Method for generating a protocol for a medical imaging device

Abstract

The invention describes a computer-implemented method for generating a protocol (Psyn) for a medical imaging device (50) installed at a specific location (3), the method comprising the following steps: providing a homomorphically encrypted large language model (MT) trained to generate a homomorphically encrypted medical imaging protocol (Esyn); obtaining patient-specific input data (33) for an intended medical imaging task; performing homomorphic encryption on the patient-specific input data (33); applying the homomorphically encrypted patient-specific input data (ESeed) to the trained homomorphically encrypted large language model (MT); and performing homomorphic encryption on the generated homomorphically encrypted medical imaging protocol (Esyn) to obtain a medical plaintext imaging protocol (Psyn).and wherein homomorphic encryption and decryption are performed using a first encryption scheme (S3) specific to the site (3) and a second encryption scheme (S5) specific to the provider (5). The invention further describes a computer-implemented method for training a large language model and a data processing arrangement (1).

Description

BackgroundIn order to perform a medical imaging procedure such that a satisfactory image is obtained, the imaging device needs to be correctly configured. The set of input parameters used during an imaging task is referred to as the "imaging protocol" or "medical imaging protocol.". A medical imaging protocol is a predefined procedure or set of policies that outlines how to perform a particular medical imaging examination. These protocols aim to standardize the imaging process to ensure consistency and reproducibility, while also ensuring that the imaging procedure provides the required diagnostic information and minimizes patient exposure to possible risks such as radiation. Examples of values that could be specified in a medical imaging protocol include the modality that references the type of imaging technology used, such as X-ray, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound, or positron emission tomography (PET). The protocol also includes specific parameters used by the imaging device. For example, in an MRI protocol, these could include the magnetic field strength, echo time, repetition time, inversion time, and flip angle. For a CT protocol, parameters could include tube current, tube voltage, rotation time, and layer thickness. Some protocols may specify the use of a contrast agent to improve the visibility of certain tissues or structures, including the type of contrast agent, the dosage, and the timing of its administration. The protocol may also provide instructions for patient positioning in the scanner, and for modalities such as MRI, the protocol may specify the use of certain sequences that provide different types of contrast between tissues.The user of a medical imaging device, for example a radiology assistant, generally requires extensive training to be familiar with the (usually many) features of the device and the various input parameters. To choose an optimal set of imaging protocol parameters for an intended imaging task, the user may refer to relevant or similar prior scans, imaging protocols, and patient results. To interpret such information correctly, the user may have collected experience over years.Even for an experienced user of the imaging device, it may be a challenge to define a correct protocol for each imaging task. An incorrect or suboptimal imaging protocol may result in the imaging task having to be repeated (assuming that the protocol deficiencies are detected) and may even have an adverse effect on patient care.To improve the step of assembling a protocol, the provider of the imaging device may provide templates, for example templates for specific imaging tasks (e.g., Hals CT, Shoulder MRI; etc.). The radiology assistant may select an appropriate protocol and make the adjustments required to adjust the protocol for the particular patient and make a counter-check to ensure that the protocol follows any relevant protocol recipe, for example a protocol recipe of that particular facility or clinic.Furthermore, the lack of standardization between different facilities and the different personnel preferences among skilled operators can result in significant variations in image quality, radiation dose, and diagnostic results, even for the same patient and imaging task. Inexperienced operators may find it difficult to compile an optimal protocol or select the most appropriate protocol from a recipe book, and the quality of the resulting diagnostic images may be suboptimal, with an adverse effect on patient care. The known approaches are therefore time-consuming and error-prone.One possible way to overcome these problems could be to build a collection of satisfactory protocols for a facility, for example, in the radiology department of a hospital. Over time, the collection could comprise many thousand protocols. A radiology assistant could search this comprehensive "protocol library" hoping to find a protocol for a particular imaging task that was created in the past for a similar patient. However, an initially submitted compilation of such a collection would require a very long time during which many thousands of imaging tasks would have to be performed on many different types of patients to create a sufficiently large database.In another scenario, some devices would share their collections of medical imaging protocols. However, this option is effectively eliminated due to the need to ensure data protection of patient data. Generally, any medical device ensures that its patient data is securely stored, for example, images and protocols may be stored in a local Picture Archiving and Communication System (PACS) that is accessible only to authorized personnel of the device or to programs installed in that device.It is therefore an object of the invention to provide a safe and simple way of obtaining a medical imaging protocol.This object is achieved by the claimed computer-implemented methods for generating a protocol for a medical imaging device and training a large language model for this purpose and by the claimed data processing arrangement.DESCRIPTION OF THE INVENTIONThe computer-implemented method of generating a medical imaging protocol of the present invention is used, for example, by a radiology assistant to generate a protocol for input to a medical imaging device installed at a particular customer site, for example, in a hospital or clinic. The computer-implemented method according to the invention comprises the following steps: providing a homomorphically encrypted large language model trained to generate a homomorphically encrypted medical imaging protocol; obtaining patient-specific input data for an intended medical imaging task; performing a homomorphic encryption on the patient-specific input data; applying the homomorphically encrypted patient-specific input data to the trained homomorphically encrypted large language model. Homomorphic decryption is then performed on the generated homomorphically encrypted medical imaging protocol to obtain a clear text medical imaging protocol. In the method of the invention, homomorphic encryption and decryption are performed using a first encryption scheme specific to the customer site and a second encryption scheme specific to the large language model provider. The site-specific and vendor-specific encryption / decryption key may be stored in a key store, for example in an authenticated gateway device, in an established manner.The computer-implemented method of training a big language model for use in the above method comprises the steps of: modifying a big language model to include a homomorphically encoded layered architecture; obtaining homomorphically encoded training data based on imaging protocols for the medical imaging device (50) installed at the particular customer site; and training the homomorphically encoded big language model on the homomorphically encoded training data to generate a homomorphically encoded medical imaging protocol. Again, homomorphic encryption and decryption are performed using the first (i.e., site-specific) encryption scheme and the second (i.e., vendor-specific) encryption scheme).An advantage of the approach of the invention is that it provides a way of using a vendor-provided machine learning tool to learn from an advantageously large collection of medical imaging protocol data while following parent data protection constraints. The homomorphically encrypted big language model is provided by the provider, which may also provide instances of the same big language model (LLM) for various other customers. The provider may also provide computing services to multiple clients. The two scheme encryption employed by the method of the invention enables data processing steps to be performed at the customer site, but also enables data processing steps to be performed off site (e.g., at the supplier site) without compromising patient data privacy. In this manner, the method of the invention makes it impossible for the supplier (or any other customer of that supplier) to access any patient data at any stage in the method of the invention. Confidential patient data remains on site (e.g., within a clinic network) and any data leaving the site is no longer readable by humans, i.e., no longer in a form that makes sense to a human reader.Furthermore, the computer-implemented method of the present invention can rapidly generate an imaging protocol for the user and can therefore help optimize imaging workflow at the customer site (e.g., a hospital, clinic, or similar facility). The user - for example a radiology assistant - only needs to provide basic information, such as patient-specific data, a descriptor of the intended imaging task, etc.The data processing arrangement according to the invention comprises the following: a means for providing a homomorphically encrypted large speech model; a model development stage which is designed for developing and training the homomorphically encrypted large speech model in order to generate a homomorphically encrypted output text from an input text; a means for obtaining a patient-specific input text for an intended medical imaging task; and a protocol generation stage which is designed for applying the patient-specific input data to the trained homomorphically encrypted large speech model in order to obtain a medical clear text imaging protocol for the intended imaging task.The data processing arrangement may be realized in the form of computer program modules which may be executed on (or on) any suitable platform(s) and which comprise program units for carrying out the steps of the computer-implemented method according to the invention. In the context of the invention, a computer program module comprising means for receiving the homomorphically encrypted vendor model and the site-specific input data that is input and means for outputting a trained large language model, referred to herein as the "model training stage" of the data processing apparatus; a computer program module comprising means for receiving the trained homomorphically encrypted large language model and the patient-specific input data and means for outputting the medical plain text imaging protocol, referred to herein as a "protocol generation stage" of the data processing apparatus. The data processing arrangement according to the invention may be a distributed system, wherein units and modules are suitably distributed over customer and supplier sites, for example.The computer-implemented method of the invention may be used by a clinician, such as a radiology assistant or a radiologist. The computer program product may be configured to present the generated imaging protocol to the user in any suitable manner (e.g., display it on the screen of a device such as a desktop computer, a tablet computer, etc.).In particular, the generated clear text medical imaging protocol may be used to control the medical imaging device installed at the particular location. In one embodiment, the medical imaging protocol may be used to manage the operations of the medical imaging device, such as initiating, pause, or ending imaging procedures. In another embodiment, the medical imaging protocol may be used to adjust the imaging parameters of the device, including, but not limited to, the intensity, frequency, and duration of the imaging signals (x-ray signals, magnetic and radio frequency signals).The object of the invention is also achieved by a computer program product comprising a computer program which can be loaded directly into the memory of a data processing arrangement and which comprises program units for performing the steps of the computer-implemented method according to the invention when the program is executed. The computer program product can be stored on a computer-readable data carrier.Particularly advantageous embodiments and features of the invention are given by the dependent claims as disclosed in the following description. Features of various claim categories may be combined as appropriate to enable further embodiments not described herein.As used in the context of the invention, the term "patient-specific input data" may refer to any data that is unique to an individual patient and that is used in the context of this patient's health care. Such data is typically used to make clinical decisions, guide treatment plans, and monitor patient progress, and can be derived from a variety of sources including medical history, physical examinations, laboratory tests, imaging studies, and patient-reported results. Examples of patient-specific input data include demographic information (including age, sex, ethnicity, and other social-fluidic factors of the patient that may affect health effects), medical history (including past and current diseases, operations, allergies, and drug intake), results of physical examinations (e.g., observations made by health providers during a physical examination, such as pulse, blood pressure, and results from the examination of various body systems), and / or laboratory test results (quantitative and qualitative data obtained from the analysis of biological samples, such as blood or urine, including blood glucose levels, cholesterol levels, and large blood image results). Any patient-specific input data used in the compilation of an imaging protocol is generally in the form of words and numbers and may also be referred to as "patient-specific input text.".In the context of the invention, it will be understood that a provider provides the customer with medical imaging devices and also with software for operating the devices. The same provider can also provide the customer with a data processing arrangement for carrying out the computer-implemented method according to the invention.The imaging modality or medical imaging device for which a generated or synthesized protocol is intended may be a CT device, an MRI device, a PET device, a single photon emission computed tomography (SPECT) device, an X-ray device, etc. The terms "medical imaging protocol", "imaging protocol", and simply "protocol" may be used interchangeably.In the context of the invention, terms such as "machine learning algorithm", "big speech model", "artificial intelligence tool", and "AI model" are synonymous and may be used interchangeably. A large speech model is an artificial neural network that is particularly well suited for applications involving text understanding or generation thereof. The machine learning algorithm is trained to generate a medical imaging protocol for an intended imaging task using only a number of patient-specific parameters. Any suitable LLM architecture may be used, for example, a long short-term memory (LSTM) network architecture. The selected LLM architecture is further modified and updated by the provider to include a homomorphically encrypted layer architecture for the highly secure privacy preserving machine learning method according to the invention. Other algorithms such as a cross entropy loss function module and an adam optimizer module may be implemented. Any algorithms required for the method according to the invention may be provided by the provider and / or by a third party.At a stage of preparation, a homomorphically encrypted version of the vendor model is generated with the customer encryption scheme and the vendor encryption scheme as previously indicated. As known to one of ordinary skill in the art, homomorphic encryption is a form of encryption that allows computations to be performed on ciphertext, thereby producing an encrypted result that, when decrypted, matches the result of operations performed on the ciphertext. This property allows third party systems to process data without having to decrypt it, thereby preserving data protection and security. The homomorphically encoded large language model is referred to herein as the "HE-LLM" (Homomorphic Encrypted Large-Language Model).In the method of the invention, training the HE LLM includes the steps of: composing a training dataset; assigning a ground truth to the training dataset; and applying the HE LLM to the training dataset to approximate the ground truth. These steps are repeated for all available training data sets and for an appropriate number of epochs until a desired level of accuracy is achieved.In the method of the invention, training data is based on medical imaging protocols previously generated at the customer premises (referred to herein as the facility). For example, the device may maintain a database of all medical imaging protocols for its imaging modalities. If necessary, the protocols are converted to a common format, such as a suitable markup language (e.g., XML). Since the accuracy of a machine learning algorithm depends to a high degree on the number of data sets with which it is trained, in a preferred embodiment of the invention the LLM is trained using at least a few hundred medical imaging protocols, more preferably a few thousand medical imaging protocols.The device may also employ medical imaging protocol recipe books to ensure device-specific standards are met in each imaging procedure. In a specific preferred embodiment of the invention, the protocol data therefore comprise any relevant recipe books for medical imaging protocols for a medical imaging device. A protocol recipe in PDF format can be scanned and digitized and converted to the same format used for the protocols.The LLM could also be trained with synthetically generated medical imaging protocols and / or protocol recipe books to improve its accuracy. It can be expected that by training the LLM on many thousands of verified or confirmed protocols used for past imaging tasks, a synthesized protocol becomes very accurate, so that the inventive approach can help minimize the dose of radiation to which a patient is exposed.To prepare a training data set, tokenization and padding are performed on each input data set. During tokenization, the text of each document is converted into a sequence of tokens and each token is assigned a number. To ensure that the tokenized sequences have the same length, padding is done as needed. In a "clean up" stage, duplicate or irrelevant records are removed. At the completion of this data preprocessing stage, the padded tokenized data sets are divided into a training subset, a validation subset and a test subset.Homomorphic encryption is then performed on each subset of the padded tokenized data using the customer encryption scheme and the provider encryption scheme as explained above. In a preferred approach, approximately 70% of the homomorphically encoded data sets are used to train the model and the remaining data sets are used in test and validation stages, as is known to one of ordinary skill in the art. Of course, the subset sizes may be adjusted according to the total number of available data sets to further improve accuracy. Only at this stage, i.e. after homomorphic encryption, are data sets allowed to leave the device. Since the data is no longer in a form readable by humans, patient data security is ensured.An embedding layer then maps the encrypted token representations to dense vector representations that preserve semantic relationships between the encrypted tokens. The resulting encoded dense vectors are passed through an encoded dense (fully connected) layer with an activation function in the form of a rectified linear unit (ReLU). The ReLU introduces nonlinearity into the model, which allows the model to capture complex relationships in the data. Softmax (normalized exponential function) activation is then applied to produce a probability distribution across multiple classes, i.e., the likelihood for each word in the vocabulary is predicted to be the next word in a sequence. The result is a probability distribution over the vocabulary.The formed model is then assembled with the homomorphically encrypted version of the categoric cross entropy loss function and the homomorphically encrypted version of the adam optimizer. Categorical cross entropy is a known loss function for multiclass classification problems that compares the predicted probability distribution to the actual distribution (typically a one-time encoded vector), thereby allowing the model to determine its training progress. The adam optimizer is a known optimization algorithm used to minimize the loss function during training.The compiled homomorphically encrypted model is then trained with the homomorphically encrypted training data and homomorphically encrypted target labels to predict the next word when a sequence of words is given. During training, the input data comprises sequences of tokens (words or subwords) from the training data and a target label is the token immediately following a sequence of the training text.Training is performed for an appropriate number of epochs, and the result is a trained homomorphically encoded model that will synthesize a homomorphically encoded medical imaging protocol when an appropriate text input is input thereto. The training may be monitored, unsupervised, or self-supervised. A supervised training procedure preferably comprises a technique of backpropagation to find the best set of parameters that will allow the HE LLM to infer ground truth from the input, in this case to infer the most suitable log entry to follow the previous log entry. The training procedure is preferably designed to achieve an advantageously low cost function. Of course, the HE-LLM may continue learning by including any verified protocol in the training dataset.The trained homomorphically encoded model may be deployed whenever a user requires a medical imaging protocol for an intended imaging task. To obtain an imaging protocol, the user prepares an initial input sequence or seed text. An input sequence may include multiple pairs of field descriptor and field value to characterize the patient (e.g., "patient ID:PD0001"; "age:45"; "previous operations:none", etc.). Preferably, the protocol generation stage performs any required tokenization and padding on the input sequence and performs homomorphic encryption on the tokenized and padded data. With this input, the trained homomorphically encrypted model (which may be executed at an out-of-location location location, on a cloud computing platform, etc.) generates homomorphically encrypted text that is returned to the user in the device. To convert the synthesized protocol to a usable, i.e., human readable, protocol, a decryption stage then applies homomorphic encryption with the customer encryption scheme and the provider encryption scheme to obtain a clear text output protocol that is then presented to the user, for example, on the monitor of a desktop computer. The fields of the synthesized protocol may be presented along with the fields of a standard protocol so that the user can easily identify any discrepancy between the values of these fields. For example, if a parameter "T1-weighted TR" of the synthesized protocol suggests a value of 500 ms and the corresponding field of the standard protocol indicates a range of 400-800 ms, the user may be sure that the suggested value is acceptable.Other objects and features of the present invention will become apparent from the following detailed descriptions taken in conjunction with the accompanying drawings. It is to be understood, however, that the drawings are designed solely for purposes of illustration and not as a definition of the limits of the invention. FIG. 1 illustrates the principle of the invention; FIG. 2 is an exemplary flow diagram of the method according to the invention; FIGS. 3 and 4 show a schematic representation of the data processing arrangement according to the invention; FIG. 5 shows an example imaging protocol synthesized using the computer-implemented method of the present invention.In the diagrams, like numbers refer to like objects throughout. Objects in the diagrams are not necessarily drawn to scale.Fig. 1 illustrates the principle of the invention. The diagram shows a facility 3, which may be a hospital, clinic, radiology practice, etc. The apparatus 3 may employ any number of medical imaging devices 50 for imaging modalities such as CT, MRI, etc. Patient data is securely stored, for example, in a PACS database 30. a provider 5 provides the apparatus 3 with software and hardware, for example, medical imaging devices 50 and software for operating the devices 50, a client-side server 51 for processing and analyzing medical images and for managing the resulting data. The provider 5 also provides software updates as needed, and may also provide data processing services on a provider-side server 51, a cloud computing service, etc.Here, the provider 5 also provides an AI model 5M. The homomorphically encrypted model ME is trained for the customer 3 using locally stored data P, C P and the trained homomorphically encrypted model MT can then be used to generate medical imaging protocols P syn for the customer, as explained below.Data security is ensured by employing encryption at some stages during the development and implementation of the computer-implemented method. Encryption is performed using an encryption scheme S3 that is unique to device 3 and an encryption scheme S5 that is unique to provider 5.Units and modules of the data processing arrangement 1 may be distributed in any suitable configuration. In the configuration of FIG. 1, a model development stage 1T for training the model 5M and a protocol generation stage 1G for generating a protocol E syn are realized at the customer premises 3. Of course, the model development stage 1T may be realized at the provider 5; likewise, the protocol generation stage 1G may be realized at the provider 5. Figs. 2 and 3 are block diagrams showing an embodiment of the data processing apparatus 1 according to the present invention.An exemplary sequence of steps performed by the method of the invention is illustrated in Figure 2. Data is collected in a first step 21, for which purpose a collector stage 11 collects a various and representative data record of medical imaging protocols P for various diseases. These protocols P originate from the local facility 3. One or more local protocol recipe books C P can also be collected. The data collected preferably includes a broad range of information such as patient demographics, medical history, scan parameters, corresponding results, etc., and preferably covers a broad range of anatomical regions and pathologies.A step of digitizing the information P, C, P may also be performed at this stage, if necessary. For example, a protocol recipe in PDF format may be scanned and digitized into an appropriate format, such as an XML markup language. The data D 21, collected in this way, comprises a collection of documents, i.e. a set of machine-readable protocols and optionally a set of machine-readable protocol recipe books.In a subsequent step 22, the collected data sets D 21 are loaded into a preprocessing module 12 which performs tokenization and padding on the input data D 21.During tokenization, the text of each document is converted into a sequence of tokens such as integers. To ensure that the tokenized sequences have the same length, padding is done as needed. The data preparation step may also include a "clean up" stage to remove duplicate entries or irrelevant entries. At this stage, consistent formatting and structure may be applied to the data. Upon completion of the preprocessing step 22, the padded tokenized data D 22 is divided into a training subset, a validation subset and a test subset, as explained above.In a subsequent step 23, the homomorphic encryption by the encryption stage 13 is applied to each subset of the padded tokenized data D 22 using a homomorphic encryption scheme S3 that is unique to the device 3 and a homomorphic encryption scheme S5 that is unique to the provider 5 to obtain a homomorphically encrypted representation of the data E 23.In a subsequent step 24, an embedding layer 14 maps each homomorphically encrypted token from the homomorphically encrypted representation E 23 to obtain a homomorphically encrypted dense vector representation E 24.In a subsequent step 25, homomorphic encryption is applied to each layer of the provider-provided HE-LLM 5M. To form the HE-LLM model, the output of an encrypted layer is fed into an encrypted dense layer with rectification linear unit (ReLU) activation. Softmax (normalized exponential function) activation is then used to predict the homomorphically encoded probability distribution across the vocabulary for the next word. The model is then assembled with an encrypted version of the categoric cross entropy loss function 54 and an encrypted version of the adam optimizer 55.In a subsequent step 26, a training stage 16 trains the assembled homomorphically encrypted model MC with homomorphically encrypted training data E 23. Using plain text as an example, an input sequence could be "no evidence of cysts or tumors.". The corresponding tokens are ["no", "hint", "on", "cyst", "or", "tumors"], and possible training sequences may be: ["no", "hint", "on"] → target: "cyst"; ["on", "cyst", "or"] → target: "tumors"; etc. In the approach according to the invention, the training of the LLM is done exclusively on homomorphically encoded data and therefore may be done "off site" without any risk of compromising patient data security. For this purpose, the homomorphically encoded training data E 23 and homomorphically encoded target labels E L are passed to a fitting function, for example a stochastic gradient descent function, a batch gradient descent function, a transfer learning function, etc. The model is trained for an appropriate number of epochs. The result of this training stage 16 is a trained homomorphically encrypted model MT.The generation and prediction of an output text are handled in a subsequent step 27 indicated by the predictor stage 17 in Fig. 4. Here, an input sequence 33 (after it is tokenized and padded as needed) provided by the user undergoes homomorphic encryption. The trained homomorphically encrypted model MT predicts the next word w pt based on the homomorphically encrypted input sequence E Seed. The seed text is updated by appending the predicted word. The process of generating predictions and updating the seed text is repeated until the desired number of words - i.e. the synthesized protocol E syn- has been generated.In a final step 28, the text E syn, obtained at the last iteration, is passed to a decryption stage 18, which applies homomorphic decryption (with the location encryption scheme S3 and the provider encryption scheme S5) to obtain a synthesized clear text output protocol P syn which is then presented to the user, for example, in tabular form.All steps of forming and training the model 5M may be done off-site using a homomorphically encoded training dataset generated locally, i.e., at the hospital or clinic facility. Similarly, the trained out-of-location model 5M may be tested and verified using homomorphically encoded test data sets and verification data sets. In this way, confidential patient data remains local or on site while offloading resource and time consuming model generation, for example, to a server farm (on site or cloud-based) provided by the provider. The trained and verified model 5M - which can be executed to an out-of-site location, such as on a cloud computing server - can then be used at any time by a user of the device who only needs to provide a basic relevant seed text that is homomorphically encrypted before exiting the device. Model 5M returns a synthesized complete imaging protocol to the device. The homomorphically encoded protocol undergoes homomorphic encryption to make it readable by humans.The protocol P syn, generated by the data generation apparatus 1 according to the invention, can be presented together with a standard protocol, so that the user can easily identify any possible discrepancy, as illustrated in Figure 5. The diagram shows a synthesized imaging protocol P syn, comprising a list of parameters 60 relevant to an imaging task, a list 61 of values as proposed in the synthesized imaging protocol P syn and a list 62 of corresponding allowed ranges as determined by the appropriate reference imaging protocol P Ref( for example a recipe book protocol). The graph shows that the values 61 in the synthesized imaging protocol P syn are all within the allowable ranges 62 of the recipe book protocol Ref.Although the present invention has been disclosed in the form of preferred embodiments and variations thereof, it should be understood that numerous additional modifications and variations could be made thereto without departing from the scope of the invention.For clarity, it should be understood that the use of "a" or "an" throughout this application does not exclude a plurality and that "comprising" does not exclude other steps or elements. Mention of a "unit" or "module" does not exclude the use of more than one unit or module. Regardless of the grammatical use of a term, individuals with male, female, or other sex identity are included in the term.

Claims

A computer-implemented method for generating a protocol (P syn) for a medical imaging device (50) installed at a specific site (3), the method comprising the steps of: - providing a homomorphically encrypted large language model (MT) trained to generate a homomorphically encrypted medical imaging protocol (E syn) ; - obtaining patient-specific input data (33) for an intended medical imaging task; - performing homomorphic encryption on the patient-specific input data (33); - applying the homomorphically encrypted patient-specific input data (E Seed) to the trained homomorphically encrypted large language model (MT); and performing homomorphic encryption on the generated homomorphically encrypted medical imaging protocol (E syn), to obtain a clear text medical imaging protocol (P syn) ; and wherein homomorphic encryption and decryption are performed using a first encryption scheme (S 3) specific to the location (3) and a second encryption scheme (S 5) specific to the provider (5).The computer-implemented method of the preceding claim, presenting the clear text medical imaging protocol (P syn) to a user (32).The computer-implemented method according to any one of the preceding claims, wherein the trained homomorphically encrypted large language model (ME) is provided by the provider (5).A computer-implemented method for training a big language model for use in the method of any one of claims 1 to 3, the method comprising the steps of: - modifying a big language model (5M) to comprise a homomorphically encrypted layered architecture; - obtaining homomorphically encrypted training data (E 23) based on imaging protocols (P) for the medical imaging device (50) installed at the particular customer site (3); and - training the homomorphically encrypted big language model (ME) on the homomorphically encrypted training data (E 23), to generate a homomorphically encrypted medical imaging protocol (E syn) ; and wherein homomorphic encryption and decryption are performed using the first encryption scheme (S 3) and the second encryption scheme (S 5).The computer-implemented method of claim 4, wherein the homomorphically encoded training data (E 23) is generated by: - obtaining protocol data (P, C P) for the medical imaging device (50) installed at the particular customer premises (3); - generating training data (D 23) from the protocol data (P, C P); and - performing homomorphic encoding on the training data (D 23) using the first encoding scheme (S3) and the second encoding scheme (S5).The computer-implemented method of claim 4 or claim 5, wherein the protocol data comprises at least a few hundred imaging protocols (P), preferably at least a few thousand imaging protocols (P), for a medical imaging device (50).The computer-implemented method of any of claims 4 to 6, wherein the protocol data comprises a series of imaging protocol recipe books (C P) for the medical imaging device (50).The computer-implemented method according to any one of claims 4 to 7, wherein the steps of the method are performed by the provider (5).The computer-implemented method according to any one of claims 1 to 3, wherein the homomorphically encrypted large language model (MT) has been trained using the computer-implemented method according to any one of claims 4 to 8.A data processing arrangement (1) adapted to perform the steps of the method according to any of claims 1 to 3, and comprising: - means for obtaining patient-specific input data (33) for an intended medical imaging task; - an encryption module (13) adapted to perform homomorphic encryption on the patient-specific input data (33); - a protocol generation stage (1G) adapted to apply the homomorphically encrypted patient-specific input data (33) to the trained homomorphically encrypted large language model (MT) to obtain a generated protocol (E syn) ; and a decryption module (18) configured to perform homomorphic decryption on the generated protocol (E syn) to obtain a clear text medical imaging protocol (P syn).The data processing arrangement according to the preceding claim, wherein the encryption module (13) and the decryption module (18) are located at the customer site (3).A data processing arrangement according to claim 10 or claim 11, wherein the protocol generation stage (1G) is provided by the provider (5).The data processing arrangement according to any one of claims 10 to 12, further comprising a model development stage (1T) provided by the provider (5), wherein the model development stage (1T) comprises preparation stages (11, 12, 13, 14) configured to prepare homomorphically encoded training data (D 23) from the collected protocol data (P, C P).Data processing arrangement according to claim 13, wherein the model development stage (1T) comprises further stages (13, 15, 16) configured to train a homomorphically encrypted version (ME) of the provider model (5M) with the training data (D 23).A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any one of claims 1 to 9.

Images