WO2021036366A1 - 标准化的人工智能自动放疗计划方法和*** - Google Patents

标准化的人工智能自动放疗计划方法和*** Download PDF

Info

Publication number
WO2021036366A1
WO2021036366A1 PCT/CN2020/091843 CN2020091843W WO2021036366A1 WO 2021036366 A1 WO2021036366 A1 WO 2021036366A1 CN 2020091843 W CN2020091843 W CN 2020091843W WO 2021036366 A1 WO2021036366 A1 WO 2021036366A1
Authority
WO
WIPO (PCT)
Prior art keywords
dose
radiotherapy
plan
model
angle
Prior art date
Application number
PCT/CN2020/091843
Other languages
English (en)
French (fr)
Inventor
李贵
***
范威阳
章桦
Original Assignee
北京连心医疗科技有限公司
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
Priority claimed from CN201910820650.3A external-priority patent/CN110415785A/zh
Priority claimed from CN201911229101.5A external-priority patent/CN111028914B/zh
Priority claimed from CN201911421531.7A external-priority patent/CN113130042B/zh
Application filed by 北京连心医疗科技有限公司 filed Critical 北京连心医疗科技有限公司
Priority to US16/977,095 priority Critical patent/US11964170B2/en
Publication of WO2021036366A1 publication Critical patent/WO2021036366A1/zh

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
    • 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
    • A61N5/1036—Leaf sequencing algorithms
    • 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
    • A61N5/1039—Treatment planning systems using functional images, e.g. PET or MRI
    • 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
    • A61N5/1048—Monitoring, verifying, controlling systems and 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
    • 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
    • A61N2005/1034—Monte Carlo type methods; particle tracking
    • 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

Definitions

  • the invention relates to the field of intelligent medical technology, in particular to a standardized artificial intelligence automatic radiotherapy planning method and a standardized artificial intelligence automatic radiotherapy planning system.
  • Tumor radiotherapy has become one of the main methods of breast cancer treatment, and it has become one of the three major methods of tumor treatment. Its key purpose is to reduce the dose deposition of surrounding normal tissues as much as possible while ensuring that the target area reaches the prescribed dose. Dosimetry verification is the main method of current clinical radiotherapy technology quality control and quality audit. Similarly, in the process of making radiotherapy plans, dose volume is also the main indicator for us to evaluate plan quality and forecast standards. However, the quality of radiotherapy plans is limited by the accumulation of experience of plan designers. The delineation of target areas of different breast types by different institutions and the equipment used in plan design are very different, and the consistency of plan quality is difficult to guarantee.
  • KBP Knowledge-based planning
  • PB-AIO Protocol-based Automatic Iterative Optimisation, protocol/template automatic iterative optimization
  • MCO Multi-Criteria Optimisation
  • the KBP-based method requires careful adjustment and optimization of the model, otherwise the tumor fitness and target coverage are not as good as the original manual plan; the predicted plan is only clinically acceptable, not necessarily optimal; based on
  • the parameters of the input template directly determine the quality of the plan. If the template parameters are not set well enough, the automatically generated plan is not as good as that made by an experienced physicist through manual optimization.
  • the use of the method is limited by the experience of the physicist; the MCO-based method is divided into a posteriori method and a priori method. The a priori method is still in the category of automation, and no AI method is used.
  • the resulting radiotherapy plan is mechanical.
  • the plan obtained by the posterior method is the Pareto optimal solution within the flux range without directly considering the optimization of the machine parameters.
  • the final plan needs to be converted.
  • the dose characteristics will change during the conversion process, especially in the case of low-density tissue on the target area. There will be obvious dose differences before and after the conversion. At this time, manual participation is required to carefully adjust the parameters.
  • the dose generation in the existing radiation therapy planning system is through dose optimization and dose calculation algorithms, which indirectly affect the dose by adjusting algorithm control parameters, dose volume constraint parameters or biological constraint parameters, editing flux or leaf sequence, etc. distributed.
  • dose optimization and dose calculation algorithm takes a long time, there are too many adjustable planning parameters, the planning adjustment strategy is not clear, and the way that indirect adjustment parameters affect the dose distribution is not intuitive.
  • the planning design process requires repeated adjustment of parameters and dose optimization. The design efficiency is not high.
  • the present invention provides a standardized artificial intelligence automatic radiotherapy planning method and system.
  • the prescription dose prediction model and exposure are added.
  • the automatic angle optimization process realizes fully automatic dose prediction, improves the efficiency and effect of dose prediction, so as to generate high-quality and fast executable radiotherapy plans with good accuracy, stability and standardization, so that it can Improve the utilization of medical software and hardware resources.
  • the present invention provides a standardized artificial intelligence automatic radiotherapy planning method, including: obtaining medical images; automatically delineating the ROI (region of interest) region of the medical image to obtain geometric anatomy Structure; Determine the prescription according to the disease type information corresponding to the medical image, the geometric anatomical structure, and the preset disease type-prescription template library; Determine the radiotherapy exposure according to the disease type information, the geometric anatomical structure and the prescription Angle; input the medical image, the geometric anatomical structure, the disease information, the prescription and the radiation angle of radiation into a dose prediction model to obtain a radiation dose distribution result; take the radiation dose distribution result as
  • a reverse optimization algorithm based on dose distribution or DVH guidance is used for optimization processing to generate an executable radiotherapy plan;
  • the executable radiotherapy plan includes a forward radiotherapy plan, a stereotactic radiotherapy plan, and an intensity-modulated radiotherapy plan, wherein the intensity-modulated radiotherapy plan
  • Radiotherapy plans include dynamic IMRT plans, static IMRT plans, volume IMRT plans and rotary
  • the radiotherapy planning method further includes: scoring the generated executable radiotherapy plan through the combination of unified prescription standards and artificial intelligence to obtain the total score of the plan evaluation; using Monte Carlo three-dimensional dose Verification technology to perform 2D or 3D Gamma analysis on the generated executable radiotherapy plan to obtain the pass rate of the Gamma analysis; automatically generate the radiotherapy plan based on the executable radiotherapy plan, the total score of the plan evaluation and the pass rate of the Gamma analysis Report; the doctor reviews the radiotherapy plan report.
  • the radiation therapy dose distribution result further includes: entering the dose editing mode when the dose editing trigger instruction is received; the cut plane graphics of the spatial dose model on the current radiotherapy image section follow the control cursor
  • the position of the control cursor is the center of the space dose model, and the trajectory of the control cursor corresponds to the movement trajectory of the action control device; the action event of the action control device is monitored and based on the action
  • the preset manipulation command corresponding to the event adjusts the dose at the center point of the space dose model; uses the dose at the center point of the space dose model as a reference to calculate the dose at the remaining points in the space dose model through an interpolation algorithm; saves and updates all
  • the dose data of each point in the space dose model is described, and the dose data of the area outside the space dose model is not modified.
  • the automatic delineation of the ROI area of the medical image to obtain the geometric anatomical structure specifically includes: automatic identification and automatic delineation of normal organs: automatic identification and delineation of various parts of the human body based on machine learning Normal organs; automatic identification and delineation of tumor sites: If the organs of the whole body are delineated, the tumors will be delineated in reverse; after the delineation of the endangered organs, the remaining part will be the tumor site; the disease type PTV (Planning Target Volume, using machine learning, The relationship between planned target volume and GTV (Gross Tumor Volume, tumor target volume), the remaining part is automatically outlined.
  • the determining the radiation angle of radiation therapy according to the disease type information, the geometric anatomical structure and the prescription specifically includes: performing machine learning on the disease type information, geometric anatomical structure and prescription of the historical case , Determine the radiation angle prediction model, and input the disease information, the geometric anatomical structure and the prescription of the current case into the radiation angle prediction model to obtain the predicted radiation angle as the radiation angle of radiation; or,
  • the organ weight of the target area of the disease marking plan calculate the cumulative value of the organ weight for each angle along the ray direction, merge the adjacent angles that meet the preset weight threshold, and use the angle that meets the weight threshold as the radiation angle; or,
  • the method for constructing the dose prediction model includes: establishing a data set based on the normalized average dose of PTV, and formulating a scoring template based on the data set; standardized naming the region of interest;
  • the medical image is divided into two-dimensional slices as the training set and the test set; the ray angle of the three-dimensional planning target area data of the training set is read, the ray angle is projected on the planning target area to obtain the network weight, and the network weight is calculated by dose
  • the algorithm performs dose calculation to obtain the beam channel; uses the U-net network or the V-net network as the generator and the Markov discriminator as the discriminator to construct the Pix2pix dose prediction model; uses the two-dimensional slice image as the generator Take the predicted dose and the original dose output by the generator as the input of the discriminator, and the discriminator outputs the judgment result; input all the two-dimensional slices of the training set into the Pix2pix dose prediction model training.
  • the radiation therapy dose distribution result is used as a reference, and an inverse optimization algorithm based on dose distribution or DVH guidance is used for optimization processing to generate an executable radiation therapy plan, which specifically includes: optimization based on flux map Algorithm to optimize the flux weight map; then, combined with the machine information of the accelerator, the leaf sequence algorithm is used to automatically generate an executable dynamic intensity-modulated radiotherapy plan; or,
  • volume IMRT plan or rotary IMRT plan Based on genetic algorithm or column generation algorithm, automatically generate volume IMRT plan or rotary IMRT plan; or,
  • the adjusting the dose at the center point of the spatial dose model according to the preset manipulation command corresponding to the action event specifically includes: when the action event is monitored, floating display of a dose adjustment indicator label; When the control cursor detects the first action event triggered by the action control device when the control cursor is in the dose adjustment indication label area, the dose value corresponding to the position of the control cursor is taken as the center point of the spatial dose model Dose; when the control cursor is on the indicator slider on the dose adjustment indicator label, it is monitored that the action control device is triggered by the second action event, then the control cursor is located when the second action event is clicked The dose value corresponding to the position is taken as the dose at the center point of the spatial dose model; when the control cursor is in the current section graphics area, it is detected that the action control device is triggered by the third action event, then the third action The action parameter of the event adjusts the dose at the center point of the spatial dose model; when the control cursor is not in the current section graphics area,
  • the radiotherapy planning method further includes: monitoring the action event of the action control device, and according to the preset control command corresponding to the action event, the radiotherapy image where the spatial dose model is located can also be turned over , And the size of the spatial dose model can be adjusted; when a fourth action event is triggered by the action control device, the size of the spatial dose model is adjusted according to the action parameters of the fourth action event.
  • the upper limit D1 and the lower limit Du of the dose adjustment at the center point of the spatial dose model are respectively:
  • Dl is the lower limit of the adjustable dose
  • Du is the upper limit of the adjustable dose
  • D0 is the point dose at the center point of the spatial dose model when the action event of the action control device is triggered
  • R is the spatial dose model
  • Dmax is the global maximum dose value of the dose data
  • n is a constant.
  • the present invention also provides a standardized artificial intelligence automatic radiotherapy planning system, which is used to realize the standardized artificial intelligence automatic radiotherapy planning method as described in any of the above technical solutions.
  • the present invention has the following beneficial effects: on the basis of the prediction model based on the geometric anatomical structure and the three-dimensional dose distribution of organs, the prescription dose prediction model and the automatic optimization process of the irradiation angle are added to realize the fully automatic dose. Prediction improves the efficiency and effect of dose prediction, thereby generating high-quality and fast-acting radiotherapy plans with good accuracy, stability and standardization, thereby improving the utilization of medical software and hardware resources.
  • the method of directly editing the dose allows the user to intuitively and directly obtain the desired dose distribution, which is faster and more intuitive than the indirect adjustment parameter affecting the dose distribution, which greatly improves the efficiency of plan design.
  • FIG. 1 is a schematic flowchart of a standardized artificial intelligence automatic radiotherapy planning method disclosed in an embodiment of the present invention
  • FIG. 2 is a schematic diagram of the image display of the Beam channel disclosed in an embodiment of the present invention.
  • FIG. 3 is a schematic diagram of the principle of a cGAN generation network disclosed in an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of the principle of the Pix2pix generation network disclosed in an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of the principle of a generator model disclosed in an embodiment of the present invention.
  • FIG. 6 is a schematic diagram of the principle of the discriminating process of the discriminator disclosed in an embodiment of the present invention.
  • FIG. 7 is a schematic diagram of the comparison between the predicted dose and the original dose disclosed in an embodiment of the present invention.
  • FIG. 8 is a schematic diagram of a comparison of data volume averages of a training set and a prediction set disclosed in an embodiment of the present invention.
  • FIG. 9 is a schematic diagram of the comparison of the difference between the average organ volume of the training data and the prediction data disclosed in an embodiment of the present invention.
  • Fig. 10 is a schematic diagram of a dice similarity coefficient curve of a typical absolute dose disclosed in an embodiment of the present invention.
  • FIG. 11 is a schematic diagram of a loss curve of iterative training disclosed in an embodiment of the present invention.
  • Figure 12 is a flow chart of data preprocessing disclosed in an embodiment of the present invention.
  • Figure 13 is a beam data statistical diagram disclosed in an embodiment of the present invention.
  • Figure 14 is a beam data screening diagram disclosed in an embodiment of the present invention.
  • FIG. 15 is an effect diagram of using a GAN network for prediction according to an embodiment of the present invention.
  • FIG. 16 is a diagram of DVH comparison results disclosed in an embodiment of the present invention.
  • Figure 17 is a cross-sectional view of a simulated head and neck tumor disclosed in an embodiment of the present invention
  • FIG. 18 is a mapping dictionary list of standard names and aliases disclosed in an embodiment of the present invention.
  • Figure 19 is a flowchart of a dose editing method disclosed in an embodiment of the present invention.
  • Fig. 20 is a schematic block diagram of the flow of a dose editing method disclosed in an embodiment of the present invention.
  • a standardized artificial intelligence automatic radiotherapy planning method includes: obtaining medical images; automatically delineating the ROI area of the medical images (such as CT and MR images) to obtain the geometric anatomical structure; Determine the prescription according to the medical image corresponding disease information, geometric anatomical structure, and preset disease-prescription template library; determine the radiation angle according to the disease information, geometric anatomical structure and prescription; combine the medical image, geometric anatomy, and disease
  • the information, prescriptions and radiation angles are input into the dose prediction model to obtain the radiation dose distribution results; using the radiation dose distribution results as a reference, the reverse optimization algorithm based on dose distribution or DVH guidance is used for optimization processing to generate an executable radiation treatment plan;
  • the implementation of radiotherapy plans includes forward radiotherapy plans, stereotactic radiotherapy plans and intensity-modulated radiotherapy plans.
  • Intensity-modulated radiotherapy plans include dynamic intensity-modulated radiotherapy plans, static intensity-modulated radiotherapy plans, volume-intensity-modulated radiotherapy plans, and rotational intensity-modulated
  • the prescription dose template and angle automatic optimization process are added to realize fully automatic dose prediction; the specific process is as follows:
  • Medical image acquisition Obtain patient images through CT machine or nuclear magnetic (MR) and store them in a preset format.
  • OIS Tumor Information System
  • MR nuclear magnetic
  • Dicom standard preferably Dicom standard
  • Organ delineation automatically delineate the acquired medical images to obtain the geometric anatomical structure, where the delineation process includes the delineation of normal organs and tissues and the delineation of tumor target areas;
  • the prescription is automatically determined according to the above outlined information and disease information; the prescription is determined by the mapping relationship between the prescription and the disease in the preset disease-prescription template library, and the disease-prescription template library needs to be defined in advance Great;
  • Angle determination use prescription, disease type, and outline information to automatically determine the irradiation angle
  • Dose prediction needs to be completed before dose prediction. Model training only requires one training. Daily use only needs to input data in a set format to complete dose prediction.
  • the dose prediction model is A dose prediction model based on Pix2pix for the patient's geometric anatomical structure and three-dimensional dose distribution of organs, the input data adopts the Dicom standard format;
  • a deep learning neural network model is established. After training an excellent radiotherapy planning data set through deep learning methods, a high-quality automatic planning model can be generated, which can be used according to The medical image input by the user is used to predict the radiotherapy plan.
  • the above-mentioned whole process is completed without human intervention, and when only a computer is occupied, the efficiency of the production of radiotherapy plans is greatly improved, thereby greatly reducing the patient's waiting time for treatment, and indirectly improving the curative effect.
  • automatically delineating the ROI area of the medical image to obtain the geometric anatomical structure specifically includes: automatic identification and automatic delineation of normal organs: automatic identification and delineation of normal organs of the human body based on machine learning; tumor; Automatic identification and delineation of parts: If the organs of the whole body are delineated, the tumor will be delineated in reverse; after the delineation of the organ at risk is completed, the remaining part will be the tumor site; the relationship between PTV and GTV expansion using machine learning is correct. The remaining part is automatically outlined.
  • the automatic target area delineation based on deep learning can realize:
  • Automatic identification and automatic delineation of normal (endangered) organs can automatically delineate various organs of the human body based on machine learning;
  • the ROI of interest for the ROI under delineation determines the ROI of interest for the ROI under delineation, and select at least one PTV and one OAR (Organ At Risk).
  • PTV Physical Broadcast
  • OAR Organic Radar At Risk
  • For each ROI perform projection within the beam angle range of 0-360 degrees, given a set of initial segments (representing the position of JAW, index is represented by i), and each segment is given a set of initial angles (index Denote by j).
  • At an angle within a segment find the minimum bounding rectangle of PTV(i,j) (treatment plan target area), and record it as block(i,j).
  • OAR the index is represented by k
  • the intersection of block (i, j) is performed at this angle, and the intersection area is obtained.
  • the three methods for determining the radiation angle of radiotherapy according to disease information, geometric anatomical structure and prescription, and the three methods specifically include:
  • the organ weights of different diseases are marked according to different diseases.
  • the tumor target area is marked as 0; the organ weight can be divided into the given organ sub-weight; the weight marking method can be used to allow the maximum exposure of the organ
  • the reciprocal of the dose is determined.
  • the larger the maximum allowable exposure dose the smaller the weight.
  • the organ weight or sub-weight accumulated value of each angle is calculated according to the preset angle interval along the ray direction. When selecting an angle that meets the weight threshold, if the number of angles is less than the preset minimum value, it defaults to the preset value, and if the angle is greater than the preset maximum value, it defaults to the maximum value.
  • the beam angle is also called the ray angle or beam direction.
  • the purpose of radiotherapy is to send a high enough dose to the planned target area to control the tumor, while at the same time ensuring that the surrounding normal tissues and organs at risk (OARs) are at an acceptable dose level to avoid damage.
  • OARs normal tissues and organs at risk
  • the setting of the beam angle affects the planned target area and the dose to the organ at risk, and has an important impact on the quality of the treatment plan. Due to the influence of the curved surface and uneven tissue of the human body on the dose distribution, it is difficult to determine the angle of the radiation field during the design of the treatment plan, making it a time-consuming, trial and error to determine the most suitable radiation field incidence direction for each patient. the process of. Therefore, considering the ray angle in the training process helps to predict a more accurate dose distribution map that is more in line with clinical requirements.
  • the method to generate the beam channel is as follows: first take out the 3D PTV data, read out the beam angle contained in the case, and project the beam angle onto the PTV to obtain the network weight of the beam channel data (here, the beam will fall within the range of the beam).
  • the area is set to 1, and the other positions are set to 0), and the high-speed dose calculation algorithm is used to directly perform dose calculation on the network weight to obtain the Beam channel.
  • the image display of the generated Beam channel is shown in Figure 2.
  • a dose prediction model based on Pix2pix is used.
  • the method for constructing the dose prediction model includes: establishing a data set based on the normalized average dose of PTV, and formulating a scoring template based on the data set; Carry out standardized naming; divide the 3D medical image into 2D slices as the training set and test set; read the ray angle of the 3D plan target area data of the training set, and project the ray angle on the plan target area to obtain the network weight.
  • the weight uses the dose calculation algorithm to calculate the dose to obtain the beam channel; the U-net or V-net network is used as the generator, and the Markov discriminator is used as the discriminator to construct the Pix2pix dose prediction model; the two-dimensional slice image is used as the generator Take the predicted dose and the original dose output by the generator as the input of the discriminator, and the discriminator outputs the judgment result; all the two-dimensional slices of the training set are input to the Pix2pix dose prediction model for training.
  • the model training involves the formulation of a scoring template, the standardization of the naming of the region of interest, and the model training process.
  • Pix2pix is a GAN-based image translation model.
  • the GAN network contains the generator G and the discriminator D, which mutually restrict each other and promote each other.
  • the image generated by G and the ground truth are handed over to D for judgment at the same time.
  • the judgment result is the probability that the generated image is a real image. If the probability is large , Which shows that the image generated by G is very close to the original image, thus deceiving D; if the judgment is false, it means that D has recognized that the generated image is quite different from the real image.
  • this embodiment takes the lead in using the Pix2pix model plus the Beam channel to predict the radiation dose.
  • Pix2pix is a GAN-based image translation model.
  • the GAN network contains generator G and discriminator D. The two restrict each other and promote each other.
  • the image generated by G and the ground truth are simultaneously handed over to D for judgment, and the judgment result is generated
  • the probability that the image is a real image If the probability is high, it means that the image generated by G is very close to the original image, which deceives D; if the judgment is false, it means that D has identified a big difference between the generated image and the real image.
  • the input of cGAN generation network G includes noise Z and condition Y, and the output generates a fake_x.
  • the input of the discrimination network D includes fake_x or real_x and condition Y, and the output is the judgment result 0 or 1, namely FAKE or REAL.
  • Pix2pix draws on the idea of cGAN.
  • cGAN When inputting the G network, it will not only input noise, but also input a condition (condition).
  • the fake images generated by the G network will be affected by specific conditions. Taking the image as the condition, the generated fake images and the condition images have a corresponding relationship, thus realizing an image-to-image translation process.
  • the input terminal of the generation network G of Pix2pix has only one condition Y, where Y is a picture imgA.
  • the U-net structure is used to generate the network G, and the input Y code is decoded into the real image imgB'.
  • the input of the discriminator is the generated image imgB' or the real image real_x(imgB) and the condition Y, and finally the image-to-image transformation is realized.
  • the input Y is a 4-channel image, including a 3-channel dose image and a beam channel.
  • the predicted dose fake_x is obtained through the U-net generator, and the original dose and the generated predicted dose are put into the discriminator together , Judge the difference between the predicted dose and the actual dose and enter the judgment result.
  • the generator in this embodiment uses an 8-level U-net to realize the mapping from image to dose.
  • the entire network structure can be regarded as a feature extraction part and an up-sampling part.
  • the input starts with 4 channels of a 256 ⁇ 256 pixel image.
  • the feature extraction part performs a 3 ⁇ 3 convolution operation for each layer, and uses a 2 ⁇ 2 maximum pooling layer to the next layer.
  • the purpose is to reduce the feature size of 256 ⁇ 256 pixels to 1 ⁇ 1 pixels.
  • the up-sampling part the same convolution kernel is used to convolve each layer of data, and when entering the next layer, the maximum pooling layer becomes 2 ⁇ 2 deconvolution, the purpose is to transform the image to the original image size.
  • the method shown in Figure 5 is used to retain the underlying features.
  • the final output image is a dose map of 256 ⁇ 256 ⁇ 1.
  • Adam algorithm is selected as the optimizer to minimize the loss function.
  • the learning rate is 2 ⁇ 10 -5 and the epochs is 100; in the second stage, the learning rate is 2e-06 and the epochs is 300.
  • This embodiment is divided into two stages of training. On the one hand, the convergence speed can be improved, and on the other hand, the training can be continued after the training is interrupted without affecting the result.
  • the discriminator uses a Markov discriminator (PatchGAN) as shown in Figure 6 to discriminate whether it is a generated picture. Because different patches can be considered to be independent of each other, the idea of PatchGAN is to allow the discriminator to distinguish between true and false for each patch with a size of N ⁇ N. Pix2pix cuts a picture into different patches of N ⁇ N size, and the discriminator judges whether each patch is true or false, and averages the results of all patches of a picture as the final discriminator output. For 256 ⁇ 256 input, when the patch size is 70 ⁇ 70, the judgment result is the best.
  • PatchGAN Markov discriminator
  • Figure 7 shows the comparison between the predicted dose image and the real dose image of the same case.
  • the left side is the predicted dose image, and the right is the original dose image. It can be seen that thanks to the fact that the generator uses the U-net network to save the underlying information, the detailed information of the predicted dose image is better preserved.
  • the difference between the receptor volumes will also have an impact on the receptor dose.
  • the MEAN value of the training set plan target volume (PTV) volume is 915.46 cm 3
  • the GTV volume MEAN value is 658.74 cm 3
  • the left lung volume MEAN value is 1001.09 cm 3
  • the right lung volume MEAN value is 1315.38 cm 3.
  • the MEAN value of the heart volume is 533.33 cm 3
  • the MEAN value of the spinal cord volume is 40.95 cm 3 .
  • the MEAN value of the target volume (PTV) of the prediction set plan is 978.1cm 3
  • the GTV volume MEAN value is 743.06cm 3
  • the left lung volume MEAN value is 981.9 cm 3
  • the right lung volume MEAN value is 1329 cm 3
  • the heart volume MEAN value is 552.3 cm 3
  • the MEAN value of spinal cord volume is 43 cm 3 .
  • the volume average of each labeled organ in the data is compared, and the calculation formula is predict mean- raw mean .
  • the PTV volume gap reached 95.1 cm 3
  • the volume gap between the lungs, heart, and spinal cord was less than 20 cm 3 .
  • the dose rate of change D2 and D98 are 0.33% and 3.55%, respectively.
  • the predicted dose is lower than the original dose, but the dose of D95
  • the average values all meet the prescription dosage requirements and meet the clinical dosage requirements.
  • the left lung volume difference is small, and although the dose change rate reaches 8.19%, it belongs to the clinically acceptable range because the base is only about 1.5Gy.
  • the right lung is a large-area receiving area, and based on the small volume difference, the dose change rates of V5, V10, and V20 are all less than 1%. Although the dose change rate of V30 reaches 12.5%, the predicted average dose is compared with the actual dose. The difference between the averages is only 3.1 Gy, which is within the clinically acceptable range.
  • the spinal cord dose difference is about 2.8 Gy, and the volume difference is 2.1 cm 3 , which is also within the clinically acceptable range.
  • Table 1 shows the comparison between the predicted average dose and the actual average dose change rate.
  • the calculation formula of HI and CI is: For these two indicators, the smaller the value of HI and the value of CI are approximately 1, the better the radiotherapy plan is made. For large organs such as the left lung and right lung, the predicted change rate is 0.12 at most. For small organs, such as the heart, although the change rate is large, the actual difference between the predicted average dose and the actual average dose is only 0.176 Gy. The difference between the prediction data set and the original data in the prediction result is within the clinically acceptable range, and the rate of change compared with the training set data is not large.
  • the similarity coefficient between the predicted dose and the actual dose is expressed by the Dice similarity coefficient. It can be seen from Figure 10 that in the absolute dose range of 3-50Gy, the Dice value under the 20Gy dose gradually rises between 0.76 and 0.86, and the Dice value in the 20-45Gy range fluctuates between 0.86 and 0.9, and the absolute dose is 45Gy. Above, the Dice value has a slow downward trend, but it remains above 0.83.
  • the similarity coefficient below the absolute dose of 20Gy is low, but although the similarity coefficient is low in the low-dose region, the rate of change and the performance on DVH meet the clinical dose requirements.
  • Dose-volume histogram is currently a widely accepted treatment plan evaluation method in three-dimensional conformal radiotherapy, which visually represents the relationship between dose and volume in the target area and normal tissues.
  • the DVH of the target area can show the uniformity of the irradiation, and the DVH of the normal tissue can provide the irradiated dose of the organ and its corresponding volume, especially for normal organs whose radiation tolerance is related to the irradiated volume. It has very important clinical significance.
  • Dose volume index as a good method to evaluate the radiotherapy plan, has the characteristic of intuitively feeling the dose change.
  • the physicist can use the eclipse software to manually adjust the weight and the dose on the DVH to monitor the dose change in real time. It can be seen directly from the image that the PTV curve and the BODY curve basically coincide, which ensures that the target area is adequately exposed to the dose and the body is kept within the clinically acceptable range.
  • FIG. 11 shows the generative model. Loss graph. The loss of the training model started at 13,800 and dropped to 4210 after 50 iterations. After 300 iterations of training, the curve converges smoothly, and the loss of 350 iterations is reduced by 40 compared with 300 iterations. Finally, the model converges to 5 after 400 iterations. There is no greater fluctuation in the loss curve after continuing training. This is the reason why 400 iterations were selected. .
  • Pix2pix is used in combination with the field angle to realize the dose prediction of the complex target area of breast cancer.
  • the dose prediction is defined as an image coloring problem, and the Pix2pix dose prediction model performs well on this problem.
  • the Pix2pix dose prediction model is convenient to use in experiments for data pairing requirements.
  • the generator in the model uses the u-net network, which retains the underlying information, and provides a guarantee for the model to predict the details of the image. The experimental results show that by comparing the change rate of the target dose volume parameter with that of the normal organ, a clinically acceptable predictive radiotherapy dose result is obtained.
  • the beam angle When training and predicting, the beam angle is used as a prior condition, the cases are labeled, and the machine learning model with conditional probability is used for training and prediction.
  • the user enters a custom beam angle combination (because of a The plan has multiple beams, so there are multiple beam angles, here is called a beam angle combination).
  • the system can automatically give a set of recommended beam angles by default.
  • the recommended value is a set of optimal beam angles found by the clustering method in machine learning based on cases of the same disease type. It is also possible to establish a prediction model based on the anatomical characteristics of historical cases and images, and automatically predict the appropriate beam angle combination based on the input CT images.
  • the model that can be referred to is the conditional generation confrontation network CGAN, and the original generation confrontation network model has been implemented.
  • CGAN schematic diagram is shown as in Fig. 3.
  • the input CT (and RS) data is preprocessed.
  • the brackets here represent two possible solutions, including RS processing or not including RS processing.
  • the CT data is rotated around a certain axis (the gantry rotation axis is recommended) according to the beam angle (and the MU weight in the existing case plan, and the same MU weight is included as an example) , Crop, position, and then synthesize.
  • the synthesis is based on MU as the weight, and the synthesized CT data is obtained by the center of gravity method. If the MU is not considered, the weight is set to 1.
  • the virtual CT is reconstructed.
  • the coordinates are rotated, and the rotated coordinates are also generated according to the center of gravity method to produce the final coordinates, and the weights refer to the synthesis method of CT.
  • the flowchart is shown in Figure 12.
  • the left image represents the data preprocessing process without considering the beam angle; the right image uses three slices for each set of CT as an example to illustrate the reconstruction process.
  • preprocessing is used to filter out plans that deviate from the main beam combination and discard them, and the collected historical case plans are screened according to the formula proposed below.
  • Screening is carried out according to the standard of 1 times the standard deviation, and the cases beyond are not used for training. The rest can be used as a training set.
  • This method reduces the impact of different beam angles on the predicted dose distribution to a certain extent, but reduces the cases that can be used for training to a certain extent.
  • An example of 20 sets of planned cases is shown in the table below. First, extract all beam angles from 20 sets of cases.
  • Each row represents an array of planned angles.
  • the length of the array is determined according to the maximum number of beams, and the number of beams for each plan is recorded in advance with a variable.
  • the beam data is counted, and the image shown in Figure 13 is obtained.
  • the radiation therapy dose distribution result is used as a reference, and an inverse optimization algorithm based on dose distribution or DVH guidance is used for optimization processing to generate an executable radiation therapy plan, which specifically includes: an optimization algorithm based on a flux map, and optimization Flux weight map; then, combined with the machine information of the accelerator, the leaf sequence algorithm is used to automatically generate an executable dynamic intensity modulated radiotherapy plan; or,
  • volume IMRT plan or rotary IMRT plan Based on genetic algorithm or column generation algorithm, automatically generate volume IMRT plan or rotary IMRT plan; or,
  • Automatic reverse planning is to use the predicted dose distribution or DVH as the reference dose distribution or reference DVH, and use a reverse optimization algorithm based on the dose distribution or DVH guidance, combined with the voxel dose unit calculated by a specific dose calculation engine, to optimize
  • the flux weight map combined with the machine information of a specific accelerator, uses the leaf sequence algorithm to automatically obtain the final treatment plan; the optimization algorithm can be implemented using a linear model or a nonlinear model.
  • the generated plan automatically calls the automatic plan evaluation-including the scores and total scores, which are provided to the doctor for approval;
  • plan optimization engine uses a series of linear objective functions to form the reverse plan optimization problem:
  • ⁇ is the weight parameter vector
  • g is the smooth objective function expression matrix
  • A is the constraint factor matrix
  • the final optimization model is a standard linear programming problem.
  • the model can optimize the value of x, and then substitute the value into the original reverse planning optimization model:
  • ⁇ is known, the above model is a standard linear programming problem, and x (Fluence Map) can also be optimized, and finally a high-quality plan can be generated.
  • the algorithm test of the automatic planning prototype proposed in the above embodiment is carried out. It can automatically optimize a plan that meets regulatory requirements without manual intervention.
  • the total time is about 5 minutes, and the only thing the user needs to input is the prescription dose requirement and the predicted three-dimensional dose data.
  • the final DVH comparison result is shown in Figure 16.
  • the solid line and the dashed line respectively represent the artificially optimized DVH and the automatically reverse optimized DVH, among which the predicted DVH used in the automatic optimization.
  • Figure 17 The constraints and target requirements of the example questions in the national standard are shown in Figure 17. It can be seen from Figure 17 that the DVH result of the automatic plan is slightly worse than that of the manual PTV, but it is also a plan that meets the requirements. And the automatic plan only takes about 5 minutes, while the manual plan takes at least an hour.
  • the method of automatic dose prediction in the present invention is:
  • the above-mentioned automatic dose prediction step can be replaced with automatic dose volume histogram (DVH) prediction:
  • DVH automatic dose volume histogram
  • the radiotherapy planning method further includes: scoring the generated executable radiotherapy plan through the combination of unified prescription standards and artificial intelligence to obtain the total score of the plan evaluation; using Monte Carlo three-dimensional dose Verification technology to perform 2D or 3D Gamma analysis on the generated executable radiotherapy plan to obtain the pass rate of the Gamma analysis; automatically generate the radiotherapy plan based on the executable radiotherapy plan, the total score of the plan evaluation and the pass rate of the Gamma analysis Report; the doctor reviews the radiotherapy plan report.
  • the generated executable radiotherapy plan is scored, and the total score of the plan evaluation is obtained; among them:
  • the prerequisite for automatic planning is to have an excellent planning database. Therefore, an algorithm and tool for automatic screening of excellent plans is needed. This tool can be used for planning screening, and can also be used for scoring after the automatic plan is generated.
  • the evaluation software is positioned as a multi-functional information and data management application, used to help doctors and physicists improve and enhance the standardization of radiotherapy plans, and used to score and screen existing plan cases, and select higher and lower scores
  • the cases were used for machine learning training and prediction testing.
  • the input of the application is the DICOM data exported by the treatment planning system (TPS), including two parts: RS and RD.
  • TPS treatment planning system
  • the output is the scoring result.
  • the scoring template required by the evaluation software is formulated by each hospital's unified rules, and different templates can be developed for different disease types or even more specific classifications. International standards and the hospital’s own internal standards can be referred to in the formulation process.
  • the dose prediction model uses the average PTV dose of 5000 cGy to standardize the plan.
  • the normalization of PTV average dose establishes a unified data set, which is more conducive to training the model, and the normalized plan has greater clinical relevance and evaluation value.
  • the production of the scoring template is based on the summary and exchange of information from RTOG-1005, physicists with more than 5 years of work experience, and radiologists.
  • the scoring template contains items as follows: PTV’s V48, V50, V53, V55, DMAX, D2, D98, HI, CI; heart’s V10, DMEAN; left lung’s V4, V5, DMEAN; right lung’s V4, V5, V8, V10, V20, V30, DMEAN and DMAX of the spinal cord.
  • the data is scored by setting the upper and lower limits of volume and volume. The closer the upper limit is, the higher the score will be, and the lower limit will be scored. At the same time, different weights are assigned to each attribute to make it more in line with doctors' prescription requirements and clinical needs.
  • the purpose of making a scoring template is as follows: 1 In the process of selecting data, there may be left and right breast errors, disease errors and other things. The wrong data is selected by formulating a scoring template to avoid affecting the accuracy of the model. 2 Normalized data is conducive to the accuracy of the training model.
  • the radiotherapy plan report is automatically generated based on the executable radiotherapy plan, the total score of the plan evaluation and the pass rate of the Gamma analysis obtained in the above embodiment; the radiotherapy plan report is then reviewed by the doctor; among them:
  • the audit report has abstract and detailed content.
  • the summary describes the pass rate of the 3D QA Gamma analysis, the total score of the plan evaluation, and the summary of the plan to be executed; the detailed introduction of each content is convenient for doctors to carefully review the automatic plan. If the approval is passed, it will be published to the medical accelerator for execution; if the approval is not passed, manual intervention is allowed to modify the plan manually.
  • the radiation therapy dose distribution result further includes: entering the dose editing mode when the dose editing trigger instruction is received; the section of the spatial dose model on the current radiotherapy image section The graph moves with the trajectory of the control cursor, where the position of the control cursor is the center of the space dose model, and the trajectory of the control cursor corresponds to the movement trajectory of the action control device; monitors the action event of the action control device, and according to the preset corresponding to the action event Set the control command to adjust the dose at the center point of the space dose model; use the dose at the center point of the space dose model as the reference to calculate the dose at the remaining points in the space dose model by interpolation; save and update the dose data at each point in the space dose model, and Do not modify the dose data in the area outside the space dose model.
  • the user can intuitively and directly obtain the desired dose distribution, which is faster and more intuitive than the indirect adjustment of the parameter influencing the dose distribution, which greatly improves the planning design. effectiveness.
  • adjusting the dose at the center point of the spatial dose model according to the preset manipulation command corresponding to the action event specifically includes: when the action event is monitored, the dose adjustment indication label is displayed floating; When adjusting the indicator label area, if the first action event is triggered by the action control device, the dose value corresponding to the position of the control cursor is used as the dose at the center point of the spatial dose model; when the indicator slides on the control cursor on the dose adjustment indicator label When the second action event is triggered by the action control device when the block is on, the dose value corresponding to the position of the control cursor when the second action event is released is used as the dose at the center point of the spatial dose model; when the control cursor is in the current section graph If the third action event is triggered by the action control device when in the area, adjust the dose at the center point of the spatial dose model with the action parameters of the third action event; if the control cursor is not in the current section graphics area, if the action control is monitored When the device is not in the current section graphics area, if the action
  • the radiotherapy planning method further includes: monitoring the action event of the action control device, and according to the preset control command corresponding to the action event, the radiotherapy image where the spatial dose model is located can also be turned over, and The size of the space dose model can be adjusted; when the fourth action event is triggered by the motion control device, the size of the space dose model is adjusted according to the action parameters of the fourth action event.
  • the spatial dose model is a sphere, cube, cuboid or ellipsoid
  • the action control device is a mouse or other human-computer interaction control device. If a mouse is used as the action control device, the first action event of the action control device can be the mouse left The button is clicked, the second action event can be the left mouse button being clicked and kept in the clicked state, the third action event can be the mouse wheel being scrolled, and the fourth action event can be the right mouse button being clicked and kept in the clicked state.
  • the dose at the center point of the dose sphere and other points in the space model dose model are updated and saved to the database after the dose is calculated, and the front-end update display is performed in the form of DVH, isodose line and/or dose volume histogram statistical data.
  • the method of dose editing in the radiation treatment planning system is described in detail. Then the method of dose editing in the radiation treatment planning system is specific include:
  • the tangent circle of the dose ball on the current radiotherapy image section moves with the trajectory of the control cursor, where the position of the control cursor is the center of the dose sphere, and the trajectory of the control cursor corresponds to the movement trajectory of the mouse;
  • the current position of the control cursor is taken as the center point position of the dose ball and the dose adjustment indicator bar is displayed at the same time.
  • the position of the dose slider is the current point dose at the center of the sphere;
  • the mouse wheel can increase and decrease the dot dose at the center of the sphere when the mouse wheel is scrolled up and down.
  • the mouse wheel can be scrolled up by n steps.
  • the scroll step is (Du-D0)/n
  • the mouse wheel is moving toward Down can be scrolled m steps
  • the rolling step length is (D0-Dl)/m, where Dl is the lower limit of the adjustable dose, Du is the upper limit of the adjustable dose, and D0 is the point at the center of the dose ball when the left mouse button is clicked Dose, n and m are constants;
  • the radius of the dose ball can still be displayed and adjusted by pressing the right mouse button and moving;
  • the front-end updates display dose data, including dose volume histogram, isodose line, DVH statistics, etc.
  • the interpolation algorithm includes linear interpolation, bilinear interpolation, cubic interpolation, bicubic interpolation, nearest neighbor interpolation, cubic convolution interpolation algorithm, natural neighbor interpolation, triangulation/linear interpolation, Chebet method, radial basis function method, multiple regression method, minimum curvature method, kriging method and distance reciprocal multiplication method.
  • the radius of the dose ball can also be adjusted when the right button of the mouse is clicked and kept in the clicked state.
  • the upper limit D1 and the lower limit Du of the dose adjustment at the center point of the spatial dose model are respectively:
  • Dl is the lower limit of the adjustable dose
  • Du is the upper limit of the adjustable dose
  • D0 is the point dose at the center point of the spatial dose model when the action event of the action control device is triggered
  • R is the characteristic parameter of the spatial dose model
  • Dmax is The global maximum dose value of the dose data
  • n is a constant.
  • the present invention also provides a standardized artificial intelligence automatic radiotherapy planning system, which is used to realize the standardized artificial intelligence automatic radiotherapy planning method as in any of the above embodiments.
  • the present invention also provides a computing device, including:
  • One or more processors are One or more processors;
  • One or more programs wherein one or more programs are stored in a memory and configured to be executed by one or more processors, and the one or more programs include instructions for realizing the above-mentioned standardized artificial intelligence automatic radiotherapy planning method .
  • the present invention also provides a computer-readable storage medium storing one or more programs, the one or more programs including instructions, which are adapted to be loaded by a memory and execute the above-mentioned standardized artificial intelligence automatic radiotherapy planning method.

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Veterinary Medicine (AREA)
  • Animal Behavior & Ethology (AREA)
  • Pathology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Radiology & Medical Imaging (AREA)
  • Surgery (AREA)
  • Urology & Nephrology (AREA)
  • Epidemiology (AREA)
  • Medical Informatics (AREA)
  • Primary Health Care (AREA)
  • Radiation-Therapy Devices (AREA)

Abstract

一种标准化的人工智能自动放疗计划方法和***，其中放疗计划方法包括：获取医学影像；对医学影像的ROI区域进行自动勾画以获取几何解剖结构；根据医学影像对应的病种信息、几何解剖结构以及预设的病种-处方模板库确定处方，并确定放疗照射角度；利用剂量预测模型，得出放疗剂量分布结果；以放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划。实现了全自动的剂量预测，提高了剂量预测的效率和效果，从而高质量、快速地生成可执行的放疗计划，具有较好的准确性、稳定性和规范性，并且能够直观、直接地编辑调整剂量分布，极大提高了计划设计效率。

Description

标准化的人工智能自动放疗计划方法和*** 技术领域
本发明涉及智能医疗技术领域，尤其涉及一种标准化的人工智能自动放疗计划方法和一种标准化的人工智能自动放疗计划***。
背景技术
肿瘤放射治疗已经成为乳腺癌肿瘤治疗主要方式之一，已成为肿瘤治疗的三大手段之一。其关键目的是在确保靶区达到处方剂量的同时，尽可能的降低周围正常组织的剂量沉积。剂量学验证是当前临床放疗技术质量控制与质量审核的主要方式。同样，我们在制作放疗计划的过程中，剂量体积也是我们评估计划质量和预测标准的主要指标。但放疗计划的质量受限于计划设计人员的经验累积，不同机构对于不同乳腺类型靶区的勾画，计划设计使用的设备等都存在很大的差异，计划质量的一致性难以保证。同时，临床计划多服从于群体化的规范标准，无法为患者提供个体化的治疗计划。如今，三维剂量分布预测模型分为基于BP(back propagation，反向传播)神经网络和基于深度卷积网络。但是基于BP神经网络的三维剂量分布预测方法需要人工手动提取特征，这导致了特征选择具有很强的主观性。而且上述方法的处方和照射角度需要有经验的医生和物理师确定的，无法做到全自动的剂量预测。
治疗计划的自动化方法有如下几种：基于KBP(Knowledge-based planning，经验知识制作计划)的方法、基于PB-AIO(Protocol-based Automatic Iterative Optimisation，协议/模板的自动迭代优化)的方法、基于MCO的方法(Multi-Criteria Optimisation，多准则优化)以及基于人工智能的自动放疗计划的方法。
但是，基于KBP的方法需要对模型仔细地调节和优化，否则肿瘤适形度，靶区覆盖率都不如原来的手动计划；预测的计划只是临床可接受的，并不一定是最优的；基于PB-AIO的自动优化方法中，输入模板的参数直接决定了计划的好坏，如果模板参数设置得不够好，那么自动生成的计划还不如有经验的物理师通过手动优化做出的，因此该方法的使用受限于物理师的经验；基于MCO的方法又分为后验方法和先验方法，先验方法尚处于自动化的范畴， 没有用到AI的方法，所得的放疗计划具有机械性，没有计划评估与三维剂量验证，无法确保计划优秀性与可靠性，后验方法得到的计划是通量范围内的帕累托最优解，而没有直接考虑机器参数的优化，最终的计划需要转换为可以治疗用的计划，而转化过程中剂量特性会发生改变，尤其是靶区上有低密度组织的案例在转化前后会出现很明显的剂量差异，这时候就需要人工参与仔细的调整参数。
此外，现有放射治疗计划***中剂量的生成均是通过剂量优化和剂量计算算法而来，通过调节算法控制参数，剂量体积约束参数或生物约束参数，编辑通量或叶片序列等间接影响剂量的分布。其中，剂量优化及剂量计算算法耗时较长，可调节的计划参数太多，计划调节策略不明确，间接调节参数影响剂量分布的方式不直观，计划设计过程需要反复调整参数和剂量优化，计划设计效率不高。
发明内容
针对上述问题中的至少之一，本发明提供了一种标准化的人工智能自动放疗计划方法和***，在基于几何解剖结构和器官三维剂量分布的预测模型基础上，增加了处方剂量预测模型和照射角度自动优化过程，实现了全自动的剂量预测，提高了剂量预测的效率和效果，从而高质量、快速地生成可执行的放疗计划，具有较好的准确性、稳定性和规范性，从而能够提高医疗软硬件资源利用率。
为实现上述目的，本发明提供了一种标准化的人工智能自动放疗计划方法，包括：获取医学影像；对所述医学影像的ROI(region of interest，感兴趣区域)区域进行自动勾画以获取几何解剖结构；根据所述医学影像对应的病种信息、所述几何解剖结构以及预设的病种-处方模板库确定处方；根据所述病种信息、所述几何解剖结构和所述处方确定放疗照射角度；将所述医学影像、所述几何解剖结构、所述病种信息、所述处方和所述放疗照射角度输入剂量预测模型中，得出放疗剂量分布结果；以所述放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划；所述可执行放疗计划包括正向放疗计划、立体定向放疗计划和调强放疗计划，其中所述调强放疗计划包括动态调强放疗计划、静态调强放 疗计划、容积调强放疗计划和旋转调强放疗计划。
在上述技术方案中，优选地，放疗计划方法还包括：通过统一的处方标准与人工智能结合，对所生成的可执行放疗计划进行评分，得到计划评估的评分总分；通过蒙特卡罗三维剂量验证技术，对所生成的可执行放疗计划进行2D或者3D Gamma分析，得到Gamma分析的通过率情况；基于可执行的放疗计划、计划评估的评分总分和Gamma分析的通过率情况自动生成放疗计划报告；医生对所述放疗计划报告进行审核。
在上述技术方案中，优选地，针对所述放疗剂量分布结果，还包括：在接收到剂量编辑触发指令时，进入剂量编辑模式；空间剂量模型在当前放射治疗图像断面上的切面图形随控制光标的轨迹移动，其中，所述控制光标所在位置为所述空间剂量模型的中心，所述控制光标的轨迹与动作操控装置的移动轨迹对应；监听所述动作操控装置的动作事件，并根据该动作事件对应的预设操控命令调整所述空间剂量模型中心点处的剂量；以所述空间剂量模型中心点的剂量为基准通过插值算法计算所述空间剂量模型内其余点的剂量；保存和更新所述空间剂量模型内各点的剂量数据，并对所述空间剂量模型以外区域的剂量数据不做修改。
在上述技术方案中，优选地，所述对所述医学影像的ROI区域进行自动勾画以获取几何解剖结构具体包括：正常器官的自动识别与自动勾画：基于机器学习自动识别和勾画人体全身的各个正常器官；肿瘤部位的自动识别与勾画：如果全身器官能进行勾画，则肿瘤采用反向勾画；将危及器官勾画完成后，剩余部分为肿瘤部位；采用机器学习的病种PTV(Planning Target Volume，计划靶区)与GTV(Gross Tumor Volume，肿瘤靶区)的外扩的关系，对剩余部分进行自动勾画。
在上述技术方案中，优选地，所述根据所述病种信息、所述几何解剖结构和所述处方确定放疗照射角度具体包括：将历史病例的病种信息、几何解剖结构和处方进行机器学习，确定照射角度预测模型，将当前病例的所述病种信息、所述几何解剖结构和所述处方输入所述照射角度预测模型以获取预测照射角度，作为所述放疗照射角度；或，
根据病种标记计划靶区的器官权重，沿射线方向计算每个角度的器官权重累加值，合并相邻满足预设权重阈值的角度，将符合权重阈值的角度作为 所述放疗照射角度；或，
确定感兴趣区域，选择至少一个计划靶区和一个危及器官，并对每个感兴趣区域进行全角度照射角度投影；在每个分段角度的每个角度上对计划靶区计算最小外接矩形，对某危及器官在该角度与对应的最小外接矩形进行交集运算，得到交集面积；对所有分段角度的交集面积进行求和，以最小的和值作为目标函数，采用非线性整数最优化算法求解得到最优分段索引和最优角度索引，作为所述放疗照射角度。
在上述技术方案中，优选地，所述剂量预测模型的构建方法包括：以归一化的PTV平均剂量建立数据集，并以该数据集制定评分模板；对感兴趣区域进行标准化命名；将三维医学影像分割为二维切片作为训练集和测试集；读取所述训练集的三维计划靶区数据的射线角度，将射线角度投影在计划靶区上得到网络权重，对该网络权重采用剂量计算算法进行剂量计算得到射束通道；以U-net网络或者V-net网络作为生成器、以马尔科夫判别器作为判别器构建Pix2pix剂量预测模型；以所述二维切片图像作为所述生成器的输入，以所述生成器输出的预测剂量和原始剂量作为所述判别器的输入，所述判别器输出判断结果；将所述训练集的所有二维切片输入所述Pix2pix剂量预测模型训练。
在上述技术方案中，优选地，所述以所述放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划，具体包括：基于通量图优化算法，优化出通量权重图；然后，结合加速器的机器信息，采用叶片序列算法，自动生成可执行动态调强放疗计划；或，
基于直接子野优化方法，自动生成可执行静态调强放疗计划；或，
基于遗传算法或者列生成算法，自动生成容积调强放疗计划或者旋转调强放疗计划；或，
正向放疗计划；或，
立体定向放疗计划。
在上述技术方案中，优选地，所述根据该动作事件对应的预设操控命令调整所述空间剂量模型中心点处的剂量具体包括：在监听到动作事件时，浮动显示剂量调节指示标签；在所述控制光标在所述剂量调节指示标签区域时 监听到所述动作操控装置被触发第一动作事件，则以所述控制光标所处位置对应的剂量值作为所述空间剂量模型中心点处的剂量；在所述控制光标在所述剂量调节指示标签上的指示滑块上时监听到所述动作操控装置被触发第二动作事件，则以解除点击第二动作事件时所述控制光标所处位置对应的剂量值作为所述空间剂量模型中心点处的剂量；在所述控制光标处于当前切面图形区域内时监听到所述动作操控装置被触发第三动作事件，则以所述第三动作事件的动作参数调整所述空间剂量模型中心点处的剂量；在所述控制光标未处于当前切面图形区域内时若监听到所述动作操控装置被触发第三动作事件，则以所述第三动作事件的动作参数对所述放射治疗图像进行翻层操作，移出该切面图形区域时保存所述第三动作事件调整的剂量。
在上述技术方案中，优选地，放疗计划方法还包括：监听动作操控装置的动作事件，根据该动作事件对应的预设操控命令还可对所述空间剂量模型所在的放射治疗图像进行翻层操作，以及可调整所述空间剂量模型的尺寸；在监听到所述动作操控装置被触发第四动作事件时，根据所述第四动作事件的动作参数调节所述空间剂量模型的尺寸。
在上述技术方案中，优选地，所述空间剂量模型中心点的剂量调节上限Dl和下限Du分别为：
Figure PCTCN2020091843-appb-000001
Figure PCTCN2020091843-appb-000002
其中，Dl为剂量可调节的下限，Du为剂量可调节的上限，D0为所述动作操控装置的动作事件被触发时所述空间剂量模型中心点处的点剂量，R为所述空间剂量模型的特征参数，Dmax为剂量数据的全局最大剂量值，n为常数。
本发明还提出一种标准化的人工智能自动放疗计划***，所述放疗计划***用于实现如上述技术方案中任一项所述的标准化的人工智能自动放疗计划方法。
与现有技术相比，本发明的有益效果为：通过在基于几何解剖结构和器官三维剂量分布的预测模型基础上，增加了处方剂量预测模型和照射角度自动优化过程，实现了全自动的剂量预测，提高了剂量预测的效率和效果，从 而高质量、快速地生成可执行的放疗计划，具有较好的准确性、稳定性和规范性，从而能够提高医疗软硬件资源利用率。此外，通过直接编辑剂量的方式让用户直观、直接地得到想要的剂量分布，相对于间接调整参数影响剂量分布的方式更加快捷直观，极大提高了计划设计效率。
附图说明
图1为本发明一种实施例公开的标准化的人工智能自动放疗计划方法的流程示意图；
图2为本发明一种实施例公开的Beam通道的图像化展示示意图；
图3为本发明一种实施例公开的cGAN生成网络的原理示意图；
图4为本发明一种实施例公开的Pix2pix生成网络的原理示意图；
图5为本发明一种实施例公开的生成器模型的原理示意图；
图6为本发明一种实施例公开的判别器判别流程的原理示意图；
图7为本发明一种实施例公开的预测剂量与原始剂量的对比示意图；
图8为本发明一种实施例公开的训练集与预测集的数据体积均值的对比示意图；
图9为本发明一种实施例公开的训练数据与预测数据的器官体积均值的差值对比示意图；
图10为本发明一种实施例公开的典型绝对剂量的dice相似性系数曲线示意图；
图11为本发明一种实施例公开的迭代训练的loss曲线示意图；
图12为本发明一种实施例公开的数据预处理流程图；
图13为本发明一种实施例公开的beam数据统计图；
图14为本发明一种实施例公开的beam数据筛选图；
图15为本发明一种实施例公开的采用GAN网络进行预测的效果图；
图16为本发明一种实施例公开的DVH比较结果图；
图17为本发明一种实施例公开的模拟头颈肿瘤横断图
图18为本发明一种实施例公开的标准名与别名的映射字典列表；
图19为本发明一种实施例公开的剂量编辑方法的流程图；
图20为本发明一种实施例公开的剂量编辑方法的流程示意框图。
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中的附图，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的一部分实施例，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做出创造性劳动的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。
下面结合附图对本发明做进一步的详细描述：
如图1所示，根据本发明提供的一种标准化的人工智能自动放疗计划方法，包括：获取医学影像；对医学影像(如CT、MR图像)的ROI区域进行自动勾画以获取几何解剖结构；根据医学影像对应的病种信息、几何解剖结构以及预设的病种-处方模板库确定处方；根据病种信息、几何解剖结构和处方确定放疗照射角度；将医学影像、几何解剖结构、病种信息、处方和放疗照射角度输入剂量预测模型中，得出放疗剂量分布结果；以放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划；可执行放疗计划包括正向放疗计划、立体定向放疗计划和调强放疗计划，其中调强放疗计划包括动态调强放疗计划、静态调强放疗计划、容积调强放疗计划和旋转调强放疗计划。
在该实施例中，具体地，在基于几何解剖结构和器官三维剂量分布的预测模型基础上，增加处方剂量模板和角度自动优化过程，实现全自动的剂量预测；具体流程如下：
(1)医学影像获取：通过CT机或者核磁(MR)获取病人影像，以预设格式存储，优选地，建立OIS(肿瘤信息***)来实现患者影像、患者信息和治疗信息的管理，优选采用Dicom标准；
(2)器官勾画：对所获得的的医学影像进行自动勾画，获取几何解剖结构，其中，勾画过程包括正常器官组织的勾画以及肿瘤靶区的勾画；
(3)处方确定：根据上述的勾画信息以及病种信息，自动确定处方；处方由预设的病种-处方模板库中处方与病种的映射关系确定，病种-处方模板库需要事先定义好；
(4)角度确定：利用处方、病种、勾画信息，自动确定照射角度；
(5)剂量预测：剂量预测前需要完成模型的训练，模型训练只需要一次训练，日常使用只需要将数据按照设定的格式输入就可以完成剂量预测，其中，优选地，该剂量预测模型是一种基于Pix2pix的患者几何解剖结构和器官三维剂量分布的剂量预测模型，输入数据采用Dicom标准格式；
(6)输出结果与显示：对于生成的预测结果，可以通过靶区受量平均变化率、DVH对比、预测剂量图像对比以及Dice相似性系数几个方面进行验证，剂量预测的结果为后续计划优化和计划质量控制提供更为充分的数据信息。
在此基础上，通过AI与放疗计划数据集结合，建立深度学习神经网络模型，通过深度学习方法对优秀的放疗计划数据集进行训练后，可产生高质量的自动计划模型，利用该模型可根据用户输入的医学影像进行放疗计划预测。上述整个过程在无人干预下完成，只占用计算机机时，大大提高了生产放疗计划的效率，从而大大减少了病人等待治疗的时间，也间接地提高了疗效。
在上述实施例中，优选地，对医学影像的ROI区域进行自动勾画以获取几何解剖结构具体包括：正常器官的自动识别与自动勾画：基于机器学习自动识别和勾画人体全身的各个正常器官；肿瘤部位的自动识别与勾画：如果全身器官能进行勾画，则肿瘤采用反向勾画；将危及器官勾画完成后，剩余部分为肿瘤部位；采用机器学习的病种PTV与GTV的外扩的关系，对剩余部分进行自动勾画。
在该实施例中，具体地，基于深度学习的自动靶区勾画，可实现：
1、正常(危及)器官的自动识别与自动勾画：可以基于机器学习自动勾画人体全身的各个器官；
2、肿瘤部位的自动识别与勾画：如果全身器官能进行勾画，则肿瘤采用反向勾画，将危及器官勾画完成后，剩余部分为肿瘤部位，采用机器学习的病种PTV与GTV的外扩的关系，对剩余部位进行自动勾画。
具体地，首先针对勾画中的ROI，确定感兴趣的ROI，至少选择一个PTV和一个OAR(Organ At Risk，危及器官)。对每一个ROI，进行0-360度的射束角度范围内投影，给定一组初始分段(表示JAW的位置，索引用i表示)，每一个分段给定一组初始的角度(索引用j表示)。在一个分段内的一个角度上，对PTV(i,j)(治疗计划靶区)求最小外接矩形，记作block(i,j)。对某一个的OAR(索引用k表示)在该分段，该角度下和block(i,j)进行求交集运算，得 到交集面积。对该分段所有的OAR与block(i,j)的交集面积求和，进而对所有分段求和，记作Sall。以最小的Sall结果为目标函数，采用非线性整数最优化算法求解得到最优的分段索引和最优的角度索引。以上步骤最终优化出最佳铅门位置和最佳照射角度，在充分照射PTV的基础上，充分减少了漏射对OAR的照射，该种算法完全从符合放疗计划制作的基本原则为出发点而提出，充分保护了正常组织。
在上述实施例中，优选地，根据病种信息、几何解剖结构和处方确定放疗照射角度的方法包括三种，该三种方法具体包括：
(一)将历史病例的病种信息、几何解剖结构和处方进行机器学习，确定照射角度预测模型，将当前病例的病种信息、几何解剖结构和处方输入照射角度预测模型获取预测照射角度，作为放疗照射角度；
(二)根据病种标记计划靶区的器官权重，沿射线方向计算每个角度的器官权重累加值，合并相邻满足预设权重阈值的角度，将符合权重阈值的角度作为放疗照射角度；
具体地，根据病种不同标记不同病种的器官权重，权重越大越重要，其中肿瘤靶区标记为0；器官权重中可以分区给定器官子权重；其中的权重标记方法，可用器官最大允许照射剂量的倒数确定，最大允许照射剂量越大，权重越小。角度权重确定过程中，沿着射线方向，按照预设的角度间隔计算每个角度的器官权重或子权重累加值。选择符合权重阈值的角度时，如果角度个数小于预设最小值，则默认为预设值，如果角度大于预设最大值则默认为最大值。
(三)确定感兴趣区域，选择至少一个计划靶区和一个危及器官，并对每个感兴趣区域进行全角度照射角度投影；在每个分段角度的每个角度上对计划靶区计算最小外接矩形，对某危及器官在该角度与对应的最小外接矩形进行交集运算，得到交集面积；对所有分段角度的交集面积进行求和，以最小的和值作为目标函数，采用非线性整数最优化算法求解得到最优分段索引和最优角度索引，作为放疗照射角度。
在上述实施例中，Beam角度又称射线角度或射束方向。放射治疗的目的是向计划靶区发送足够高的剂量以控制肿瘤，而与此同时要保证其周围的正常组织和危及器官(OARs)处于可接受的剂量水平，以免受到损伤。在精准治 疗中，Beam角度的设置影响计划靶区和危及器官所受到的剂量，对治疗计划的优劣有重要的影响。由于人体曲面和不均匀组织对剂量分布的影响，给治疗计划设计时确定射野角度带来了困难，使得针对每个病人都制定最合适射野入射方向变成了一个费时的、反复试错的过程。因此，在训练过程中考虑射线角度有助于预测出更加准确且更符合临床要求的剂量分布图。
生成Beam通道的方法如下：首先取出3维PTV数据，读取出病例中包含的Beam角度，将射野角度投影到PTV上得到beam通道数据的网络权重(在这里将落在射野范围内的区域设为1，其他位置设为0)，采用高速的剂量计算算法对该网络权重直接进行剂量计算从而得到Beam通道。生成的Beam通道的图像化展示如图2所示。
在上述实施例中，优选地，采用基于Pix2pix的剂量预测模型，剂量预测模型的构建方法包括：以归一化的PTV平均剂量建立数据集，并以该数据集制定评分模板；对感兴趣区域进行标准化命名；将三维医学影像分割为二维切片作为训练集和测试集；读取训练集的三维计划靶区数据的射线角度，将射线角度投影在计划靶区上得到网络权重，对该网络权重采用剂量计算算法进行剂量计算得到射束通道；以U-net网络或者V-net网络作为生成器、以马尔科夫判别器作为判别器构建Pix2pix剂量预测模型；以二维切片图像作为生成器的输入，以生成器输出的预测剂量和原始剂量作为判别器的输入，判别器输出判断结果；将训练集的所有二维切片输入Pix2pix剂量预测模型训练。
在该实施例中，具体地，模型训练中涉及到制定评分模板、感兴趣区域命名的标准化及模型训练过程。其中，Pix2pix是一种基于GAN的图像翻译模型。在GAN网络中包含着生成器G和判别器D，两者相互制约相互促进，G生成的图像与ground truth同时交给D来判别，判别结果为生成图像是真实图像的概率，如果概率很大，说明G生成的图像很接近原始图像，从而欺骗了D；如果判别为假，说明D识别出了生成图像与真实图像有较大区别。在上述G与D的博弈过程中，两者都学习了经验，G生成的假图越来越真，D判别的结果越来越正确。当D已经无法判别G生成的图是真是假时，得到了一套训练好的剂量预测模型。优选地，剂量预测模型迭代训练至收敛曲线达到预设收敛值时完成训练。
具体地，本实施例率先使用了Pix2pix模型加Beam通道实现放疗剂量的 预测。Pix2pix是一种基于GAN的图像翻译模型，在GAN网络中包含着生成器G和判别器D，两者相互制约相互促进，G生成的图像与ground truth同时交给D来判别，判别结果为生成图像是真实图像的概率，如果概率很大，说明G生成的图像很接近原始图像，从而欺骗了D；如果判别为假，说明D识别出了生成图像与真实图像有较大区别。在上述G与D的博弈过程中，两者都学习了经验，G生成的假图越来越真，D判别的结果越来越正确。当D已经无法判别G生成的图是真是假时，我们得到了一套训练好的生成模型。
如图3所示，cGAN生成网络G的输入包括噪声Z和条件Y，输出生成一个fake_x。判别网络D的输入包括fake_x或real_x和条件Y，输出为判断结果0或1即FAKE或REAL。
如图4所示，Pix2pix借鉴了cGAN的思想，在输入G网络的时候不止会输入噪声，还会输入一个条件(condition)，G网络生成的fake images会受到具体的condition的影响。将图像作为condition，生成的fake images与condition images有对应关系，从而实现了一个image-to-image translation的过程。具体地，Pix2pix的生成网络G的输入端只有一个条件Y，这里的Y为一张图片imgA。生成网络G用到的是U-net结构，输入的Y编码再解码成真实图像imgB’。判别器的输入为生成图像imgB’或真实图像real_x(imgB)和条件Y，最终实现图像到图像的变换。
在本实施例当中，输入Y为4通道图像，其中包括3通道的dose image和1个beam通道，通过U-net生成器后得到预测剂量fake_x，原始剂量与生成的预测剂量共同放入判别器，判断预测剂量与真实剂量之间的差异并输入判断结果。
下面详细介绍本实施例中使用的生成器与判别器结构。
如图5所示，本实施例中的生成器使用一个8级层次的U-net，实现了图像到剂量的映射。整个网络结构可以看作特征提取部分和上采样部分。输入从256×256像素图像的4个通道开始。特征提取部分每层进行一次3×3卷积操作，到下一层时使用2×2最大池化层，其目的是减少256×256像素的特征尺寸到1×1像素。在上采样部分使用同样卷积核对每层数据进行卷积操作，而进入下一层时，最大池化层变为2×2反卷积，目的是将图像变换会原始图像大小。而为了保留底层信息不丢失，保留图像细节信息，使用如图5 所示的方法将底层特征保留。最终输出图像为256×256×1的剂量图。
在训练阶段，选择Adam algorithm作为优化器来最小化损失函数。本实施例中将训练设置为两个阶段，Adam parametersβ 1＝0.55，β 2＝0.999。第一阶段学习率(learning rate)为2×10 -5，epochs为100；第二阶段为学习率为2e-06，epochs为300。本实施例分成两个阶段训练，一方面可以提升收敛速度，另一方面可以中断训练后继续训练，不会影响结果。
判别器采用如图6所示的马尔可夫判别器(PatchGAN)，判别是否是生成的图片。因为不同的patch之间可以认为是相互独立的，所以PatchGAN的思想是让判别器对图像的每个大小为N×N的patch做真假判别。Pix2pix对一张图片切割成不同的N×N大小的patch，判别器对每一个patch做真假判别，将一张图片所有patch的结果取平均作为最终的判别器输出。对于256×256的输入，patch大小为70×70时，判断结果最好。
如图7所示，作为Pix2pix模型的一个典型预测示例，图7展示了同一个病例预测剂量图像与真实剂量图像的对比，左侧为预测剂量图像，右侧为原始剂量图像。可见，得益于生成器使用了U-net网络保存了底层信息，预测剂量图像细节信息保存较好。
在本发明中，以乳腺癌靶区放射治疗的剂量预测过程为例，在进行剂量学评估中，受体体积间的差异对受体剂量也会产生影响。由图8可以看出，训练集计划靶区(PTV)体积MEAN值为915.46cm 3，GTV体积MEAN值为658.74cm 3，左肺体积MEAN值为1001.09cm 3，右肺体积MEAN值为1315.38cm 3，心脏体积MEAN值为533.33cm 3，脊髓体积MEAN值为40.95cm 3。预测集计划靶区(PTV)体积MEAN值为978.1cm 3，GTV体积MEAN值为743.06cm 3，左肺体积MEAN值为981.9cm 3，右肺体积MEAN值为1329cm 3，心脏体积MEAN值为552.3cm 3，脊髓体积MEAN值为43cm 3。
如图9所示，在上述实施例中对数据中各标记器官的体积均值进行比较，计算公式为predict mean-raw mean。其中PTV体积差距达到95.1cm 3，双肺、心脏以及脊髓体积差距均小于20cm 3。将体积差值与表1剂量变化率相结合得出，对于体积差值较大的PTV区域，剂量变化率D2与D98分别为0.33％和3.55％，预测剂量低于原始剂量，但D95的剂量均值均达到了处方剂量要求，符合临床剂量需要。左肺体积差值较小，剂量变化率虽然达到8.19％但因为基 数仅为1.5Gy左右，故属于临床可接受范围。右肺属于大面积受量区域，且在体积差值较小的基础上，V5,V10,V20的剂量变化率均小于1％，虽然V30剂量变化率达到12.5％，但预测剂量均值与实际剂量均值的差值仅为3.1Gy，属于临床可接受范围。脊髓剂量差值约为2.8Gy，体积差值为2.1cm 3，同样属于临床可接受范围。
分析表1数据可以看出，预测结果的剂量相较于实际计划剂量普遍偏小，但均满足了D95符合处方剂量要求，因此我们认为部分训练数据的剂量还有下降的空间。模型充分考虑危机器官的体积与受量，学习并提出了更佳的剂量分布结果。
表1.预测剂量均值与真实剂量均值变化率对比
Figure PCTCN2020091843-appb-000003
表1展示了预测剂量均值与真实剂量均值变化率对比，其中HI与CI的计算公式为：
Figure PCTCN2020091843-appb-000004
对于这两种指标，HI的值越小和CI的值约接近1，说明放疗计划制作的越好。对于左肺、右肺等大体积器官的预测变化率最大为0.12，而对于小体积器官，如心脏，变化率虽大，但预测剂量均值与真实剂量均值数据实际差量仅为0.176Gy。预测数据集在预测结果上与原始数据差距在临床可接受范围内，与训练集数据变化率相比幅度 不大。
预测剂量与实际剂量的相似性系数使用Dice相似性系数表示。由图10看出，在3～50Gy的绝对剂量区间内，20Gy剂量下的Dice值在0.76～0.86之间逐渐上升，在20～45Gy区间Dice值在0.86～0.9之间浮动，绝对剂量在45Gy以上时，Dice值有缓慢下降趋势，但都保持在0.83之上。通过分析数据我们发现，右侧乳腺癌中，低剂量区域主要分布在脊髓，左肺和心脏，受量在30Gy以下，高剂量区域主要分布在右肺以及PTV。这种复杂靶区存在剂量分布不均匀，剂量跨度大的特点。因此在20Gy绝对剂量以下的相似性系数较低，但低剂量区域虽然相似性系数较低，变化率以及DVH上的表现都符合临床剂量要求。
剂量体积直方图(DVH)是目前三维适形放疗中被广泛接受的治疗计划评估方法,直观表示靶区和正常组织中剂量和体积的关系。靶区的DVH能显示照射的均匀度,正常组织的DVH能提供该器官受照的剂量及其相应的体积,特别是对放射耐受性与受照体积有关的正常器官有非常重要的临床意义。剂量体积指标作为评估放疗计划优良的方法，具有直观感受剂量变化的特点。在放疗计划制作过程中，物理师通过eclipse软件可以在DVH上手动通过调整权重与受量实时监看剂量变化情况。通过图像可以直观看出PTV曲线、BODY曲线基本重合，这保证了靶区照射剂量充分以及身体受量保持在临床可接受范围。
总的来说，在本实施例中对120例训练样本、共大约12000张训练图像进行了总epochs为400次的训练，这个过程在gtx1080显卡上运行了4×24h，图11显示了生成模型的loss图。训练模型的损失开始为13800，在经过50次迭代后降loss低到了4210。训练迭代到300次后曲线收敛平稳，350次较300次损失降低了40，最终在400次迭代后模型收敛到5，继续训练后损失曲线并无更大波动，这也是选择400次迭代的原因。
综上，在上述实施例中，使用了Pix2pix结合射野角度实现了乳腺癌复杂靶区的剂量预测。在本实施例中将剂量预测定义为一个图像着色问题，Pix2pix剂量预测模型在这个问题上表现良好。其次是Pix2pix剂量预测模型对于数据成对的要求在实验中使用便利。最后，模型中的生成器使用了u-net网络，保留了底层信息，为模型预测图像的细节提供了保证。实验结果表明，通过对 比靶区剂量体积参数变化率和正常器官的剂量体积参数变化率，得到了临床可接受的预测放疗剂量结果。
此外，在构造考虑beam角度的自动剂量预测模型过程中，如果能考虑beam角度不同引起的剂量分布的差异，可以提高训练和预测准确度；解决病例来源不足的问题；该算法也是解决自动剂量预测普适性的核心问题。可能的解决方案是预测和训练都考虑beam角度。具体来讲还可以有如下几种方案：
方案一：
训练和预测的时候将beam角度作为一个先验条件，对病例进行标注，采用带有条件概率的机器学习模型进行训练和预测，在进行预测的时候用户输入一个自定义的beam角度组合(因为一个计划有多个beam，所以是多个beam角度，这里称作一个beam角度组合)即可。如果用户不进行beam角度的输入，***默认可以自动给出一组推荐beam角度。该推荐值是根据同样病种病例，采用机器学习中的聚类方法找到的一组最佳beam角度。也可以按照根据历史病例和影像的解剖学特征，建立起预测模型，根据输入的CT图像，自动预测出合适的beam角度组合。可以参考的模型有条件生成对抗网络CGAN，目前已经实现的是原始的生成对抗网络模型。CGAN原理图如图3所示。
方案二：
训练和预测之前对输入的CT(和RS)进行数据预处理。这里的括号代表有两种可能的解决思路，包含RS处理或者不包含RS的处理。以包含RS的处理为例，对CT按照beam角度(和现有的病例计划中的MU权重，同样以包含MU权重为例)对CT数据围绕某一个轴(推荐使用gantry旋转的轴)进行旋转、裁切、部位，然后合成，合成是以MU为权重，采用重心法得到合成的CT数据，如果不考虑MU，则全部设置为权重为1。最后重建出虚拟CT。对RS，则是对坐标进行旋转，对旋转的坐标同样按照重心法产生最终的坐标，权重参考CT的合成方式。流程图如图12所示，左图代表不考虑beam角度的数据预处理过程；右图以每套CT有三个切片为例，说明重建的过程。
方案三：
针对不同的beam角度的病例计划，采用预处理的方式筛选掉偏离主要beam组合的计划弃之不用，筛选出按照如下提出的公式对收集到的历史病例计划进行筛选。
Figure PCTCN2020091843-appb-000005
按照1倍标准偏差的标准进行筛选，超出的病例不用于训练。剩下的即可以作为训练集。这种方式一定程度上降低了beam角度不同对预测剂量分布影响，但是一定程度上减少了可以用于训练的病例。以20套计划病例举例如下表所示。首先从20套病例中提取所有的beam角度。
Figure PCTCN2020091843-appb-000006
其中每一行代表一个计划的角度数组。数组长度按照beam数目最多的定，每个计划的beam数目用一个变量提前记录下来。首先对beam数据进行统计，得到如图13所示的图像。
根据beam个数的分布情况，得知治疗用的比较多的是8个beam。频数为14。这14个计划作为筛选下来的计划，应用筛选公式。得到如图14所示的结果；即，三条横线从上到下分别代表平均值偏大1倍标准差、平均值和平均值偏小1倍标准差。这里认为介于1倍标准差范围内的为beam角度接近的案例，于是可以排除掉2个案例，剩余的12套计划案例可以用于训练。
那么，根据上述实施例提供的考虑beam角度的自动剂量预测模型，在进行自动剂量预测过程中，通过自动评估进行病例收集，针对该病种分类(发病早晚期、处方剂量、癌变部位等)采用统一标准模板进行病例筛选，然后将CT，勾画，以及剂量一起训练，产生预测模型，从而进行剂量预测，预测 的输入为CT和基于CT的自动勾画(RS)，预测的输出结果为每个CT切片图像对应的切片剂量或者由此进而产生的DVH。以针对头颈案例采用GAN网络进行预测的一个效果图，如图15所示，可以看出在PTV部分的预测结果略微好于原始结果。
在上述实施例中，优选地，以放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划，具体包括：基于通量图优化算法，优化出通量权重图；然后，结合加速器的机器信息，采用叶片序列算法，自动生成可执行动态调强放疗计划；或，
基于直接子野优化方法，自动生成可执行静态调强放疗计划；或，
基于遗传算法或者列生成算法，自动生成容积调强放疗计划或者旋转调强放疗计划；或，
正向放疗计划；或，
立体定向放疗计划。
具体的：
自动逆向计划就是将上述预测的剂量分布或DVH作为参考剂量分布或参考DVH，采用基于剂量分布或DVH引导的逆向优化算法，结合由特定的剂量计算引擎的计算的体素剂量单元，，优化出通量权重图，结合特定加速器的机器信息，采用叶片序列算法，自动得到最终的治疗计划；其中优化算法可以采用线性模型，也可以采用非线性模型实现。
所产生的计划自动调用自动计划评估——包括各项得分以及总分，提供给医生批准使用；
以自动优化模型的线性模型为例：
计划优化引擎采用一系列线性目标函数来构成逆向计划优化问题：
为每个OAR设定剂量最大值目标函数，剂量平均值目标函数，等效均匀性目标函数；
最大值目标函数：
Figure PCTCN2020091843-appb-000007
平均值目标函数：
Figure PCTCN2020091843-appb-000008
等效均匀性目标函数：
Figure PCTCN2020091843-appb-000009
为每个PTV设定剂量最大值目标函数，偏离处方剂量程度目标函数；
最大值目标函数：
Figure PCTCN2020091843-appb-000010
偏离处方剂量程度目标：
Figure PCTCN2020091843-appb-000011
Figure PCTCN2020091843-appb-000012
确保计划可执行性，设定Fluence Map平滑约束目标函数；
平滑目标函数：
Figure PCTCN2020091843-appb-000013
整个逆向计划优化问题可以线性表达成如下公式：
Figure PCTCN2020091843-appb-000014
Figure PCTCN2020091843-appb-000015
Figure PCTCN2020091843-appb-000016
……
……
……
其中γ i β i κ i β t φ t 
Figure PCTCN2020091843-appb-000017
就是各个目标函数的权重参数，也就是物理师需要反反复复调整的参数，恰好逆向计划优化模型能根据AI剂量预测模型预测的剂量分布数据优化出这些参数，也就是AI模型从历史病例中学习到了如何设计计划。把上述逆向计划优化问题表达成矩阵的形式：
Figure PCTCN2020091843-appb-000018
subject to Ax≥b，
x≥0.
α  为权重参数向量
C   为目标函数表达矩阵
x   为决策变量
g   为平滑目标函数表达矩阵
A   为约束条件因子矩阵
b   为约束边界
下面是原逆向问题的对偶问题：
Figure PCTCN2020091843-appb-000019
subject to C′α+g≥A′p，
p≥0.
p是原问题约束条件的对偶变量
所以要想优化出α可以把原问题变成一个优化绝对对偶间隙的形式：
Figure PCTCN2020091843-appb-000020
subject to C′α+g≥A′p，
α≥0，p≥0.
Figure PCTCN2020091843-appb-000021
是各个目标函数的值，这些值可以通过AI剂量预测引擎学习得到；
Figure PCTCN2020091843-appb-000022
对一个临床计划来说是一个常量，对优化问题没有贡献，可以舍去。最后问题变成：
Figure PCTCN2020091843-appb-000023
subject to C′α+g≥A′p，
α≥0，p≥0.
显然最后的优化模型是一个标准的线性规划问题，模型能优化x的值， 再把值代入到最原始的逆向计划优化模型里面：
Figure PCTCN2020091843-appb-000024
subject to Ax≥b，
x≥0.
α已知，上面模型又是一个标准的线性规划问题，也就能优化出x(Fluence map),最后生成高质量的计划。
在本发明中，以国家标准YY/T0889(与AAPM TG 119例题一致)标准中的头颈案例，对上述实施例提出的自动计划原型进行了算法测试。无需人工干预的基础上能够自动优化出符合法规要求的计划。用时总计在5分钟左右，用户唯一需要输入的是处方剂量要求和预测的三维剂量数据。最终的DVH比较结果如图16所示。实线和虚线分别代表人工优化的DVH和自动逆向优化的DVH，其中自动优化采用的预测DVH。
国家标准中的例题的约束和目标要求如图17所示。由图17可以看出自动计划给出的DVH结果比人工做的化在PTV的部分稍微差一点，但是也是符合要求的计划。而且自动计划只需要大概5分钟左右，而人工计划至少要一个小时。
进一步地，本发明中自动剂量预测的方法为：
a)病例收集和自动筛选；
b)构造考虑射束角度的自动剂量预测模型；
c)病例的人工智能(AI)训练；
d)病例的自动剂量预测。
上述的自动剂量预测步骤可替换成，自动剂量体积直方图(DVH)预测：方法1：基于机器学习的DVH预测：
a)病例收集和筛选；
b)构造考虑射束角度的自动剂量预测模型；
c)病例的人工智能(AI)训练；
d)病例的DVH预测。
方法2：基于统计方法的DVH预测：
a)病例收集和筛选；
b)病例的DVH统计；
c)病例的DVH预测。
方法3：基于模板的DVH预测
a)利用预设的模板生成最初的目标约束项；
b)算法自动添加约束项、调节权重；
c)算法自动添加辅助器官；
d)病例的DVH预测。
在上述实施例中，优选地，放疗计划方法还包括：通过统一的处方标准与人工智能结合，对所生成的可执行放疗计划进行评分，得到计划评估的评分总分；通过蒙特卡罗三维剂量验证技术，对所生成的可执行放疗计划进行2D或者3D Gamma分析，得到Gamma分析的通过率情况；基于可执行的放疗计划、计划评估的评分总分和Gamma分析的通过率情况自动生成放疗计划报告；医生对放疗计划报告进行审核。
具体地，通过统一的处方标准与人工智能(Artificial Intelligence，简称AI)结合，对所生成的可执行放疗计划进行评分，得到计划评估的评分总分；其中：
1、针对特定病种命名的规范化
为了保证训练预处理数据的时候能自动提取勾画结构别名为统一名称，需要制定标准名与别名的映射字典列表，如图18所示；
2、自动计划评估
自动计划的前提是具有优秀的计划数据库。因此，需要一个自动筛选优秀计划算法与工具。该工具既可以给计划筛选使用，也可以给自动计划产生后进行评分使用。
评估软件定位于多功能的信息和数据管理应用，用于帮助医生和物理师对放疗计划进行改善和提升规范性，用于对现有的计划案例进行评分筛选，选出分数较高和较低的案例分别用于机器学习训练和预测测试。应用程序的输入为治疗计划***(TPS)导出的DICOM数据，包括RS和RD两部分。输出为评分结果。评估软件所需的评分模板都是每个医院统一规则自己制定，针对不同的病种甚至更具体的分类，可以制定不同的模板。制定过程中可以参考国际标准和医院自己的内部标准。
在本发明中，为了对所有计划之间进行平等的比较，剂量预测模型使用PTV平均剂量5000cGy来规范化计划。PTV平均剂量归一化建立了一个统一的数据集，该数据集更有利于训练模型，归一化的计划具有更大的临床相关性和评估价值。评分模板的制作是依据RTOG-1005、5年以上工作经验的物理师以及放射科医生三方信息的总结和交流制定。评分模板包含项目如下：PTV的V48,V50,V53,V55,DMAX,D2,D98,HI,CI；心脏的V10,DMEAN；左肺的V4,V5,DMEAN；右肺的V4,V5,V8,V10,V20,V30,DMEAN以及脊髓的DMAX。通过设置体积和受量的上下限对数据进行评分，越接近上限分数越高，超过下限才有得分，同时对每个属性分配不同权重，使其更加符合医生处方要求与临床需要。制作评分模板的目的有以下几点：1在挑选数据的过程中可能会出现左右***错误，病种错误等事情的发生，通过制定评分模板选出错误数据，避免影响模型精度。2规范化的数据有利于训练模型的准确性。
具体地，通过蒙特卡罗三维剂量验证技术，对所生成的可执行放疗计划进行2D或者3D Gamma分析，得到Gamma分析的通过率情况；其中：所产生的计划自动使用基于蒙卡QA的3D Gamma分析；给出Gamma分析的通过率情况，用于给医生最后批准计划做参考，为精确放疗做推进。
基于上述实施例中得到的可执行的放疗计划、计划评估的评分总分和Gamma分析的通过率情况自动生成放疗计划报告；再由医生对放疗计划报告进行审核；其中：
审核报告具有摘要与详情内容。摘要描述了3D QA的Gamma分析的通过率情况，计划评估的评分总分，以及准备执行计划的摘要；详细为每个内容的详细介绍，方便医生对该自动计划仔细审核。如果批准通过，则发布到医用加速器上准备执行；如果审批不通过，则允许人工介入进行手动修改计划。
如图19所示，在上述实施例中，优选地，针对放疗剂量分布结果，还包括：在接收到剂量编辑触发指令时，进入剂量编辑模式；空间剂量模型在当前放射治疗图像断面上的切面图形随控制光标的轨迹移动，其中，控制光标所在位置为空间剂量模型的中心，控制光标的轨迹与动作操控装置的移动轨迹对应；监听动作操控装置的动作事件，并根据该动作事件对应的预设操控 命令调整空间剂量模型中心点处的剂量；以空间剂量模型中心点的剂量为基准通过插值算法计算空间剂量模型内其余点的剂量；保存和更新空间剂量模型内各点的剂量数据，并对空间剂量模型以外区域的剂量数据不做修改。
在该实施例中，通过直接利用编辑工具编辑放疗剂量的方式，让用户直观、直接地得到想要的剂量分布，相对于间接调整参数影响剂量分布的方式更加快捷直观，极大提高了计划设计效率。
在上述实施例中，优选地，根据该动作事件对应的预设操控命令调整空间剂量模型中心点处的剂量具体包括：在监听到动作事件时，浮动显示剂量调节指示标签；在控制光标在剂量调节指示标签区域时监听到动作操控装置被触发第一动作事件，则以控制光标所处位置对应的剂量值作为空间剂量模型中心点处的剂量；在控制光标在剂量调节指示标签上的指示滑块上时监听到动作操控装置被触发第二动作事件，则以解除点击第二动作事件时控制光标所处位置对应的剂量值作为空间剂量模型中心点处的剂量；在控制光标处于当前切面图形区域内时监听到动作操控装置被触发第三动作事件，则以第三动作事件的动作参数调整空间剂量模型中心点处的剂量；在控制光标未处于当前切面图形区域内时若监听到动作操控装置被触发第三动作事件，则以第三动作事件的动作参数对放射治疗图像进行翻层操作，移出该切面图形区域时保存第三动作事件调整的剂量。
在上述实施例中，优选地，放疗计划方法还包括：监听动作操控装置的动作事件，根据该动作事件对应的预设操控命令还可对空间剂量模型所在的放射治疗图像进行翻层操作，以及可调整空间剂量模型的尺寸；在监听到动作操控装置被触发第四动作事件时，根据第四动作事件的动作参数调节空间剂量模型的尺寸。
其中，空间剂量模型为球体、立方体、长方体或椭球体，动作操控装置为鼠标或其他人机交互的操控装置，若采用鼠标为动作操控装置，则动作操控装置的第一动作事件可为鼠标左键被点击，第二动作事件可为鼠标左键被点击被保持点击状态移动，第三动作事件可为鼠标滚轮被滚动，第四动作事件可为鼠标右键被点击被保持点击状态移动。
优选地，将剂量球中心点及空间模型剂量模型内其他点的剂量计算后，更新保存至数据库，并以DVH、等剂量线和/或剂量体积直方图统计数据的方 式进行前端更新显示。
如图20所示，具体地，以鼠标作为动作操控装置、以球体的空间剂量模型为例，对放射治疗计划***中剂量编辑的方法进行详细描述，则放射治疗计划***中剂量编辑的方法具体包括：
1.在点击剂量球编辑按钮，接收到剂量编辑触发指令时，进入剂量编辑模式；
2.剂量球在当前放射治疗图像断面上的切面圆随控制光标的轨迹移动，其中，控制光标所在位置为剂量球的中心，控制光标的轨迹与鼠标的移动轨迹对应；
3.监听鼠标的动作事件，在鼠标的滚轮滚动时对放射治疗图像进行翻层操作，在鼠标的右键被点击且保持点击状态移动时调节剂量球的半径；
4.在鼠标的左键被点击时以控制光标当前位置作为剂量球的中心点位置同时显示剂量调节指示条，剂量滑块所在位置即为当前球心处的点剂量；
5.剂量编辑的方式有以下三种，不同的鼠标事件完成不同的功能：
1)控制光标在切面圆内时，鼠标滚轮上下滚动时可增大和减小球心处的点剂量，鼠标滚轮向上可滚动n步，滚动步长为(Du-D0)/n，鼠标滚轮向下可滚动m步，滚动步长为(D0-Dl)/m，其中，Dl为剂量可调节的下限，Du为剂量可调节的上限，D0为鼠标左键点击时剂量球中心点处的点剂量，n、m为常数；
2)鼠标左键点击剂量调节指示条上的滑块并拖动，松开鼠标后球心处的剂量调整为滑块所在位置的剂量值；
3)鼠标左键点击剂量调节指示条上时，滑块移到点击的位置，并且球心处的剂量调整为滑块所在位置的剂量值；
6.剂量球内剂量分布的重新调整：剂量球球心的剂量编辑确定后，剂量球内空间各点的剂量由剂量球边缘与球心处剂量通过插值算法得到；
7.剂量球位置确定后依然可以通过按住鼠标右键移动来显示和调整剂量球的半径；
8.控制光标移出切面圆时，滚动鼠标滚轮可实现图像翻层功能；
9.剂量的保存与更新：球内剂量分布调整后自动将编辑后的剂量保存到数据库；
10.前端更新显示剂量数据，包括剂量体积直方图，等剂量线，DVH统计数据等。
在上述实施例中，优选地，插值算法包括线性插值、双线性插值、三次插值、双三次插值、最近邻插值、立方卷积插值算法、自然邻点插值法、三角网/线形插值法、谢别德法、径向基本函数法、多元回归法、最小曲率法、克里金法和距离倒数乘方法。
优选地，在剂量球的中心点位置确定后，在鼠标的右键被点击且保持点击状态移动时，同样能够调整剂量球的半径。除直接的直观方式编辑剂量外，还可通过调整积分DVH曲线或调整微分DVH曲线的方式修改某些感兴趣体积内的剂量；通过编辑等剂量线的方式修改剂量分布；通过剂量云的笔刷等方式修改剂量分布；在剂量的三维视图下修改剂量分布等方式，对剂量进行编辑。
在上述实施例中，优选地，空间剂量模型中心点的剂量调节上限Dl和下限Du分别为：
Figure PCTCN2020091843-appb-000025
Figure PCTCN2020091843-appb-000026
其中，Dl为剂量可调节的下限，Du为剂量可调节的上限，D0为动作操控装置的动作事件被触发时空间剂量模型中心点处的点剂量，R为空间剂量模型的特征参数，Dmax为剂量数据的全局最大剂量值，n为常数。
本发明还提出一种标准化的人工智能自动放疗计划***，放疗计划***用于实现如上述实施例中任一项的标准化的人工智能自动放疗计划方法。
本发明还提供一种计算设备，包括：
一个或多个处理器；
存储器；以及
一个或多个程序，其中一个或多个程序存储在存储器中并被配置为由一个或多个处理器执行，该一个或多个程序包括用于实现上述标准化的人工智能自动放疗计划方法的指令。
本发明还提供一种存储一个或多个程序的计算机可读存储介质，该一个或多个程序包括指令，指令适于由存储器加载并执行上述标准化的人工智能 自动放疗计划方法。
以上仅为本发明的优选实施例而已，并不用于限制本发明，对于本领域的技术人员来说，本发明可以有各种更改和变化。凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种标准化的人工智能自动放疗计划方法，其特征在于，包括：
    获取医学影像；
    对所述医学影像的ROI区域进行自动勾画以获取几何解剖结构；
    根据所述医学影像对应的病种信息、所述几何解剖结构以及预设的病种-处方模板库确定处方；
    根据所述病种信息、所述几何解剖结构和所述处方确定放疗照射角度；
    将所述医学影像、所述几何解剖结构、所述病种信息、所述处方和所述放疗照射角度输入剂量预测模型中，得出放疗剂量分布结果；
    以所述放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划；
    所述可执行放疗计划包括正向放疗计划、立体定向放疗计划和调强放疗计划，其中所述调强放疗计划包括动态调强放疗计划、静态调强放疗计划、容积调强放疗计划和旋转调强放疗计划。
  2. 根据权利要求1所述的标准化的人工智能自动放疗计划方法，其特征在于，还包括：
    通过统一的处方标准与人工智能结合，对所生成的可执行放疗计划进行评分，得到计划评估的评分总分；
    通过蒙特卡罗三维剂量验证技术，对所生成的可执行放疗计划进行2D或者3D Gamma分析，得到Gamma分析的通过率情况；
    基于可执行的放疗计划、计划评估的评分总分和Gamma分析的通过率情况自动生成放疗计划报告；
    医生对所述放疗计划报告进行审核。
  3. 根据权利要求1所述的标准化的人工智能自动放疗计划方法，其特征在于，针对所述放疗剂量分布结果，还包括：
    在接收到剂量编辑触发指令时，进入剂量编辑模式；
    空间剂量模型在当前放射治疗图像断面上的切面图形随控制光标的轨迹移动，其中，所述控制光标所在位置为所述空间剂量模型的中心，所述控制光标的轨迹与动作操控装置的移动轨迹对应；
    监听所述动作操控装置的动作事件，并根据该动作事件对应的预设操控命令调整所述空间剂量模型中心点处的剂量；
    以所述空间剂量模型中心点的剂量为基准通过插值算法计算所述空间剂量模型内其余点的剂量；
    保存和更新所述空间剂量模型内各点的剂量数据，并对所述空间剂量模型以外区域的剂量数据不做修改。
  4. 根据权利要求1所述的标准化的人工智能自动放疗计划方法，其特征在于，所述对所述医学影像的ROI区域进行自动勾画以获取几何解剖结构具体包括：
    正常器官的自动识别与自动勾画：基于机器学习自动识别和勾画人体全身的各个正常器官；
    肿瘤部位的自动识别与勾画：如果全身器官能进行勾画，则肿瘤采用反向勾画；将危及器官勾画完成后，剩余部分为肿瘤部位；采用机器学习的病种PTV与GTV的外扩的关系，对剩余部分进行自动勾画；
    所述根据所述病种信息、所述几何解剖结构和所述处方确定放疗照射角度具体包括：
    将历史病例的病种信息、几何解剖结构和处方进行机器学习，确定照射角度预测模型，将当前病例的所述病种信息、所述几何解剖结构和所述处方输入所述照射角度预测模型以获取预测照射角度，作为所述放疗照射角度；或，
    根据病种标记计划靶区的器官权重，沿射线方向计算每个角度的器官权重累加值，合并相邻满足预设权重阈值的角度，将符合权重阈值的角度作为所述放疗照射角度；或，
    确定感兴趣区域，选择至少一个计划靶区和一个危及器官，并对每个感兴趣区域进行全角度照射角度投影；在每个分段角度的每个角度上对计划靶区计算最小外接矩形，对某危及器官在该角度与对应的最小外接矩形进行交集运算，得到交集面积；对所有分段角度的交集面积进行求和，以最小的和值作为目标函数，采用非线性整数最优化算法求解得到最优分段索引和最优角度索引，作为所述放疗照射角度。
  5. 根据权利要求1所述的标准化的人工智能自动放疗计划方法，其特征在于，所述剂量预测模型的构建方法包括：
    以归一化的PTV平均剂量建立数据集，并以该数据集制定评分模板；
    对感兴趣区域进行标准化命名；
    将三维医学影像分割为二维切片作为训练集和测试集；
    读取所述训练集的三维计划靶区数据的射线角度，将射线角度投影在计划靶区上得到网络权重，对该网络权重采用剂量计算算法进行剂量计算得到射束通道；
    以U-net网络或者V-net网络作为生成器、以马尔科夫判别器作为判别器构建Pix2pix剂量预测模型；
    以所述二维切片图像作为所述生成器的输入，以所述生成器输出的预测剂量和原始剂量作为所述判别器的输入，所述判别器输出判断结果；
    将所述训练集的所有二维切片输入所述Pix2pix剂量预测模型训练。
  6. 根据权利要求1所述的标准化的人工智能自动放疗计划方法，其特征在于，所述以所述放疗剂量分布结果作为参考，采用基于剂量分布或DVH引导的逆向优化算法进行优化处理，生成可执行放疗计划，具体包括：
    基于通量图优化算法，优化出通量权重图；然后，结合加速器的机器信息，采用叶片序列算法，自动生成可执行动态调强放疗计划；或，
    基于直接子野优化方法，自动生成可执行静态调强放疗计划；或，
    基于遗传算法或者列生成算法，自动生成容积调强放疗计划或者旋转调强放疗计划；或，
    正向放疗计划；或，
    立体定向放疗计划。
  7. 根据权利要求3所述的标准化的人工智能自动放疗计划方法，其特征在于，所述根据该动作事件对应的预设操控命令调整所述空间剂量模型中心点处的剂量具体包括：
    在监听到动作事件时，浮动显示剂量调节指示标签；
    在所述控制光标在所述剂量调节指示标签区域时监听到所述动作操控装置被触发第一动作事件，则以所述控制光标所处位置对应的剂量值作为所述空间剂量模型中心点处的剂量；
    在所述控制光标在所述剂量调节指示标签上的指示滑块上时监听到所述动作操控装置被触发第二动作事件，则以解除点击第二动作事件时所述控制光标所处位置对应的剂量值作为所述空间剂量模型中心点处的剂量；
    在所述控制光标处于当前切面图形区域内时监听到所述动作操控装置被触发第三动作事件，则以所述第三动作事件的动作参数调整所述空间剂量模型中心点处的剂量；
    在所述控制光标未处于当前切面图形区域内时若监听到所述动作操控装置被触发第三动作事件，则以所述第三动作事件的动作参数对所述放射治疗图像进行翻层操作，移出该切面图形区域时保存所述第三动作事件调整的剂量。
  8. 根据权利要求3所述的标准化的人工智能自动放疗计划方法，其特征在于，还包括：
    监听动作操控装置的动作事件，根据该动作事件对应的预设操控命令还可对所述空间剂量模型所在的放射治疗图像进行翻层操作，以及可调整所述空间剂量模型的尺寸；
    在监听到所述动作操控装置被触发第四动作事件时，根据所述第四动作事件的动作参数调节所述空间剂量模型的尺寸。
  9. 根据权利要求3或7所述的标准化的人工智能自动放疗计划方法，其特征在于，所述空间剂量模型中心点的剂量调节上限Dl和下限Du分别为：
    Figure PCTCN2020091843-appb-100001
    Figure PCTCN2020091843-appb-100002
    其中，Dl为剂量可调节的下限，Du为剂量可调节的上限，D0为所述动作操控装置的动作事件被触发时所述空间剂量模型中心点处的点剂量，R为所述空间剂量模型的特征参数，Dmax为剂量数据的全局最大剂量值，n为常数。
  10. 一种标准化的人工智能自动放疗计划***，其特征在于，所述放疗计划***用于实现如权利要求1至9中任一项所述的标准化的人工智能自动放疗计划方法。
PCT/CN2020/091843 2019-08-29 2020-05-22 标准化的人工智能自动放疗计划方法和*** WO2021036366A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/977,095 US11964170B2 (en) 2019-08-29 2020-05-22 Standardized artificial intelligence automatic radiation therapy planning method and system

Applications Claiming Priority (6)

Application Number Priority Date Filing Date Title
CN201910820650.3A CN110415785A (zh) 2019-08-29 2019-08-29 人工智能引导放疗计划的方法及***
CN201910820650.3 2019-08-29
CN201911229101.5 2019-12-04
CN201911229101.5A CN111028914B (zh) 2019-12-04 2019-12-04 人工智能引导的剂量预测方法与***
CN201911421531.7A CN113130042B (zh) 2019-12-31 2019-12-31 放射治疗计划***中剂量编辑的方法
CN201911421531.7 2019-12-31

Publications (1)

Publication Number Publication Date
WO2021036366A1 true WO2021036366A1 (zh) 2021-03-04

Family

ID=74685559

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/091843 WO2021036366A1 (zh) 2019-08-29 2020-05-22 标准化的人工智能自动放疗计划方法和***

Country Status (2)

Country Link
US (1) US11964170B2 (zh)
WO (1) WO2021036366A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113658149A (zh) * 2021-08-23 2021-11-16 上海联影医疗科技股份有限公司 计算机可读存储介质以及计算设备
CN113658168A (zh) * 2021-08-25 2021-11-16 上海联影医疗科技股份有限公司 指定剂量区的获取方法、***、终端及存储介质
CN117011245A (zh) * 2023-07-11 2023-11-07 北京医智影科技有限公司 融合mr信息指导ct的直肠癌肿瘤区自动勾画方法及装置
CN117323584A (zh) * 2023-10-18 2024-01-02 迈胜医疗设备有限公司 用于放射治疗的计划调整方法、放射治疗***及相关装置
EP4316581A1 (en) * 2022-08-05 2024-02-07 Varian Medical Systems Inc Radiotherapy dose parameter computation techniques
CN117653937A (zh) * 2024-01-31 2024-03-08 四川大学华西医院 一种分离放疗中剂量学效应的方法、***和存储介质
WO2024068167A1 (en) * 2022-09-27 2024-04-04 Siemens Healthineers International Ag Method and apparatus to modify dose values during radiation treatment planning

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4019084A1 (en) * 2020-12-22 2022-06-29 Koninklijke Philips N.V. Planning apparatus for planning a radiation therapy
US20230290465A1 (en) * 2022-03-10 2023-09-14 Varian Medical Systems, Inc. Artificial intelligence boosting dose calculations
US20230315272A1 (en) * 2022-03-30 2023-10-05 Varian Medical Systems, Inc. Graphical user interface control device for radiation therapy treatment planning
CN116779173B (zh) * 2023-08-24 2023-11-24 北京大学第三医院（北京大学第三临床医学院） 一种基于人工智能的放射治疗剂量预测***和方法
CN117031136B (zh) * 2023-09-22 2024-03-29 北京中成康富科技股份有限公司 毫米波治疗仪的检测方法、装置、设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105930636A (zh) * 2016-03-29 2016-09-07 中北大学 一种自动确定目标函数权重的放射治疗方案优化***
CN108717866A (zh) * 2018-04-03 2018-10-30 陈辛元 一种预测放疗计划剂量分布的方法、装置、设备及存储介质
CN109771843A (zh) * 2017-11-10 2019-05-21 北京连心医疗科技有限公司 云放射治疗计划评估方法、设备与存储介质
CN110415785A (zh) * 2019-08-29 2019-11-05 北京连心医疗科技有限公司 人工智能引导放疗计划的方法及***
CN111028914A (zh) * 2019-12-04 2020-04-17 北京连心医疗科技有限公司 人工智能引导的剂量预测方法与***

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7369645B2 (en) * 2004-06-21 2008-05-06 Derek Graham Lane Information theoretic inverse planning technique for radiation treatment
JP2020525093A (ja) * 2017-06-22 2020-08-27 リフレクション メディカル， インコーポレイテッド 生物学的適合放射線療法のためのシステムおよび方法
CN107403201A (zh) 2017-08-11 2017-11-28 强深智能医疗科技（昆山）有限公司 肿瘤放射治疗靶区和危及器官智能化、自动化勾画方法
CN109801696A (zh) 2017-11-17 2019-05-24 北京连心医疗科技有限公司 一种人工智能的云放疗计划方法、设备、存储介质和***

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105930636A (zh) * 2016-03-29 2016-09-07 中北大学 一种自动确定目标函数权重的放射治疗方案优化***
CN109771843A (zh) * 2017-11-10 2019-05-21 北京连心医疗科技有限公司 云放射治疗计划评估方法、设备与存储介质
CN108717866A (zh) * 2018-04-03 2018-10-30 陈辛元 一种预测放疗计划剂量分布的方法、装置、设备及存储介质
CN110415785A (zh) * 2019-08-29 2019-11-05 北京连心医疗科技有限公司 人工智能引导放疗计划的方法及***
CN111028914A (zh) * 2019-12-04 2020-04-17 北京连心医疗科技有限公司 人工智能引导的剂量预测方法与***

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113658149A (zh) * 2021-08-23 2021-11-16 上海联影医疗科技股份有限公司 计算机可读存储介质以及计算设备
CN113658149B (zh) * 2021-08-23 2024-05-17 上海联影医疗科技股份有限公司 计算机可读存储介质以及计算设备
CN113658168A (zh) * 2021-08-25 2021-11-16 上海联影医疗科技股份有限公司 指定剂量区的获取方法、***、终端及存储介质
CN113658168B (zh) * 2021-08-25 2024-04-16 上海联影医疗科技股份有限公司 指定剂量区的获取方法、***、终端及存储介质
EP4316581A1 (en) * 2022-08-05 2024-02-07 Varian Medical Systems Inc Radiotherapy dose parameter computation techniques
WO2024068167A1 (en) * 2022-09-27 2024-04-04 Siemens Healthineers International Ag Method and apparatus to modify dose values during radiation treatment planning
CN117011245A (zh) * 2023-07-11 2023-11-07 北京医智影科技有限公司 融合mr信息指导ct的直肠癌肿瘤区自动勾画方法及装置
CN117011245B (zh) * 2023-07-11 2024-03-26 北京医智影科技有限公司 融合mr信息指导ct的直肠癌肿瘤区自动勾画方法及装置
CN117323584A (zh) * 2023-10-18 2024-01-02 迈胜医疗设备有限公司 用于放射治疗的计划调整方法、放射治疗***及相关装置
CN117323584B (zh) * 2023-10-18 2024-03-29 迈胜医疗设备有限公司 用于放射治疗计划调整的电子设备、放射治疗***及相关装置
CN117653937A (zh) * 2024-01-31 2024-03-08 四川大学华西医院 一种分离放疗中剂量学效应的方法、***和存储介质
CN117653937B (zh) * 2024-01-31 2024-04-19 四川大学华西医院 一种分离放疗中剂量学效应的方法、***和存储介质

Also Published As

Publication number Publication date
US11964170B2 (en) 2024-04-23
US20230128148A1 (en) 2023-04-27

Similar Documents

Publication Publication Date Title
WO2021036366A1 (zh) 标准化的人工智能自动放疗计划方法和***
US10765888B2 (en) System and method for automatic treatment planning
CN111028914B (zh) 人工智能引导的剂量预测方法与***
CN109069858B (zh) 一种放射治疗***及计算机可读存储装置
WO2022142770A1 (zh) 放射治疗自动计划***、自动计划方法及计算机程序产品
US10346593B2 (en) Methods and systems for radiotherapy treatment planning
US10342994B2 (en) Methods and systems for generating dose estimation models for radiotherapy treatment planning
US20040165696A1 (en) Systems and methods for global optimization of treatment planning for external beam radiation therapy
US20140275706A1 (en) Systems and methods for determining and delivering radiation treatment plans
CN112770811A (zh) 使用深度学习引擎进行放射疗法治疗计划的方法和***
US20230020911A1 (en) Methods and systems for adaptive radiotherapy treatment planning using deep learning engines
US11282192B2 (en) Training deep learning engines for radiotherapy treatment planning
CN113891742B (zh) 用于基于连续深度学习的放射疗法治疗规划的方法和***
US11786204B2 (en) Automatically-registered patient fixation device images
US11013936B2 (en) Methods and systems for generating dose estimation models for radiotherapy treatment planning
CN114344740A (zh) 用于使用深度学习引擎进行自适应放射治疗计划的方法和***
WO2020257216A1 (en) Methods and systems for quality-aware continuous learning for radiotherapy treatment planning
CN107666940A (zh) 选择射束几何结构的方法
CN115938587A (zh) 放疗计划的验证方法、装置、计算机设备和存储介质
CN113272909A (zh) 针对基于深度迁移学习的放射疗法处理计划的方法和***
Duprez The application of a deep convolution neural network for the automated delineation of the target and organs at risk in High Dose Rate Cervical Brachytherapy

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: 20856238

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: 20856238

Country of ref document: EP

Kind code of ref document: A1