施例2的右肺的模型预测剂量体积直方图和临床实际剂量体积直方图效果对比图。图中，上边界是位置在上的曲线，下边界是位置在下的曲线。
图9为本发明实施例2的左肺的模型预测剂量体积直方图和临床实际剂量体积直方图效果对比图。图中，上边界是位置在上的曲线，下边界是位置在下的曲线。
具体实施方式
为使本发明实施例的目的、技术方案和优点更加清楚，下面将结合本发明实施例，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
本实施例公开一种食管放射治疗计划中风险器官剂量学评估方法，包括以下步骤：
步骤一、从收集食管癌病人的VMAT计划数据信息，本实施例通过包括使用基于Matlab的放射学研究平台CERR包，从原始放疗数据中提取出 CT图像、结构轮廓图像；
步骤二、对食管癌VMAT计划数据进行计算，得到距离目标直方图(DTH)，并通过距离目标直方图构建距离目标直方图的几何特征向量；
DTH根据结构轮廓图像和CT图像(CT图像为切片图，剖面视图，每一张剖面图都有3mm厚度，代表z轴高度，剖面图中显示有水平几何信息)中表征的危险器官以及计划靶区的几何结构信息，计算危险器官和计划靶区的空间几何距离以及相应距离对应的重叠体积，绘制出DTH。
通过计算危险器官的体积百分数来建立距离目标直方图，
Figure PCTCN2020077844-appb-000007
其中v i OAR表示危险器官的第i个体素；v k PTV表示计划靶区的第k个体素；S PTV为计划靶区的体素集合；r(v i OAR，PTV)为危险器官的体素到计划靶区表面的欧式距离；
使用距离目标直方图来表示通过计划靶区轮廓等距外扩或者得到不同距离的外扩或者收缩计划靶区轮廓，计算危险器官与各外扩计划靶区轮廓的重叠区域的体积百分数；利用不同距离上的重叠体积百分数表示计划靶区和危险器官的几何关系。特别的，当距离值为负值时，以表示计划靶区对危险器官的侵入。
本实施例从距离目标直方图的曲线x轴等距离选取50个坐标点((x i,y i)i∈1,2,…,50)；用选取的这50个离散坐标点来表达出这条曲线；每个坐标点包含体积分数值y i(i∈1,2,…,50)和相应的距离值x i(i∈1,2,…,50)；y i表示第i个坐标的体积分数值，y i(i∈1,2,…,50)为向量元素，用以构建50维距离目标直方图的几何特征向量；x i表示第i个坐标的距离值；
步骤三、采用自动编码器中的编码层对步骤二中的距离目标直方图的几何特征向量进行降维，得到降维后的几何特征向量即距离目标直方图降维特征向量；
如图4所示，本实施例的自动编码器包括4个编码层和3个的解码层构成。图中的数字指的是编码器或者解码器某一层中的神经元个数。
每个编码层通过一个完全连接层缩小来压缩输入的维数，此层中每个神经元的激活函数为：
P(h i＝1|v i)＝sigmoid(c i+W iv i)
其中v i表示第i层编码层中作为输入的经i-1次降维后的特征向量，h i表示第i层编码层中作为输出的经i次降维后的特征向量，h i,v i∈{0,1},W i表示第i层编码层的权重矩阵；c i是第i层编码层的偏差值；这里h i的维数小于v i，因此编码层能减少输入特征向量的维数；
解码层是一个与编码层相反的过程，它通过增加输入特征向量的维数来重建原始输入的特征向量；
P(v j＝1|h j)＝sigmoid(b j+W jh j)
其中h j表示第j层解码层中作为输入的经j-1次升维后的特征向量，v j表示第i层解码层中作为输出的经j次升维后的特征向量，h j,v j∈{0,1},W j表示第j层解码层的权重矩阵，b j是第j层解码层的偏差值，这里v j的维数大于h j的维数，因此解码层可以发挥特征重构的作用。
步骤四、建立深度置信网络模型，迭代直至其收敛，完成深度置信网络模型的训练；
首先搜集训练用食管癌病人的VMAT计划数据，作为训练集数据，并 从训练集数据中提取出训练用CT图像、训练用结构轮廓图像，计算训练用剂量-体积直方图(DVH)、训练用距离目标直方图；
本实施例的DVH使用基于Matlab的放射学研究平台CERR包，直接使用内置函数提取。也可以从剂量图中提取剂量信息，以及相应位置处的体积信息，绘制DVH，DVH的x轴代表危险器官接受剂量的剂量百分比，y轴代表体积百分数，表示该体积接受的剂量等于或者大于x轴标明的剂量。
本发明从计算训练用剂量-体积直方图(DVH)的曲线x轴等距离选取50个坐标点((n j,m j)j∈1,2,…,50)；用选取的这50个离散坐标点来表达出这条曲线；每个坐标点包含体积分数值m j(j∈1,2,…,50)和相应的剂量分数值n j(j∈1,2,…,50)；m j表示第j个坐标的体积分数值，m j(j∈1,2,…,50)为向量元素，用以构建50维训练用剂量-体积直方图的剂量特征向量；n j表示第j个坐标的剂量分数值；
训练集数据中的训练用距离目标直方图通过步骤三降维后的训练用距离目标直方图的三维几何特征向量dt＝(dt_1,dt_2,dt_3)作为深度置信网络输入，其中dt_1,dt_2,dt_3分别表示降维后的三维几何特征向量的第一个维度的成分，第二个维度的成分和第三个维度的成分；通过步骤三降维后的训练用剂量-体积直方图的剂量特征向量dv＝(dv_1,dv_2,dv_3)；dv_1,dv_2,dv_3分别表示降维后的三维剂量特征向量的第一个维度的成分，第二个维度的成分和第三个维度的成分。
深度置信网络包括若干个玻尔兹曼机堆叠形成的，本实施例采用三个玻尔兹曼机。训练过程为先对每一个玻尔兹曼机进行预训练，再把预训练好的玻尔兹曼机堆叠起来作为深度置信网络的整体进行网络微调训练；
首先对每一个玻尔兹曼机的结构进行预训练。每一个玻尔兹曼机结构包括相应的可见层v和隐藏层h。
玻尔兹曼机在训练过程中的迭代步骤为：首先将可见层中的降维后的距离目标直方图的几何特征向量映射到隐藏层中，然后用隐藏层中的向量重构新的距离目标直方图的几何特征向量，再将重构的距离目标直方图的几何特征向量映射到隐藏层中，这个过程表示为一个循环，每个玻尔兹曼机在预训练过程中重复三次循环更新玻尔兹曼机可见层与隐藏层中的权重矩阵；
在深度置信网络的每一层中，目标损失函数是：
Figure PCTCN2020077844-appb-000008
式中dt为降维后的训练用距离目标直方图的几何特征向量，也是每一个玻尔兹曼机结构的可见层的输入向量。式中
Figure PCTCN2020077844-appb-000009
表示降维后的训练用距离目标直方图的几何特征向量dt到隐藏层特征向量g的均方误差，α s为实验中经验确定的约束系数，函数Y表示玻尔兹曼机结构中可见层到隐藏层的过程，其能量函数为：
Figure PCTCN2020077844-appb-000010
其中θ＝{W ij,a i,b j}是玻尔兹曼机的参数，他们均为实数。其中W ij表示可见向量与隐藏向量之间的权重，可见层共有可见向量(降维后的训练用距离目标直方图特征的向量dt)单元i个，v i表示可见层可见向量的第i个单元，隐藏层共有隐藏向量单元j个，h j表示隐藏层隐藏向量的第j个单元；a i表示第i个可见向量单元的偏置，b j表示第j个隐藏向量单元的偏置；
Figure PCTCN2020077844-appb-000011
表示由隐藏向量g提供的输入与重构成新的距离目标直方图的 几何特征向量之间的均方误差，Y′是Y的相反过程，表示用隐藏层中的向量重构新的距离目标直方图的几何特征向量；α r||W|| 1表示通过非零的惩罚因子来保证式中的稀疏性表达，||W|| 1表示对深度置信网络每一层的权重矩阵的L1正则化，α r为稀疏系数；
每一个玻尔兹曼机结构的可见层的可见向量i的个数是不同的，因为最后要把三个玻尔兹曼机结构堆叠在一起，所以第一层的隐藏层的向量单元个数j是第二层的可见层的向量单元个数，即第一层的输出个数为第二层的输入个数。
本实施例对于第一个玻尔兹曼机结构，可见层的可见向量单元个数为30个，隐藏层的隐藏向量单元个数为25个；同理，对于第二个玻尔兹曼机结构，可见层的可见向量单元个数维25个，隐藏层的隐藏向量单元个数为10个；以此类推第三个玻尔兹曼机结构，可见层的可见向量单元个数维10个，隐藏层的隐藏向量单元个数为5个。
每一个玻尔兹曼机的训练过程通过迭代循环，当其目标损失函数小于0.05，则趋于稳定收敛而结束训练过程同时保存玻尔兹曼机的参数值，得到预训练完成的玻尔兹曼机模型，若干个预训练完成的玻尔兹曼机模型叠加成预训练完成的深度置信网络。
对于预训练好的三个玻尔兹曼机组成的深度置信网络模型，重新把降维后的训练用距离目标直方图的几何特征向量输入到该深度置信网络模型中进行微调训练，损失函数使用均方误差损失函数:
Figure PCTCN2020077844-appb-000012
当均方误差损失函数损失值小于0.05，则趋于稳定收敛，保存参数值， 完成深度置信网络模型的训练。
其中函数D表示三个玻尔兹曼机结构堆叠成的经过预训练的深度置信网络模型；dt为降维后的训练用距离目标直方图的几何特征向量；dv_i为采用步骤三降维后的训练用剂量-体积直方图的剂量特征向量的第i个成分；本实施例降维后的训练用剂量-体积直方图的剂量特征向量为三维剂量-体积直方图的特征向量，故存在三个成分，三个成分需要建立三个不同的深度置信网络，三个深度置信网络的预测输出对应的一组降维后的剂量特征向量。
步骤五、降维后的几何特征向量通过训练后的深度置信网络非线性地拟合剂量特征向量和降维后几何特征向量之间的相关性，预测得到降维后的剂量特征向量；
降维后的几何特征向量和降维后的剂量特征向量之间的非线性相关性由三个非线性函数构成三个深度置信网络模型来拟合函数；降维后三维的剂量特征向量和降维后三维的几何特征向量之间的非线性相关性为：
Figure PCTCN2020077844-appb-000013
F i表示第i个由三个玻尔兹曼机结构组成的深度置信网络模型，对应输出为第i个剂量特征向量的dv_i；由于本实施例降维后的剂量特征维度为三维，故i∈[1,3]；通过三个深度置信网络模型得到一组三维的剂量特征向量dvp＝(dvp_1,dvp_2,dvp_3)。
本发明dvp的获得存在于验证用或者临床用食管癌病人的VMAT计划数据信息的处理。
步骤六、通过步骤三中的自动编码器中的解码层把预测得到的降维后的三维剂量特征向量dvp＝(dvp_1,dvp_2,dvp_3)重构回原始维度(50维)的剂量特征向量，最终得到预测危险器官的剂量-体积直方图。
进一步，优选地，通过本发明方法得到的剂量-体积直方图与现有技术得到的剂量-体积直方图进行比对，进一步验证本发明方法的精确性。如达不到设定要求，则进一步训练深度置信网络模型。
通过解码层重构的50维剂量特征向量，把重构的50维剂量特征向量作为新的剂量-体积直方图的坐标点的纵坐标，即m’ j(j∈1,2,…,50),把剂量分数值n j(j∈1,2,…,50)作为坐标点的横坐标，得到新的50个坐标点((n j,m’ j)j∈1,2,…,50)；连接这50个坐标点绘制的曲线即为模型预测的DVH曲线(预测危险器官的剂量-体积直方图)。
实施例2
本实施例与上述实施例的区别在于，共收集食管癌VMAT计划80例，平均分成8份，每份10例计划。在第一组实验中选择第一份作为测试集数据，其余作为训练集数据，第二组实验中选择第二份作为测试集数据，其余作为训练集数据，以此类推，完成八组实验，计算八组实验的平均值，作为最后的实验结果。
数据分类和预处理：
从收集的优质食管癌VMAT计划中计算得出病人的危险器官对应的剂量信息，本实施例中目标危险器官分别是左肺，右肺，心脏，脊髓。分别计算出相对应的剂量-体积直方图。每一种危险器官得到80组DVH。通过等剂量间隔采样DVH曲线中的50个坐标点，每个坐标点包含剂量分数值 和相应的体积分数值。选择体积分数值来构建一个50维的DVH特征向量。
从收集的优质食管癌VMAT计划中分别计算出病人的目标危险器官(左肺，右肺，心脏，脊髓)的距离目标直方图，用来描述放射靶区与危险器官之间的几何关系。每一种危险器官得到80组DTH。从DTH曲线中选择具有相等距离间隔的50个坐标点。每个坐标点包含体积分数值和相应的距离值。选择体积分数值来构建50维DTH特征向量。
本发明在对于新患者预测OAR的DVH时，首先计算OAR的DTH，其次可以通过对DTH曲线进行采样来建立50维DTH特征向量，然后通过自动编码器将该特征向量简化为三维特征向量。其相应的三维DVH特征向量可以用相应的深度置信网络模型映射得到。最后，使用自动编码器中的解码层来重建DVH特征向量，最终得到预测的OAR的DVH。
综上所述，本发明是深度学习技术的一次有效应用，相对于提取几何特征的线性模型和传统的机器学习方法来预测食管癌放疗剂量学特征而言，根据OAR与PTV的几何关系进行食管放射治疗计划的OAR的剂量的自动评估。实验证明，该模型方法能够实现准确的DVH预测并且可以为食管治疗计划提供接近最优的参数，这可以显著缩短制定食管癌放疗计划时间以达到减轻物理放疗师的负担。
需要说明的是，在本文中，如若存在第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且 还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
以上实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的精神和范围。

Claims (10)

  1. 一种食管放射治疗计划中风险器官剂量学评估方法，其特征在于，包括以下步骤：
    步骤一、收集食管癌病人的VMAT计划数据信息，包括提取出的CT图像、结构轮廓图像；
    步骤二、对食管癌VMAT计划数据信息计算，得到距离目标直方图；
    步骤三、采用自动编码器结构中的编码层对步骤二中的距离目标直方图的几何特征向量进行降维，得到降维后的几何特征向量；
    步骤四、建立深度置信网络模型，迭代直至其收敛，完成深度置信网络模型的训练；
    步骤五、降维后的几何特征向量通过训练后的深度置信网络非线性地拟合剂量特征向量和降维后几何特征向量之间的相关性，得到与降维后的几何特征向量的维度相同的剂量特征向量；
    步骤六、通过自动编码器结构的解码层将从步骤五的剂量特征向量重构，得到与降维前的几何特征向量的维度相同的剂量特征向量，最终得到预测危险器官的剂量-体积直方图。
  2. 根据权利要求1所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，所述步骤二中通过计算风险器官的体积百分数来建立距离目标直方图，
    Figure PCTCN2020077844-appb-100001
    其中v i OAR表示危险器官的第i个体素；v k PTV表示计划靶区的第k个体素；S PTV为计划靶区的体素集合；r(v i OAR，PTV)为危险器官的体素到计划靶区表面的欧式距离。
  3. 根据权利要求2所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，从距离目标直方图的曲线x轴等距离选取n个坐标点((x i,y i)i∈1,2,…,n)；用选取的这n个离散坐标点来表达出这条曲线；每个坐标点包含体积分数值y i(i∈1,2,…,n)和相应的距离值x i(i∈1,2,…,n)；y i表示第i个坐标的体积分数值，y i(i∈1,2,…,n)为向量元素，用以构建n维距离目标直方图的几何特征向量；x i表示第i个坐标的距离值。
  4. 根据权利要求1所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，编码器由若干个编码层和若干个的解码层构成；
    每个编码层通过一个完全连接层缩小来压缩输入的维数，此层中每个神经元的激活函数为：
    P(h i＝1|v i)＝sigmoid(c i+W iv i)
    其中v i表示第i层编码层中作为输入的经i-1次降维后的特征向量，h i表示第i层编码层中作为输出的经i次降维后的特征向量，h i,v i∈{0,1},W i表示第i层编码层的权重矩阵；c i是第i层编码层的偏差值；这里h i的维数小于v i，
    解码层是一个与编码层相反的过程，它通过增加输入特征向量的维数来重建原始输入的特征向量；
    P(v j＝1|h j)＝sigmoid(b j+W jh j)
    其中h j表示第j层解码层中作为输入的经j-1次升维后的特征向量，v j表示第i层解码层中作为输出的经j次升维后的特征向量，h j,v j∈{0,1},W j表示第j层解码层的权重矩阵，b j是第j层解码层的偏差值，这里v j的维数大于h j的维数。
  5. 根据权利要求1所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，所述步骤四中的深度置信网络模型包括若干个玻尔兹曼机堆叠形成的；训练过程为先对每一个玻尔兹曼机进行预训练，再把预训练好的玻尔兹曼机堆叠起来作为深度置信网络的整体进行网络微调训练。
  6. 根据权利要求5所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，搜集训练用食管癌病人的VMAT计划数据，作为训练集数据，并从训练集数据中提取出训练用CT图像、训练用结构轮廓图像，计算训练用剂量-体积直方图、训练用距离目标直方图；并计算得到训练用剂量-体积直方图的剂量特征向量、训练用距离目标直方图的几何特征向量；并通过自动编码器结构的解码层得到降维后的训练用剂量-体积直方图的剂量特征向量、降维后的训练用距离目标直方图的几何特征向量；
    对每一个玻尔兹曼机的结构进行预训练；每一个玻尔兹曼机结构包括相应的可见层v和隐藏层h；
    玻尔兹曼机在训练过程中的迭代步骤为：首先将可见层中的降维后的距离目标直方图的几何特征向量映射到隐藏层中，然后用隐藏层中的向量重构新的距离目标直方图的几何特征向量，再将重构的距离目标直方图的几何特征向量映射到隐藏层中，这个过程表示为一个循环，每个玻尔兹曼机在预训练过程中重复三次循环更新玻尔兹曼机可见层与隐藏层中的权重矩阵；
    在深度置信网络的每一层中，目标损失函数是：
    Figure PCTCN2020077844-appb-100002
    式中dt为降维后的训练用距离目标直方图的几何特征向量，也是每一个玻尔兹曼机结构的可见层的输入向量；式中
    Figure PCTCN2020077844-appb-100003
    表示降维后的训练用距离目标直方图的几何特征向量dt到隐藏层特征向量g的均方误差，α s为实验中经验确定的约束系数，函数Y(dt)表示玻尔兹曼机结构中可见层到隐藏层的过程，其能量函数为：
    Figure PCTCN2020077844-appb-100004
    其中θ＝{W ij,a i,b j}是玻尔兹曼机的参数，其中W ij表示可见向量与隐藏向量之间的权重，可见层共有可见向量单元i个，隐藏层共有隐藏向量单元j个，a i表示第i个可见向量单元的偏置，b j表示第j个隐藏向量单元的偏置；
    Figure PCTCN2020077844-appb-100005
    表示由隐藏向量g提供的输入与重构成新的距离目标直方图的几何特征向量之间的均方误差，Y′(g)是Y的相反过程，表示用隐藏层中的向量重构新的距离目标直方图的几何特征向量；α r||W|| 1表示通过非零的惩罚因子来保证式中的稀疏性表达，||W|| 1表示对深度置信网络每一层的权重矩阵的L1正则化，α r为稀疏系数；
    每一个玻尔兹曼机的训练过程通过迭代循环，当其目标损失函数小于设定阈值，则趋于稳定收敛而结束训练过程同时保存玻尔兹曼机的参数值，得到预训练完成的玻尔兹曼机模型，若干个预训练完成的玻尔兹曼机模型叠加成预训练完成的深度置信网络。
  7. 根据权利要求6所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，重新把降维后的训练用距离目标直方图的几何特征向量输入到所述预训练好的深度置信网络模型中进行微调训练，损失函数使用均方误差损失函数:
    Figure PCTCN2020077844-appb-100006
    当均方误差损失函数损失值小于设定阈值，则趋于稳定收敛，保存参数值，完成深度置信网络模型的训练；
    其中函数D(dt)表示三个玻尔兹曼机结构堆叠成的经过预训练的深度置信网络模型；dt为降维后的训练用距离目标直方图的几何特征向量；dv_i为采用步骤三降维后的训练用剂量-体积直方图的剂量特征向量的第i个成分；i个成分需要建立i个不同的深度置信网络，i个深度置信网络的预测输出对应的一组降维后的剂量特征向量。
  8. 根据权利要求1所述的食管放射治疗计划中风险器官剂量学评估方法，其特征在于，所述步骤六通过解码层重构n维剂量特征向量，把重构的n维剂量特征向量作为新的剂量-体积直方图的坐标点的纵坐标，即m’ j(j∈1,2,…,n)，把剂量分数值n j(j∈1,2,…,n)作为坐标点的横坐标，得到新的n个坐标点((n j,m’ j)j∈1,2,…,n)；连接这n个坐标点绘制的曲线即为预测危险器官的剂量-体积直方图。
  9. 一种食管放射治疗计划中风险器官剂量学评估***，其特征在于，包括，
    编码层，用以对距离目标直方图的几何特征向量进行降维，得到降维后的几何特征向量；
    深度置信网络模型，降维后的几何特征向量通过训练后的深度置信网络非线性地拟合剂量特征向量和降维后几何特征向量之间的相关性，得到与降维后的几何特征向量的维度相同的剂量特征向量；
    解码层，用以将剂量特征向量重构，得到与降维前的几何特征向量的 维度相同的剂量特征向量。
  10. 根据权利要求9所述的食管放射治疗计划中风险器官剂量学评估***，其特征在于，深度置信网络模型包括若干个玻尔兹曼机堆叠形成的；训练过程为先对每一个玻尔兹曼机进行预训练，再把预训练好的玻尔兹曼机堆叠起来作为深度置信网络的整体进行网络微调训练。
PCT/CN2020/077844 2019-04-01 2020-03-04 食管放射治疗计划中风险器官剂量学评估方法、评估*** WO2020199836A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/437,845 US11515025B2 (en) 2019-04-01 2020-03-04 Dosimetry assessment method and system for organ at risk in esophageal radiotherapy plan

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910256615.3 2019-04-01
CN201910256615.3A CN109859817B (zh) 2019-04-01 2019-04-01 食管放射治疗计划中风险器官剂量学评估方法、评估***

Publications (1)

Publication Number Publication Date
WO2020199836A1 true WO2020199836A1 (zh) 2020-10-08

Family

ID=66902760

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077844 WO2020199836A1 (zh) 2019-04-01 2020-03-04 食管放射治疗计划中风险器官剂量学评估方法、评估***

Country Status (3)

Country Link
US (1) US11515025B2 (zh)
CN (1) CN109859817B (zh)
WO (1) WO2020199836A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109859817B (zh) * 2019-04-01 2022-12-09 安徽大学 食管放射治疗计划中风险器官剂量学评估方法、评估***
CN110604550B (zh) * 2019-09-24 2022-06-21 广州医科大学附属肿瘤医院 一种肿瘤放疗后正常组织器官并发症预测模型的建立方法
CN115702437A (zh) * 2020-06-09 2023-02-14 索尼半导体解决方案公司 信号处理装置及信号处理方法
RU2744748C2 (ru) * 2020-09-03 2021-03-15 Федеральное государственное бюджетное учреждение «Национальный медицинский исследовательский центр радиологии» Министерства здравоохранения Российской Федерации (ФГБУ «НМИЦ радиологии» Минздрава России) Способ лечения плоскоклеточного рака головы и шеи
CN113256436A (zh) * 2021-07-02 2021-08-13 平安科技（深圳）有限公司 车险赔付预提示方法、装置、设备及存储介质
CN117717719B (zh) * 2023-09-07 2024-07-23 山东卓业医疗科技有限公司 一种在粒子放射剂量模型中计算靶区dvh的方法
CN117877721A (zh) * 2024-03-12 2024-04-12 天津市肿瘤医院（天津医科大学肿瘤医院） 一种个性化甲状腺癌放疗后内分泌治疗管理***

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104751498A (zh) * 2013-12-31 2015-07-01 深圳市医诺智能科技发展有限公司 一种放疗计划呈现方法及***
CN105358219A (zh) * 2013-05-21 2016-02-24 瓦里安医疗***国际股份公司 用于自动产生剂量预测模型以及作为云服务的疗法治疗计划的***和方法
CN107441637A (zh) * 2017-08-30 2017-12-08 南方医科大学 调强放疗计划中三维剂量分布的预测方法及其应用
CN109166613A (zh) * 2018-08-20 2019-01-08 北京东方瑞云科技有限公司 基于机器学习的放射治疗计划评估***及方法
US20190051398A1 (en) * 2017-08-09 2019-02-14 Varian Medical Systems, Inc. Aggregation of artificial intelligence (ai) engines
CN109859817A (zh) * 2019-04-01 2019-06-07 安徽大学 食管放射治疗计划中风险器官剂量学评估方法、评估***

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102306239B (zh) * 2011-07-22 2014-11-05 上海赛提菲克医疗器械有限公司 基于锥形束ct图像ct值校正技术的放疗剂量评估和优化方法
WO2014068435A2 (en) * 2012-10-29 2014-05-08 Koninklijke Philips N.V. Automatic optimal imrt/vmat treatment plan generation.
EP3384959B1 (en) * 2013-06-18 2019-11-06 Duke University Systems and methods for specifying treatment criteria and treatment parameters for patient specific radiation therapy planning
US11495355B2 (en) * 2014-05-15 2022-11-08 The Johns Hopkins University Method, system and computer-readable media for treatment plan risk analysis
CN105447330B (zh) * 2015-12-30 2019-01-08 上海联影医疗科技有限公司 调强放射治疗的权重调整方法和装置
CN105854191B (zh) * 2016-04-26 2018-09-21 中国科学院合肥物质科学研究院 一种放射治疗中三维剂量验证***及验证方法
US11517768B2 (en) * 2017-07-25 2022-12-06 Elekta, Inc. Systems and methods for determining radiation therapy machine parameter settings
US11615873B2 (en) * 2018-08-31 2023-03-28 The Board Of Regents Of The University Of Texas System Deep learning based dosed prediction for treatment planning and quality assurance in radiation therapy

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105358219A (zh) * 2013-05-21 2016-02-24 瓦里安医疗***国际股份公司 用于自动产生剂量预测模型以及作为云服务的疗法治疗计划的***和方法
CN104751498A (zh) * 2013-12-31 2015-07-01 深圳市医诺智能科技发展有限公司 一种放疗计划呈现方法及***
US20190051398A1 (en) * 2017-08-09 2019-02-14 Varian Medical Systems, Inc. Aggregation of artificial intelligence (ai) engines
CN107441637A (zh) * 2017-08-30 2017-12-08 南方医科大学 调强放疗计划中三维剂量分布的预测方法及其应用
CN109166613A (zh) * 2018-08-20 2019-01-08 北京东方瑞云科技有限公司 基于机器学习的放射治疗计划评估***及方法
CN109859817A (zh) * 2019-04-01 2019-06-07 安徽大学 食管放射治疗计划中风险器官剂量学评估方法、评估***

Also Published As

Publication number Publication date
CN109859817B (zh) 2022-12-09
US11515025B2 (en) 2022-11-29
US20220101974A1 (en) 2022-03-31
CN109859817A (zh) 2019-06-07

Similar Documents

Publication Publication Date Title
WO2020199836A1 (zh) 食管放射治疗计划中风险器官剂量学评估方法、评估***
US11386557B2 (en) Systems and methods for segmentation of intra-patient medical images
CN109272443B (zh) 一种基于全卷积神经网络的pet与ct图像配准方法
CN112546463B (zh) 基于深度神经网络的放疗剂量自动预测方法
CN110354406A (zh) 一种放射治疗的三维剂量预测方法和***
CN111028914A (zh) 人工智能引导的剂量预测方法与***
WO2022011616A1 (zh) 一种通过影像组学特征判断癌症治疗反应的方法及***
Kieselmann et al. Cross‐modality deep learning: contouring of MRI data from annotated CT data only
CN110378881B (zh) 一种基于深度学习的肿瘤定位***
CN110141801A (zh) 面向近距离粒子内放疗手术智能剂量规划方法
CN104867104B (zh) 基于xct图像非刚度配准的目标鼠解剖结构图谱获取方法
CN109949352A (zh) 一种基于深度学习的放疗影像靶区勾画方法及勾画***
WO2022095167A1 (zh) 剂量确定方法及装置
Wang et al. DeepEC: An error correction framework for dose prediction and organ segmentation using deep neural networks
US11724126B2 (en) Methods and systems for quality-aware continuous learning for radiotherapy treatment planning
Johnston et al. Dose-volume-based evaluation of convolutional neural network-based auto-segmentation of thoracic organs at risk
CN109671499B (zh) 一种直肠毒性预测***构建方法
CN113499091B (zh) 一种患者体表和体内肿瘤运动相关性及肿瘤内部动度的预测方法和***
CN107480445A (zh) 基于隐马尔可夫模型的肿瘤临床靶区侵犯概率计算方法
CN112837782B (zh) 一种基于深度学习的放疗三维剂量预测方法
US20240001149A1 (en) Methods and systems for dose distribution prediction
CN111888665A (zh) 一种基于自适应修正对抗网络的三维剂量分布预测方法
CN113178242B (zh) 一种基于耦合式生成对抗网络的自动计划优化***
CN116563402A (zh) 跨模态mri-ct图像合成方法、***、设备和介质
JP7378694B2 (ja) デジタルヒューマン技術に基づく肺葉分割方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20783138

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20783138

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20783138

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22/04/2022)

122 Ep: pct application non-entry in european phase

Ref document number: 20783138

Country of ref document: EP

Kind code of ref document: A1