Device for controlling a medical system by word input from a user

Abstract

The invention relates to a device (1) for controlling a medical technology system of a given type by word inputs from a user, the device (1) comprising: - a machine-learning LLM system (2) comprising a number of large language models (L), wherein the LLM system (2) is trained to derive probabilities (W) for desired settings of the medical-technical system after receiving a user intention (B) in the form of words as input data (E), and to output output data (A) based on the probabilities (W) for settings, - an input interface (3) designed to receive the user intention (B) in the form of words, and output input data (E) to the LLM system (2), - an output interface (4) designed to output output data (A) of the LLM system (2). Furthermore, the invention comprises a training method, a control method of a medical technology system, a control device and a medical technology system.

Description

[0001] The invention relates to a device for controlling a medical-technical system of a predetermined type by word inputs from a user, a training method for training an LLM system for such a device, a control method of a medical-technical system, a control device for a medical-technical system and a medical-technical system.

[0002] Medical technology systems are becoming increasingly complex and are gradually gaining many new functions and configuration options. This applies to both measurement systems, such as CT or MRI systems, and systems for evaluating measurement data, such as diagnostic systems or image reconstruction systems. This can lead to the situation where only a technical expert has an overview of the entire functionality of such a system, and a clinical user often only uses basic functions.

[0003] To date, there is no intuitive and easily understandable way to explain to a clinical user all the available functions of a complex medical technology system, such as a complex modular medical diagnostic software. In this regard, there is often insufficient time for intensive training, as it is usually not just the function itself that matters, but also its placement in a medical context. In short: for each setting option, it would be necessary to know which setting might be advantageous for which medical cases.

[0004] It is an object of the present invention to provide a device for controlling a medical technology system by word inputs from a user, a training method for training an LLM system for such a device, a control method for a medical technology system, a control device for a medical technology system, and a medical technology system that avoids the disadvantages described above. In particular, an object of the invention is to enable a translation of user intentions with regard to the functionality of a medical technology system and its application.

[0005] This object is achieved by a device according to patent claim 1, a training method according to patent claim 7, a control method of a medical technology system according to patent claim 11, a control device according to patent claim 12 and a medical technology system according to patent claim 13.

[0006] A device according to the invention serves to control a medical technology system of a given type through verbal input from a user. It comprises the following components: - a machine-learning LLM system comprising a number of large-language models, wherein the LLM system is trained to derive probabilities for desired settings of the medical-technical system after receiving a user intention in the form of words as input data and to output data based on the probabilities for settings, - an input interface designed to receive the user's intention in the form of words and output input data to the LLM system, - an output interface designed to output output data from the LLM system.

[0007] The device can be used to directly control the medical technology system by outputting control data or control commands for this system. The control data can also be data for presetting, and the control commands can be sent from a control unit of the system. For example, the device can compile a pulse sequence for an MRI system, and an MRI examination can be performed by the control unit of the MRI system using this pulse sequence.

[0008] The device operates with verbal input from a user. In a simple case, this could be readable text entered via a keyboard, but it is preferred that the device be configured to "understand" spoken words or convert them into machine-readable text.

[0009] The heart of the device is a machine-learning-capable LLM system (the abbreviation LLM stands for "large language model"). This term implies that the LLM system includes at least one large language model. It can consist of this LLM, but also of a combination of the LLM with other machine-learning models or other LLMs. Preferred architectures are described below.

[0010] Large-language models are already known in the state of the art, e.g., "Chat-GPT." A large-language model is characterized by its ability to achieve general language understanding and general language generation. An LLM is a weighted prediction model that can predict subsequent words or statements for a given text based on calculated probabilities.

[0011] The LLM system is specifically trained to derive probabilities for desired settings of the medical technology system after receiving the user's intention in the form of words (input data). This means that the LLM system interprets the input words and uses them to determine how the medical technology system should be configured to fulfill the user's intention.

[0012] The output data can be direct information about settings, e.g. control data or numerical values ​​for setting parameters of the medical technology system. They can also be direct control data or instructions for the medical technology system. It is also possible, however, that the output data are instructions for operating the medical technology system. Basically, the output data should contain information for a user or for the medical technology system that is based on the probabilities for settings of the medical technology system calculated by the LLM system. For a CT system, this could be data for setting the beam energy and the acquisition mode, for an MRI system a pulse sequence or information for a sequence of acquisition modes or contrasts.In the case of an image reconstruction unit, the output data could include information on computational operations or filters, and in the case of diagnostic systems, information on the display of images or an automated search for image elements.

[0013] It should be noted that the LLM system does not necessarily need to specify the derived settings of the medical device system separately (provided the settings are not part of the output data). The derived settings can also be used directly to generate the output data. It is only necessary that the LLM system must know the settings at some point during processing in order to generate the appropriate output data.

[0014] In order for the LLM system to receive input data and output data, the device comprises an input interface and an output interface.

[0015] The input interface is used to receive the user's intent. It can be a simple data interface for receiving written text, but it can also be highly complex, allowing it to understand spoken text. Such input interfaces are already known in the state of the art and are already used in households, for example, in mobile phones or consumer electronics. Essentially, all that is important is that the input interface provides the user's intent to the LLM system in an input data format with which it has been trained.

[0016] The output interface is another data interface that outputs the LLM system's output data. It is preferably designed to send the output data directly to a medical device system of the specified type, e.g., setting data, control data, or control commands. Alternatively, it is designed to send the output data to a user, e.g., instructions for operation or settings.

[0017] The device can be used for all possible medical technology systems, but must then be tailored to the system in question. If it is tailored to a specific type of medical technology system, then it can be used for all medical technology systems of that type (i.e., with the same functions). This can be achieved by specifically training the LLM system on medical technology systems of that type. The device is particularly advantageous for medical technology measurement systems, e.g., MRI systems, CT systems, PET systems, X-ray systems, tomosynthesis systems, ultrasound systems, endoscopy systems, or for image reconstruction systems or diagnostic systems.

[0018] A training method according to the invention serves to train an LLM system for a device according to the invention. It comprises the following steps: a) providing a plurality of training data sets, wherein a training data set comprises a user intention in the form of words for a specific use of a medical technology system as input data and data on settings of the medical technology system for this specific use as ground truth data sets, b) Entering the input data into the LLM system and setting parameters of the LLM system based on the respective ground truth data set, c) Repeat step b) with the provided training data sets.

[0019] In principle, training methods for LLMs are known in the state of the art. However, training an LLM specifically for a specific task depends on the training datasets used. Each training dataset includes input data that is fed into the LLM to be trained and output data associated with the input data that reflects the desired output, i.e., a ground truth. For clarity, the data that specifies the desired output data is referred to below as the "ground truth dataset."

[0020] As part of the training process, a variety of training datasets are provided. The input data can come from everyday clinical practice, such as log files, user input, examination reports, or recorded speech samples from a clinical user.

[0021] The ground truth dataset can include, for example, data from user manuals, interface documentation (software and / or hardware), system documentation, software source code, configuration documentation, parameter documentation, fed scripting, or even log files. Importantly, the training dataset includes a user intent in the form of words for a specific use of a medical device system as input data, and the ground truth dataset includes data on the medical device system's settings for this specific use.

[0022] For training, the input data from the training datasets is then fed sequentially into the LLM system. This then produces output data, which is then compared with the corresponding ground-truth dataset. The internal parameters of the LLM system are then adjusted based on this comparison. This can be done multiple times, with the same input data being entered each time and parameters being adjusted until the output data matches the ground-truth dataset to a certain degree.

[0023] This step is then repeated with all training datasets. It is certainly possible to process training datasets again in a modified form, as described in more detail below.

[0024] It is certainly possible to use an already pre-trained LLM (with a general vocabulary) in the LLM system.

[0025] This has the advantage that even general sentence structures can be recognized, which simplifies training. A control method according to the invention for a medical technology system with a device according to the invention comprises the following steps: - Entering words into the device's input interface, - Forwarding the input data of the input interface to the LLM system, whereby the LLM system is designed to generate control data for the medical technology system from input data, - Output of the control data generated by the LLM system via the output interface.

[0026] With a sufficiently trained LLM system in the device, the steps after input can be carried out fully automatically.

[0027] A control device according to the invention for a medical technology system comprises a device according to the invention.

[0028] A medical-technical system according to the invention comprises a control device according to the invention and is designed in particular for medical-technical measurements and / or medical-technical examinations. It is preferably a radiological system, an image reconstruction system, or a diagnostic system.

[0029] The invention can be implemented in particular in the form of a computer unit with suitable software. For this purpose, the computer unit can, for example, have one or more cooperating microprocessors or the like. In particular, it can be implemented in the form of suitable software program parts in the computer unit. A largely software-based implementation has the advantage that even computer units already in use can be easily retrofitted with a software or firmware update in order to operate in the manner according to the invention. In this respect, the object is also achieved by a corresponding computer program product with a computer program that can be loaded directly into a memory device of a computer unit, with program sections in order to carry out all steps of the method according to the invention when the program is executed in the computer unit.Such a computer program product may, in addition to the computer program, include additional components such as documentation and / or additional components, including hardware components such as hardware keys (dongles, etc.) for using the software.

[0030] A computer-readable medium, such as a memory stick, a hard disk or another portable or permanently installed data storage device, on which the program sections of the computer program that can be read and executed by a computer unit are stored, can be used for transport to the computer unit and / or for storage on or in the computer unit.

[0031] Further, particularly advantageous embodiments and developments of the invention emerge from the dependent claims and the following description, wherein the claims of one claim category can also be developed analogously to the claims and description parts to form another claim category and, in particular, individual features of different embodiments or variants can be combined to form new embodiments or variants.

[0032] A preferred device additionally comprises a checking unit which is designed to check whether the highest setting probability differs from the other setting probabilities by more than a predetermined threshold. If this is not the case, a request for further information is issued to the user. This function is very advantageous in complex medical technology systems, since for an expressed user intention there could possibly be several possible setting options which differ fundamentally from one another. The fact that there are several is recognized by the fact that not a single setting stands out from the others due to its probability, but rather several settings are probable. The request for further information can, in the simplest case, be "I need more information".In a preferred complex device, the LLM system may issue a query specifically aimed at filtering the settings that may be considered.

[0033] For this purpose, the LLM system preferably additionally comprises a learning model, in particular a large language model, which has been trained to output a request for further information to a user from a plurality of predefined settings of the medical technology system, preferably in the form of words. The LLM can be the same as described above, i.e., the one that also outputs the setting options. However, it is preferred that a different LLM be used, one that has been specifically trained to output a query formulated in words when certain possible settings are present.

[0034] For example, the request "Take an MRI scan" is very general, and many settings are possible. An initial query might be "Which body area should be scanned?" Another might be "Which contrasts should be acquired?" or "What body area should be examined for?" However, it might also ask whether the same examination should be performed as last time.

[0035] A preferred device is characterized in that the LLM system additionally comprises a learning model, in particular a large language model, which has been trained to output instructions on how to operate an input interface of a medical technology system of the specified type so that it is controlled according to the specified settings. Thus, a user is explained what settings should be made in the system and how. This is particularly advantageous for systems that cannot be controlled externally or can only be controlled with difficulty.

[0036] A preferred device is characterized in that the LLM system additionally comprises a learning model, in particular a large language model, which has been trained to output control data for controlling a medical technology system of the specified type. This allows the system to be controlled directly and automatically using commands or, for example, to provide an MRI system with a pulse sequence for an examination.

[0037] A preferred device is characterized in that the LLM system additionally comprises a learning model, in particular a large language model, which has been trained to output setting data for presetting a medical technology system of the specified type. This allows a suitable presetting to be made automatically, and a user can operate the system with a standardized control system using the specific presettings.

[0038] A preferred device is characterized in that the LLM system has been trained with medical-technical texts and / or examination instructions as input data and a ground truth based on documentation of the specified type of medical-technical system (possibly domain-specific). The medical-technical texts are preferably texts from medical-technical textbooks and / or clinical guidelines. The documentation is preferably at least one document from the group of operating instructions, manuals, software interface documentation, configuration documentation, parameter documentation, API documentation, Fed scripting, log files, and software source code and relates to the specified type of medical-technical system.

[0039] In a preferred embodiment, the large language model of the LLM system includes an output vocabulary corresponding to the settings of the medical device system of the specified type. This does not necessarily mean words of a human language, but can also be abstract parameters, states, commands, or values.

[0040] It should be noted that an LLM is designed to calculate probabilities for settings. It does not perform an objective test to determine whether these settings are semantically or technically plausible. Furthermore, while the determined settings for frequently occurring standard situations are often correct, settings for rare special cases can be subject to frequent errors. If the output vocabulary is reduced to the settings available in the system, potential sources of error are also reduced. Furthermore, the entire system can be trained more specifically. The LLM system preferably uses a predefined medical technology vocabulary as input vocabulary. This also serves to reduce potential errors.

[0041] Preferably, the LLM system includes an additional large-language model that uses the specified medical vocabulary as the output vocabulary for queries. The preferred option for a query to the user was already described above. This, too, can contain errors. Using the specific medical vocabulary when generating the query can eliminate errors that might occur with larger, more general vocabularies.

[0042] The LLM system preferably comprises an additional large-language model, which includes a specific output vocabulary containing setting data, control data, or control commands of the medical-technical system, in particular system-understandable commands or setting data. If the settings are known to the LLM system or were selected based on calculated probabilities, it is not necessarily guaranteed that a medical-technical system will understand automatically generated control commands or presetting data. The LLM described here is specifically designed to convert the settings determined by the LLM system into output data that the medical-technical system understands. To this end, the additional LLM specifically uses system-understandable commands or setting data.

[0043] In general, it is preferred that the LLM system has multiple LLMs, with different large language models having different output vocabularies.

[0044] A preferred device is characterized by the fact that the large language model of the LLM system is designed to generate a feature vector as an intermediate form of the user's intent. Therefore, no settings are determined here, but rather a general indication of what the user actually wants is provided. This could be supported by "intermediate training," in which the user intent is used as input data in the form of supervised training, and a semantic interpretation of the user intent is used as ground truth.

[0045] This approach has the advantage that the LLM system can now be trained on a variety of medical device types. The user can specify that they want to perform an MRI scan or request an image analysis, and the LLM system reflects this in the feature vector.

[0046] The preferred LLM system in this regard has an architecture in which additional machine-learning models, particularly large-language models, access this feature vector. These models can be specifically trained for individual types of medical technology systems using the method described above (if necessary together with the LLM, which generates the feature vector). For example, one model is trained to control an image reconstruction unit, one to preset a CT system, another to generate pulse sequences for an MRI system, and yet another to output requests to a user. Each of these models accesses the feature vector and determines whether it needs to be activated at all, and if so, which settings need to be made.

[0047] It should be noted that the verification unit mentioned above can access the feature vector directly. If multiple possibilities for a user's intention already exist, it is likely that there will be multiple setting options later. Therefore, the verification unit preferentially derives the need for a query from the feature vector by checking whether there are multiple possible interpretations of the user's intention. This can be derived from the probabilities of the feature vector by checking whether the entry with the highest probability differs from the probabilities of the other entries by a specified value.

[0048] This allows for simple training of type-specific models, and any additional models for other medical technology systems can also be integrated and trained. Each of these models has, of course, been trained to generate probabilities for desired settings for a specific medical technology system and / or user requests from the feature vector.

[0049] According to a preferred training method, the LLM system is trained with a plurality of audit logs from the same type of medical technology system. Audit logs are known in the art and include a plurality of system parameter values. In particular, the audit logs include user inputs and parameters related to medical technology processes, in particular measurements and / or examinations.

[0050] User inputs are used as input data, and the parameters (relevant to the settings) serve as ground truth for monitoring the output of the LLM system. The audit logs can provide large amounts of training data sets.

[0051] During a training process, the LLM system is preferably provided with parts from ground-truth datasets. The LLM system is essentially helped to recognize the correct settings by being given some of them beforehand. The training is then designed so that the LLM system supplements further elements of the corresponding ground-truth dataset. Therefore, especially at the beginning, it no longer has to "guess" all of the settings, but only some. Preferably, changing parts from a ground-truth dataset (a training dataset) are provided several times, and the LLM system is given reinforcing feedback when it suggests the correct parts of the ground-truth dataset as additions.

[0052] For example, if a ground truth dataset contains the values ​​A, B and C for three settings in addition to the input data, the training dataset could be used for multiple training runs. In each case, the input data is entered and the output is compared with the ground truth dataset. However, in this particular training run, A and B are already given in one training run and only C needs to be guessed. In other training runs, A and C are given (and B is guessed) and B and C are given (and A is guessed). Then only A or B or C could be given in each case and the missing values ​​can be guessed. Finally, all three values ​​could be guessed. In this way, one training dataset can be used for multiple training runs.

[0053] However, a training data set with several different input data can also be used multiple times for training. In a preferred training method, a large number of input data is generated for a training dataset using a ground-truth dataset. The ground-truth dataset represents a user intent or a specific application of the medical technology system. However, this user intent can be formulated in different ways. Therefore, a large number of different input data is generated in the form of different descriptions for the respective application. This can be done, in particular, using a large language model that has been trained to vary descriptions. This allows multiple training datasets to be created from the single training dataset and used for training.

[0054] For example, anatomy-specific or disease-specific scan sequences or AI algorithm analyses based on specific characteristics are created as ground truth. Then, for example, scripts are hand-crafted to fulfill important higher-level intents. From these, a large number of additional scripts are then machine-generated from the blueprints to create a large, potentially enormous training set for the domain-specific large language model. The user intent is paired with the corresponding set of script variants and used as training sets.

[0055] Components of the invention are preferably present as a “cloud service.” Such a cloud service is used to process data, in particular by means of artificial intelligence, but can also be a service based on conventional algorithms or a service in which human evaluation takes place in the background. In general, a cloud service (hereinafter also referred to as “cloud”) is an IT infrastructure in which, for example, storage space or computing power and / or application software is made available via a network. Communication between the user and the cloud takes place via data interfaces and / or data transmission protocols. In the present case, it is particularly preferred that the cloud service provides both computing power and application software.

[0056] Within the scope of a preferred method, data obtained within the scope of the invention is provided to the cloud service via the network. This cloud service comprises a computing system that generally does not include the user's local computer. The method can be implemented using a command constellation in a network. The data calculated in the cloud is later sent back to the user's local computer via the network.

[0057] The invention is explained in more detail below with reference to the accompanying figures using exemplary embodiments. In the various figures, identical components are provided with identical reference numerals. The figures are generally not to scale. They show: Fig. 1 a device for controlling a medical technology system by word inputs of a user, Fig. 2 a device for controlling three medical technology systems, Fig. 3 a training method for training an LLM system for a device according to the invention.

[0058] Fig. Figure 1 shows a device 1 for controlling a medical technology system S2 of a given type using verbal input from a user. The device 1 comprises a machine-learning-capable LLM system 2, an input interface 3, and an output interface 4.

[0059] The user (left) expresses a user intention B in the form of words (indicated by a speech bubble). A specific examination is to be performed using the MRI system S2.

[0060] The input interface 3 is used to receive the spoken user intention B. It is designed to convert the user's words into a text that acts as input data to the LLM system 2.

[0061] The LLM system 2 could simply consist of a large-language model L. However, in this example, it is more complex. First, it comprises a large-language model 2, into which the input data is fed and processed into a feature vector V. The feature vector V represents an intermediate form that reflects the user's intent, which has been derived from the words of user intent B by the large-language model L.

[0062] Another machine-learning model 7 directly accesses this feature vector V. It was trained to generate probabilities W for desired settings of the medical system S2 in the form of control commands S from the feature vector V.

[0063] The device also includes a verification unit 5, which is designed to check whether there are multiple possible types of settings. This is already derived from the feature vector V by checking whether there are multiple probable interpretations of the user's intention B.

[0064] If there are several possible interpretations, the LLM system 2 includes an additional machine-learning model 6, which has been trained to output a request F for further information to the user in the form of words from a plurality of predefined options for settings of the medical technology system S2.

[0065] The output interface 4 is used to output output data A of the LLM system 1 to the medical technology system S2 and, if necessary, to output a request F to the user.

[0066] Fig. Figure 2 shows a device for controlling three medical systems S1, S2, S3, namely an image reconstruction system S1, an MRI system S2 and a system for automated interventions S3. For this purpose, the device 1 can be Fig. 1 can be modified by having two additional machine-learning models, which have been specifically trained for the other medical technology systems S1, S3, access the feature vector V.

[0067] Fig. 3 shows a training procedure for training an LLM system 1 as it is e.g. in Fig. 1 is shown.

[0068] In step I, a plurality of training data sets T is provided. Each training data set T comprises a user intention B in the form of words, e.g. in the form of a text block, for a specific use of a medical technology system S1, S2, S3 (see Fig.2). In addition, a training dataset T includes data D on settings of the medical technology system S1, S2, S3 for this specific use as ground truth datasets D. For example, audit logs from the same type of medical technology system can be used for training data. A training dataset T can be varied and used multiple times, e.g., by providing a constantly changing portion of the ground-truth dataset D as guidance or by representing the user intent B multiple times, each time using different words.

[0069] In step II, the user intention B is input as input data E into the LLM system 2 of the device 1, and its output data A is compared with the respective ground truth data set D. Based on this comparison, parameters of the LLM system 2 are adjusted.

[0070] Step II is then repeated with the provided training data T.

[0071] Finally, it should be noted once again that the invention described in detail above merely represents exemplary embodiments which can be modified in a variety of ways by a person skilled in the art without departing from the scope of the invention. Furthermore, the use of the indefinite articles “a” or “an” does not exclude the possibility that the features in question may be present in multiple instances. Likewise, terms such as “unit” do not exclude the possibility that the components in question consist of several interacting sub-components which may also be spatially distributed. The term “a number” should be read as “at least one”. Regardless of the grammatical gender of a particular term, this includes persons with male, female or other gender identities.

Claims

[1] Device (1) for controlling a medical-technical system (S1, S2, S3) of a predetermined type by word inputs from a user, the device (1) comprising: - a machine-learning LLM system (2) comprising a number of large language models (L), wherein the LLM system (2) is trained to derive probabilities (W) for desired settings of the medical-technical system (S1, S2, S3) after receiving a user intention (B) in the form of words as input data (E), and to output output data (A) based on the probabilities (W) for settings, - an input interface (3) designed to receive the user intention (B) in the form of words, and output input data (E) to the LLM system (2), - an output interface (4) designed to output output data (A) of the LLM system (2). [2] Device according to claim 1, comprising a checking unit (5) which is designed to check whether the highest probability (W) for settings differs by more than a predetermined limit value (G) from the other probabilities (W) for settings, and if not, to output a request for further information to the user, preferably wherein the LLM system (2) additionally comprises a learning model (6), in particular a large language model, which has been trained to output a request (F) for further information to a user from a plurality of predetermined possibilities for settings of the medical technology system (S1, S2, S3), preferably in the form of words. [3] Device according to one of the preceding claims, wherein the LLM system (2) additionally comprises a machine-learning model (7), in particular a large language model, which has been trained to - to issue instructions on how an input interface of a medical technology system (S1, S2, S3) of the specified type must be operated so that it is controlled according to the specified settings and / or - to output control data (S) for controlling a medical technology system (S1, S2, S3) of the specified type, - Output setting data for presetting a medical technology system (S1, S2, S3) of the specified type. [4] Device according to one of the preceding claims, wherein the LLM system (2) has been trained with medical-technical texts and / or instructions for examinations as input data (E) and a ground truth based on documentation of the predetermined type of medical-technical system (S1, S2, S3), preferably, the medical-technical texts are texts from medical-technical textbooks and / or clinical guidelines and / or Preferably, the documentation comprises at least one document from the group of operating instructions, manual, software interface documentation, configuration documentation, parameter documentation, API documentation, Fed scripting, log files and software source code relating to the specified type of medical technology system (S1, S2, S3). [5] Device according to one of the preceding claims, wherein the large language model (L) of the LLM system (2) comprises an output vocabulary corresponding to the settings of the medical-technical system (S1, S2, S3) of the predetermined type, preferably wherein the LLM system (2) comprises a predetermined medical-technical vocabulary as input vocabulary, preferably wherein the LLM system comprises a further large language model (L) which comprises the predetermined medical-technical vocabulary as output vocabulary for queries, and / or which comprises a specific output vocabulary which comprises setting data, control data or control commands (S) of the medical-technical system (S1, S2, S3). [6] Device according to one of the preceding claims, wherein the large language model (L) of the LLM system (2) is designed to generate a feature vector (V) as an intermediate form of the user's intention, and a plurality of further machine-learning models (6, 7), in particular large language models (L), which have been trained to generate probabilities (W) for desired settings for different medical technology systems (S1, S2, S3) and / or requests (F) to a user from the feature vector (V). [7] Training method for training an LLM system (2) for a device (1) according to one of the preceding claims, comprising the steps: a) Providing a plurality of training data sets (T), wherein a training data set (T) comprises a user intention (B) in the form of words for a specific use of a medical technology system (S1, S2, S3) as input data (E) and data (D) on settings of the medical technology system (S1, S2, S3) for this specific use as ground truth data sets (D), b) Input of the input data (E) into the LLM system (2) and setting of parameters of the LLM system (2) based on the respective ground truth data set (D), c) Repeat step b) with the provided training data sets (T). [8] Training method according to claim 7, wherein the LLM system (2) is trained with a plurality of audit logs of the same type of medical system (S1, S2, S3), wherein the audit logs comprise user inputs and parameters relating to medical processes, in particular measurements and / or examinations, and wherein the user inputs are used as input data (E) and the parameters serve as ground truth for monitoring the output of the LLM system (2). [9] Training method according to claim 7 or 8, wherein parts from ground truth data sets (D) are specified to the LLM system (2) and the training is designed so that the LLM system (2) supplements further elements of the corresponding ground truth data set (D), wherein changing parts are specified several times from a ground truth data set (D) and the LLM system (2) is given reinforcing feedback when it suggests the correct parts of the ground truth data set (D) as a supplement. [10] Training method according to one of claims 7 to 9, wherein for a training data set with a ground truth data set (D) for this ground truth data set (D) a user intention (B) is generated as a plurality of different input data (E) in the form of different paraphrases for the application in question, in particular by means of a large language model which has been trained to vary paraphrases. [11] Control method of a medical technology system (S1, S2, S3) with a device (1) according to one of claims 1 to 6, comprising the steps: - entering words into the input interface (3) of the device (1), - forwarding the input data (E) of the input interface (3) to the LLM system (2), wherein the LLM system (2) is designed to generate control data (S) for the medical technology system (S1, S2, S3) from input data (E), - Output of the control data (S) generated by the LLM system (2) via the output interface (4). [12] Control device for a medical technology system (S1, S2, S3) comprising a device (1) according to one of claims 1 to 6. [13] Medical-technical system (S1, S2, S3) comprising a control device according to claim 12, wherein the medical-technical system is preferably a radiological system (S2), an image reconstruction system (S1) or a diagnostic system and is designed in particular for medical-technical measurements and / or medical-technical examinations. [14] A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the training method according to any one of claims 7 to 10. [15] A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the training method according to any one of claims 7 to 10.

Images