A magnetic resonance scan parameter determination system based on a large language model

Abstract

The application relates to the technical field of magnetic resonance examination, in particular to a magnetic resonance scanning parameter determination system based on a large language model, which comprises the following modules: a first acquisition module, which obtains examination items and personal information of a patient to be examined; a description module, which constructs scanning description information according to the examination items and the personal information of the patient; a model interaction module, which inputs the scanning description information into an external large language model to obtain a scanning scheme fed back by the large language model; and a second acquisition module, which establishes scanning intermediate data based on an echo sequence; and the model interaction module inputs the scanning intermediate data into the large language model to obtain an adjusted scanning scheme. The beneficial effect is that, by introducing the large language model, the scanning scheme is well constructed in real time according to the processed scanning description information, and the large language model is fed back according to the scanning intermediate image in the scanning process to obtain the adjusted scanning scheme, so that a better scanning effect is achieved.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging (MRI) technology, and more specifically to an MRI scanning parameter determination system based on a large-scale language model. Background Technology

[0002] MRI is a commonly used medical computed tomography method that uses the magnetic resonance phenomenon to obtain electromagnetic signals from the human body and reconstruct information about the body. This technology utilizes the principle of nuclear magnetic resonance, based on the different attenuations of released energy in different structural environments within a substance. By detecting the emitted electromagnetic waves through an applied gradient magnetic field, the location and type of atomic nuclei that make up the object can be determined, allowing for the creation of an image of the object's internal structure. Since two-thirds of the human body's weight is water, and the water content in organs and tissues varies, many pathological processes cause changes in the form of water, which can be reflected in MRI images.

[0003] Current MRI examination procedures employ pre-designed scanning protocols for different scanned areas and corresponding conditions. These protocols utilize various scanning sequences to achieve optimal scanning results for the areas to be scanned. Each scanning sequence is configured with different scanning parameters. During the actual scan, the physician selects the appropriate scanning sequence according to the pre-designed protocol and adjusts the MRI equipment parameters based on their personal experience to achieve the best imaging results.

[0004] However, in actual implementation, the inventors found that, due to the different physical conditions of different patients, the above-mentioned fixed scanning parameters often cannot achieve good scanning results during the scanning process. They usually need to be adjusted based on the doctor's personal experience, which makes the adjustment process highly dependent on the doctor's experience and may lead to unsatisfactory examination results. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, a magnetic resonance imaging (MRI) scan parameter determination system based on a large-scale language model is provided.

[0006] The specific technical solution is as follows:

[0007] A magnetic resonance imaging (MRI) scan parameter determination system based on a large language model includes:

[0008] The first acquisition module acquires the examination items and the patient's personal information from the patient to be examined.

[0009] A description module is connected to the first acquisition module, and the description module constructs scanning description information based on the items to be examined and the patient's personal information;

[0010] The model interaction module is connected to the description module and an external large language model. The model interaction module inputs the scanning description information into the external large language model to obtain the scanning plan fed back by the large language model.

[0011] The scanning scheme includes multiple scanning steps and scanning parameters corresponding to each scanning step;

[0012] The second acquisition module acquires echo sequences during the scanning process and establishes intermediate scanning data based on the echo sequences.

[0013] The model interaction module is connected to the second acquisition module. The model interaction module inputs the intermediate scanning data into the large language model to obtain the adjusted scanning plan.

[0014] On the other hand, the first acquisition module includes:

[0015] An interface module is provided, which connects to an external hospital information system and receives the electronic medical record of the patient to be examined from the hospital information system.

[0016] A field extraction module, which is connected to the interface module, extracts the items to be examined and the patient's personal information from the electronic medical record according to field attributes.

[0017] On the other hand, the description module includes:

[0018] The project identification module determines the part to be inspected based on the item to be inspected.

[0019] A template matching module, which is connected to the project identification module, calls an inspection description template according to the part to be inspected;

[0020] A filling module, which is connected to the template matching module, fills the examination description template according to the patient's personal information to obtain the scan description information.

[0021] On the other hand, the model interaction module includes:

[0022] A semantic recognition module receives the scan description information and extracts keywords and keyword description information corresponding to the keywords according to the context of the scan description information.

[0023] The model input module acquires multiple description groups contained in the scan description information and inputs the description groups sequentially into the large language model.

[0024] The verification module is connected to the semantic recognition module and the model input module respectively. After the description group is input into the model input module, the verification module receives the feedback content of the large language model.

[0025] The verification module matches the feedback content according to the keywords and the keyword description information, and controls the model input module to re-input the description group when a match cannot be found;

[0026] A model output module is connected to the model input module. The model output module receives the feedback content, assembles it to obtain the scanning plan, and outputs it.

[0027] On the other hand, the semantic recognition module uses a semantic recognition model to recognize the scanning description information, and the semantic recognition model includes:

[0028] A long short-term memory network, wherein the long short-term memory network sequentially receives the input scan description information;

[0029] A word frequency extraction module is connected to the long short-term memory network. The word frequency extraction module segments the scanned description information into words and counts word frequencies.

[0030] The keyword acquisition module is connected to the word frequency extraction module. After traversing the scanned description information, the keyword acquisition module acquires the keywords according to the word frequency.

[0031] The association module is connected to both the Long Short-Term Memory Network and the keyword acquisition module. The association module extracts information from the Long Short-Term Memory Network based on the associated words to obtain the keyword description information.

[0032] On the other hand, the verification module includes:

[0033] A keyword determination module, which matches the corresponding keywords according to the currently input description group;

[0034] A word vector construction module, which is connected to the keyword determination module, constructs a first word vector based on the keyword description information of the matched keywords, and constructs a second word vector based on the feedback content;

[0035] The comparison module is connected to the word vector construction module. The comparison module compares the first word vector and the second word vector and controls the model input module based on the comparison result.

[0036] On the other hand, the second acquisition module includes:

[0037] An image reconstruction module receives the echo sequence and performs image reconstruction on the echo sequence to obtain an intermediate scan image;

[0038] An image recognition module, connected to the image reconstruction module, identifies the intermediate scan image to obtain at least one marker point;

[0039] A point description module is connected to the image recognition module. The point description module constructs point description information corresponding to each of the marker points according to the marker points and the scanning parameters in the echo sequence.

[0040] A discrimination module is connected to the point description module. The discrimination module compares the point description information with the scanning plan to output a discrimination result.

[0041] A feedback module is connected to the discrimination module, and the feedback module controls the model interaction module according to the discrimination result.

[0042] On the other hand, the image recognition module is implemented using a morphological model, which includes:

[0043] A feature extraction network extracts high-dimensional features from the scanned intermediate image;

[0044] A feature matching module is connected to the feature extraction network, and the feature matching module matches the label features from the high-dimensional features;

[0045] A prediction module, which is connected to the feature matching module, predicts the location of the marker point from the marker features.

[0046] The above technical solution has the following advantages or beneficial effects:

[0047] To address the issue that existing MRI examination methods may lead to poor examination results due to limitations in physician experience, this solution introduces a large-scale language model. Based on the processed scan description information, it achieves better real-time construction of the scan plan. Furthermore, by feeding back intermediate scan images to the large-scale language model during the scan process, an adjusted scan plan is obtained, thereby achieving better scan results. Attached Figure Description

[0048] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0049] Figure 1 This is an overall schematic diagram of an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the first acquisition module in an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram illustrating the modules described in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the model interaction module in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the semantic recognition model in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of the verification module in an embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of the second acquisition module in an embodiment of the present invention;

[0056] Figure 8 This is a schematic diagram of the morphological model in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0059] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0060] This invention includes:

[0061] A magnetic resonance imaging (MRI) scan parameter determination system based on a large language model, such as Figure 1 As shown, it includes:

[0062] First acquisition module 1: First acquisition module 1 acquires the examination items and personal information of the patient to be examined;

[0063] Description module 2 is connected to the first acquisition module 1. Description module 2 constructs scanning description information based on the items to be examined and the patient's personal information.

[0064] Model interaction module 3 connects to description module 2 and an external large language model. Model interaction module 3 inputs the scanning description information into the external large language model to obtain the scanning plan fed back by the large language model.

[0065] The scanning plan includes multiple scanning steps and scanning parameters corresponding to each scanning step;

[0066] The second acquisition module 4 acquires the echo sequence during the scanning process and establishes intermediate scanning data based on the echo sequence;

[0067] The model interaction module 3 is connected to the second acquisition module 4. The model interaction module 3 inputs the intermediate scanning data into the large language model to obtain the adjusted scanning plan.

[0068] Specifically, addressing the issue that existing scanning schemes are limited by the doctor's experience, potentially leading to unsatisfactory scanning results, this embodiment acquires the examination items and personal information of the patient through the first acquisition module 1. This information is then described by the description module 2, constructing scanning description information to characterize the patient's condition and examination items. Subsequently, this information is input into an external large-scale language model, which constructs the scanning scheme, including the sequential scanning steps and the scanning parameters to be used at each step. The doctor then performs an MRI scan based on this scheme and sends the received echo sequences to the second acquisition module 4, enabling the second acquisition module 4 to construct intermediate scanning data. Upon reaching a predetermined scanning node, the model interaction module 3 inputs the intermediate scanning data into the large-scale language model to obtain an adjusted scanning scheme. These modules enable the determination and automatic adjustment of the scanning scheme, resulting in better scanning performance. The large-scale language model can be an existing model implemented on a public cloud, or a locally deployed model constructed using a specific medical dataset, depending on actual needs.

[0069] In one embodiment, such as Figure 2 As shown, the first acquisition module 1 includes:

[0070] Interface module 11 connects to an external hospital information system and receives the electronic medical records of the patients to be examined from the hospital information system.

[0071] Field extraction module 12, connected to interface module 11, extracts the examination items and patient personal information from the electronic medical record according to field attributes.

[0072] Specifically, to achieve better adaptability, in this embodiment, an interface module 11 and a field extraction module 12 are configured in the first acquisition module 1. The first acquisition module 1 interfaces with the hospital information system to directly obtain the incoming electronic medical records from the hospital information system. These electronic medical records may have different formats due to limitations imposed by different hospital information system providers. Subsequently, the field extraction module 12 iterates through multiple fields, identifies their attributes, and thereby filters out the items to be examined and the patient's personal information.

[0073] In one embodiment, such as Figure 3 As shown, module 2 includes:

[0074] Project identification module 21 determines the part to be inspected according to the project to be inspected;

[0075] Template matching module 22 is connected to project identification module 21. Template matching module 22 calls the inspection description template according to the part to be inspected.

[0076] Filling module 23 is connected to template matching module 22. Filling module 23 fills the examination description template according to the patient's personal information to obtain the scan description information.

[0077] Specifically, to enable the large language model to better identify the relevant examination information, in this embodiment, an item recognition module 21 is configured in the description module 2. This module can identify the items to be examined, thereby obtaining the parts of the patient that need to undergo MRI examination, such as the head and chest cavity. Subsequently, the template matching module 22 calls the corresponding examination description template according to the parts to be examined, which contains pre-compiled text content. Then, the filling module 23 fills in the patient's personal information, including previous test results, height, weight, etc., which enables the large language model to better determine the physical condition of the patient to be examined, thereby adjusting the existing scanning plan and achieving better scanning results.

[0078] In one embodiment, such as Figure 4 As shown, the model interaction module 3 includes:

[0079] The semantic recognition module 31 receives the scanning description information and extracts the keywords and corresponding keyword description information according to the context of the scanning description information.

[0080] Model input module 32 obtains multiple description groups contained in the scanned description information and inputs the description groups into the large language model in sequence;

[0081] The verification module 33 is connected to the semantic recognition module 31 and the model input module 32 respectively. After the model input module 32 inputs the description group, the verification module 33 receives the feedback content of the large language model.

[0082] The verification module 33 matches the feedback content according to the keywords and keyword description information, and controls the model input module to re-enter the description group when a match cannot be found;

[0083] Model output module 34 is connected to model input module 32. Model output module 34 receives feedback content, assembles it to obtain a scanning plan, and outputs it.

[0084] Specifically, to ensure a more accurate scanning scheme from the large language model, in this embodiment, before inputting the scanning description information into the model, the semantic recognition module 31 pre-identifies the keywords and keyword descriptions. Then, the model input module 32, following a pre-programmed input order, groups the scanning description information and inputs it sequentially into the large language model, obtaining feedback from the model. When the large language model outputs feedback content, the verification module 33 receives the feedback content and pre-matches it against the keywords and keyword descriptions that should be included in the current description group, thereby determining whether the large language model has correctly output the scanning scheme. If yes, the model output module 34 receives and stores the feedback content; otherwise, the model input module 32 re-inputs the content to adjust the model's output. After all description groups have been input, the model output module 34 concatenates all stored feedback content to obtain the scanning scheme for output.

[0085] In one embodiment, the semantic recognition module 31 uses the semantic recognition model 31A to recognize the scanning description information, such as... Figure 5 As shown, the semantic recognition model 31A includes:

[0086] Long Short-Term Memory Network 31A1 receives the input scan description information sequentially;

[0087] The word frequency extraction module 31A2 is connected to the long short-term memory network 31A1. The word frequency extraction module 31A2 segments the scanned description information into words and counts the word frequency.

[0088] Keyword acquisition module 31A3 is connected to word frequency extraction module 31A2. After traversing and scanning the description information, keyword acquisition module 31A3 acquires keywords according to word frequency.

[0089] The association module 31A4 is connected to the long short-term memory network 31A1 and the keyword acquisition module 31A3 respectively. The association module 31A4 extracts the long short-term memory network 31A1 according to the associated words to obtain the keyword description information.

[0090] Specifically, to achieve better recognition results, in this embodiment, a Long Short-Term Memory (LSTM) network 31A1 is set in the semantic recognition model 31A to store directed text sequences and recognize the text information therein; then, the word frequency extraction module 31A2 performs word segmentation according to the text sequence stored in the LSM network 31A1 to obtain several word segmentation results, and statistically analyzes the word segmentation results to obtain word frequencies; then, the keyword acquisition module 31A3 obtains keywords according to the statistically obtained word frequencies, and the keyword acquisition module 31A3 extracts keyword description information from the LSM network 31A1 according to the contextual relationship of the keywords.

[0091] In one embodiment, such as Figure 6 As shown, the verification module 33 includes:

[0092] Keyword determination module 331: The keyword determination module 331 matches corresponding keywords according to the currently input description in groups;

[0093] The word vector construction module 332 is connected to the keyword determination module 331. The word vector construction module 332 constructs a first word vector based on the keyword description information of the matched keywords, and constructs a second word vector based on the feedback content.

[0094] The comparison module 333 is connected to the word vector construction module 332. The comparison module 333 compares the first word vector and the second word vector, and controls the model input module 32 according to the comparison result.

[0095] Specifically, to achieve better verification results, in this embodiment, the keyword determination module 331 pre-matches the keywords contained in the current description group according to the input description group. Subsequently, the word vector construction module 332 constructs its word vector as the first word vector based on the keyword description information corresponding to the keyword and semantic recognition technology. Correspondingly, the received feedback content is also identified to construct the second word vector. Based on the first word vector and the second word vector, the comparison module 333 can calculate their angle and determine whether it is within the matching range. If it is, it indicates that the feedback content contains the corresponding information. If not, the model input module 32 is controlled to re-input the content to make the feedback content accurate.

[0096] In one embodiment, such as Figure 7 As shown, the second acquisition module 4 includes:

[0097] Image reconstruction module 41 receives the echo sequence and performs image reconstruction on the echo sequence to obtain the intermediate scan image;

[0098] Image recognition module 42 is connected to image reconstruction module 41. Image recognition module 42 recognizes intermediate scan images to obtain at least one marker point.

[0099] Point description module 43 is connected to image recognition module 42. Point description module 43 constructs point description information corresponding to each marker point according to the marker point and the scanning parameters in the echo sequence.

[0100] The discrimination module 44 is connected to the point description module 43. The discrimination module 44 compares the point description information with the scanning plan to output the discrimination result.

[0101] Feedback module 45 is connected to discrimination module 44, and feedback module 45 controls model interaction module according to discrimination result.

[0102] Specifically, to achieve better adjustment of the scanning scheme, in this embodiment, the image reconstruction module 41 performs image reconstruction on the real-time received echo sequence. Subsequently, the image is input into the image recognition module 42 for image recognition, thereby identifying the landmark points of the scanning area. Based on these landmark points and scanning parameters, such as duration and depth, the landmark point description module 43 constructs landmark point description information, and the discrimination module 44 determines whether the correct landmark point has appeared in the current scanning step, thus forming a discrimination result. Based on this discrimination result, the feedback module 45 can optionally control whether the model interaction module 3 submits intermediate scanning images for the large language model to use as a reference, thereby adjusting the scanning scheme.

[0103] In one embodiment, the image recognition module 42 is implemented using a morphological model 42A, such as Figure 8 As shown, morphological model 42A includes:

[0104] Feature extraction network 42A1 extracts high-dimensional features from scanned intermediate images;

[0105] Feature matching module 42A2 is connected to feature extraction network 42A1. Feature matching module 42A2 obtains the label feature from high-dimensional features.

[0106] The prediction module 42A3 is connected to the feature matching module 42A2, and the prediction module 42A3 predicts the location of the marker point from the marker features.

[0107] Specifically, to achieve better recognition results, a morphology module 42A is introduced in this embodiment for recognition. In this model, a feature extraction network 42A1 composed of a series of convolutional layers extracts high-dimensional features from the scanned intermediate image. Then, the feature matching module 42A2 matches the high-dimensional features to obtain the marker features, and then the prediction module 42A3 predicts the position of the marker point, thus achieving a better recognition result.

[0108] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A large language model based magnetic resonance scan parameter determination system, characterized in that, include: The first acquisition module acquires the examination items and the patient's personal information from the patient to be examined. A description module is connected to the first acquisition module, and the description module constructs scanning description information based on the items to be examined and the patient's personal information; The model interaction module is connected to the description module and an external large language model. The model interaction module inputs the scanning description information into the external large language model to obtain the scanning plan fed back by the large language model. The scanning scheme includes multiple scanning steps and scanning parameters corresponding to each scanning step; The second acquisition module acquires echo sequences during the scanning process and establishes intermediate scanning data based on the echo sequences. The model interaction module is connected to the second acquisition module, and the model interaction module inputs the intermediate scanning data into the large language model to obtain the adjusted scanning plan. The second acquisition module includes: An image reconstruction module receives the echo sequence and performs image reconstruction on the echo sequence to obtain an intermediate scan image; An image recognition module, connected to the image reconstruction module, identifies the intermediate scan image to obtain at least one marker point; A point description module is connected to the image recognition module. The point description module constructs point description information corresponding to each of the marker points according to the marker points and the scanning parameters in the echo sequence. A discrimination module is connected to the point description module. The discrimination module compares the point description information with the scanning plan to output a discrimination result. A feedback module is connected to the discrimination module, and the feedback module controls the model interaction module according to the discrimination result.

2. The magnetic resonance scanning parameter determination system according to claim 1, characterized in that, The first acquisition module includes: An interface module is provided, which connects to an external hospital information system and receives the electronic medical record of the patient to be examined from the hospital information system. A field extraction module, which is connected to the interface module, extracts the items to be examined and the patient's personal information from the electronic medical record according to field attributes.

3. The magnetic resonance scanning parameter determination system according to claim 1, characterized in that, The description module includes: The project identification module determines the part to be inspected based on the item to be inspected. A template matching module, which is connected to the project identification module, calls an inspection description template according to the part to be inspected; A filling module, which is connected to the template matching module, fills the examination description template according to the patient's personal information to obtain the scan description information.

4. The magnetic resonance scanning parameter determination system according to claim 1, characterized in that, The model interaction module includes: A semantic recognition module receives the scan description information and extracts keywords and keyword description information corresponding to the keywords according to the context of the scan description information. The model input module acquires multiple description groups contained in the scan description information and inputs the description groups sequentially into the large language model. The verification module is connected to the semantic recognition module and the model input module respectively. After the description group is input into the model input module, the verification module receives the feedback content of the large language model. The verification module matches the feedback content according to the keywords and the keyword description information, and controls the model input module to re-input the description group when a match cannot be found; A model output module is connected to the model input module. The model output module receives the feedback content, assembles it to obtain the scanning plan, and outputs it.

5. The magnetic resonance scanning parameter determination system according to claim 4, characterized in that, The semantic recognition module uses a semantic recognition model to recognize the scan description information. The semantic recognition model includes: A long short-term memory network, wherein the long short-term memory network sequentially receives the input scan description information; A word frequency extraction module is connected to the long short-term memory network. The word frequency extraction module segments the scanned description information into words and counts word frequencies. The keyword acquisition module is connected to the word frequency extraction module. After traversing the scanned description information, the keyword acquisition module acquires the keywords according to the word frequency. The association module is connected to both the long short-term memory network and the keyword acquisition module. The association module extracts information from the long short-term memory network based on the keywords to obtain the keyword description information.

6. The magnetic resonance scanning parameter determination system according to claim 4, characterized in that, The verification module includes: A keyword determination module, which matches the corresponding keywords according to the currently input description group; A word vector construction module, which is connected to the keyword determination module, constructs a first word vector based on the keyword description information of the matched keywords, and constructs a second word vector based on the feedback content; The comparison module is connected to the word vector construction module. The comparison module compares the first word vector and the second word vector and controls the model input module based on the comparison result.

7. The magnetic resonance scanning parameter determination system according to claim 1, characterized in that, The image recognition module is implemented using a morphological model, which includes: A feature extraction network extracts high-dimensional features from the scanned intermediate image; A feature matching module is connected to the feature extraction network, and the feature matching module matches the label features from the high-dimensional features; A prediction module, which is connected to the feature matching module, predicts the location of the marker point from the marker features.

Images