Large language model generation of lexical pulse sequence data

Abstract

A medical system (100, 300, 500) is disclosed that includes a memory (110) storing machine executable instructions (120) and a domain-specific large language model (122). The large language model is configured to output lexical pulse sequence data (126) in response to receiving object data (124). The lexical pulse sequence data is configured as lexical elements selected from a library of pulse sequence lexical elements (322). The medical system also includes a computing system (104). Execution of the machine executable instructions causes the computing system to receive (200) lexical pulse sequence data in response to inputting object data into a domain-specific large language model.

Description

TECHNICAL FIELD

[0001] The present invention relates to magnetic resonance imaging, in particular to automatic generation of pulse sequence data for configuring a magnetic resonance imaging system. BACKGROUND

[0002] As part of a process for generating images of a patient’s body, a magnetic resonance imaging (MRI) system or scanner uses a large static magnetic field to align the nuclear spins of atoms. This large static magnetic field is referred to as the B0 field or main magnetic field. By controlling gradient magnetic fields and radio frequency pulses, the nuclear spins can be manipulated to produce radio frequency signals that can be sampled or measured as k-space data. The k-space data can then be reconstructed into a magnetic resonance image that can image the internal anatomy of a subject. The time-dependent control of the radio frequency signals and radio frequency pulses and the sampling of k-space data are described in a pulse sequence setup. Correctly configuring a magnetic resonance imaging system requires a lot of training and experience. SUMMARY

[0003] The invention provides in independent claims a medical system, a method, a computer program, a method of training a domain-specific large language model, and a domain-specific large language model. Embodiments are given in dependent claims.

[0004] In one aspect, a medical system is disclosed. The medical system comprises a memory storing machine executable instructions and a domain-specific large language model. The domain-specific large language model is configured to output tokenized pulse sequence data in response to receiving subject data. The tokenized pulse sequence data is configured as a token selected from a library of pulse sequence tokens. The medical system further comprises a computing system. Execution of the machine executable instructions causes the computing system to receive the tokenized pulse sequence data in response to inputting the subject data into the domain-specific large language model.

[0005] In another aspect, a method is disclosed. The method comprises receiving tokenized pulse sequence data in response to inputting subject data into a domain-specific large language model. The domain-specific large language model is configured to output the tokenized pulse sequence data in response to receiving the subject data. The tokenized pulse sequence data is configured as a token selected from a library of pulse sequence tokens.

[0006] In another aspect, a computer program is disclosed. The computer program comprises machine executable instructions and a domain-specific large language model, wherein both are configured to be executed by a computing system. The large language model is configured to output tokenized pulse sequence data in response to receiving object data. Execution of the machine executable instructions causes the computing system to receive the tokenized pulse sequence data in response to inputting the object data into the large language model. The tokenized pulse sequence data is configured to select tokens from a pulse sequence token library.

[0007] In another aspect, a method of training a domain-specific large language model is disclosed. The method comprises receiving a base large language model. The method further comprises providing a pre-trained large language model by training the base large language model with medical documents containing pulse sequence command specifications. Providing the pre-trained large language model comprises generating a preliminary tokenized pulse sequence library. The method further comprises receiving site-specific medical records. The site-specific medical records comprise object medical data paired with historical pulse sequence commands. The method further comprises generating a pulse sequence token library by excluding portions of the preliminary tokenized pulse sequence library that are not present in the site-specific medical records. The method further comprises providing a domain-specific large language model by training the pre-trained large language model using the site-specific medical records. Training the pre-trained large language model comprises at least partially tokenizing the site-specific medical records with the pulse sequence token library.

[0008] In another aspect, a domain-specific large language model trained according to the method of training a domain-specific large language model is disclosed.

[0009] In some examples, the domain-specific large language model can be a data structure. BRIEF DESCRIPTION OF DRAWINGS

[0010] In the following, the preferred embodiments of the present application will be described by way of example only and with reference to the accompanying drawings, in which:

[0011] Figure 1 An example of a medical system is illustrated.

[0012] Figure 2 is a flowchart illustrating a method of a medical system using Figure 1 .

[0013] Figure 3 An example of a medical system is illustrated.

[0014] Figure 4 is a flowchart illustrating a method of a medical system using Figure 3 .

[0015] Figure 5 An example of a medical system is illustrated.

[0016] Figure 6 is a flowchart illustrating a method of a medical system using Figure 5 a domain-specific large language model.

[0017] Figure 7 A method of using a domain-specific large language model is illustrated.

[0018] Figure 8 A method of training a domain-specific large language model is illustrated.

[0019] Figure 9 An example of a graphical user interface is illustrated. List of Reference Signs 100 medical system 102 computer 104 computing system 106 hardware interface 108 user interface 110 memory 120 machine-executable instructions 122 domain-specific large language model 124 object data 126 tokenized pulse sequence data 128 pulse sequence template database 130 database query 132 selected pulse sequence template 134 pulse sequence command 200 receiving, in response to inputting object data into a domain-specific large language model, tokenized pulse sequence data 202 generating, using a pulse sequence token library, a pulse sequence command from the tokenized pulse sequence data 204 receiving, by querying a pulse sequence template database with at least a portion of the tokenized pulse sequence data, a selected pulse sequence template 206 generating, by inputting the tokenized pulse sequence data into the selected pulse sequence template, a pulse sequence command 300 medical system 302 magnetic resonance imaging system 304 magnet 306 bore of the magnet 308 imaging zone 309 field of view 310 magnetic field gradient coil 312 magnetic field gradient coil power supply 314 radio frequency coil 316 transceiver 318 object 320 object support 322 pulse sequence token library 324 k-space data 326 magnetic resonance image 400 acquiring k-space data by controlling a magnetic resonance imaging system with pulse sequence commands 500 medical system 520 thought chain data 522 translated pulse sequence data 524 selected magnetic resonance imaging protocol 526 decision module 600 receiving thought chain data from a domain-specific large language model upon receiving tokenized pulse sequence data 602 providing translated pulse sequence data by translating the tokenized pulse sequence data using the pulse sequence token library as text associated with a respective plurality of alternative magnetic resonance imaging protocols 604 providing the thought chain data for the respective plurality of alternative magnetic resonance imaging protocols 606 receiving a selection of a selected magnetic resonance imaging protocol in response to providing translated pulse sequence data as text associated with a respective plurality of alternative magnetic resonance imaging protocols and as a plurality of alternative magnetic resonance imaging protocols and in response to providing respective thought chain data, wherein the pulse sequence commands are generated for the selected magnetic resonance imaging protocol 700 begin 702 retrieve patient details, medical history, medical reports 704 retrieve guideline information and written requests 706 build input prompts for domain-specific large language model 708 execute domain-specific large language prompt completion 710 extract tokenized pulse sequence data from output of domain-specific large language model 712 extract thought chain data from output of domain-specific large language model 714 current reasoning (thought chain data) and protocol selection from tokenized pulse sequence data 716 multiple protocol suggestions? 718 let user or algorithm select protocol 720 is user satisfied with protocol? 722 send protocol definition to magnetic resonance imaging scanner 724 end 726 Request user to input prompt to add 800 Pre-training portion 802 Fine-tuning portion 804 Optional thought chain prompt engineering portion 806 Base large language model 808 Medical document 810 Pre-trained model 812 Site-specific medical record 814 Hospital’s own patient records and MRI sequences 816 Domain-specific large language model 900 Interactive API 902 Button for retrieving medical records 904 Record summary with highlights 906 Button for retrieving radiology information system (RIS) data 908 Additional text input area 910 Protocol chatbot 912 Selector for protocol 914 Share button DETAILED DESCRIPTION

[0020] In these drawings, like-numbered elements are equivalents or perform similar functions. If the function is equivalent, an element discussed in a later drawing can not necessarily be discussed in a previous drawing.

[0021] In an example, the medical system includes a memory that stores machine executable instructions and a domain-specific large language model. As used herein, a domain-specific large language model encompasses a large language model. The label domain-specific is to indicate that the large language model has been trained for a specific use.

[0022] A large language model (LLM) as used herein encompasses a neural network architecture, typically consisting of transformers (encoders and decoders) with self-attention layers and residual connections, which have been trained using unlabelled text using self-supervised or semi-supervised learning. Typically, LLMs are trained using billions of words. For example, LLMs can be trained in an autoregressive fashion, where given a piece of text, the model can predict the next word (token) or words (tokens). Another training mode is where a word or token in a sentence is missing and then the LLM predicts the missing word or token. Both types of LLMs can be configured to be used in a so-called prompt paradigm, where a text query or statement is input into the LLM and the LLM outputs a completion or sentence. The LLMs described herein are configured to operate in the prompt paradigm. Example LLMs are GPT-3, GPT-4, BERT, LLaMA, etc. LLMs can be trained for specific tasks using reinforcement learning or reinforcement learning through human feedback (RLHF). The output of an existing LLM can be adjusted using fine-tuning. In fine-tuning, a new set of weights can be trained on specific data that connects the last layers of the language model. Typically, this is done by freezing other weights in the neural network (except the final output layer), so that only the final output and format are affected.

[0023] The domain-specific large language model is configured to output tokenized pulse sequence data in response to receiving object data. The tokenized pulse sequence data contains pulse sequence parameters and / or specifications encoded in a numerical form. The object data comprises data or metadata describing an object. The tokenized pulse sequence data can be converted into a language description of a pulse sequence protocol or directly into parameters for configuring a set of pulse sequence commands for acquiring k-space data.

[0024] The tokenized pulse sequence data is configured to be a token selected from a pulse sequence token library. The pulse sequence token library can vary in different examples. In one case, the pulse sequence token library can contain identifiers of specific sequence protocols or identifiers of individual parameters for customizing a pulse sequence. The domain-specific large language model has been configured to output tokenized pulse sequence data selected from this pulse sequence token library. Thus, it will not output arbitrary descriptions or pulse sequence parameters but is limited to descriptions or parameters in the pulse sequence token library. This can be, for example, a method of controlling the output of the domain-specific large language model to match the output of one or more specific clinical sites.

[0025] The medical system also includes a computing system. Execution of the machine executable instructions causes the computing system to receive tokenized pulse sequence data in response to inputting the subject data into the domain-specific large language model. This example can be beneficial because it can provide a means of generating tokenized pulse sequence data that is tailored to the particular subject data. The restriction of selecting the tokenized pulse sequence data from the pulse sequence token library can have the advantage of forcing the domain-specific large language model to provide valid tokenized pulse sequence data that can be useful in a clinical setting.

[0026] In another example, execution of the machine executable instructions also causes the computing system to generate pulse sequence commands from the tokenized pulse sequence data using the pulse sequence token library. The pulse sequence commands are configured to control a magnetic resonance imaging system to acquire k-space data. This example can be beneficial because it can provide a method of fully automating magnetic resonance imaging system operation. For example, it can be used to automatically generate pulse sequence commands or can be used as an assistant to a magnetic resonance imaging system operator.

[0027] In another example, execution of the machine executable instructions also causes the computing system to display the pulse sequence commands using a user interface. Execution of the machine executable instructions also causes the computing system to receive pulse sequence command modification data in response to displaying the pulse sequence commands using the user interface. Execution of the machine executable instructions also causes the computing system to modify the pulse sequence commands using the pulse sequence modification data. Execution of the machine executable instructions also causes the computing system to train the domain-specific large language model using the subject data and the modified pulse sequence commands.

[0028] In this example, two things can be achieved. First, a method or assistant for generating pulse sequence commands in a particular clinical setting is provided. The reception of pulse sequence command modification data can for example help to fine-tune or adjust the pulse sequence commands. The pulse sequence command modification data can for example be received from an operator or from additional algorithms or machine learning components. The modified pulse sequence commands and the subject data can also be fed into the domain-specific large language model to provide additional training. As the pulse sequence commands are updated, they are used to improve the performance of the domain-specific large language model in generating tokenized pulse sequence data.

[0029] In another example, the medical system also includes a magnetic resonance imaging system. Execution of the machine executable instructions also causes the computing system to acquire the k-space data by controlling the magnetic resonance imaging system with the pulse sequence commands. This example can be beneficial because it can provide a method of automating or assisting a magnetic resonance imaging system in acquiring k-space data.

[0030] In another example, the medical system further includes a pulse sequence template database storing pulse sequence templates each configured to generate a pulse sequence command. The pulse sequence templates may, for example, be pulse sequences for a particular magnetic resonance imaging protocol with typical values inserted. In other examples, the templates can be algorithms for generating a pulse sequence command upon receiving values for specified pulse sequence parameters. Execution of the machine executable instructions further cause the computing system to receive a selected pulse sequence template by querying the pulse sequence template database with at least a portion of the tokenized pulse sequence data. This query can be made in different ways in different examples. In one example, the tokenized pulse sequence data can contain an identifier for selecting the selected pulse sequence template. In other examples, the tokenized pulse sequence data can have a particular format or sequence of commands specified in the tokenized pulse sequence database that is recognized and then used to select the selected pulse sequence template. Execution of the machine executable instructions further cause the computing system to generate the pulse sequence command by inputting the tokenized pulse sequence data into the selected pulse sequence template. This can include using the tokenized pulse sequence data to modify various pulse sequence parameters of the selected pulse sequence template.

[0031] In another example, the domain-specific large language model is further configured to output thought chain data describing the tokenized pulse sequence data. The thought chain data is data for explaining or providing information about why the large language model had a particular output. This is known in the art and is advantageous for improving the quality of the domain-specific large language model. It can also be used by a human or other algorithm to evaluate the various choices of the domain-specific large language model output. Execution of the machine executable instructions further cause the computing system to receive the thought chain data from the domain-specific large language model upon receiving the tokenized pulse sequence data.

[0032] In another example, the tokenized pulse sequence data specifies a plurality of alternative magnetic resonance imaging protocols. The domain-specific large language model is configured to provide the thought chain data for each of the plurality of alternative magnetic resonance imaging protocols. Execution of the machine executable instructions further cause the computing system to provide translated pulse sequence data as text associated with the respective plurality of alternative magnetic resonance imaging protocols by translating the tokenized pulse sequence data using the pulse sequence token library. Execution of the machine executable instructions further cause the computing system to provide thought chain data for the respective plurality of alternative magnetic resonance imaging protocols.

[0033] Execution of the machine executable instructions further cause the computing system to receive a selection of a selected magnetic resonance imaging protocol in response to providing the translated pulse sequence data as text associated with a respective plurality of alternative magnetic resonance imaging protocols and in response to providing respective thought chain data for each of a plurality of alternative magnetic resonance imaging protocols. Generating the pulse sequence commands for the selected magnetic resonance imaging protocol can be beneficial in this example because it can provide a way to select a selected magnetic resonance imaging protocol from a large number of magnetic resonance imaging protocols generated from a domain-specific large language model. This can be achieved, for example, by receiving a selection from a user interface or another software component.

[0034] In another example, the memory further comprises a decision module configured to provide a selection of the selected magnetic resonance imaging protocol in response to receiving the translated pulse sequence data and the thought chain data. Execution of the machine executable instructions further cause the computing system to receive the selected magnetic resonance imaging protocol in response to inputting the translated pulse sequence data and the thought chain data into the decision module. This can be beneficial because it can provide a way to automatically select a selected magnetic resonance imaging protocol. The decision module can be implemented in different ways. For example, it can use an additional large language model to provide this decision process. The additional large language model can be trained, for example, by recording a large number of selections of human operators and utilizing it to train this additional large language model.

[0035] In another example, the tokenized pulse sequence data comprises a magnetic resonance imaging protocol identifier, a pulse sequence parameter, and combinations thereof. The magnetic resonance imaging protocol identifier can be used, for example, to identify pulse sequence commands for a particular magnetic resonance imaging protocol and / or an anatomical region to be imaged. The pulse sequence parameter can be, for example, various parameters for altering or customizing a pulse sequence according to a particular clinical situation.

[0036] In another example, the pulse sequence parameter comprises any of a pulse sequence repetition time, an echo time, a specification of a field of view, a flip angle, a k-space sampling method or pattern, an inversion time, a pulse shape definition (e.g. a radio frequency pulse), a gradient waveform of a magnetic field gradient, and combinations thereof. This example can be beneficial because these different parameters can help to customize a particular set of pulse sequences according to a particular clinical situation.

[0037] In another example, the subject data includes any of the following: clinical inquiry, clinical referral, subject medical record, prior written medical report, current medical condition of the patient, current medical condition of the patient, patient data describing mobility and communication abilities, patient data describing ability to perform breath hold, patient metadata, patient age, patient body mass index, patient age, patient gender, language spoken by the patient, intravenous access data, and combinations thereof.

[0038] In some examples, the subject data can also contain anatomical information, such as the location of anatomical markers on the subject’s surface or in the survey scan.

[0039] In another example, the computer program includes machine executable instructions and further includes a domain-specific large language model, each executed by a computing system. The large language model is configured to output tokenized pulse sequence data in response to receiving subject data. Execution of the machine executable instructions causes the computing system to receive the tokenized pulse sequence data in response to inputting the subject data into the large language model. The tokenized pulse sequence data is configured as a token selected from a library of pulse sequence tokens.

[0040] In another example, there is a method of training a domain-specific large language model. The method includes receiving a base large language model. As used herein, a base large language model includes a large language model trained using general knowledge and data. The base large language model can be, for example, a widely available large language model such as ChatGPT-4 or Alpaca model. The method further includes providing a pre-trained large language model by training the base large language model with medical documents containing pulse sequence command specifications. The pulse sequence command specifications are data explicitly describing the structure of the pulse sequence. The medical literature can be, for example, journal articles discussing specific pulse sequence commands, textbooks describing the structure and function of pulse sequence commands, and historical records and data of previously used pulse sequence commands used in clinical settings. For example, the medical files can include written descriptions of the use of these various pulse sequence commands and the like. Providing the pre-trained large language model includes generating a preliminary library of tokenized pulse sequences.

[0041] Foundation large language models often have a system for tokenizing input text to them. Tokenization can be based on assigning tokens to individual letters, word fragments, or entire words. However, these foundation large language models do not have specific tokens for specific medical documents. A preliminary tokenization pulse sequence library can be an additional token library generated during the foundation large language model training process with medical documents including or containing pulse sequence command specifications. The method of training a domain-specific large language model also includes receiving site-specific medical records. For example, these can include medical records used or generated at a specific clinical site. The site-specific medical records include subject medical data paired with historical pulse sequence commands. For example, at a specific clinic or hospital, there can be a series of medical records containing data about pulse sequences used and the condition of the subject paired with this pulse sequence data.

[0042] The method also includes generating a pulse sequence token library by excluding portions not present in the site-specific medical records in the preliminary tokenization pulse sequence library. For example, this can be accomplished by using the preliminary tokenization pulse sequence library to tokenize the site-specific medical records. Tokens not selected in this tokenization process can be excluded from the resulting token library. This can have the effect of preventing the use of preliminary tokens in the preliminary tokenization pulse sequence library when they are not present in the medical records of a specific site. The method also includes providing a domain-specific large language model by training a pre-trained large language model using the site-specific medical records. Training the pre-trained large language model includes at least partially tokenizing the site-specific medical records with the pulse sequence token library. This example can have the benefit of providing a domain-specific large language model adapted to generate tokenized pulse sequence data tailored to a specific clinical site and can be used to configure a specific magnetic resonance imaging system.

[0043] In another example, there is a domain-specific large language model trained according to the method of training a domain-specific large language model.

[0044] Figure 1An example of a medical system 100 is illustrated. The medical system 100 is shown to include a computer 102. The computer 102 can represent one or more computers 102 located at one or more locations. The computer 102 includes a computing system 104. The computing system 104 can represent one or more computing systems or computing cores also located at one or more locations. The computer 102 is further shown to contain an optional hardware interface 106 and an optional user interface 108, both in communication with the computing system 104. The hardware interface 106 can enable the computing system 104 to control other components of the medical system 100, if present. It can also enable data exchange with a network or other components of the medical system 100. The user interface 108 can enable an operator or user to control the operation and functionality of the medical system 100. The medical system 100 is further shown to include a memory 110. The memory 110 is intended to represent various types of memory that can be accessed by the computing system 104. In one example, the memory 110 is a non-transitory storage medium.

[0045] The memory 110 is shown to contain machine executable instructions 120. The machine executable instructions 120 can enable the computing system 104 to perform various computations and data operations as well as image processing tasks. The machine executable instructions 120 can also be used to or enable the computing system 104 to control other components of the medical system 100 via the optional hardware interface 106, for example. The memory 110 is further shown to contain a domain-specific large language model 122. The domain-specific large language model is configured to output tokenized pulse sequence data 126 in response to receiving object data 124. The tokenized pulse sequence data 126 can be limited to tokens present in a pulse sequence token library. The object data 124 can be received via a network connection or from another database, for example. In some examples, the object data 124 can be received in part from the user interface 108. The tokenized pulse sequence data 126 is shown to be stored in the memory 110 and received from the domain-specific large language model 122 upon input of the object data 124.

[0046] Memory 110 is also shown to contain an optional pulse sequence template database 128. The pulse sequence template database 128 contains templates that can be used to provide examples or fillable algorithms for generating pulse sequences according to a specific or selected magnetic resonance imaging protocol. Memory 110 is also shown to contain a database query 130. Database query 130 can be reconstructed or constructed, for example, from lexicalized pulse sequence data 126. In some examples, lexicalized pulse sequence data 126 may contain identifiers that are directly used as database query 130. In other examples, fingerprints or command sequences found in lexicalized pulse sequence data 126 are used to construct database query 130. Database query 130 is used to query the pulse sequence template database 128 and receives, in response, an selected pulse sequence template 132.

[0047] The selected pulse sequence template 132 is then modified using lexicalized pulse sequence data 126 to generate optional pulse sequence commands 134. These pulse sequence commands 134 can be used to control the magnetic resonance imaging system to acquire k-space data. Utilizing these optional features, the medical system 100 includes a control system that can be used to automatically configure the magnetic resonance imaging system to acquire k-space data using only the object data 124.

[0048] Figure 2 The illustrated operation is shown. Figure 1 A flowchart of a method for a medical system 100. In step 200, in response to inputting object data 124 into a domain-specific large language model 122, lexicalized pulse sequence data 126 is received. Steps 202, 204, and 206 are optional. In step 202, a pulse sequence lexicon is used to generate a pulse sequence command 134 based on the lexicalized pulse sequence data 126. For example, the pulse sequence lexicon is used to provide specific pulse sequence parameters from the individual lexicons present in the lexicalized pulse sequence data. In this particular example, the generation of the pulse sequence is indirect. Step 202 can be performed by performing the illustrated steps 204 and 206. In step 204, a selected pulse sequence template 132 is received by querying the pulse sequence template database 128 130 using at least a portion of the lexicalized pulse sequence data 126.

[0049] In step 206, the pulse sequence commands 134 are generated by inputting the tokenized pulse sequence data 126 into the selected pulse sequence template 132. In some cases, this can include first converting the tokenized pulse sequence data 126 into individual data or numerical data, which is then used to modify the selected pulse sequence template 132. In other examples, the selected pulse sequence template 132 can be an algorithm that is configured to directly receive the tokenized pulse sequence data 126. For example, it can be a fillable form, which upon receiving a particular token, the selected pulse sequence template automatically modifies or generates the pulse sequence commands.

[0050] Figure 3 Another example of a medical system 300 is illustrated. Figure 3 The medical system 300 in Figure 1 The medical system 100 in The magnetic resonance imaging system 302 comprises a magnet 304. The magnet 304 is a superconducting cylindrical type magnet with a bore 306 through it. Split cylindrical magnets and so-called open magnets can also be used. A split cylindrical magnet is similar to a standard cylindrical magnet except that the cryostat has been split into two parts to allow access to the isoplanar of the magnet, so that the magnet can be used in conjunction with charged particle beam therapy for example. An open magnet has two magnet parts, one on top of the other, with a space in between large enough to receive a subject. The arrangement of the two parts is similar to the arrangement of Helmholtz coils. Open magnets are popular because the subject is less confined. Inside the cryostat of the cylindrical magnet there is a collection of superconducting coils.

[0051] Within the bore 306 of the cylindrical magnet 304, there is an imaging zone 308 in which the magnetic field is strong and uniform enough to perform magnetic resonance imaging. A field of view 309 within the imaging zone 308 is shown. Measured k-space data is acquired for the field of view 309. The region of interest can be similar to the field of view 309, or it can be a sub-volume of the field of view 309. A subject 318 is shown supported by a subject support 320 so that at least a portion of the subject 318 is within the imaging zone 308 and the field of view 309. In this example, the head of the subject 318 is positioned within the field of view 309. Thus, the measured k-space data to be acquired is for performing a head or brain scan.

[0052] The magnet's bore 306 also contains an assembly of magnetic field gradient coils 310, which are used to acquire measured k-space data for spatial encoding of magnetic spins within the imaging region 308 of the magnet 304. The magnetic field gradient coils 310 are connected to a magnetic field gradient coil power supply 312. The magnetic field gradient coils 310 are intended to be representative. Typically, the magnetic field gradient coils 310 comprise three sets of independent coils for spatial encoding in three orthogonal spatial directions. The magnetic field gradient power supply provides current to the magnetic field gradient coils. The current supplied to the magnetic field gradient coils 310 is time-controlled and can be either slanted or pulsed.

[0053] Adjacent to the imaging region 308 is an RF coil 314, which is used to manipulate the orientation of the magnetic spins within the imaging region 308 and to receive RF transmissions from spins also located within the imaging region 308. The RF antenna may comprise multiple coil elements. The RF antenna may also be referred to as a channel or antenna. The RF coil 314 is connected to an RF transceiver 316. The RF coil 314 and the RF transceiver 316 may be replaced by separate transmit and receive coils, as well as separate transmitters and receivers. It is to be understood that the RF coil 314 and the RF transceiver 316 are representative. The RF coil 314 is intended to also represent a dedicated transmit antenna and a dedicated receive antenna. Similarly, the transceiver 316 may also represent separate transmitters and receivers. The RF coil 314 may also have multiple receive / transmit elements, and the RF transceiver 316 may have multiple receive / transmit channels. The transceiver 316 and the gradient controller 312 are shown as a hardware interface 106 connected to the computer system 102.

[0054] The memory is also shown as containing a pulse sequence lexicon library 322. In this case, the pulse sequence lexicon library 322 is used to transform the lexiconized pulse sequence data 126 into pulse sequence commands 134. In some examples, it can be used... Figure 2 The method is illustrated in the figure. Memory 110 is also shown as containing k-space data 324 acquired by controlling the magnetic resonance imaging system 100 using pulse sequence commands 134. Memory 110 is also shown as containing a k-space image 326 reconstructed from the imaging magnetic resonance data 324.

[0055] Figure 4 The illustrated operation is shown. Figure 3 The flowchart of the method for the medical system 300. Steps 200 and 202 are as follows. Figure 2 As shown, this is performed. In step 400, k-space data 134 is acquired by controlling the magnetic resonance imaging system 302 using pulse sequence commands 324. Although not shown in the figure, the method can continue by reconstructing the magnetic resonance image 326 from the k-space data 324. In some instances, Figure 1 and Figure 3The features of the medical system 500 can be combined. Similarly, Figure 2 The steps illustrated in the medical system 500 can also be combined with Figure 4 The steps illustrated in the medical system 500 can also be combined with

[0056] Figure 5 Another example of a medical system 500 is illustrated. In this example, the domain-specific large language model 122 is further configured to output thought chain data 520 while outputting tokenized pulse sequence data 126. Furthermore, in this example, the tokenized pulse sequence data 126 is output for multiple respective pulse sequence protocols. In this example, the domain-specific large language model 122 outputs several different suggestions, and provides thought chain data 520 for each suggestion. This thought chain data 520 can be used to help select a selected magnetic resonance imaging protocol 524. In this example, the pulse sequence token library 322 is used to generate translated pulse sequence data 522 from the tokenized pulse sequence data 126. The thought chain data 520 and the translated pulse sequence data 522 are provided individually for each suggestion and used to select the selected magnetic resonance imaging protocol 524.

[0057] The selected magnetic resonance imaging protocol 524 then has its respective tokenized pulse sequence data 126 for generating pulse sequence commands 134. In some examples, the data can be presented on a user interface, and then a user can select the selected magnetic resonance imaging protocol 524. This process can also be automated. The memory 110 is further shown as containing an optional decision module 526 that receives the thought chain data 520 and the translated pulse sequence data 522 for each suggestion, and then outputs the selected magnetic resonance imaging protocol 524. The optional decision module 526 can be implemented, for example, using an additional large language model configured for this purpose.

[0058] Figure 6 A flowchart illustrating a method of operating Figure 5 the medical system 500 is shown. Step 200 is as Figure 2 and Figure 4As shown in the middle. In step 600, upon receiving tokenized pulse sequence data 126, thought chain data 520 is received from the domain-specific large language model 122. In this case, the tokenized pulse sequence data 126 contains several different alternatives. In step 602, by translating the tokenized pulse sequence data labeled using the pulse sequence token library 322, translated pulse sequence data 522 is provided as text with a corresponding plurality of alternative magnetic resonance imaging protocols. In step 604, the thought chain data 520 is provided for each of the plurality of alternative magnetic resonance imaging protocols. In step 606, in response to providing the translated pulse sequence data 522 and the corresponding thought chain data 520, a selection of a selected magnetic resonance imaging protocol 524 is received. Next in step 202, pulse sequence commands 134 are generated from the tokenized pulse sequence data, in this case, the portion corresponding to the selected magnetic resonance imaging protocol 524 and the pulse sequence token library 322.

[0059] Figure 7 A supplementary method is illustrated that can be performed in conjunction with the methods shown in Figure 2 、 Figure 4 and Figure 6 . The method starts in step 700. In step 702, various subject data 124 is retrieved. This can include patient details, medical history, and medical reports, for example. In step 704, additional subject data 124 is retrieved. This can include additional directive information and written requests. In step 706, an input prompt is constructed for the domain-specific large language model 122. In step 708, the prompt completion for the domain-specific large language model 122 is performed. In step 710, protocol definitions for tokenized pulse sequence data 126 are extracted. In step 712, thought chain data 520 is also extracted from the output of the domain-specific large language model 122. In step 714, the translated pulse sequence data 522 and the thought chain data 520 for each of the plurality of suggestions are presented to an operator.

[0060] Step 716 is a decision block and proposes multiple protocols in the translated pulse sequence data 522. If the answer is yes, then in step 718 the user or decision module 526 selects a particular magnetic resonance imaging protocol 524. If the answer is no in block 716, then the method proceeds to block 720. The method also reaches block 720 after block 718 is executed. In step 720, the user is prompted if they are satisfied with the selected magnetic resonance imaging protocol 524. If the answer is yes, then the method proceeds to step 722 where the protocol definition is sent to the magnetic resonance imaging system 302. This can include generating pulse sequence commands 134. If the user is not satisfied with the selected magnetic resonance imaging protocol 524 in block 720, then the system can prompt the operator for additional data from the user. This can be used, for example, to add additional data to add to the prompt to the domain-specific large language model 122. The method then returns to step 706.

[0061] Figure 8 A flowchart illustrating one method of training a domain-specific large language model 122 is shown. This can be divided into three parts. There is a pre-training portion 800, a fine-tuning portion 802, and an optional thought chain data prompt engineering 804.

[0062] In pre-training 800, a base model 806 is received and trained on medical documents 808. This provides a pre-trained model 810. The medical documents 808 can include, for example, magnetic resonance imaging (MRI) sequences, variable settings, protocol names, a combination of medical and public data. The pre-training 808 can start with pre-processing the input by tokenizing the text input into machine-understandable numbers. While tokenization is not new for traditional text data, tokenizing MRI sequences while preserving their clinical meaning is novel. One approach is to split each sequence name by a delimiter, such as a hyphen or underscore, and then apply 1 to N-gram tokenization. For example, a sequence name of “T2W SSH Cor_trigg” would yield 7 tokens: “T2W”, “SSH”, “Cor”, “trigg”, “T2W SSH”, “T2W SSH Cor”, and “T2W SSH Cortrigg”. A more complex approach is to add the sequence’s variables and their values into the scope of consideration, such as TR (repetition time), TE (echo time), field of view, flip angle, etc. Although different MRI manufacturers have different naming conventions for protocols, the settings of the scan variables identify the sequence, and the order of the sequences defines the MRI protocol. By tokenizing the pre-formatted variable and value pairs while remembering the mapping of the protocol name to the sequence and its settings, the pre-processing step integrates the MRI sequence language into readable clinical information. The output of this step is a set of tokens, a mix of vocabulary.

[0063] Next, depending on the frequency of use in the input, a weighting scheme can be applied to assign numbers to all the tokens, resulting in a matrix with all the tokens as columns and their normalized weights as values. This matrix can then be used as input to the base model, whose initial parameters are updated during training, for example via backpropagation. The output of the pre-training is a set of parameters with optimized values. The related model 810 can have learned the semantic relationships of the tokens, their order, and the structure, adapting to the medical domain related to MRI protocols. For example, in the semantic space of tokens, the sequence term magnetic resonance cholangiopancreatography or MRCP will be close to abdomen or liver metastasis.

[0064] In the fine-tuning part 802, the hospital's own patient records and magnetic resonance imaging sequences 814 are used to generate a site-specific medical record 812. It contains data about individual patients, including patient data and magnetic resonance imaging protocols. It is then used to train the pre-trained model into a fine-tuned model 122.

[0065] For the task of "transforming clinical input into a list of MRI sequences", fine-tuning can use the pre-trained model and specialized structured data input to guide the pre-trained model to solve the target task. Depending on the user's needs, this module can also be adapted to handle multiple tasks, such as summarizing clinical input and then recommending MRI protocols. The input data for this module should be limited to a specific clinical entity, such as a specific hospital, so that the fine-tuned model can give the most appropriate output that best fits the hospital's own language usage.

[0066] For the translation task, the module can use a sub-module of data pairing, which takes the hospital's own clinical data and MRI protocols (i.e. protocol names, lists of sequences, and variable value pairs for each sequence) and converts them into pairs, for example in the form of (original text, MRI protocol). These pairs will then be used as input to adjust the pre-trained model. During fine-tuning, the values of the pre-trained model's parameters will be further updated to adapt to the translation task, so that when given clinical information combined with patient information and diagnostic purposes, a list of MRI sequences will be generated.

[0067] To make the output of the fine-tuned model consistent with human references, the model can be updated continuously through human feedback, for example, based on the positive or negative responses given by human users to the recommended MRI protocols, the model updates its parameter values to adjust its alignment with human references. To this end, a reward model can be used to punish or reward the model to trigger or automate the adjustment.

[0068] Once the fine-tuned model 122 is ready, it can be selectively refined using the think chain data prompt engineering 804. During the training of the base model 806, the medical documents 808 can be tokenized into a preliminary tokenized impulse sequence library. When generating the paired data set 812, there can be a tokenization step, and it can be noted which of the preliminary labeled impulse sequence library are not used. These labels can then be excluded, and the impulse sequence token library can be generated according to the exclusion. During the training of the pre-trained model 810 into the fine-tuned model 122, only the impulse sequence token library 322 is used.

[0069] After fine-tuning the model, the process of prompt engineering with think chain mechanism can be further established to improve the output of the model. This step is to instruct the model to not only output a list of MRI sequences, but also the reasons. This can use many instructions in the form of "clinical information -> reason -> MRI protocol" as input to the fine-tuned model. In this example, the method in this example can be applicable to the clinical radiology scenario, focusing on the specific task of recommending MRI protocols.

[0070] In some examples, the fine-tuned model 122 can be self-upgraded via continuous human feedback. This can be achieved through an interactive application. The following Figure 9 illustrates the user interface of a possible application. The example can be able to capture the user's input, solve it and output options with a probability exceeding a threshold. When the recommendation is not satisfactory, the model model 122 is adjusted according to the user's input and outputs an improved answer. In addition, it is also able to adapt to the changes or absence of input, based on which it outputs selectable answers to the user. Through interaction with the user, the backend model should be able to further align with human preferences, but maintain a certain degree of generalization to avoid overfitting to human input.

[0071] Example prompts (data input to the domain-specific large language model 122) can be: We know the following details about the patient to be examined: <patient details from RIS> The referring physician requires the following: <referral details and reports> Please suggest a suitable protocol definition for the required MRI examination. Give step-by-step reasoning for your proposal. Example prompt completion (output of the domain-specific large language model 122): For this case, I know that the patient has no claustrophobia and from the patient's details and history I infer that the patient is able to lie still and easily follow instructions. Based on the referral, it is clear that anatomical and physiological images are required. Therefore, I recommend the following protocol: (think chain data example) [protocol start] Investigation T1W axial T2W axial DWI [protocol end] (Example of a selected MR protocol)

[0072] Several specific examples are given below:

[0073] In the first example, an example prompt is used as input to a domain-specific large language model, and outputs a selection of an MR protocol to be performed. In this case, the operation of the MR system is not automatic, but rather provides assistance to an operator. As shown in the example above, the prompt (object data 124) input to the domain-specific large language model 122 is: We know the following details about the patient to be examined: <patient details from RIS> The referring physician requires the following: <referral details and reports> Please suggest a suitable protocol definition for the required MRI examination. Give step-by-step reasoning for your proposal. The final output of the domain-specific large language model 122 is then: For this case, I know that the patient has no claustrophobia and from the patient's details and history I infer that the patient is able to lie still and easily follow instructions. Based on the referral, it is clear that anatomical and physiological images are required. Therefore, I recommend the following protocol: (Example of thought chain data) [protocol start] A1 A2 A3 A4 [protocol end] (Example of a selected MR protocol)

[0074] The first part of this output is thought chain data, and the values A1, A2, A3, and A4 are tokens representing the selection of mediators MR acquisition to be performed: survey, T1W axial, T2W axial, DWI. In this case, the pulse sequence lexicon 322 can be used to convert these values A1, A2, A3, and A4 to text descriptions survey, T1W axial, T2W axial, DWI. Alternatively, these values can be used to retrieve templates for the pulse sequences, or can be used to trigger pre-programmed configurations of the magnetic resonance imaging system.

[0075] In the second example, the tokenized pulse sequence template data provides more detailed information, and can be used to automatically configure the magnetic resonance imaging system. In this example, the input prompt (object data) 124 can include: Object metadata (object height, weight, gender) Identification of the anatomy to be imaged Request from the physician for a clinical trial, or Geometric position of the object with respect to the magnetic resonance imaging system, which can be provided, for example, by fiducial markers on the object, the use of an anatomic key point locator using a camera, or via an investigation scan with identified anatomic landmarks. The geometric position records the position of the object into the MR system. Possibly the electronic patient history from before. Please suggest a pulse sequence for the required MRI examination. Give step-by-step reasoning for your proposal. Based on the object data input and the physician's requirements, a diffusion-weighted image (DWI) is suggested. The word tokens for the selection and specification of the appropriate pulse sequence are as follows. A3 B7 4E AE 41 28 98 E1

[0076] For example, an example output can be as follows: Figure 8 Figure 9 Figure 1

[0077] The value A3 B7 4E AE 41 28 98 E1 is a code, where A3 is used to query the pulse sequence template database 128 for a diffusion weighted image pulse sequence. In this case, a pre-made template for a DWI pulse sequence is retrieved. The value B7 4E AE 41 28 98 E1 is then used to customize or change the pulse sequence for the specific acquisition: B7 specifies the slice selection, B7 4E AE specifies the field of view, 41 specifies the configuration of the RF pulses, 28 specifies the configuration of the diffusion gradient pulses, 98 specifies the echo time (TE), and E1 specifies the pulse repetition time. In this example, the details output by the domain-specific large language model 122 are very detailed and can be used to configure the MR system. The domain-specific large language model 122 is able to do this because it was trained as shown using site-specific medical records 812. The historical subject data and previously performed MR pulse sequences contained in the site-specific medical records 814 were used to prepare the site-specific medical records 812 and provide these details during training. ​

[0078] ​ An example of an interactive user interface that can be used to interact with the medical system shown in FIGS. ​ , 3 or 5 is illustrated. The system can function as a chatbot when additional user input is required. For this application, the domain-specific large language model can have received additional training to interact with the user in cases where the protocol definition is ambiguous to accurately ask for the required details to make a better decision regarding the protocol. The figure also includes an example of two different protocols that the user can select from.

[0079] ​In this example, the user interface 900 includes a button 902 for retrieving a medical record, which can be equivalent to retrieving the subject data 124. There are additional buttons to retrieve information from the RIS 906, with an additional text entry box 908 in which additional data describing the subject can be entered. The medical system can also use a chatbot and provide a protocol chatbot 910 to ask the operator additional questions. The large language model can be configured to ask the physician questions to provide at least a portion of the subject data 124, rather than entering the text directly. The chatbot 912 can also have an area in which the operator can select different protocols 912. The link labeled 912 serves as a selector for magnetic resonance imaging protocols. Once a protocol is selected, there can also be an additional share button 914 that enables the operator to share the selection of the medical protocol with a colleague.

[0080] It should be understood that one or more of the foregoing examples or embodiments of the present application can be combined, provided the combined embodiments are not mutually exclusive.

[0081] As will be appreciated by those skilled in the art, the present application can be embodied as an apparatus, method, or computer program product. Accordingly, aspects of the present application can take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects that can all generally be referred to herein as a "circuit," "module" or "system." Furthermore, aspects of the present application can take the form of a computer program product embodied in one or more computer readable medium(s) having computer executable code embodied thereon.

[0082] Any combination of one or more computer readable medium can be utilized. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A "computer-readable storage medium" as used herein encompasses any tangible storage medium which can store instructions which are executable by a processor of a computing device or computing system. The computer-readable storage medium can be referred to as a computer-readable non-transitory storage medium. The computer-readable storage medium can also be referred to as a computer-readable tangible storage medium. In some embodiments, a computer-readable storage medium can also be able to store data which is able to be accessed by the computing system or a processor of the computing device. Examples of computer- readable storage media include, but are not limited to: a floppy disk, a magnetic hard disk drive, a solid state hard disk, flash memory, a USB thumb drive, Random Access Memory (RAM), Read Only Memory (ROM), an optical disk, a magneto-optical disk, and the register file of the processor of a computer system. Examples of optical disks include Compact Disks (CDs) and Digital Versatile Disks (DVDs), such as CD-ROM, CD-RW, CD-R, DVD-ROM, DVD-RW, or DVD-R disks. The term computer readable-storage medium also refers to various types of recording media capable of being accessed by the computer device via a network or communication link. For example, a data can be retrieved over a modem, over the internet, or over a local area network. Computer executable code embodied on a computer readable medium can be for implementing various aspects of the present disclosure as described in the detailed description.

[0083] A computer readable signal medium can include a propagated data signal with computer executable code embodied in an analog or digital format within baseband or as part of a carrier wave. Such a propagated signal can take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium can be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate or transport program for use by or in connection with an instruction execution system, apparatus, or device.

[0084] "Computer memory" or "memory" is an example of a computer readable storage medium. Computer memory is any memory accessible by a computing system directly. "Computer storage" or "storage" is another example of a computer readable storage medium. Computer storage is any non-volatile computer readable storage medium. In some embodiments, computer storage can also be computer memory, or vice versa.

[0085] A "computing system" as used herein encompasses an electronic component able to execute a program or machine executable instruction or computer executable code. References to computing system include references to a single computing system or multiple computing systems or processing cores in a computing system. For example, a computing system can be a multi-core processor system. References to a computing system also include references to a collection of one or more computing systems or processing cores working together. The term computing system should also be interpreted to possibly include a collection of computing devices or a network of computing devices each including a processor or processing core. Machine executable instructions or computer executable code can be executed by the one or more computing systems or processing cores that can be within the same computing device or even across multiple computing devices.

[0086] Machine executable instructions or computer executable code can include instructions or programs that cause a processor or other computing system to perform a task or tasks. Computer executable code or instructions can be written in any combination of one or more programming languages, including an object- oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer executable code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0087] The computer executable code can also be loaded onto a computer or computing system to cause a series of operations to be performed on the computer or computing system to produce a computer implemented process such as the process of claim 1. In other scenarios, the machine executable code can be encoded in a computer readable medium that can be read by a computer system to produce a series of operations performed on the computer system to generate a computer implemented process.

[0088] Aspects of the application are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. Also, it will be understood that when an element or layer is referred to as being "on" another element or substrate, it can be directly on the other element or substrate or intervening elements can also be present. In embodiments where flowchart diagrams, flowcharts or block diagrams are used to describe aspects of the application, it will be understood that each block of the flowchart, flowchart or block diagram, and combinations of blocks in the flowchart, flowchart or block diagram, can be implemented by computer program instructions. Such instructions can be implemented as "software" or "firmware" applications stored in

[0089] These machine executable instructions or computer program instructions can also be stored in a computer readable medium that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the computer readable medium produce an article of manufacture including instructions which implement the function / act specified in the flowchart and / or block diagram block or blocks.

[0090] The machine executable instructions or computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which run on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0091] A "user interface" as used herein is an interface which allows a user or operator to interact with a computer or computer system. A "user interface" can also be referred to as a "human interface device." A user interface can provide information or data to the operator and / or receive information or data from the operator. A user interface can enable input from an operator to be received by the computer and can provide output from the computer to the user. In other words, the user interface can allow an operator to control or manipulate a computer and the interface can allow the computer to indicate the effects of the operator's control or manipulation. Display of data or information on a display or graphical user interface is an example of providing information to an operator. Receiving data or information from an operator through a keyboard, mouse, trackball, touchpad, pointing stick, graphics tablet, joystick, gamepad, webcam, headset, pedals, wired glove, remote control, and accelerometer are all examples of user interface components that enable receiving information or data from an operator.

[0092] A "hardware interface" as used herein encompasses an interface that enables a computing system of a computer system to interact with or control an external computing device and / or apparatus. The hardware interface can allow the computing system to send control signals or instructions to the external computing device and / or apparatus. The hardware interface can also enable the computing system to exchange data with the external computing device and / or apparatus. Examples of hardware interfaces include, but are not limited to: a universal serial bus, an IEEE 1394 port, a parallel port, an IEEE 1284 port, a serial port, an RS-232 port, an IEEE-488 port, a Bluetooth connection, a wireless local area network connection, a TCP / IP connection, an Ethernet connection, a control voltage interface, a MIDI interface, an analog input interface, and a digital input interface.

[0093] A "display" or "display device" as used herein encompasses an output device or user interface suitable for displaying images or data. The display can output visual, audio, and or tactile data. Examples of displays include, but are not limited to: computer monitors, television screens, touchscreens, tactile electronic displays, Braille screens,

[0094] A "display" or "display device" as used herein encompasses an output device or user interface suitable for displaying images or data. The display can output visual, audio, and or tactile data. Examples of displays include, but are not limited to: computer monitors, television screens, touchscreens, tactile electronic displays, Braille screens,

[0095] While the application has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The application is not limited to the disclosed embodiments.

[0096] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a study of the drawings, the disclosure, and the claims. In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality. A single processor or other unit can fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. A computer program can be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but can also be distributed in other forms such as via the Internet or other wired or wireless telecommunication systems. Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A medical system (100, 300, 500) comprising: a memory (110) storing machine executable instructions (120) and a domain specific large language model (122), wherein the large language model is configured to output tokenized pulse sequence data (126) in response to receiving object data (124), wherein the tokenized pulse sequence data is configured to select tokens from a pulse sequence token library (322), wherein the tokenized pulse sequence data includes pulse sequence parameters; and a computing system (104), wherein execution of the machine executable instructions causes the computing system to: receive (200) the tokenized pulse sequence data in response to inputting the object data into the domain specific large language model; and generate (202) pulse sequence commands (134) from the tokenized pulse sequence data using the pulse sequence token library, wherein the pulse sequence commands are configured to control a magnetic resonance imaging system (302) to acquire k-space data (324).

2. The medical system of claim 1, wherein, execution of the machine executable instructions further cause the computing system to: display the pulse sequence commands using a user interface; receive pulse sequence command modification data in response to displaying the pulse sequence commands using the user interface; modify the pulse sequence commands using the pulse sequence modification data; and train the domain specific large language model using the object data and the modified pulse sequence commands. the medical system further comprises the magnetic resonance imaging system, wherein execution of the machine executable instructions further cause the computing system to acquire (400) the k-space data by controlling the magnetic resonance imaging system with the pulse sequence commands.

3. The medical system of claim 1 or 2, wherein, the medical system further comprises a pulse sequence template database (128) storing pulse sequence templates configured to generate pulse sequence commands, wherein execution of the machine executable instructions further cause the computing system to:

4. The medical system of claim 1, 2, or 3, wherein, receive (204) a selected pulse sequence template by querying the pulse sequence template database with at least a portion of the tokenized pulse sequence data; and generate (206) the pulse sequence commands by modifying the selected pulse sequence template with the tokenized pulse sequence data. the domain specific large language model is further configured to output thought chain data (520) describing the tokenized pulse sequence data, wherein execution of the machine executable instructions further cause the computing system to receive (600) the thought chain data from the domain specific large language model when receiving the tokenized pulse sequence data.

5. The medical system of any one of claims 2 to 5, wherein, the tokenized pulse sequence data specifies a plurality of alternative magnetic resonance imaging protocols, wherein the domain specific large language model is configured to provide the thought chain data for each of the plurality of alternative magnetic resonance imaging protocols, wherein execution of the machine executable instructions further cause the computing system to:

6. The medical system of claim 6, wherein, ​ providing (602) translated pulse sequence data (522) by translating the tokenized pulse sequence data using the pulse sequence token library as text associated with a respective plurality of alternative magnetic resonance imaging protocols; providing (604) the thought chain data for the respective plurality of alternative magnetic resonance imaging protocols; and receiving (606) a selection of a selected magnetic resonance imaging protocol (524) in response to providing the translated pulse sequence data as text associated with the respective plurality of alternative magnetic resonance imaging protocols and in response to providing the respective thought chain data, wherein the pulse sequence commands are generated for the selected magnetic resonance imaging protocol.

7. The medical system of claim 7, wherein, the memory further comprises a decision module (526) configured to provide the selection of the selected magnetic resonance imaging protocol in response to receiving the translated pulse sequence data and the thought chain data, wherein execution of the machine executable instructions further causes the computing system to receive the selected magnetic resonance imaging protocol in response to inputting the translated pulse sequence data and the thought chain data into the decision module.

8. The medical system of any one of the following, wherein, the tokenized pulse sequence data comprises a magnetic resonance imaging protocol identifier.

9. The medical system of any one of the preceding claims, wherein, the pulse sequence parameters comprise any of: a pulse sequence repetition time, an echo time, a specification of a field of view, a flip angle, a k-space sampling method, an inversion time, a pulse shape definition, a gradient waveform, and combinations thereof.

10. The medical system of any one of the preceding claims, wherein, the subject data comprises any of: a clinical inquiry, a clinical referral, a subject medical record, a prior written medical report, a current medical condition of a patient, a mental condition of a patient, patient data describing mobility and communication abilities, patient data describing ability to perform breath holds, patient metadata, patient age, patient body mass index, patient age, patient gender, a spoken language, IV access data, and combinations thereof.

11. The medical system of any one of the preceding claims, wherein, the domain-specific large language model (122) is trained according to a training method, wherein the training method comprises: receiving a base large language model (806); providing a pre-trained large language model (810) by training the base large language model with medical documents (808) comprising pulse sequence command specifications, wherein providing the pre-trained large language model comprises generating a preliminary tokenized pulse sequence library; receiving site-specific medical records (812), wherein the site-specific medical records comprise subject medical data paired with historical pulse sequence commands; generating a pulse sequence token library (322) by excluding portions of the preliminary tokenized pulse sequence library that are missing in the site-specific medical records; providing the domain-specific large language model (122) by training the pre-trained large language model with the site-specific medical records, wherein training the pre-trained large language model comprises at least partially tokenizing the site-specific medical records with the pulse sequence token library.

12. A computer-implemented method comprising: receiving (200), in response to inputting object data (124) into a domain-specific large language model (122), tokenized pulse sequence data, wherein the domain-specific large language model is configured to output, in response to receiving object data, tokenized pulse sequence data, wherein the tokenized pulse sequence data is configured to select tokens from a pulse sequence token library, wherein the tokenized pulse sequence data includes pulse sequence parameters; and generating (202), using the pulse sequence token library, a pulse sequence command (134) from the tokenized pulse sequence data, wherein the pulse sequence command is configured to control a magnetic resonance imaging system (302) to acquire k-space data (324).

13. The computer-implemented method of claim 12, wherein the domain-specific large language model (122) is trained according to a training method, wherein, The training method includes: receiving a base large language model (806); providing a pre-trained large language model (810) by training the base large language model with medical documents (808) including pulse sequence command specifications, wherein providing the pre-trained large language model includes generating a preliminary tokenized pulse sequence library; receiving site-specific medical records (812), wherein the site-specific medical records include subject medical data paired with historical pulse sequence commands; generating a pulse sequence token library (322) by excluding portions of the preliminary tokenized pulse sequence library that are missing in the site-specific medical records; providing the domain-specific large language model (122) by training the pre-trained large language model with the site-specific medical records, wherein training the pre-trained large language model includes at least partially tokenizing the site-specific medical records with the pulse sequence token library.

14. A computer program comprising machine executable instructions (120) and a domain specific large language model (122) for execution by a computing system (104), wherein, The large language model is configured to output, in response to receiving object data, tokenized pulse sequence data (126), wherein the tokenized pulse sequence data includes pulse sequence parameters; wherein execution of the machine executable instructions causes the computing system to: receive (200), in response to inputting the object data into the domain-specific large language model, the tokenized pulse sequence data, wherein the tokenized pulse sequence data is configured to select tokens from a pulse sequence token library (322), and generate (202), using the pulse sequence token library, a pulse sequence command (134) from the tokenized pulse sequence data, wherein the pulse sequence command is configured to control a magnetic resonance imaging system (302) to acquire k-space data (324).

15. The computer program of claim 14, wherein, The domain-specific large language model (122) is trained according to a training method, wherein the training method includes: receiving a base large language model (806); providing a pre-trained large language model (810) by training the base large language model with medical documents (808) including pulse sequence command specifications, wherein providing the pre-trained large language model includes generating a preliminary tokenized pulse sequence library; receiving site-specific medical records (812), wherein the site-specific medical records include subject medical data paired with historical pulse sequence commands; generating a pulse sequence token library (322) by excluding portions of the preliminary tokenized pulse sequence library that are missing from the site-specific medical record; providing the domain-specific large language model (122) by training the pre-trained large language model with the site-specific medical record, wherein training the pre-trained large language model includes at least partially tokenizing the site-specific medical record with the pulse sequence token library.

16. A method (122) of training a domain-specific large language model, wherein the method comprising: receiving a base large language model (806); providing a pre-trained large language model (810) by training the base large language model with medical documents (808) including pulse sequence command specifications, wherein providing the pre-trained large language model includes generating a preliminary tokenized pulse sequence library; receiving a site-specific medical record (812), wherein the site-specific medical record includes subject medical data paired with historical pulse sequence commands; generating a pulse sequence token library (322) by excluding portions of the preliminary tokenized pulse sequence library that are missing from the site-specific medical record, wherein the pulse sequence token library includes pulse sequence parameters; providing the domain-specific large language model (122) by training the pre-trained large language model with the site-specific medical record, wherein training the pre-trained large language model includes at least partially tokenizing the site-specific medical record with the pulse sequence token library.

17. A domain-specific large language model (122) trained according to claim 16.

Images