Method for converting a detector image of a first detector of a radiation source into a detector image of a second detector of the radiation source

Abstract

Method for converting a detector image of a first detector of a radiation source (1) into a detector image of a second detector of the radiation source (1) by an artificial intelligence, AI, (12), wherein the AI ​​(12) was initially trained with data pairs, the training of the AI ​​comprising the following steps:- generating (20) a control plan segment of a control plan,- generating (21) a first detector image for the control plan segment by simulating the first detector, wherein the first detector is an inaccurate detector,- generating (22) a second detector image for the control plan segment by simulating the second detector, wherein the second detector is an accurate detector,- training (23) the AI ​​with the first and second detector images as a data pair.

Description

[0001] The invention relates to a method for converting a result of a first radiophysical process of a radiation source into a result of a second radiophysical process of the radiation source.

[0002] This method is used, for example, in radiation therapy. To irradiate a tumor, a treatment plan or targeting plan is typically created, specifying which part of the tumor should be irradiated and with what dose. Using this treatment plan, it is possible to perform measurements on a phantom. However, the phantom does not correspond to the patient, so the actual dose received by the patient can differ from the measurements taken on the phantom. Because these two radiation physics processes differ, there is no simple way for the physician to verify the treatment plan.

[0003] Furthermore, especially in the field of radiation therapy, additional problems arise that necessitate the conversion from one result to another.

[0004] The object of the invention is to create a method for transferring such results.

[0005] According to the invention, an artificial intelligence (AI) is used for this purpose, wherein the AI ​​was initially trained with data pairs, each data pair representing a result of the first radiation physics process and a result of the second radiation physics process.

[0006] In this way, it is possible to map two radiation physics processes onto each other. Thus, in the example mentioned at the beginning, a measurement in a phantom can be converted into a measurement on a patient. This makes it easier and faster to check and validate a control plan.

[0007] In one version, the first and / or the second radiation physics process represents an ideal of a result. Here, the ideal reflects a result into which another result is to be transformed.

[0008] In one embodiment, the second radiophysical process models the irradiation of a patient, with the result representing a dose distribution in the patient.

[0009] In an alternative implementation, the second radiation physics process replicates the measurement in a high-precision detector. Such a detector could, for example, be a diamond detector.

[0010] In one interpretation, the first radiation physics process represents an erroneous reality of a result. Such an erroneous reality could, for example, be a measurement result from a less precise detector.

[0011] In one embodiment, the first and / or the second radiation physics process simulates the irradiation of a phantom, with the result representing a dose distribution in the phantom.

[0012] In one implementation, either the first or the second result is obtained through a measurement at a radiation source. In this way, real measurement data can be transformed into other measurement data, just as it would have been measured in the other application.

[0013] In one version, the AI ​​features at least one artificial neural network.

[0014] In one interpretation, a result is a dose distribution or a detector image. A result can also be a measurement curve or another result based on a radiation physics process, for example, an intensity distribution.

[0015] In one configuration, the radiation source is designed to emit ionizing radiation or particle radiation. The radiation source can alternatively or additionally be an X-ray source, a computed tomography scanner, or a linear accelerator.

[0016] In one implementation, generating a data pair for training the AI ​​includes the following steps: Creating a control plan segment of a Control plan, Generating an initial result for the Control plan segment through representation of the first radiation physics process, Generating a second result for the Control plan segment through representation of the second radiation physics process, Training the AI ​​with the first and second results as a data pair.

[0017] Crucially, all training data, or at least the vast majority of it, must be artificially generated. Within the control plan segment, all available parameters can be artificially and / or randomly generated. This allows for the creation of a virtually unlimited amount of training data, enabling the AI ​​to produce highly accurate results.

[0018] In one version, a control plan segment includes a position of the linear accelerator, an aperture configuration, an irradiation intensity, and other parameters of the accelerator.

[0019] In one implementation, a representation is generated through simulation, where the simulation is based on physical and / or stochastic principles. For example, a Monte Carlo simulation can be used. Depending on the application, however, other simulation principles, such as ray tracing, finite element analysis, or AI-supported methods, could also be used.

[0020] In one implementation, a representation is generated by converting the representation of the other radiation physics process. For example, a simulated measurement result from a high-precision detector (second process) can be converted into a measurement result from a less precise detector using known algorithms to obtain a representation of the first process. The algorithm might simply consist of reducing the resolution, i.e., the number of measurements. However, the algorithm could also be based on physical or stochastic principles.

[0021] In one version, the transfer process includes the following steps: - Measurement (30) of an initial result for a real control plan segment, - Input (31) of the measurement result into an AI (12) to convert it into a second measurement result.

[0022] In one implementation, a control plan is divided into several control plan segments. The conversion process is performed for each control plan segment, and the converted results of the control plan segments are combined into an overall result. Control plans are typically divided into segments. Therefore, during training and conversion, it is simpler to consider individual control plan segments. With such training data, only one control plan segment can be converted at a time.

[0023] The invention comprises the use of an AI for converting a dose distribution measured on a phantom into a dose distribution on a patient, in particular in a method according to one of the preceding claims, wherein the AI ​​was initially trained with data pairs, each data pair comprising a simulated dose distribution on the phantom and a simulated dose distribution on the patient.

[0024] The invention also includes the use of an AI to improve the image resolution of an image generated by a detector, wherein the AI ​​was initially trained with data pairs, each data pair comprising a simulated image of a first, in particular accurate, detector and a simulated or calculated image of a second detector.

[0025] The invention is explained in more detail below with reference to exemplary embodiments and the accompanying drawings.

[0026] It shows: Fig. 1. A first arrangement for training an AI, Fig. 2 an application of the after Fig. 1 trained AI, Fig. 3 a second arrangement for training an AI, Fig. 4 an application of the after Fig. 3 trained AIs, Fig. 5. A flowchart of a procedure for training an AI, and Fig. 6 a flowchart for converting a first result into a second result according to the invention.

[0027] The Fig. Figure 1 schematically shows a first arrangement for training an AI for a method according to the invention. Fig. Figure 1 schematically and exemplarily shows a linear accelerator 1 as a radiation source and a patient table 2. The linear accelerator has a particle beam 3.

[0028] For irradiation, a control plan is typically created for linear accelerator 1. This control plan is usually divided into individual control plan segments. Each control plan segment contains, for example, the position of the linear accelerator, the particle energy, the irradiation duration, and the aperture configuration. Aperture 4 of the linear accelerator, in particular, influences the beam shape and thus the dose distribution.

[0029] In Fig. Figure 3 shows three different aperture configurations 4 as examples, representing three different control plan segments. Each of these control plan segments is generated randomly, with each parameter, including the aperture configuration 4, being generated randomly.

[0030] To generate a data pair, a result of a radiation physics process is determined starting from one of these control plan segments. In this example, a measurement result from a high-precision detector is simulated (second process). Such a detector could be, for example, a diamond detector. The simulation of the measurement result can be performed, for example, using a Monte Carlo simulation.

[0031] The result of the first radiation physics process is a measurement from a less precise but faster detector. This can be done either through simulation or by calculating the result from the previously determined result of the second process. Algorithms are available for this purpose, but conversion is only possible from the precise result to the less precise result.

[0032] A data pair consisting of the first and second results is used to train an AI. This AI could, for example, be an artificial neural network.

[0033] In Fig. Figure 1 shows an example measurement curve in which the solid line 5 corresponds to the simulated curve of the precise detector and the dashed line 6 to the curve of the less precise detector. A data pair therefore contains both curves.

[0034] In practice, according to Fig. 2. A three-dimensional measurement was performed in a real linear accelerator 1 using an ionization chamber 7 in a water phantom 8. Such an ionization chamber 5 has only a low resolution, but enables very fast measurements. Especially in everyday clinical practice, it is crucial that a linear accelerator is not occupied and thus blocked for a long time by a single measurement.

[0035] With the help of the according Fig. With 1 trained AI, it is now possible to convert the dashed measurement curve 6 of the inaccurate ionization chamber into a measurement result of a highly accurate detector according to the solid line 5.

[0036] Based on the Fig. 3 and Fig. Section 4 describes a second application of the invention. Fig. Section 3 describes a second arrangement for training an AI. As in Fig. 1. Here too, control plan segments are initially generated, which are then used in the Fig. 3 are represented by the three aperture configurations 4.

[0037] In addition, artificial CT data is generated here using artificially created patients. The artificial patient can, for example, be randomly composed of different density distributions that have, for instance, different absorption rates.

[0038] First, a result is simulated using the control plan segment, which would have been determined in a measurement.

[0039] For a generated control plan segment and an artificial patient, i.e., 3D CT image 9, a radiation physics process is determined as a second result. In this example, a result that would have been measured on this artificial patient is simulated.

[0040] For example, a Monte Carlo simulation can be used for the simulation.

[0041] Below the two CT images 9 are in the Fig. Three diagrams of the two simulated measurement curves are shown.

[0042] Here too, the two curves form part of a data pair that is used to train the AI.

[0043] In this way, a large number of data pairs can be generated for a large number of different patients.

[0044] In the application after Fig. 4 will now be as in Fig. 2. A measuring device 8 is irradiated. With the help of AI, the measurement result can now be transferred to the real CT image of the real patient in order to determine the actual dose that would actually reach the patient.

[0045] In this way, a control plan segment can be easily validated.

[0046] As mentioned above, a control plan consists of individual control plan segments. The AI ​​is trained based on these individual control plan segments. Therefore, the AI ​​can only process individual control plan segments. A complete picture can only be created after processing the individual control plan segments by combining the individual processed results. Fig. Figure 5 shows a flowchart of a procedure for training an AI.

[0047] In a first step, a control plan segment of a control plan is generated. This control plan segment contains, for example, the position of the linear accelerator, the particle energy, the irradiation duration, and the aperture configuration. All individual parameters of the control plan segment can be generated randomly. In some applications, artificial patient data can also be generated. This artificial patient data can be generated, for example, by randomly concatenating density distributions. The artificial patient data can be in the form of artificial CT data (computed tomography data).

[0048] In a second step 21, a result is generated for the control plan segment created above by a representation of the second radiation physics process. This representation can, for example, be a simulation of the radiation physics process. In the examples mentioned above, this corresponds to the exact measured values ​​or the measured values ​​at the patient.

[0049] In a third step 22, a result for the control plan segment is generated by representing the first radiation physics process. This representation can also be generated by simulation or by converting the previously determined result of the second process. In the examples mentioned above, this corresponds to the result of the inaccurate detector or the measured values ​​at the phantom.

[0050] In a fourth step, the AI ​​is trained with many data pairs.

[0051] Typically, the preceding steps must be repeated for each data pair. A data pair then consists of the determined representations of the two radiation physics processes. In one execution, a data pair can also contain additional information. More than one data pair can be obtained from a single pair of representations of radiation physics processes in one execution.

[0052] The advantage now is that almost all data is artificially generated, and therefore a very large number of data pairs can be created to train the AI.

[0053] The Fig. Figure 6 shows a flowchart for converting a first result into a second result according to the invention.

[0054] In a first step, 31, an initial result of a radiation physics process is determined. This can be a measurement with an ionization chamber according to Fig. 2 or Fig. 4. Whereby the results are determined separately for each control plan segment.

[0055] In a second step 32, each of these first results of a control plan segment is converted into a second result by a according to Fig. Five trained, suitable AIs were selected. It is important to ensure that an AI is used that has been trained with data relevant to the application. For example, if a phantom measurement is to be performed after Fig. 4. In patient data, a [further step] must also be taken. Fig. Three trained AIs will be used.

[0056] Since all control plan segments are transferred separately, in a third step 33, all individual transferred results are combined into an overall result. Reference symbol list 1 linear accelerator 2 patient tables 3 particle beam 4 Aperture / Aperture configuration 5 measurement curve accurate 6. Measurement curve inaccurate 7 Ionization chamber 8 Water Phantom 9 CT data 10 Patient measurement curve 11 Measurement curve Phantom 12 AI 20-23 training steps 30-32 application steps

Claims

[1] Method for converting a result of a first radiation-physical process of a radiation source (1) into a result of a second radiation-physical process of the radiation source (1) by an artificial intelligence, AI, (12), wherein the AI (12) was initially trained with data pairs, each data pair representing a result of the first radiation-physical process (5) and a result of the second radiation-physical process (6). [2] Method according to claim 1, characterized by that the first and / or the second radiation-physical process represents an ideal result. [3] Method according to one of the preceding claims, characterized by that the second radiation-physical process depicts the irradiation of a patient, with the result representing a dose distribution in the patient. [4] Method according to one of the preceding claims, characterized bythat the first radiation-physical process depicts an error-prone reality of a result. [5] Method according to one of the preceding claims, characterized by that the first and / or the second radiation-physical process depicts the irradiation of a phantom, the result being a dose distribution in the phantom. [6] Method according to one of the preceding claims, characterized by that the first result and / or the second result was obtained by a measurement on a radiation source (1). [7] Method according to one of the preceding claims, characterized by that the AI (12) has at least one artificial neural network. [8] Method according to one of the preceding claims, characterized by that a result is a dose distribution or a detector image. [9] Method according to one of the preceding claims, characterized bythat the radiation source is designed to emit ionizing radiation or particle radiation, and / or is an X-ray source, a computer tomograph or a linear accelerator (1). [10] Method according to one of the preceding claims, characterized by that generating a data pair for training the AI involves the following steps: - generating (20) a control plan segment of a control plan, - generating (21) a first result for the control plan segment by representing the first radiation-physical process, - generating (22) a second result for the control plan segment by representing the second radiation-physical process, - Training (23) the AI with the first and second result as a data pair. [11] Method according to one of the preceding claims, characterized bythat a control plan segment comprises artificial measurement data, an artificial computed tomography, an artificial patient and / or an artificial irradiation plan. [12] Method according to one of the preceding claims, characterized by that a representation is generated by simulation, wherein the simulation is based on physical and / or stochastic principles, in particular wherein a Monte Carlo simulation is used for the simulation. [13] Method according to one of the preceding claims, characterized by that a representation is generated by converting the representation of the other radiation-physical process. [14] Method according to one of the preceding claims, characterized by that the transfer includes the following steps: - measuring (30) a first result for a real control plan segment, - Input (31) of the measurement result into a KI (12) for conversion into a second measurement result. [15] Method according to one of the preceding claims, characterized by that a control plan is divided into several control plan segments, wherein the transfer (31) is carried out for each control plan segment and the transferred results of the control plan segments are combined (32) to form an overall result. [16] Use of an AI for converting a dose distribution measured on a phantom into a dose distribution on a patient, in particular in a method according to one of the preceding claims, wherein the AI was initially trained with data pairs, each data pair comprising a simulated dose distribution on the phantom and a simulated dose distribution on the patient. [17] Use of an AI for improving the image resolution of an image generated by a detector, in particular in a method according to one of the preceding claims 1 to 15, wherein the AI was initially trained with data pairs, each data pair comprising a simulated image of a first, in particular inaccurate, detector and a simulated or calculated image of a second, in particular accurate, detector.

Images