CN110085298B - Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on deep network learning - Google Patents
Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on deep network learning Download PDFInfo
- Publication number
- CN110085298B CN110085298B CN201910341687.8A CN201910341687A CN110085298B CN 110085298 B CN110085298 B CN 110085298B CN 201910341687 A CN201910341687 A CN 201910341687A CN 110085298 B CN110085298 B CN 110085298B
- Authority
- CN
- China
- Prior art keywords
- layer
- dimensional
- dose
- contour
- feature
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61N—ELECTROTHERAPY; MAGNETOTHERAPY; RADIATION THERAPY; ULTRASOUND THERAPY
- A61N5/00—Radiation therapy
- A61N5/10—X-ray therapy; Gamma-ray therapy; Particle-irradiation therapy
- A61N5/103—Treatment planning systems
- A61N5/1031—Treatment planning systems using a specific method of dose optimization
-
- 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
- G06N3/08—Learning methods
-
- 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
- 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
- G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
-
- 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
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61N—ELECTROTHERAPY; MAGNETOTHERAPY; RADIATION THERAPY; ULTRASOUND THERAPY
- A61N5/00—Radiation therapy
- A61N5/10—X-ray therapy; Gamma-ray therapy; Particle-irradiation therapy
- A61N5/103—Treatment planning systems
- A61N2005/1041—Treatment planning systems using a library of previously administered radiation treatment applied to other patients
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Public Health (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Medical Informatics (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Pathology (AREA)
- Primary Health Care (AREA)
- Epidemiology (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Databases & Information Systems (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Urology & Nephrology (AREA)
- Radiology & Medical Imaging (AREA)
- Animal Behavior & Ethology (AREA)
- Veterinary Medicine (AREA)
- Surgery (AREA)
- Radiation-Therapy Devices (AREA)
Abstract
The invention provides a prediction method of intensity modulated radiotherapy plan three-dimensional dose distribution based on deep learning. The method comprises the following steps: collecting effective intensity modulated radiotherapy plan data to form a case database; extracting three-dimensional anatomical contour features of the region of interest of each patient from a case database; dividing a three-dimensional anatomical structure contour of a region of interest of a patient into a plurality of two-dimensional contour slice images; extracting the dose characteristics of each patient from a case database and dividing the dose characteristics into a plurality of two-dimensional dose slice distribution maps; building a depth convolution network, inputting a two-dimensional contour slice image of a patient and a corresponding two-dimensional dose slice distribution map, and obtaining a correlation model between the contour feature and the dose feature of the anatomical structure through model training; predicting a three-dimensional dose distribution of the new patient using the trained correlation model. By using the method of the invention, the incidence relation between the anatomical structure characteristic and the dose characteristic can be effectively obtained, and the accuracy of dose prediction is improved.
Description
Technical Field
The invention relates to the technical field of intelligent radiotherapy, in particular to a prediction method of three-dimensional dose distribution in an intensity modulated radiotherapy plan based on deep network learning.
Background
The tumor radiotherapy has unique advantages, and is one of the main means in tumor therapy proposed by the world health organization, and the main aim of the tumor radiotherapy is to reduce the dose deposition of surrounding normal tissues as much as possible and improve the local control rate of tumors while ensuring that a target region can reach a specific dose. Intensity modulated radiotherapy, i.e. intensity modulated radiation therapy, is one kind of three-dimensional conformal radiotherapy, and requires the dose intensity in the radiation field to be regulated according to certain requirements. The distribution of the radiotherapy dosage is consistent with the shape of the target area, and the target area receives uniformly distributed high dosage, thereby ensuring the killing of tumor cells and improving the treatment effect of radiotherapy.
In the prior art, the quality of a radiotherapy plan often depends on the knowledge level and experience accumulation of plan designers, and it is difficult to ensure high quality of all plans. Meanwhile, the clinical plans are subject to uniform standard standards, which results in that the plan design cannot meet the difference between individual patients to reach the optimal plan under the condition of meeting the clinical standard. Quality control and quality assurance are also an extremely important process in precision radiotherapy, wherein dose verification of IMRT (intensity modulated radiation therapy) as a main mode of quality control and quality audit in current clinical radiotherapy technology is a key step for ensuring that the dose received by a patient reaches a standard.
Research shows that the reachable dose of a patient is particularly related to the geometric anatomical structure of the patient, and the dose prediction can be realized by mining potential cases and constructing a correlation model of the geometric anatomical structure between organs and corresponding radiotherapy plan dosimetry information, so that the standard is provided for the dosimetry verification and quality control, the individualized specific requirement of the patient is met, and meanwhile, a foundation is provided for the radiotherapy automation.
"A planning evaluation tool for a pro state adaptive IMRT based on a machine learning", published in Med Phys publication by zhu et al in 2011, describes a method for learning a dose volume histogram DVH of an Organ At Risk (OAR) from anatomical data using a support vector machine model, which describes the spatial relationship of the Organ At Risk (OAR) and a target volume (PTV) from distance and volume. However, the dose characteristics selected by the method are extracted two-dimensional information, and the representation situation of dose distribution on a three-dimensional space cannot be sufficiently reflected. In 2017, singing et al in CN107441637A ("a method for predicting three-dimensional dose distribution in intensity modulated radiotherapy planning and its application") proposed a method for predicting three-dimensional dose distribution by extracting voxel point anatomical features and dose features and using neural network. However, the anatomical features in the method are manually extracted through clinical experience, the relation between the anatomical structure and the dose deposition cannot be completely expressed, and different organs at risk need to be repeatedly modeled for many times, so that the method is complicated in clinical application.
Accordingly, there is a need for improvements in the prior art to provide an effective method for predicting the three-dimensional dose distribution of intensity modulated radiation therapy plans.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a three-dimensional dose distribution prediction method suitable for intensity modulated radiotherapy planning so as to solve the problems that the description of anatomical information is incomplete, a plurality of regions of interest cannot be predicted simultaneously and the like in the prior art.
According to a first aspect of the present invention, a method for predicting three-dimensional dose distribution in an intensity modulated radiotherapy plan based on deep learning is provided. The method comprises the following steps:
step S1: collecting effective intensity modulated radiotherapy plan data to form a case database;
step S2: extracting three-dimensional anatomical contour features of the region of interest of each patient, including target region contours, torso contours, and organ-at-risk contours, from the case database;
step S3: dividing the three-dimensional anatomical structure contour of the region of interest of the patient into a plurality of two-dimensional contour slice images according to the required image size;
step S4: extracting the dose characteristics of each patient from the case database, registering according to the size of a required image and dividing into a plurality of two-dimensional dose slice distribution maps;
step S5: building a depth convolution network, inputting a two-dimensional profile slice diagram of a patient and a corresponding two-dimensional dose slice distribution diagram, learning a mapping relation between the profile characteristic and the dose characteristic of the anatomical structure through model training, and performing cross validation on the model to obtain a correlation model between the profile characteristic and the dose characteristic of the anatomical structure;
step S6: predicting a three-dimensional dose distribution of the new patient using the trained correlation model.
In one embodiment, the step S2 includes:
the extracted three-dimensional anatomical structure contour of the region of interest is a three-dimensional binary matrix, and a matrix value of 1 represents that a voxel of the region of interest exists therein.
In one embodiment, in the steps S3 and S4, the three-dimensional anatomical structure contour feature and the dose feature of the patient are divided from different directions of the three-dimensional CT image, including a transverse plane, a coronal plane and a sagittal plane, when divided into two-dimensional slices according to the resolution size of the CT image; and rotational translation is used to enhance the slice number.
In one embodiment, the step S5 includes:
step S51: importing an open-source deep learning framework TensorFlow in Python;
step S52: constructing a deep convolution network by using a TensorFlow framework, wherein the number of feature maps of an input layer is N, and the feature maps represent the extracted contour features of the N regions of interest; the number of the characteristic graphs of the output layer is 1, and the characteristic graphs represent a dose distribution graph; a symmetrical 5-layer U-shaped network is connected between the input layer and the output layer, the left side of the U shape is a continuous down-sampling part, the down-sampling operation after each layer is completed by a cavitation layer power layer and a relu layer, and the size of the feature map of the upper layer is reduced by half; the right side of the U shape is a continuous up-sampling part, the up-sampling operation after each layer is completed by an upsample layer contribution layer and a relu layer, and the feature diagram of the upper layer is restored to be equal to the size of the left symmetrical layer; each layer of the U-shaped structure is composed of a plurality of denseblocks; two symmetrical sides of the U-shaped structure are connected by a catenate operation; selecting an optimized objective function as an MSE algorithm;
step S53: and inputting the extracted two-dimensional contour slice image and the corresponding two-dimensional dose slice distribution map into the constructed depth convolution network, training the correlation model of the two, and evaluating the performance of the model through cross validation.
In one embodiment, in the step S52, each layer of the U-shaped structure includes a plurality of denseblocks, and the structure of the denseblocks is:
the denseblock is composed of a convolution structure A and a connection structure B, wherein the A comprises a convolution layer, an activation layer, a bantchnorm layer and a dropout layer, the B is composed of a concatenate layer, and an input feature graph after passing through the structure A is connected with an input feature graph through the structure B to serve as output;
the left side and the right side of the first layer of the U-shaped structure are respectively provided with a denseblock, the right side is connected with 4 convolution structures A behind the denseblock to obtain output layer data, the 2 nd layer and the 3 rd layer on the left side are provided with two denseblocks, the 4 th layer and the 5 th layer are provided with three denseblocks, and the right side is identical to the left side in structure.
In one embodiment, in the step S53, the association model is trained according to the following steps:
step S531: collecting clinical plans of tumors of the same type in a case database, and extracting three-dimensional anatomical structure contour characteristics and dose characteristics of a patient of each plan;
step S532: randomly screening 80% of the plans in the collected case database as training data of the model, and using the rest 20% as verification data;
step S533: respectively inputting three-dimensional anatomical contour feature maps and dose feature maps of different regions of interest of a training plan into an established deep convolution network, determining model parameters and training parameters, and training to obtain an association model between an anatomical structure and dose features by using MSE as a loss function;
step S534: inputting contour feature maps of different regions of interest of the verification plan into the trained correlation model, and obtaining a corresponding dose feature distribution map;
step S535: calculating the mean square error of the verification data to evaluate the training effect of the model;
step S536: repeating the steps S532 to S535 for a plurality of times, and selecting the model with the best training effect as the correlation model.
In one embodiment, in step S533, the determined model parameters include convolution kernel size, activation function, dropout rate, and the determined training parameters include learning rate, optimization algorithm, and batch-size.
In one embodiment, in said step S6, the three-dimensional dose distribution of the new patient is predicted using a correlation model:
step S61: extracting two-dimensional contour slice characteristics of a new patient according to the training data slice mode;
step S62: inputting two-dimensional contour slice characteristics of a new patient into the correlation model to calculate a corresponding dose characteristic distribution map;
step S63: and arranging and integrating the dose characteristic distribution map to obtain a three-dimensional dose distribution map of the new patient.
According to a second aspect of the present invention, a system for predicting a three-dimensional dose distribution in an intensity modulated radiotherapy plan based on deep learning is provided. The system comprises:
case database: the system is used for storing effective intensity modulated radiotherapy plan sample data;
an anatomical feature extraction module: extracting three-dimensional anatomical contour features of a region of interest of each patient including a target region contour, a torso contour, a contour of an organ at risk from the case database;
an anatomical feature processing module: the three-dimensional anatomical structure contour of the region of interest of the patient is divided into a plurality of two-dimensional contour slice images according to the required image size;
a dose feature extraction module: the two-dimensional dose slice distribution map is used for extracting dose characteristics of each patient from the case database, registering according to the size of a required image and dividing into a plurality of two-dimensional dose slice distribution maps;
a model training module: the system is used for building a deep convolution network, inputting a two-dimensional contour slice diagram of a patient and a corresponding two-dimensional dose slice distribution diagram, learning the mapping relation between the contour characteristic and the dose characteristic of the anatomical structure through model training, and performing cross validation on the model to obtain an association model between the contour characteristic and the dose characteristic of the anatomical structure;
a dose prediction module: for predicting a three-dimensional dose distribution of a new patient using the trained correlation model.
Compared with the prior art, the invention has the advantages that: the interesting contour map is used as an anatomical structure expression mode, so that the defect of incomplete information extraction by manpower is avoided; the three-dimensional dose distribution of the whole irradiation area can be predicted so as to present more dose information; the model is trained using a deep learning approach, improving the accuracy of predicting dose distribution based on anatomical features.
Drawings
The invention is illustrated and described only by way of example and not by way of limitation in the scope of the invention as set forth in the following drawings, in which:
FIG. 1 is a flowchart of a method for predicting three-dimensional dose distribution in a deep learning-based intensity modulated radiation therapy plan, according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a three-dimensional dose distribution prediction method in a deep learning-based intensity modulated radiation therapy plan according to an embodiment of the present invention;
FIG. 3 is a structural distribution diagram of nasopharyngeal carcinoma slices in different directions, which is a cross section, a coronal plane and a sagittal plane from left to right;
FIG. 4 is a block diagram of a deep convolutional network according to one embodiment of the present invention;
FIG. 5 is a graph comparing an actual dose volume histogram DVH of test data with a predicted dose volume histogram DVH according to an embodiment of the present invention;
wherein: input (Input); output (Output); FeatureMap (feature map); convolution; concatenate (in series); upsample (Upsample); maxporoling (max pooling); batch normalization.
Detailed Description
In order to make the objects, technical solutions, design methods, and advantages of the present invention more apparent, the present invention will be further described in detail by specific embodiments with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
In all examples shown and discussed herein, any particular value should be construed as merely illustrative, and not as a limitation. Thus, other examples of the exemplary embodiments may have different values.
Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail but are intended to be part of the specification where appropriate.
According to one embodiment of the invention, a method for predicting three-dimensional dose distribution of intensity modulated radiotherapy plan based on deep network learning is provided, and prediction of three-dimensional dose distribution is realized by predicting dose distribution of each layer of CT image of a patient through deep learning. Briefly, the method comprises, in its entirety, two parts: obtaining a correlation model through deep learning according to the correlation between the three-dimensional anatomical structure characteristics of the patient and the dose characteristics in the intensity modulated radiotherapy plan; dose characteristics of the new patient are predicted based on the obtained correlation model to provide a clinical reference. Specifically, referring to fig. 1 and 2, the method includes the steps of:
and step S110, collecting effective intensity modulated radiotherapy plan data to form a case database.
The case database is used for storing cases of different parts of different patients, and the shared case database can be established by utilizing the existing resources.
And step S120, extracting three-dimensional anatomical structure characteristics and dose characteristics of the patient from the case database, and constructing a training sample set.
In an embodiment of the invention, the set of training samples constructed comprises a correspondence between three-dimensional anatomical features and dose features.
The three-dimensional anatomical contours include target volume (PTV) contours, trunk contours, contours of individual organs at risk, including, for example, trunk contours, spinal cord contours, brainstem contours, parotid contours, temporal lobe contours, crystalloid contours, chiasmatic contours, optic nerve contours, laryngeal contours, pituitary contours, lobe contours, and the like.
The dose characteristic is a dose distribution corresponding to a three-dimensional anatomical contour, the unit of dose value of which is cGy (centigray).
And step S130, training the deep learning network based on the training sample set to obtain an association model between the anatomical structure characteristic and the dose characteristic.
In this step, a deep learning network is constructed, training is performed based on a training sample set, a correlation model between anatomical structure features and dose features is obtained, and after the model training is completed, optimized model parameters, such as weights of layers, sizes of convolution kernels, and the like, can be obtained.
It should be understood that, when training the model, the directly input may be not the three-dimensional anatomical contour feature, but the processed data, for example, converting the three-dimensional anatomical contour into a corresponding two-dimensional anatomical contour slice image, wherein the slice mode includes a transverse plane, a coronal plane, a sagittal plane, and the like.
Step S140, predicting the three-dimensional dose distribution of the new patient using the correlation model.
Similar to the data processing mode in the training process, for a new patient, extracting anatomical information, such as the contour of the trunk, spinal cord, brainstem, parotid gland, temporal lobe, crystal, visual cross, optic nerve, throat, pituitary and brain lobe, and extracting two-dimensional anatomical contour characteristics of the new patient according to a training data slice mode; inputting the two-dimensional anatomical contour features into the correlation model, and calculating a corresponding dose feature distribution map; and arranging and integrating the dose distribution map to obtain a three-dimensional dose distribution map of the new patient.
The following is a nasopharyngeal carcinoma example to further illustrate the principles of the present invention.
Specifically, this example performs dose prediction for nasopharyngeal carcinoma, and selects IMRT plans of 32 nasopharyngeal carcinoma patients for model training.
First, three-dimensional anatomical contour features and dose features of the patient are extracted using MATLAB, wherein the anatomical contour features include trunk contour, spinal cord contour, brainstem contour, parotid gland contour, temporal lobe contour, crystalline contour, visual cross contour, optic nerve contour, laryngeal contour, pituitary contour, and brain lobe contour. The dose is characterized by the corresponding dose distribution, the unit of dose value being cGy.
Then, the three-dimensional features are divided into two-dimensional slices in a coronal plane slicing mode, the slicing direction can be selected to be any direction, and in nasopharyngeal carcinoma data, the model is best in the coronal plane slicing mode (fig. 3 shows the slicing modes in three different directions, namely a transverse plane, a coronal plane and a sagittal plane from left to right); processing the slices into 128 x 128 size by cropping and dimensionality reduction; performing rotation and translation data enhancement on the slice;
and finally, establishing a deep convolutional network in a Python environment by using a TensorFlow deep learning framework, and training a correlation model of anatomical features and dose features, wherein the specific process of model training is as follows:
a TensorFlow deep learning library is introduced into Python, and a network is built by using sentences in the deep learning library, wherein in the embodiment, the network structure is shown in FIG. 4 and is formed by combining a classical network U-net and a denseNet. The data size of the network input layer is 128 × 10, wherein 10 represents that 10 channels sequentially correspond to the extracted 10 interesting region profiles; the network output layer data size is 128 x 1, representing the corresponding dose distribution of the anatomical distribution layer. And a U-shaped structure with 5 layers is arranged between the input layer and the output layer. The left side of the U-shaped structure is a downsampling part, the left side of the U-shaped structure is a continuous downsampling part, downsampling operation after each layer is completed by a cavitation layer enabling layer and an activation layer (activation layer), and the size of a feature map of the upper layer is reduced by half; the right side of the U shape is a continuous up-sampling part, the up-sampling operation after each layer is completed by an upsample layer contribution layer and a relu layer, and the feature diagram of the upper layer is restored to be equal to the size of the left symmetrical layer; each layer of the U-shaped structure is composed of a plurality of denseblocks; two symmetrical sides of the U-shaped structure are connected in a jumping mode through a concatenate operation; the denseblock consists of a convolution structure A and a connection structure B, wherein the A consists of a convolution layer followed by an activation layer, a batchnorlysis layer and a dropout layer, and the B consists of a bicarbonate layer. The feature graph of the input feature graph after passing through the structure A is connected with the input feature graph through the structure B and then is used as output; the left side of the first layer of the U-shaped structure is provided with a convolution structure A, the right side is connected with 4 convolution structures A after the denseblock to obtain output layer data, the 2 nd layer and the 3 rd layer on the left side are provided with two denseblocks, and the 4 th layer and the 5 th layer are provided with three denseblocks.
In this embodiment, the loss function of the network is mean square error MSE, the size of the convolution kernel is given in the figure, the activation function of the activation layer adopts a relu function, the dropout rate is in a direct proportion relation with the number of convolution kernels of the layer, the minimum value is 0, the maximum value is 0.20, the learning rate is set to 0.0005, the optimization algorithm adopts Adam algorithm, and the batch-size is 32.
In the training process, 80% of slice data are randomly selected each time and input into the constructed network to serve as a training data training model, and 5 examples serve as verification data sets. And (5) training times in total, and selecting a group with the minimum error as a correlation model for successful training by calculating the Mean Square Error (MSE).
After the model training is successful, for a new patient, extracting the three-dimensional anatomical contour information which is the same as the training process, namely the contour of the trunk, the spinal cord, the brainstem, the parotid gland, the temporal lobe, the crystal, the visual cross, the optic nerve, the throat, the pituitary and the brain lobe. And slicing from the coronal plane direction, and processing the sliced data into two-dimensional data of 128 × 128 size. And inputting the anatomical features into the trained correlation model to obtain a predicted dose characteristic value of the new patient.
To verify the reliability of the model, five test data were predicted and the mean absolute error (as shown in equation (1)) and the maximum absolute error based on the prescribed dose were calculated for each region of interest, thereby evaluating the accuracy of the correlation model, with the model verification results shown in table 1:
where | is an absolute value operation, DclinIs the actual dose value, D, of a single voxelpredIs the predicted dose value, D, of a single voxelpresIs the prescription dose value, n is the number of voxels for a single patient;
table 1: model verification results
Fig. 5 shows a comparison of predicted and actual doses, where dose value (dose) is plotted on the abscissa, volume percentage (volume) is plotted on the ordinate, comparing the actual Dose Volume Histogram (DVH) and the predicted Dose Volume Histogram (DVH) of the region of interest of the new patient, including BrainSterm (brainstem), PTV (target), Lobe (brain Lobe), Lens (crystal), Body (trunk), Parotid (Parotid), Spinalcord (spinal cord), Larynx (Larynx), etc. according to the actual dose and the predicted dose, respectively, as can be seen from FIG. 5(a) and FIG. 5(b), for each region of interest, the curve fit of the predicted dose to the actual dose is higher, that is, the dose distribution of different regions can be accurately predicted by the trained correlation model, for example, in fig. 5(a), curves C1 and C2 are the actual dose and the predicted dose of the brainstem region, respectively.
Correspondingly, the embodiment of the invention also provides a prediction system of three-dimensional dose distribution in the intensity modulated radiotherapy plan based on deep learning. The system comprises: the case database is used for storing effective intensity modulated radiotherapy plan data; an anatomical feature extraction module for extracting three-dimensional anatomical structure contour features of the region of interest of each patient from the case database, including a target region contour, a trunk contour, and a contour of an organ at risk; the anatomical feature processing module is used for dividing the three-dimensional anatomical structure contour of the region of interest of the patient into a plurality of two-dimensional contour slice images according to the size of the required image; the dose characteristic extraction module is used for extracting the dose characteristic of each patient from the case database, registering according to the size of a required image and dividing into a plurality of two-dimensional dose slice distribution maps; the model training module is used for building a deep convolution network, inputting a two-dimensional profile slice image of a patient and a corresponding two-dimensional dose slice distribution map, learning the mapping relation between the profile characteristic and the dose characteristic of the anatomical structure through model training, and performing cross validation on the model to obtain an association model between the profile characteristic and the dose characteristic of the anatomical structure; a dose prediction module to predict a three-dimensional dose distribution of the new patient using the trained correlation model.
It should be noted that, although the steps are described in a specific order, the steps are not necessarily performed in the specific order, and in fact, some of the steps may be performed concurrently or even in a changed order as long as the required functions are achieved.
The present invention may be a system, method and/or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions embodied therewith for causing a processor to implement various aspects of the present invention.
The computer readable storage medium may be a tangible device that retains and stores instructions for use by an instruction execution device. The computer readable storage medium may include, for example, but is not limited to, an electronic memory device, a magnetic memory device, an optical memory device, an electromagnetic memory device, a semiconductor memory device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a Static Random Access Memory (SRAM), a portable compact disc read-only memory (CD-ROM), a Digital Versatile Disc (DVD), a memory stick, a floppy disk, a mechanical coding device, such as punch cards or in-groove projection structures having instructions stored thereon, and any suitable combination of the foregoing.
Having described embodiments of the present invention, the foregoing description is intended to be exemplary, not exhaustive, and not limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen in order to best explain the principles of the embodiments, the practical application, or improvements made to the technology in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.
Claims (7)
1. A prediction method of three-dimensional dose distribution in an intensity modulated radiotherapy plan based on deep learning is characterized by comprising the following steps:
step S1: collecting effective intensity modulated radiotherapy plan data to form a case database;
step S2: extracting three-dimensional anatomical contour features of the region of interest of each patient, including target region contours, torso contours, and organ-at-risk contours, from the case database;
step S3: dividing the three-dimensional anatomical structure contour of the region of interest of the patient into a plurality of two-dimensional contour slice images according to the required image size, wherein the division is carried out from different directions of the three-dimensional image, including a transverse plane, a coronal plane and a sagittal plane, and the slice number is enhanced by using rotary translation;
step S4: extracting the dose characteristics of each patient from the case database, registering according to the size of a required image and dividing into a plurality of two-dimensional dose slice distribution maps;
step S5: building a deep convolutional network, inputting a three-channel two-dimensional contour slice image of a section mode of a patient and a corresponding two-dimensional dose slice distribution map, learning a mapping relation between the contour feature and the dose feature of the anatomical structure through model training, and performing cross validation on the model to obtain an association model between the contour feature and the dose feature of the anatomical structure;
step S6: predicting a three-dimensional dose distribution of the new patient using the trained correlation model;
wherein the step S5 includes:
step S51: importing an open-source deep learning framework TensorFlow in Python;
step S52: constructing a deep convolution network by using a TensorFlow framework, wherein the number of feature maps of an input layer is N, and the feature maps represent the extracted contour features of the N regions of interest; the number of the characteristic graphs of the output layer is 1, and the characteristic graphs represent a dose distribution graph; a symmetrical 5-layer U-shaped network is connected between the input layer and the output layer, the left side of the U shape is a continuous down-sampling part, the down-sampling operation after each layer is completed by a cavitation layer power layer and a relu layer, and the size of the feature map of the upper layer is reduced by half; the right side of the U shape is a continuous up-sampling part, the up-sampling operation after each layer is completed by an upsample layer contribution layer and a relu layer, and the feature diagram of the upper layer is restored to be equal to the size of the left symmetrical layer; each layer of the U-shaped structure is composed of a plurality of denseblocks; two symmetrical sides of the U-shaped structure are connected in a jumping manner by series connection operation; selecting an optimized objective function as an MSE algorithm;
step S53: inputting the extracted two-dimensional profile slice image and the corresponding two-dimensional dose slice distribution map into the constructed depth convolution network, training the correlation model of the two-dimensional profile slice image and the corresponding two-dimensional dose slice distribution map, and evaluating the performance of the model through cross validation;
in step S52, each layer of the U-shaped structure includes a plurality of denseblocks, and the structure of the denseblocks is:
the denseblock is composed of a convolution structure A and a connection structure B, wherein the A comprises a convolution layer, an activation layer, a bantchnorm layer and a dropout layer, the B is composed of a concatenate layer, and an input feature graph after passing through the structure A is connected with an input feature graph through the structure B to serve as output;
the left side and the right side of the first layer of the U-shaped structure are respectively provided with a denseblock, the right side is connected with 4 convolution structures A behind the denseblock to obtain output layer data, the 2 nd layer and the 3 rd layer on the left side are provided with two denseblocks, the 4 th layer and the 5 th layer are provided with three denseblocks, and the right side is identical to the left side in structure.
2. The method for predicting three-dimensional dose distribution in intensity modulated radiotherapy planning based on deep learning of claim 1, wherein said step S2 comprises:
the extracted three-dimensional anatomical structure contour of the region of interest is a three-dimensional binary matrix, and a matrix value of 1 represents that a voxel of the region of interest exists therein.
3. The method for predicting three-dimensional dose distribution in intensity modulated radiotherapy planning based on deep learning of claim 1, wherein in step S53, the correlation model is trained according to the following steps:
step S531: collecting clinical plans of tumors of the same type in a case database, and extracting three-dimensional anatomical structure contour characteristics and dose characteristics of a patient of each plan;
step S532: randomly screening 80% of the plans in the collected case database as training data of the model, and using the rest 20% as verification data;
step S533: respectively inputting three-dimensional anatomical contour feature maps and dose feature maps of different regions of interest of a training plan into an established deep convolution network, determining model parameters and training parameters, and training to obtain an association model between an anatomical structure and dose features by using MSE as a loss function;
step S534: inputting contour feature maps of different regions of interest of the verification plan into the trained correlation model, and obtaining a corresponding dose feature distribution map;
step S535: calculating the mean square error of the verification data to evaluate the training effect of the model;
step S536: repeating the steps S532 to S535 for a plurality of times, and selecting the model with the best training effect as the correlation model.
4. The method for predicting three-dimensional dose distribution in intensity modulated radiotherapy planning based on deep learning of claim 3, wherein in step S533, the determined model parameters comprise convolution kernel size, activation function, dropout rate, and the determined training parameters comprise learning rate, optimization algorithm, and batch-size.
5. The method for predicting three-dimensional dose distribution in deep-learning-based intensity modulated radiation therapy planning as claimed in claim 1, wherein the step S6 of predicting three-dimensional dose distribution of new patient by using correlation model comprises the following steps:
step S61: extracting two-dimensional contour slice characteristics of a new patient according to the training data slice mode;
step S62: inputting two-dimensional contour slice characteristics of a new patient into the correlation model to calculate a corresponding dose characteristic distribution map;
step S63: and arranging and integrating the dose characteristic distribution map to obtain a three-dimensional dose distribution map of the new patient.
6. A system for predicting three-dimensional dose distribution in an intensity modulated radiotherapy plan based on deep learning, comprising:
case database: the system is used for storing effective intensity modulated radiotherapy plan sample data;
an anatomical feature extraction module: extracting three-dimensional anatomical contour features of a region of interest of each patient including a target region contour, a torso contour, a contour of an organ at risk from the case database;
an anatomical feature processing module: for dividing the three-dimensional anatomical contour of a region of interest of a patient into a number of two-dimensional contour slice images according to a desired image size, wherein the division is made from different directions of the three-dimensional image, including the transverse plane, coronal plane, sagittal plane, and rotational translation is used to enhance the slice number;
a dose feature extraction module: the two-dimensional dose slice distribution map is used for extracting dose characteristics of each patient from the case database, registering according to the size of a required image and dividing into a plurality of two-dimensional dose slice distribution maps;
a model training module: the system is used for building a deep convolution network, inputting a three-channel two-dimensional contour slice image of a patient in a section mode and a corresponding two-dimensional dose slice distribution map, learning a mapping relation between the contour feature and the dose feature of the anatomical structure through model training, and performing cross validation on the model to obtain an association model between the contour feature and the dose feature of the anatomical structure;
a dose prediction module: for predicting a three-dimensional dose distribution of a new patient using the trained correlation model;
wherein the model training module performs:
importing an open-source deep learning framework TensorFlow in Python;
constructing a deep convolution network by using a TensorFlow framework, wherein the number of feature maps of an input layer is N, and the feature maps represent the extracted contour features of the N regions of interest; the number of the characteristic graphs of the output layer is 1, and the characteristic graphs represent a dose distribution graph; a symmetrical 5-layer U-shaped network is connected between the input layer and the output layer, the left side of the U shape is a continuous down-sampling part, the down-sampling operation after each layer is completed by a cavitation layer power layer and a relu layer, and the size of the feature map of the upper layer is reduced by half; the right side of the U shape is a continuous up-sampling part, the up-sampling operation after each layer is completed by an upsample layer contribution layer and a relu layer, and the feature diagram of the upper layer is restored to be equal to the size of the left symmetrical layer; each layer of the U-shaped structure is composed of a plurality of denseblocks; two symmetrical sides of the U-shaped structure are connected in a jumping manner by series connection operation; selecting an optimized objective function as an MSE algorithm;
inputting the extracted two-dimensional profile slice image and the corresponding two-dimensional dose slice distribution map into the constructed depth convolution network, training the correlation model of the two-dimensional profile slice image and the corresponding two-dimensional dose slice distribution map, and evaluating the performance of the model through cross validation;
each layer of the U-shaped structure comprises a plurality of denseblocks, and the structure of each denseblock is as follows:
the denseblock is composed of a convolution structure A and a connection structure B, wherein the A comprises a convolution layer, an activation layer, a bantchnorm layer and a dropout layer, the B is composed of a concatenate layer, and an input feature graph after passing through the structure A is connected with an input feature graph through the structure B to serve as output;
the left side and the right side of the first layer of the U-shaped structure are respectively provided with a denseblock, the right side is connected with 4 convolution structures A behind the denseblock to obtain output layer data, the 2 nd layer and the 3 rd layer on the left side are provided with two denseblocks, the 4 th layer and the 5 th layer are provided with three denseblocks, and the right side is identical to the left side in structure.
7. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910341687.8A CN110085298B (en) | 2019-04-26 | 2019-04-26 | Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on deep network learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910341687.8A CN110085298B (en) | 2019-04-26 | 2019-04-26 | Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on deep network learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110085298A CN110085298A (en) | 2019-08-02 |
| CN110085298B true CN110085298B (en) | 2022-02-01 |
Family
ID=67416862
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910341687.8A Active CN110085298B (en) | 2019-04-26 | 2019-04-26 | Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on deep network learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110085298B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111028914B (en) * | 2019-12-04 | 2024-03-22 | 北京连心医疗科技有限公司 | Artificial intelligence guided dose prediction method and system |
| CN110739043A (en) * | 2019-09-17 | 2020-01-31 | 平安科技（深圳）有限公司 | Radiotherapy plan recommendation method, radiotherapy plan recommendation device, radiotherapy plan recommendation equipment and storage medium |
| CN110960349B (en) * | 2019-12-12 | 2024-06-18 | 朱登峰 | Prediction method for diopter adjustment value of SMILE refractive surgery |
| CN112037885B (en) * | 2020-09-07 | 2022-05-17 | 平安科技（深圳）有限公司 | Dose prediction method, device, computer equipment and storage medium in radiotherapy planning |
| CN113327681A (en) * | 2020-10-30 | 2021-08-31 | 重庆市璧山区人民医院 | Tumor radiotherapy plan automatic design method based on generating type confrontation network |
| EP4011442B1 (en) * | 2020-12-09 | 2024-05-08 | RaySearch Laboratories AB | Method and computer program product for dose map prediction for radiotherapy treatment planning |
| CN112635023B (en) * | 2020-12-16 | 2024-03-29 | 福建医科大学附属第一医院 | Method for generating dose prediction model of nasopharyngeal carcinoma, dose prediction method and device |
| CN112837782B (en) * | 2021-02-04 | 2024-07-02 | 北京大学第三医院（北京大学第三临床医学院） | Radiotherapy three-dimensional dose prediction method based on deep learning |
| CN113658168B (en) * | 2021-08-25 | 2024-04-16 | 上海联影医疗科技股份有限公司 | Method, system, terminal and storage medium for acquiring designated dose zone |
| CN113989349B (en) * | 2021-10-25 | 2022-11-25 | 北京百度网讯科技有限公司 | Image generation method, training method of image processing model, and image processing method |
| CN114155934B (en) * | 2021-12-10 | 2024-05-14 | 中南大学 | Tumor intensity modulated radiotherapy dosage prediction method based on deep learning |
| CN114596934A (en) * | 2022-05-10 | 2022-06-07 | 四川省肿瘤医院 | Cervical cancer brachytherapy dose prediction system |
| CN117423426A (en) * | 2023-09-21 | 2024-01-19 | 江西省肿瘤医院(江西省第二人民医院、江西省癌症中心) | Method, device and equipment for constructing treatment plan dose prediction model |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1504246A (en) * | 2002-11-29 | 2004-06-16 | 陈超敏 | Production of three-dimensional physical compensator using section cutting method to realize physical compensation radiotherapy |
| CN107441637A (en) * | 2017-08-30 | 2017-12-08 | 南方医科大学 | The intensity modulated radiation therapy Forecasting Methodology of 3-dimensional dose distribution and its application in the works |
| US20180240235A1 (en) * | 2017-02-23 | 2018-08-23 | Zebra Medical Vision Ltd. | Convolutional neural network for segmentation of medical anatomical images |
| CN109448006A (en) * | 2018-11-01 | 2019-03-08 | 江西理工大学 | A kind of U-shaped intensive connection Segmentation Method of Retinal Blood Vessels of attention mechanism |
| CN109635862A (en) * | 2018-12-05 | 2019-04-16 | 合肥奥比斯科技有限公司 | Retinopathy of prematurity plus lesion classification method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109377470A (en) * | 2018-03-20 | 2019-02-22 | 任昊星 | A kind of heart disease risk forecasting system |
| AU2018101317A4 (en) * | 2018-09-07 | 2018-10-11 | Chen, Guoyi Mr | A Deep Learning Based System for Animal Species Classification |
-
2019
- 2019-04-26 CN CN201910341687.8A patent/CN110085298B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1504246A (en) * | 2002-11-29 | 2004-06-16 | 陈超敏 | Production of three-dimensional physical compensator using section cutting method to realize physical compensation radiotherapy |
| US20180240235A1 (en) * | 2017-02-23 | 2018-08-23 | Zebra Medical Vision Ltd. | Convolutional neural network for segmentation of medical anatomical images |
| CN107441637A (en) * | 2017-08-30 | 2017-12-08 | 南方医科大学 | The intensity modulated radiation therapy Forecasting Methodology of 3-dimensional dose distribution and its application in the works |
| CN109448006A (en) * | 2018-11-01 | 2019-03-08 | 江西理工大学 | A kind of U-shaped intensive connection Segmentation Method of Retinal Blood Vessels of attention mechanism |
| CN109635862A (en) * | 2018-12-05 | 2019-04-16 | 合肥奥比斯科技有限公司 | Retinopathy of prematurity plus lesion classification method |
Non-Patent Citations (4)
| Title |
|---|
| U-Net: Convolutional Networks for Biomedical Image Segmentation;Olaf Ronneberger 等;《International Conference on Medical Image Computing & Computer-assisted Intervention》;20150518;附图1，第2页 * |
| 基于深度卷积网络的放疗计划三维剂量分布预测方法研究;郭芙彤;《中华医学会第十五次全功放射肿瘤治疗学学术会议》;20181111;第138页 * |
| 基于神经网络学习方法的放疗计划三维剂量分布预测;孔繁图 等;《南方医科大学学报》;20180630;说明书第685-686页 * |
| 郭芙彤.基于深度卷积网络的放疗计划三维剂量分布预测方法研究.《中华医学会第十五次全功放射肿瘤治疗学学术会议》.2018, * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110085298A (en) | 2019-08-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110085298B (en) | Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on deep network learning | |
| US11386557B2 (en) | Systems and methods for segmentation of intra-patient medical images | |
| US11896847B2 (en) | Adversarial prediction of radiotherapy treatment plans | |
| Babier et al. | Knowledge‐based automated planning with three‐dimensional generative adversarial networks | |
| US11944463B2 (en) | Pseudo-CT generation from MR data using a feature regression model | |
| EP3787744B1 (en) | Radiotherapy treatment plan modeling using generative adversarial networks | |
| US11100647B2 (en) | 3-D convolutional neural networks for organ segmentation in medical images for radiotherapy planning | |
| RU2720440C1 (en) | Image segmentation method using neural network | |
| AU2016339009B2 (en) | Pseudo-CT generation from MR data using tissue parameter estimation | |
| US11964170B2 (en) | Standardized artificial intelligence automatic radiation therapy planning method and system | |
| CN111028914B (en) | Artificial intelligence guided dose prediction method and system | |
| Rusanov et al. | Deep learning methods for enhancing cone‐beam CT image quality toward adaptive radiation therapy: A systematic review | |
| CN110415785A (en) | The method and system of artificial intelligence guidance radiotherapy planning | |
| CN108766563A (en) | Radiotherapy prediction of result method and system based on dosage group | |
| CN111462916B (en) | Method and device for predicting dose volume histogram of organ at risk of radiotherapy plan | |
| CN110021399A (en) | A kind of automatic design method of radiotherapy treatment planning | |
| Babier et al. | The importance of evaluating the complete automated knowledge-based planning pipeline | |
| Wang et al. | DeepEC: An error correction framework for dose prediction and organ segmentation using deep neural networks | |
| CN111888665B (en) | Construction method of three-dimensional dose distribution prediction model based on adaptive countermeasure network | |
| Babier et al. | Knowledge-based automated planning with three-dimensional generative adversarial networks | |
| CN116206729A (en) | Intensity modulated radiotherapy plan three-dimensional dose distribution prediction method based on convolutional neural network | |
| US20220241614A1 (en) | Fluence map prediction and treatment plan generation for automatic radiation treatment planning | |
| WO2022047637A1 (en) | Automatic beam modeling based on deep learning | |
| Spezialetti et al. | Using deep learning for fast dose refinement in proton therapy | |
| WO2024168504A1 (en) | Accelerated dose calculation using deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |