Method for generating protocol for medical imaging device, method for training large language model and data processing device

Abstract

A computer-implemented method of generating a protocol (Psyn) for a medical imaging device installed at a specific location, a computer-implemented method of training a large language model and a data processing device are described. A method of generating a protocol for a medical imaging device comprises: providing a homomorphic encrypted large language model (MT) trained to generate a homomorphic encrypted medical imaging protocol (Esyn); obtaining patient-specific input data for an intended medical imaging task; performing homomorphic encryption on the patient-specific input data; applying the homomorphic encrypted patient-specific input data (Eseed) to the trained homomorphic encrypted large language model (MT); and performing homomorphic decryption on the generated homomorphic encrypted medical imaging protocol (Esyn) to obtain a plain text medical imaging protocol (Psyn); wherein homomorphic encryption and decryption are performed using a location-specific first encryption scheme (S3) and a vendor-specific second encryption scheme (S5).

Description

Technical Field

[0001] The present disclosure relates to a method of generating a protocol for a medical imaging device, a method of training a large language model used in the method, and a data processing device. Background Art

[0002] In order to perform a medical imaging process so that a satisfactory image is obtained, the imaging device must be correctly configured. A set of input parameters used during an imaging task is referred to as an "imaging protocol" or "medical imaging protocol". A medical imaging protocol is a predefined process or a set of guidelines that outlines how to perform a specific medical imaging examination. These protocols are intended to standardize imaging processing to ensure consistency and reproducibility, while also ensuring that the imaging process provides the necessary diagnostic information and minimizes the patient's exposure to potential risks such as radiation. Examples of values ​​that can be specified in a medical imaging protocol include modality, which refers to 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 specific parameters may include magnetic field strength, echo time, repetition time, inversion time, and flip angle. For a CT protocol, parameters may include tube current, tube voltage, rotation time, and slice thickness. Some protocols may specify the use of contrast agents (including the type, dose, and timing of their administration) to enhance the visibility of specific tissues or structures. The protocol may also provide instructions for positioning the patient in the scanner, and for modalities such as MRI, the protocol may specify the use of specific sequences that provide different types of contrast between tissues.

[0003] Users of medical imaging devices, such as radiographic technicians, typically require extensive training in order to become familiar with the (usually numerous) features and various input parameters of the device. To select an optimal set of imaging protocol parameters for an intended imaging task, the user may refer to relevant or similar previous scans, imaging protocols, and patient results. To correctly interpret such information, the user may have acquired years of experience.

[0004] Even for experienced imaging device users, defining the correct protocol for each imaging task can be challenging. Incorrect or suboptimal imaging protocols may result in the need to repeat the imaging task (assuming a protocol defect is detected) and may even adversely affect patient care.

[0005] To improve the protocol compilation step, the vendor of the imaging device can provide templates, for example, for specific imaging tasks (e.g., neck CT, shoulder MRT, etc.). The radiographer can select the appropriate protocol and make the necessary adjustments to make the protocol appropriate for the specific patient and cross-validate to ensure that the protocol complies with any relevant protocol manuals, such as the protocol manual for that specific site or clinic.

[0006] Furthermore, the lack of standardization between different sites and different personal preferences between experienced operators may lead to significant differences in image quality, radiation dose and diagnostic results, even for the same patient and the same imaging task. Inexperienced operators may find it difficult to assemble the best protocol or select the most appropriate protocol from a manual, and the quality of the resulting diagnostic images may be suboptimal, which has a negative impact on patient care. Thus, the known method is time consuming and prone to error.

[0007] One possible way to overcome these problems could be to build a satisfactory set of protocols for a site, such as in a radiology department of a hospital. Over time, this set could include thousands of protocols. Radiographers could search this comprehensive "library of protocols" with the expectation of finding protocols for a specific imaging task that were created in the past for similar patients. However, compiling such a set of initial submissions would take a very long time, during which thousands of imaging tasks would need to be performed on many different types of patients in order to build a sufficiently large database.

[0008] In another scenario, several sites will share their collection of medical imaging protocols. However, this option is effectively ruled out due to the need to ensure patient privacy. Typically, any medical site ensures that its patient data is stored securely, for example, images and protocols may be stored in a local PACS (Picture Archiving and Communication System) that is only accessible to authorized site staff or programs installed in the site. Summary of the invention

[0009] It is therefore an object of the present invention to provide a safe and straightforward way of obtaining medical image protocols.

[0010] This object is achieved by the claimed computer-implemented method of generating a protocol for a medical imaging device and training a large language model for this purpose, and by the claimed data processing device.

[0011] For example, the computer-implemented method of generating a medical imaging protocol of the present invention is used by a radiographer to generate a protocol for input to a medical imaging device installed at a specific customer location, such as in a hospital or clinic. The computer-implemented method of the present invention includes 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 homomorphic encryption on the patient-specific input data; and applying the homomorphically encrypted patient-specific input data to the trained homomorphically encrypted large language model. Then, homomorphic decryption is performed on the generated homomorphically encrypted medical imaging protocol to obtain a plain text medical imaging protocol. In the method of the present invention, homomorphic encryption and decryption are performed using a first encryption scheme specific to the customer location and a second encryption scheme specific to the vendor of the large language model. The location-specific and vendor-specific encryption / decryption keys can be stored in a key store in an established manner, for example, in an authenticated gateway device.

[0012] The computer-implemented method of the present invention for training a large language model for use in the above method includes the following steps: modifying the large language model to include a homomorphic encryption layer architecture; obtaining homomorphic encrypted training data based on an imaging protocol for a medical imaging device installed at a specific customer site; and training the homomorphically encrypted large language model on the homomorphically encrypted training data to generate a homomorphically encrypted medical imaging protocol. In addition, homomorphic encryption and decryption are performed using a first (i.e., site-specific) encryption scheme and a second (i.e., vendor-specific) encryption scheme.

[0013] The advantage of the method of the present invention is that it provides a way to learn from a large and advantageous medical imaging protocol data set using a machine learning tool provided by a vendor, while complying with patient data privacy constraints. The homomorphically encrypted large language model is provided by a vendor, which can also provide instances of the same large language model (LLM) to various other customers. The vendor can also provide data processing services to multiple customers. The dual-scheme encryption deployed by the method of the present invention allows data processing steps to be performed at the customer site, but also allows data processing steps to be performed off-site (e.g., at the vendor site) without compromising patient data privacy. In this way, the method of the present invention makes it impossible for the vendor (or any other customer of the vendor) to access any patient data at any stage in the method of the present invention. Sensitive patient data is retained on-site (e.g., within the clinic network), and any data that leaves the site is no longer human-readable, that is, no longer in a form that makes sense to a human reader.

[0014] Furthermore, the computer-implemented method of the present invention can quickly generate imaging protocols for users and thus can help optimize the imaging workflow at a customer location (e.g., a hospital, clinic, or the like.) The user (e.g., a radiographer) only needs to provide basic information, such as patient-specific data, a descriptor of the intended imaging task, etc.

[0015] The data processing device of the present invention includes: a device for providing a homomorphically encrypted large language model; a model development stage, which is configured to develop and train the homomorphically encrypted large language model to create a homomorphically encrypted output text based on input text; a device for obtaining patient-specific input text for an intended medical imaging task; and a protocol generation stage, which is configured to apply the patient-specific input data to the trained homomorphically encrypted large language model to obtain a plain text medical imaging protocol for the intended imaging task.

[0016] The data processing device can be implemented in the form of a computer program module that can run on any suitable platform and includes program units for performing the steps of the computer-implemented method of the present invention. In the context of the present invention, a computer program module having a device for receiving input of a homomorphically encrypted supplier model and site-specific input data and a device for outputting a trained large language model is referred to herein as the "model training phase" of the data processing device; a computer program module having a device for receiving a trained homomorphically encrypted large language model and patient-specific input data and a device for outputting a plain text medical imaging protocol is referred to herein as the "protocol generation phase" of the data processing device. The data processing device of the present invention can be a distributed system in which, for example, units and modules are appropriately distributed across customer sites and supplier sites.

[0017] The computer-implemented method of the present invention can be used by a clinician, for example, a radiographic technician or a radiologist. The computer program product can be configured to present the generated imaging protocol to the user in any suitable manner (e.g., displayed on a screen of a device such as a desktop computer, a tablet computer, etc.).

[0018] In particular, the generated plain text medical imaging protocol can be used to control a medical imaging device installed at a specific location. In one embodiment, the medical imaging protocol can be used to manage the operation of the medical imaging device, for example, to start, pause or terminate an imaging process. In another embodiment, the medical imaging protocol can be used to adjust the imaging parameters of the device, including but not limited to the intensity, frequency and duration of the imaging signal (x-ray signal, magnetic signal and radio frequency signal).

[0019] The object of the present invention is also achieved by a computer program product having a computer program, which is directly loadable into a memory of a data processing device and which comprises a program element for performing the steps of the computer-implemented method of the present invention when the program is executed. The computer program product can be stored on a computer-readable data carrier.

[0020] As disclosed in the following description, particularly advantageous embodiments and features of the invention are given by the dependent claims. Features of different claim categories may be combined as appropriate to give further embodiments not described herein.

[0021] As used in the context of the present invention, the term "patient-specific input data" may refer to any data that is unique to an individual patient and used in the context of the patient's healthcare management. Such data is typically used to inform clinical decisions, guide treatment plans, and monitor patient progress, and may be obtained from a variety of sources, including medical history, physical examination, laboratory tests, imaging studies, and patient-reported results. Examples of patient-specific input data include demographic information (including the patient's age, gender, race, and other socioeconomic factors that may affect health outcomes), medical history (which covers past and current illnesses, surgeries, allergies, and medication use), physical examination findings (e.g., observations made during a physical examination by a healthcare provider, e.g., heart rate, blood pressure, and findings from examinations of various body systems), and / or laboratory test results (quantitative and qualitative data obtained from analysis of biological samples, e.g., blood or urine, including values ​​such as blood glucose levels, cholesterol levels, and complete blood count results). Any patient-specific input data used in the compilation of an imaging protocol is typically in the form of words and numbers, and may also be referred to as "patient-specific input text."

[0022] In the context of the present invention, a supplier should be understood as supplying a medical imaging device to a customer and also supplying the customer with software for operating the device. The same supplier may also provide the customer with a data processing device for performing the computer-implemented method of the present invention.

[0023] The imaging modality or medical imaging device intended for the generated or synthesized protocol 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.

[0024] In the context of the present invention, expressions such as "machine learning algorithm", "big language model", "artificial intelligence tool" and "AI model" should be understood as synonyms and can be used interchangeably. A big language model is an artificial neural network that is particularly suitable for applications involving text understanding or text generation. The machine learning algorithm is trained to generate a medical imaging protocol for the intended imaging task using only multiple patient-specific parameters. Any suitable LLM architecture can be used, for example, a long short-term memory (LSTM) network architecture. The selected LLM architecture is further modified and updated by the vendor to include a homomorphic encryption layer architecture for a highly secure privacy-preserving machine learning method according to the present invention. Additional algorithms such as a cross entropy loss function module and an Adam optimizer module can be implemented. Any algorithm required for the method of the present invention may be provided by the vendor and / or a third party.

[0025] In the preparation phase, a homomorphic encrypted version of the vendor model is generated using the client encryption scheme and the vendor encryption scheme as described above. As known to those skilled in the art, homomorphic encryption is a form of encryption that allows computations to be performed on ciphertext, thereby generating an encrypted result that, when decrypted, matches the result of the operation performed on the plaintext. This feature allows third-party systems to process data without the need to decrypt the data, thereby protecting data privacy and data security. The homomorphic encrypted large language model is referred to as "HE-LLM" hereinafter.

[0026] In the method of the present invention, training the HE-LLM comprises the following steps: compiling a training data set; assigning a ground truth value to the training data set; and applying the HE-LLM to the training data set so as to approximate the ground truth value. These steps are repeated for all available training data sets and for a suitable number of cycles until the desired level of accuracy has been achieved.

[0027] In the method of the present invention, the training data is based on medical imaging protocols that have been previously created at the customer location (referred to herein as a site). For example, a site may maintain a database of all medical imaging protocols for its imaging modalities. If necessary, the protocols are converted into a common format, such as a suitable markup language (e.g., XML). Since the accuracy of a machine learning algorithm depends largely on the number of data sets it is trained on, in a preferred embodiment of the present invention, at least hundreds of medical imaging protocols, more preferably thousands of medical imaging protocols, are used to train the LLM.

[0028] The site may also deploy a medical imaging protocol manual to ensure that site-specific standards are followed during each imaging procedure. Therefore, in a particularly preferred embodiment of the present invention, the protocol data includes any relevant medical imaging protocol manuals for the medical imaging device. The protocol manuals in PDF format may be scanned and digitized and converted into the same format for the protocol.

[0029] The LLM can also be trained using synthetically generated medical imaging protocols and / or protocol manuals to improve the accuracy of the LLM. By training the LLM on thousands of validated or approved protocols used for past imaging tasks, it is expected that the synthesized protocols are very accurate, so that the method of the present invention can help minimize the radiation dose to which the patient is exposed.

[0030] To prepare the training dataset, tokenization and padding are performed on each input dataset. 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 performed as needed. In the "cleaning" stage, duplicate or irrelevant datasets are removed. At the completion of this data preprocessing stage, the padded tokenized dataset is divided into training, validation, and test subsets.

[0031] Homomorphic encryption is then performed on each subset of the padded tokenized data using the customer encryption scheme and the vendor encryption scheme as described above. In a preferred approach, approximately 70% of the homomorphic encrypted data set is used to train the model, and the remaining data set is used for the testing phase and the validation phase, as will be known to those skilled in the art. Of course, the subset size can be adjusted based on the total number of available data sets to further improve accuracy. Only at this stage, i.e., after homomorphic encryption, can the data set leave the premises. Since the data is no longer in a human-readable form, the security of the patient data is ensured.

[0032] The embedding layer then maps the encrypted token representation into a dense vector representation that preserves the semantic relationships between the encrypted tokens. The resulting encrypted dense vector is passed through an encrypted dense (fully connected) layer with an activation function in the form of a rectified linear unit (ReLU). ReLU introduces nonlinearity into the model, allowing the model to capture complex relationships in the data. A softmax (normalized exponential function) activation is then applied to produce a probability distribution over multiple categories, i.e., predicting the likelihood that each word in the vocabulary is the next word in the sequence. The result is a probability distribution over the vocabulary.

[0033] The constructed model is then compiled using a homomorphically encrypted version of the categorical cross entropy loss function and a homomorphically encrypted version of the Adam optimizer. Categorical cross entropy is a known loss function for multi-class classification problems that compares the predicted probability distribution to the actual distribution (usually a one-time encoded vector), allowing the model to determine its training progress. The Adam optimizer is a known optimization algorithm for minimizing the loss function during training.

[0034] The compiled homomorphic encryption model is then trained using homomorphic encrypted training data and homomorphic encrypted target labels to predict the next word given a sequence of words. During training, the input data consists of a sequence of tokens (words or subwords) from the training data, and the target label is the token that immediately follows the training text sequence.

[0035] Training is performed for a suitable number of cycles, and the result is a trained homomorphic encryption model that will synthesize a homomorphically encrypted medical imaging protocol when fed with appropriate textual input. Training can be supervised, unsupervised, or self-supervised. The supervised training process preferably includes back-propagation techniques for finding the best set of parameters that will enable the HE-LLM to infer ground truth from the input, in which case the most appropriate protocol entry is inferred to follow the previous protocol entry. The training process is preferably configured to achieve advantageous low-cost functionality. Of course, the HE-LLM can continue to learn by including any verified protocol in the training data set.

[0036] The trained homomorphic encryption model can then be deployed whenever a user needs a medical imaging protocol for an intended imaging task. To obtain an imaging protocol, the user prepares an initial input sequence or seed text. The input sequence may include multiple pairs of field descriptors and field values ​​to characterize a patient (e.g., "PatientId: PD0001"; "Age: 45"; "Previous surgery: None", etc.). Preferably, the protocol generation phase performs any necessary tokenization and padding on the input sequence, and performs homomorphic encryption on the tokenized and padded data. Using this input, the trained homomorphic encryption model (which can be run at an off-site location, on a cloud computing platform, etc.) creates a homomorphic encrypted text, which is returned to the user at the site. In order to convert the synthesized protocol into a usable (i.e., human-readable) protocol, the decryption phase then applies homomorphic decryption using a customer encryption scheme and a vendor encryption scheme to obtain a plain text output protocol, which is then presented to the user, for example, on a monitor of a desktop computer. The fields of the synthesized protocol can be presented together with the fields of the standard protocol so that the user can easily identify any differences between the values ​​of these fields. For example, if the parameter "T1-weighted TR" of the synthesized protocol suggests a value of 500 ms, and the corresponding field of the standard protocol gives a range of 400 ms to 800 ms, the user may determine that the suggested value is acceptable. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Other objects and features of the present invention will become apparent from the following detailed description considered in conjunction with the accompanying drawings. It should be understood, however, that the drawings are designed only for the purpose of illustration and not as a definition of the limits of the present invention.

[0038] Figure 1 The principle of the present invention is shown;

[0039] Figure 2 is an exemplary flow chart of the method of the present invention;

[0040] Figure 3 and Figure 4 A schematic diagram showing a data processing device of the present invention is shown;

[0041] Figure 5 An exemplary imaging protocol synthesized using the computer-implemented method of the present invention is shown.

[0042] In the drawings, like reference numerals refer to like objects throughout. Objects in the drawings are not necessarily drawn to scale. DETAILED DESCRIPTION

[0043] Figure 1The principle of the present invention is shown. The figure shows a site 3, which can be a hospital, a clinic, a radiology clinic, etc. The site 3 can deploy any number of medical imaging devices 50 for imaging modalities, such as CT, MRT, etc. Patient data is securely stored in, for example, a PACS database 30. The supplier 5 supplies software and hardware to the site 3, such as medical imaging devices 50 and software for operating the devices 50, client servers 51 for processing and analyzing medical images, and for managing the resulting data. The supplier 5 also provides software updates as needed, and can also provide data processing services on the supplier-side server 51, cloud computing services, etc.

[0044] Here, the supplier 5 also provides an AI model 5M. Using the locally stored data P, C P The homomorphic encryption model ME is trained for site 3, and then, as will be explained below, the trained homomorphic encryption model MT can be used to generate a medical imaging protocol P for a customer. syn .

[0045] Data security is ensured by deploying encryption at several stages during the development and implementation of the computer-implemented method.Encryption is performed using an encryption scheme S3 unique to site 3 and an encryption scheme S5 unique to supplier 5.

[0046] The units and modules of the data processing apparatus 1 may be distributed in any suitable configuration. Figure 1 In the configuration, the model development phase 1T for training the model 5M and the model development phase 1T for generating the protocol E syn The agreement generation phase 1G of is implemented at the site 3 (customer location). Of course, the model development phase 1T can be implemented at the supplier 5; similarly, the agreement generation phase 1G can be implemented at the supplier 5. Figure 2 and Figure 3 A block diagram is shown for explaining an exemplary embodiment of a data processing device 1 according to the present invention.

[0047] Figure 2 An exemplary sequence of steps performed by the method of the present invention is shown in FIG. Data collection is performed in a first step 21. To this end, a collector stage 11 collects a diverse and representative data set of medical imaging protocols P for various medical conditions. These protocols P originate from a local site 3. One or more local protocol manuals C P The collected data preferably includes a wide range of information, such as patient demographics, clinical history, scan parameters, corresponding results, etc., and preferably covers a wide range of anatomical regions and pathologies.

[0048] If necessary, the information P, C PA digitization step can also be performed at this stage. For example, a protocol manual in PDF format can be scanned and digitized into a suitable format, such as XML markup language. The data collected in this way 21 A collection of documents is included, namely a dataset of machine-readable protocols and optionally a dataset of machine-readable protocol manuals.

[0049] In the subsequent step 22, the collected data set D 21 is loaded into the preprocessing module 12, which processes the input data D 21 Tokenization and padding are performed. During tokenization, the text of each document is converted into a sequence of tokens, e.g., integers. To ensure that the tokenized sequences have the same length, padding is performed as needed. The data preprocessing step may also include a "cleaning" phase to remove duplicate or irrelevant entries. In this phase, a consistent format and structure may be applied to the data. At the completion of the preprocessing step 22, the padded tokenized data D is as described above. 22 It is divided into training subset, validation subset and test subset.

[0050] In a subsequent step 23, the encryption phase 13 applies homomorphic encryption to the padded tokenized data D 22 For each subset of E, use the homomorphic encryption scheme S3 unique to site 3 and the homomorphic encryption scheme S5 unique to supplier 5 to obtain data E 23 Homomorphic encryption representation of .

[0051] In the subsequent step 24, the embedding layer 14 extracts the homomorphically encrypted representation E 23 Map each homomorphically encrypted token to obtain a homomorphically encrypted dense vector representation E 24 .

[0052] In the subsequent step 25, homomorphic encryption is applied to each layer of the HE-LLM 5M provided by the vendor. To build the HE-LLM model, the output of the encrypted layer is fed into an encrypted dense layer with a rectified linear unit (ReLU) activation. Then, a softmax (normalized exponential function) activation is applied to predict the homomorphically encrypted probability distribution of the next word over the vocabulary. The model is then compiled using an encrypted version of the categorical cross entropy loss function 54 and an encrypted version of the Adam optimizer 55.

[0053] In the subsequent step 26, the training phase 16 utilizes the homomorphically encrypted training data E 23Train the compiled homomorphically encrypted model MC. Using plain text as an example, the input sequence could be “no evidence of cysts or tumors”. The corresponding tokens are [“no”, “evidence”, “of”, “cysts”, “or”, “tumors”], and possible training sequences could be [“no”, “evidence”, “of”] -> target: “cysts”; [“of”, “cysts”, “or”] -> target: “tumors”; etc. In the method of the present invention, training the LLM is performed exclusively on homomorphically encrypted data, and can therefore be performed “off-site” without any risk of compromising the security of patient data. To this end, the homomorphically encrypted training data E 23 and homomorphically encrypted target label E L is passed to a fitting function, such as a stochastic gradient descent function, a batch gradient descent function, a transfer learning function, etc. The model is trained for a suitable number of cycles. The output of this training phase 16 is the trained homomorphic encryption model MT.

[0054] Figure 4 The predictor stage 17 in indicates that the generation and prediction of the output text is handled in the subsequent step 27. Here, the input sequence 33 provided by the user is homomorphically encrypted (after being tokenized and padded as needed). The trained homomorphic encryption model MT is based on the homomorphically encrypted input sequence E seed Predict the next word w pt The seed text is updated by appending the predicted words. The process of generating predictions and updating the seed text is repeated until the desired number of words (i.e., the synthesized protocol E) is reached. syn ) has been generated.

[0055] In the final step 28, the text E obtained at the final iteration is syn Passed to the decryption stage 18, which applies homomorphic decryption (using the site's encryption scheme S3 and the vendor's encryption scheme S5) to obtain the plain text composite output protocol P syn , which is then presented to the user, for example, in tabular form.

[0056] All steps of building and training model 5M using a homomorphically encrypted training data set generated locally, i.e., at a hospital or clinic site, can be performed off-site. Similarly, the trained model 5M can be tested and verified off-site using a homomorphically encrypted test data set and a validation data set. In this way, sensitive patient data is retained locally or on-site, while resource and time-intensive model generation is outsourced, for example, to a server cluster (on-site or cloud-based) provided by a supplier. Then, users of the site can use the trained and verified model 5M (which can be run at an off-site location, such as on a cloud computing server) at any time. The user only needs to provide the homomorphically encrypted basic related seed text before leaving the site. Model 5M returns the synthesized complete imaging protocol to the site. The homomorphically encrypted protocol is homomorphically decrypted to make it human-readable.

[0057] The protocol P generated by the data processing device 1 of the present invention syn can be presented together with the standard protocol so that the user can easily identify any potential differences, such as Figure 5 The figure shows the synthesized imaging protocol P syn , which includes a parameter list 60 related to the imaging task, as in the synthesized imaging protocol P syn The values ​​presented in the table 61 and by appropriate reference imaging protocols P ref The figure shows the synthesized imaging protocol P syn The value 61 in the manual protocol P ref The allowed range is 62.

[0058] Although the present invention has been disclosed in the form of preferred embodiments and variations thereon, it will be understood that numerous additional modifications and variations could be made thereto without departing from the scope of the invention.

[0059] For the sake of clarity, it should be understood that the use of "a" or "an" throughout this application does not exclude a plurality, and "comprising" does not exclude other steps or elements. Reference to a "unit" or "module" does not exclude the use of more than one unit or module. Regardless of the grammatical term usage, individuals with masculine or feminine identities are included in the terms.

Claims

1. A method for generating a protocol (P) for a medical imaging device (50) syn ), wherein the medical imaging device (50) is installed at a specific location (3), the method comprising the following steps: - Provides a protocol trained to generate homomorphically encrypted medical imaging protocols (E syn )’s homomorphic encryption large language model (MT); - obtaining patient-specific input data (33) for an intended medical imaging task; - performing homomorphic encryption on the patient-specific input data (33); - The homomorphically encrypted patient-specific input data (E seed ) is applied to a trained homomorphically encrypted large language model (MT); as well as -Generated homomorphic encryption medical imaging protocol (E syn ) performs homomorphic decryption to obtain the plain text medical imaging protocol (P syn ); and wherein, Homomorphic encryption and decryption are performed using a first encryption scheme (S3) specific to a particular location (3) and a second encryption scheme (S5) specific to a provider (5).

2. The computer-implemented method of claim 1 , presenting the plain text medical imaging protocol (P syn ).

3. A computer-implemented method according to any one of the preceding claims, wherein: The trained homomorphically encrypted large language model (MT) is provided by the supplier (5).

4. A computer-implemented method of training a large language model for use in the method according to any one of claims 1 to 3, the method comprising the steps of: - Modified the large language model (5M) to include a homomorphic encryption layer architecture; - obtaining homomorphically encrypted training data (E) based on an imaging protocol (P) for a medical imaging device (50) installed at a specific location (3) 23 ); as well as - in the homomorphically encrypted training data (E 23 ) to train a homomorphically encrypted large language model (MT) to generate a homomorphically encrypted medical imaging protocol (E syn ); And wherein, homomorphic encryption and decryption are performed using a first encryption scheme (S3) and a second encryption scheme (S5).

5. The computer-implemented method of claim 4, wherein: The homomorphically encrypted training data (E 23 ) is generated by the following steps: - obtaining protocol data (P, C) for a medical imaging device (50) installed at the specific location (3); P ); - According to the protocol data (P, C P ) Generate training data (D 23 ); as well as - using the first encryption scheme (S3) and the second encryption scheme (S5) to encrypt the training data (D 23 ) performs homomorphic encryption.

6. A computer-implemented method according to claim 4 or claim 5, wherein: The protocol data includes at least several hundreds of imaging protocols (P), preferably at least several thousands of imaging protocols (P) for the medical imaging device (50).

7. A computer-implemented method according to any one of claims 4 to 6, wherein: The protocol data includes a plurality of imaging protocol manuals (C P ).

8. A computer-implemented method according to any one of claims 4 to 7, wherein: The steps of the method are performed by the supplier (5).

9. The computer-implemented method of any one of claims 1 to 3, wherein: The homomorphically encrypted large language model (MT) is trained using a computer-implemented method according to any one of claims 4 to 8.

10. A data processing device (1) adapted to perform the steps of the method according to any one of claims 1 to 3, and comprising: - means for obtaining patient-specific input data (33) for a desired medical imaging task; - an encryption module (13) adapted to perform homomorphic encryption on said patient-specific input data (33); a protocol generation phase (1G) configured 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 ); as well as - a decryption module (18) adapted to decrypt said generated protocol (E syn ) performs homomorphic decryption to obtain the plain text medical imaging protocol (P syn ).

11. The data processing device according to claim 10, wherein: The encryption module (13) and the decryption module (18) are located at a specific location (3).

12. The data processing apparatus according to claim 10 or claim 11, wherein: The protocol generation stage (1G) is provided by a supplier (5).

13. The data processing device according to any one of claims 10 to 12, further comprising a model development phase (1T) provided by a supplier (5), the model development phase (1T) comprising a module configured to generate a model based on the collected protocol data (P, C P ) Prepare homomorphic encrypted training data (D 23 )’s preparation stage (11, 12, 13, 14).

14. The data processing apparatus according to claim 13, wherein: The model development phase (IT) includes a process configured to utilize the training data (D 23 ) Additional stages (13, 15, 16) of training a homomorphically encrypted version (ME) of the vendor model.

15. A computer program product comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 9.

Images