INTELLIGENT OPTIMIZATION SETTING ADJUSTMENT FOR RADIOTHERAPY TREATMENT PLANNING USING PATIENT GEOMETRY INFORMATION AND ARTIFICIAL INTELLIGENCE

Abstract

By using the Al module, the method of the present invention calculates, i.e. predicts, the dependency C.sub.i (p.sub.i) of a radiotherapy (RT) quality criterion C, from an adjustment of such a radiotherapy planning parameter p.sub.i. In this way, the decision making process in RT treatment plan optimization is streamlined by prediction of promising settings of one or more radiotherapy planning parameters p, before the actual time intensive iterative optimization process is carried out. This is achieved by applying an Al module, which has been trained to predict the specific behaviour of the dose optimization algorithm, i.e. the optimizer, with respect to geometric patient data, dose prescription and treatment indication data. Thus, a computer-implemented medical method of predicting a dependency C.sub.i (p.sub.i) of a radiotherapy (RT) quality criterion C.sub.i from an adjustment of a radiotherapy planning parameter p, is presented. The method comprises the following steps of providing geometric patient data geometrically describing an area of a patient, which is to be irradiated according to a radiotherapy treatment plan (step S1), providing dose prescription data and treatment indication data for said patient (step S2), and predicting with a trained Artificial Intelligence (Al) module the dependency C.sub.i (p.sub.i) of the radiotherapy quality criterion C.sub.i from the radiotherapy planning parameter p, when adjusting said radiotherapy planning parameter p.sub.i, thereby using the geometric patient data, the dose prescription data and the treatment indication data as input for the Al module (step S3).

Claims

1. A computer-implemented medical method of predicting a dependency C.sub.i (p.sub.i) of a radiotherapy (RT) quality criterion C.sub.i from an adjustment of a radiotherapy (RT) planning parameter p.sub.i, the method comprising the following steps: providing geometric patient data geometrically describing an area of a patient, which is to be irradiated according to a radiotherapy treatment plan (step S1), providing dose prescription data and treatment indication data for said patient (step S2), predicting with a trained Artificial Intelligence (AI) module the dependency C.sub.i (p.sub.i) of the radiotherapy quality criterion C.sub.i from the radiotherapy planning parameter p.sub.i when adjusting said radiotherapy planning parameter p.sub.i, thereby using the geometric patient data, the dose prescription data and the treatment indication data as input for the AI module (step S3), and wherein the radiotherapy (RT) quality criterion C.sub.i is at least one of RT delivery time, complexity of RT quality assurance (QA), a RT dose constraint, a RT volume constraint regarding a volume within the patient, a conformity index describing how well an irradiated area correlates with a planning target volume (PTV), a gradient index describing how quickly a RT radiation dose decreases with increasing distance to a target volume, and any combination thereof.

2. The method according to claim 1, further comprising: determining a plurality of anchor points based on the predicted dependency C.sub.i (p.sub.i) of the radiotherapy quality criterion C.sub.i from an adjustment of the radiotherapy planning parameter p.sub.i, wherein each anchor point is a suggested value of said radiotherapy planning parameter p.sub.i for pre-calculating a dose distribution.

3. The method according to claim 2, wherein the determination of the anchor points uses a comparison of the predicted dependency C.sub.i (p.sub.i) with a predefined threshold.

4. The method according to claim 2, wherein the step of predicting the dependency C.sub.i (p.sub.i) comprises: determining whether within a predefined range of values of said radiotherapy planning parameter p.sub.i a change of a gradient of the dependency C.sub.i (p.sub.i) is above or below a predefined threshold, determining a number of the plurality of anchor points depending on whether the change of the gradient of the dependency C.sub.i (p.sub.i) is above or below the predefined threshold.

5. The method according to claim 4, wherein a higher number of anchor points are determined when the change of the gradient of the dependency C.sub.i (p.sub.i) is above the threshold as compared to a lower number of anchor points that are determined if the change of the gradient is lower than the predefined threshold.

6. The method according to claim 2, further comprising: pre-calculating a RT dose distribution for each of the determined anchor points, and approximating a RT dose distribution, and preferably a value for the RT quality criterion C.sub.i, for a value of the radiotherapy planning parameter p.sub.i, which value is not an anchor point, thereby using results of the pre-calculation of the RT dose distribution for the anchor points.

7. The method according to claim 1, wherein the radiotherapy quality criterion C.sub.i is a qualitative parameter indicative for a radiotherapy dose distribution.

8. (canceled)

9. The method according to claim 1, wherein the radiotherapy planning parameter p.sub.i is at least one of a weighting between a RT target and an organ at risk (OAR), a parameter describing normal tissue sparing, a parameter describing a degree of RT modulation, and any combination thereof.

10. The method according to claim 1, further comprising: determining an amended adjustment range of the RT planning parameter p.sub.i based on the predicted dependency Ci (pi), and suggesting the amended adjustment range to a user of an RT treatment planning software.

11. The method according to claim 10, wherein the step of suggesting the amended adjustment range to a user further comprises: adapting a graphical user interface (GUI) of the RT treatment planning software such that the amended adjustment range of the RT planning parameter p.sub.i is displayed to the user.

12. The method according to claim 10, further comprising: performing a full optimization of the RT treatment plan with the amended adjustment range of the RT planning parameter p.sub.i by using an RT planning optimization algorithm.

13. The method according to claim 1, wherein the steps of (S1) to (S3) are repeated for a plurality of radiotherapy quality criteria C.sub.i and a plurality of radiotherapy planning parameters p.sub.i.

14. The method according to claim 1, further comprising extracting geometrical features of the area of the patient, which is to be irradiated by radiotherapy, from at least one medical image and/or from organ structures segmented therefrom.

15. The method according to claim 14, further comprising providing said extracted geometrical features as an m-dimensional patient vector of numerical values, and using the m-dimensional patient vector as an input for the AI module.

16. The method according to claim 1, further comprising: providing for a plurality of current values of RT planning parameters p.sub.j, and considering the current values of RT planning parameters p.sub.j when predicting the dependency C.sub.i (p.sub.i) with the trained AI module.

17. A non-transitory computer medium comprising instructions, which, when running on at least one processor of at least one computer, causes the at least one processor to perform the steps of: provide geometric patient data geometrically describing an area of a patient, which is to be irradiated according to a radiotherapy treatment plan; provide dose prescription data and treatment indication data for said patient; predict with a trained Artificial Intelligence (AI) module a dependency C.sub.i (p.sub.i) of a radiotherapy quality criterion C.sub.i from a radiotherapy planning parameter p.sub.i when adjusting said radiotherapy planning parameter p.sub.i, thereby using the geometric patient data, the dose prescription data and the treatment indication data as input for the AI module; and wherein the radiotherapy (RT) quality criterion Ci is at least one of RT delivery time, complexity of RT quality assurance (QA), a RT dose constraint, a RT volume constraint regarding a volume within the patient, a conformity index describing how well an irradiated area correlates with a planning target volume (PTV), a gradient index describing how quickly a RT radiation dose decreases with increasing distance to a target volume, and any combination thereof.

18. (canceled)

19. A radiotherapy treatment system comprising one or more processors and associated memory operably coupled with the one or more processors, wherein the memory stores instructions that, in response to execution of the instructions by the one or more processors, causes the one or more processor to perform the steps of: providing geometric patient data geometrically describing an area of a patient, which is to be irradiated according to a radiotherapy treatment plan; providing dose prescription data and treatment indication data for said patient; predicting with a trained Artificial Intelligence (AI) module a dependency C.sub.i (p.sub.i) of a radiotherapy quality criterion C.sub.i from a radiotherapy planning parameter p.sub.i when adjusting said radiotherapy planning parameter p.sub.i, thereby using the geometric patient data, the dose prescription data and the treatment indication data as input for the AI module; and wherein the radiotherapy (RT) quality criterion Ci is at least one of RT delivery time, complexity of RT quality assurance (QA), a RT dose constraint, a RT volume constraint regarding a volume within the patient, a conformity index describing how well an irradiated area correlates with a planning target volume (PTV), a gradient index describing how quickly a RT radiation dose decreases with increasing distance to a target volume, and any combination thereof; a radiation treatment apparatus comprising a treatment beam source and a patient support unit; wherein the one or more processors are operably coupled to the radiation treatment apparatus for issuing a control signal to the radiation treatment apparatus for controlling at least one of: the operation of the treatment beam source or the position of the patient support unit.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0133] In the following, the invention is described with reference to the appended figures which give background explanations and represent exemplary embodiments of the invention. The scope of the invention is however not limited to the specific features disclosed in the context of the figures, wherein

[0134] FIG. 1 shows a flowchart illustrating steps of a method of predicting a dependency C.sub.i (p.sub.i) of a radiotherapy (RT) quality criterion C.sub.i from an adjustment of a RT planning parameter p.sub.i according to an exemplary embodiment of the invention;

[0135] FIG. 2 schematically shows three different sliders/RT planning parameter p.sub.i modifiable by user via a typical graphical user interface of a RT treatment planning system;

[0136] FIG. 3 schematically shows an example on how a specific RT quality criterion C.sub.i might change when one slider is moved through the whole value range without changing the remaining parameters;

[0137] FIG. 4 schematically shows that the number of anchor points needed for a good approximation is depending on the range of values achieved for a C.sub.i, which is used in another exemplary embodiment of the invention;

[0138] FIG. 5 schematically shows a method of predicting with a trained AI module the dependency C.sub.i (p.sub.i) and of determining a plurality of anchor points based on the predicted dependency C.sub.i (p.sub.i) of the RT quality criterion C.sub.i according to an exemplary embodiment of the invention;

[0139] FIG. 6 schematically shows a RT treatment system including a RT treatment planning system according to another exemplary embodiment of the invention;

[0140] FIG. 7 schematically shows a method step of generating training data for the AI module according to another exemplary embodiment of the invention.

[0141] The figures are schematic only and not true to scale. In principle, identical or like parts, elements and/or steps are provided with identical or like reference symbols in the figures.

DESCRIPTION OF EMBODIMENTS

[0142] FIG. 1 shows a flow chart illustrating the basic steps of the method of predicting a dependency C.sub.i (p.sub.i) of a radiotherapy (RT) quality criterion C.sub.i from an adjustment of a radiotherapy planning parameter p.sub.i, according to an exemplary embodiment and/or according to the first aspect.

[0143] The method of FIG. 1 is a computer-implemented medical method and comprises in step S1 the provision of geometric patient data geometrically describing an area of a patient, which is to be irradiated according to a radiotherapy treatment plan. Further, the provision of dose prescription data and treatment indication data for said patient is comprised as step S2. Predicting the dependency C.sub.i (p.sub.i) of the RT quality criterion C.sub.i from the RT planning parameter p.sub.i when adjusting said RT planning parameter p.sub.i, thereby using the geometric patient data, the dose prescription data and the treatment indication data as input for the AI module forms step S3.

[0144] If desired, the steps of S1 to S3 are repeated for a plurality of radiotherapy quality criteria C.sub.i and a plurality of radiotherapy planning parameters p.sub.i. Furthermore, it should be noted that in this and every other embodiment described herein the following additional steps may be comprised by the method. Providing for a plurality of current values, i.e. fixed values, of RT planning parameters p.sub.i, and considering the current values of RT planning parameters p.sub.j when predicting the dependency Ci (pi) with the trained AI module.

[0145] Typically the “geometric patient data” comprise the organ at risk (OAR) volume, the planning target volume (PTV), the distance OAR-PTV, OAR concavity/convexity, PTV concavity/convexity, overlap histogram (OVH) data, and others. In an embodiment, the geometric patient data are obtained from image sets like CT or MR, or the like. In this embodiment organ structures are segmented from these images and the geometrical features are extracted and used as the geometric patient data. As described before, these “geometric patient data” can be provided according to an exemplary embodiment in the form of a “patient vector”, in which also the dose prescription data and the treatment indication data for said patient are contained. Such a “patient vector” will be described in the context of FIG. 7 in more detail. Furthermore, typically the “dose prescription data and treatment indication data” comprise one or more of the PTV dose prescription, the PTV maximum dose, the PTV minimum dose, the OAR maximum dose, an OAR dose objective, a list of OARs, a number of fractions, a highlighted, a special OAR, and others.

[0146] The presented method of predicting a dependency C.sub.i (p.sub.i) will save valuable time and calculation capacity in the overall process of finding the final treatment plan, which will be used to irradiate and treat the patient.

[0147] If, for example, the predicted dependency of C.sub.i (p.sub.i) is nearly flat, i.e. the value of C.sub.i does not change or does not significantly change upon a variation of p.sub.i in a certain range. Such an example can be seen in FIG. 4. Than this result of the method of the present invention can be used to limit the range of the parameter p.sub.i, in which the subsequent time intensive iterative optimization process is accurately carried out by the optimizer. Moreover, so called “anchor points”, i.e. specific values of p.sub.i, can be determined with the method of FIG. 1. Said “anchor points” can be understood as selected values of p.sub.i, which are used for the subsequent final time intensive iterative optimization process, instead of the entire range of p.sub.i. In case of a dependency of C.sub.i (p.sub.i), where C.sub.i strongly varies upon a variation of p.sub.i in a certain range, than this result of the method of the method can be used for the final optimization process in that this range should be taken into account, and not neglected. Also this scenario will be elucidated further in the context of the embodiment of FIG. 4. Aspects about the training data of the trained AI module have been described before and are supplemented by the disclosure about FIG. 7.

[0148] In the following, a further detailed embodiment of the method of FIG. 1 is explained by means steps S1 and S3 as explained before and by means of additional method steps.

[0149] According to this further embodiment, in a first step a database of patient geometry data, dose prescription and treatment indication data and slider position configurations with characteristics of resulting treatment plans including dependencies C.sub.i (p.sub.i) is created. An embodiment of this database creation is shown in FIG. 7. In a further step the AI module is trained to predict slider positions, which create the biggest or minimum “trade-offs” in crucial dosimetric features C.sub.i of resulting treatment plans. For the current patient case of interest, geometric information from this patient case is extracted, typically by feature segmentation from medical images like CT or MRI images. Based on the geometric patient data, dose prescription and treatment indication data values of p.sub.i are determined which are most value to later on efficiently generate the final RT treatment plan. Thus, calculations are avoided in the subsequent optimization by the optimizer, which make us of values of p.sub.i, which cost calculation time but do not or not significantly change the outcome of the RT treatment plan. Afterwards, full optimizations of the adjustment range of the predictions instead carrying out the optimization in the whole possible range of p.sub.i. As a final step, the user may choose the final treatment plan, that is to be applied to the RT treatment system to initiate the treatment.

[0150] FIG. 2 schematically shows three different sliders/RT planning parameter p.sub.i modifiable by user via a typical graphical user interface of a RT treatment planning system. The radiation dose distribution used for patient treatment is acquired by a computerized optimization algorithm, as has been explained before. Characteristics of this dose distribution can be influenced by adjustment of a number of optimization parameters p.sub.i via slider controls prior to optimization. Each single one of these parameters influences a whole group of various aspects of the optimization process. These ‘composite’ parameters are chosen in such a way that each parameter describes an important aspect of the optimization result, e.g. the trade-off between irradiating as much of the target volume as possible vs. putting emphasis on sparing the healthy tissue surrounding the target. Another example is: simple QA and fast delivery vs. an optimal dose distribution at the cost of higher delivery time and a more challenging QA process. An illustration on how these parameters can be adjusted in a graphical user interface is shown in FIG. 2.

[0151] FIG. 3 schematically shows an example on how a specific RT quality criterion C.sub.i might change when one slider is moved through the whole value range without changing the remaining parameters. Typically, the resulting 3D dose distribution is evaluated by a number of quality criteria C.sub.i. Such a criterion can e.g. be a dose/volume constraint defined in the dose prescription or the conformity index (CI), or the gradient index (GI), as has been described in detail hereinbefore. Even though the value composition of the parameter values is initialized in way that gives a good effectivity for most cases, the direct behaviour of a specific C.sub.i depends on so many aspects (geometry of the patient, dose prescription and dosimetric optimization targets, the nature of the optimizer itself) that an analytical a priori statement on the behaviour of a specific C.sub.i with respect to changes in a specific slider cannot be made. However, the present invention overcomes this disadvantage as was explained before and is made even clearer with the following disclosure.

[0152] FIG. 4 shows an approximation resolution for two different RT quality criteria C.sub.i and C.sub.2. In particular, FIG. 4 schematically shows that the number of anchor points 40, 41 needed for a good approximation is depending on the range of values achieved for a C.sub.i. This insight is used in another exemplary embodiment of the invention. In said embodiment a plurality of anchor points is determined based on the predicted dependency C.sub.i (p.sub.i) of the radiotherapy quality criterion C.sub.i from an adjustment of the radiotherapy planning parameter p.sub.i. Therein each anchor point 40, 41 is a suggested value of said radiotherapy planning parameter p.sub.i for pre-calculating a dose distribution. Thus, the user may start the optimizer to do a full optimizations with only the predicted “anchor points” 40, 41 instead of the complete slider range. In other words, with this embodiment the slider positions with the most interesting “trade-offs” are predicted. The anchor points may cover the entire range of the dependency C.sub.i (p.sub.i) as is shown in both diagrams of FIG. 4. But they may also be the upper and lower limit of a partial range of the parameter p.sub.i.

[0153] FIG. 5 schematically shows a method of predicting with a trained AI module the dependency C.sub.i (p.sub.i) and the method also determines a plurality of anchor points based on the predicted dependency C.sub.i (p.sub.i) of the RT quality criterion C.sub.i according to an exemplary embodiment of the invention. This embodiment builds on FIG. 1 and comprises additional further features. As can be gathered from FIG. 5, the patient vector 50 comprises geometric patient data, dose prescription data and treatment indication data 50a. The patient vector 50 also describes, which radiotherapy planning parameter p.sub.i shall be varied/modified and defines fixed positions or values of the remaining radiotherapy planning parameter p.sub.j, see reference sign 50b. Thus, in this embodiment a plurality of fixed values of RT planning parameters p.sub.j are provided and the fixed values of RT planning parameters p.sub.j are considered by the trained AI module when predicting the dependency C.sub.i (p.sub.i). The patient vector 50 also defines, which RT quality criterion C.sub.i is selected by the user. This patient vector 50 may be used as input for the trained AI module/model 51, which predicts the values of RT quality criterion C.sub.i in the possible range of p.sub.i. The AI module also determines the number of anchor points, thereby using a comparison of the predicted values of RT quality criterion C.sub.i/the predicted dependency C.sub.i (p.sub.i) with a predefined threshold.

[0154] If for example, within a given range of p.sub.i defined by values p.sub.i,1 and p.sub.i,2 the difference between the maximum and minimum values of C.sub.i in this range is above this predefined threshold value than a relatively high number of anchor points is determined by the presented method. In this case the range of p.sub.i defined by values p.sub.i,1 and p.sub.i,2 is worth of “exploring” by calculating subsequently a full optimization with the optimizer for the relatively high number of anchor points. If, however, within a given range of p.sub.i defined by values p.sub.i,1 and p.sub.i,2 the maximum and minimum values of C.sub.i in this range is below this predefined threshold value than a relatively low number of anchor points is determined by the presented method. In other words, due to the dependency C.sub.i (p.sub.i) predicted by the AI module 51, it is determined by the presented method that this range between p.sub.i,1 and p.sub.i,2 a detailed exploration in the sense of a full calculation/optimization of the treatment plan in this range is not needed.

[0155] Similar to the previously mentioned embodiment, in a further embodiment it is determined whether within a predefined range of values of said radiotherapy planning parameter p.sub.i a change of a gradient of the dependency C.sub.i (p.sub.i) is above or below a predefined threshold. Furthermore, the number of the plurality of anchor points is determined depending on whether the change of the gradient of the dependency C.sub.i (p.sub.i) is above or below the predefined threshold.

[0156] FIG. 6 schematically shows a RT treatment system 60 including a RT treatment planning system 61 according to another exemplary embodiment of the invention. The radiotherapy treatment system 60 comprises besides the RT treatment planning system 61, a radiation treatment apparatus 66 comprising a treatment beam source 67 and a patient support unit 68. The computer 62 of the RT treatment planning system is operably coupled to the radiation treatment apparatus 66 for issuing a control signal to the radiation treatment apparatus for controlling at least one of the operation of the treatment beam source 67 or and the position of the patient support unit 68. Furthermore, the program 64, which is running on the computer 62 or when loaded onto the computer, causes the computer to perform the method steps of the method according to the first aspect of the present invention. The computer 62 comprises also a program storage medium 63 on which the program is stored, and also comprises the AI module 65. The AI module may be a separate calculation unit within the computer or may be realized as a program/algorithm running on the computer 62.

[0157] FIG. 7 schematically shows a method step of providing/generating training data 71 for the AI module 65 from several previous patient cases 72 according to another exemplary embodiment of the invention. The training data 71 may form together a training database 70. In particular, the AI module can be trained with a high number of training data 71 that are optimizations which are calculated from other patients' data 72. The optimization with the classical RT treatment planning optimizer is shown in FIG. 7 with step “Optimization” 73. The training data generally comprise training input data 71a, which comprise at least geometric patient data as well as the dose prescription data and the treatment indication data. The training data 71 also comprise training output data 71b, which comprise at least a part of a dependency C.sub.i (p.sub.i) of a RT quality criterion C.sub.i from an adjustment of a RT planning parameter p.sub.i. In the embodiment shown in FIG. 7 the quality criterion is exemplarily shown as “DVH Criterion” and the modified planning parameter p.sub.i is the first slider of the three sliders exemplarily shown in FIG. 7.

[0158] For a selection of patient cases a specific treatment setup 72a, including dose prescription data and treatment indication data, and a list 72b of quality criteria C.sub.i can be selected. From the medical image sets (CT, MR, etc.) and the segmented organ structures geometrical features 72c are extracted, wherein these geometrical features are embodiments of the geometric patient data as used in the method of the present invention. If desired, each patient case can be described as an m-dimensional vector 72 of numerical values. Such geometrical features may typically be one or more of the following parameters: the organ at risk (OAR) volume, the planning target volume (PTV), the distance OAR-PTV, OAR concavity/convexity, PTV concavity/convexity, overlap histogram (OVH) data, and others.

[0159] For each vector 72 dose optimizations 73 for all different parameter combinations can be performed. While in each case the complete resulting dose distribution could be saved, it is much more memory effective to only save the list of the resulting values of the quality criteria C.sub.i.

[0160] The AI module can then be trained with the plurality of training data 71 originating from several different patient cases. The training input data 71a, as was described before, are provided to the AI module as input. Also the corresponding dependencies C.sub.i (p.sub.i), that were previously calculated by the optimization with the classical RT treatment planning optimizer in step 73, are provided such that the AI module learns existing correlations between these three kinds of input data and the dependencies C.sub.i (p.sub.i). Thus, due to the training the AI module can then predict whether changes in particular sliders/parameters p.sub.i are likely to have a significant impact on important characteristics C.sub.i of the optimization results for new patient data in the form of new geometric patient data and new dose prescription and treatment indication data. This may of course be done for one or several RT quality criteria C.sub.i and one or several RT treatment planning parameters p.sub.i.

[0161] By using the AI module, the method of the present invention calculates, i.e. predicts, the dependency C.sub.i (p.sub.i) of a radiotherapy (RT) quality criterion C.sub.i from an adjustment of such a radiotherapy planning parameter p.sub.i. In this way, the decision making process in RT treatment plan optimization is streamlined by prediction of promising settings of one or more radiotherapy planning parameters p.sub.i before the actual time intensive iterative optimization process is carried out. This is achieved by applying an AI module, which has been trained to predict the specific behaviour of the dose optimization algorithm, i.e. the optimizer, with respect to geometric patient data, dose prescription and treatment indication data.