Normative quality control method, device and storage medium for magnetic resonance imaging sequences

Abstract

The application discloses a method and device for checking the compliance of a magnetic resonance imaging sequence and a storage medium. The method comprises: obtaining a set of magnetic resonance pulse sequences to be checked for compliance, and dividing all sequences in the set into standard sequences, known non-standard sequences and unknown non-standard sequences; inputting the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model, and outputting the standard sequences corresponding to all non-standard sequences; comparing the output standard sequences and the inherent standard sequences in the set with a list of standard sequences required for the current examination item, and determining whether there is a missed scanning sequence according to the comparison result, and giving an early warning when a missed scanning sequence is found. The application directly performs instant and lightweight intelligent analysis and standardization on the sequence name text generated by the device, and realizes the sequence compliance checking and missed scanning early warning immediately after the scanning is completed.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of medical data management, in particular to a specification quality control method and device for a magnetic resonance imaging sequence and a storage medium. BACKGROUND

[0002] As a common non-invasive imaging diagnostic method in clinic, magnetic resonance (MR) examination plays an irreplaceable role in disease screening, diagnosis and efficacy evaluation due to its high-resolution imaging advantage for soft tissues. The integrity of the key diagnostic sequence directly determines the comprehensiveness of the imaging diagnostic information, and is a core prerequisite for ensuring the accuracy of clinical diagnosis and treatment.

[0003] For example, the application with the publication number CN111695273A and the title of a magnetic resonance scanning sequence simulation platform is applied to the field of electronic information technology. In order to solve the problem that there is no scanning sequence simulation platform in the prior art that can observe the designed scanning sequence from the sequence design interface, so that the sequence designer does not need to use the magnetic resonance scanning sequence simulation platform of the magnetic resonance imaging system. The simulation platform is realized by a sequence computer and a main control device. The main control device completely generates digital radio frequency pulse waveform data and actual gradient waveform data on a digital logic chip, and then the digital acquisition module acquires and uploads the transmit pulse waveform data and the actual gradient waveform data to the sequence computer for observation. The application can help the sequence designer to determine whether the designed scanning sequence is consistent with the expected one through the simulation results of the simulation platform. It is convenient and timely to correct design errors.

[0004] For example, the application with the publication number CN110321101A and the title of a magnetic resonance imaging sequence instruction distribution method and device includes: setting a block inter-division line vertically through a plurality of time axes of a magnetic resonance imaging sequence, segmenting the magnetic resonance imaging sequence into a plurality of event blocks based on the block inter-division line, each event block containing pulse signals on a plurality of time axes; setting a block intra-division line vertically through the plurality of time axes inside each event block according to the shape of the pulse combination in each event block, segmenting each event block into a plurality of sub-event blocks based on the block intra-division line, each sub-event block containing a single pulse type; retrieving a sub-event block from the plurality of sub-event blocks, translating the retrieved sub-event block into a hardware instruction based on the pulse type contained in the retrieved sub-event block; and distributing the hardware instruction to a hardware instruction writer corresponding to the hardware instruction. The difficulty of event analysis is significantly reduced, and the analysis time is reduced.

[0005] However, in actual clinical application, especially in the MR examination process of primary medical institutions, the omission of key diagnostic sequences often occurs (for example, the B1000 image of DWI is missed in the head scan), which seriously affects the development of diagnosis and treatment. The problem is mainly caused by three core factors: first, the device model is diverse, and different manufacturers of MR devices have significant differences in hardware configuration, software system and functional support; second, the sequence naming specification is not uniform, and the same key diagnostic sequence often uses different naming methods in different brand devices, lacking a unified industry standard; third, the operation level of technicians is different, and the professional training level and clinical operation experience of technicians in primary medical institutions are uneven, and the cognition and execution standard of key diagnostic sequences are insufficient.

[0006] To investigate such missing scan problems, the existing technology considers that the device model is diverse and the sequence naming specification is not uniform, and the original sequence name (text information) directly output from the device cannot be directly and uniformly understood and compared, so it generally relies on relatively uniform generated image data as the analysis object. The usual practice is: after the patient completes the MR examination, all the image data generated by the examination are transmitted from the MR device to an independent audit server, and the image content is periodically reviewed and checked by manual audit or image recognition technology based on convolutional neural network (CNN), to infer the executed scan sequence and judge whether there is omission.

[0007] However, this mode has multiple inherent bottlenecks, which inevitably makes it a post-examination and non-real-time audit: first, due to the safety and stability of the device, the hospital MR device usually does not allow the installation of third-party real-time verification programs, and the verification program can only be deployed on an external server. Second, the MR device usually does not support real-time streaming of image data during scanning, and can only transmit the entire examination project of a large amount of image data (usually with a large data volume) to the outside once after all sequence acquisition is completed, which itself is time-consuming. Finally, after transmission is completed, loading and intelligent recognition analysis of a large amount of image data also requires considerable computing time. The significant delay caused by the above data transmission and image processing makes "periodic review" the only feasible mode.

[0008] The above-mentioned missing scan problem detection method by periodically reviewing image data has an essential defect: the discovery of the missing scan problem has a serious time lag. Since the patient has left the examination room and cannot return in time for re-scanning, the current examination fails, and the clinical diagnosis information is incomplete. This not only prolongs the diagnosis cycle and increases the burden of the patient's second visit to the doctor, but also seriously reduces the diagnosis and treatment efficiency and patient satisfaction.

[0009] The existing MR compliance quality control scheme in the prior art can only perform post review and audit due to the dependence on image data and the restriction of data transmission and processing delay, and cannot identify sequence compliance and warn potential missing scan risks during the scanning process or immediately after the scanning is completed. SUMMARY

[0010] Embodiments of the present disclosure provide a specification quality control method, device and storage medium of a magnetic resonance imaging sequence, to at least solve the technical problem that the existing MR compliance quality control scheme in the prior art can only perform post review and audit due to the dependence on image data and the restriction of data transmission and processing delay, and cannot identify sequence compliance and warn potential missing scan risks during the scanning process or immediately after the scanning is completed.

[0011] According to an aspect of an embodiment of the present disclosure, a specification quality control method of a magnetic resonance imaging sequence is provided, comprising: acquiring a set of magnetic resonance pulse sequences to be checked for compliance, and dividing all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences and unknown non-standard sequences; for each known non-standard sequence, acquiring a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores mapping relationships between all known non-standard sequences and corresponding standard sequences, and word vectors of each known non-standard sequence; for each unknown non-standard sequence, determining target known non-standard sequences co-occurring with all known non-standard sequences in historical inspection records containing all known non-standard sequences in the set of magnetic resonance pulse sequences, and determining a word vector of the unknown non-standard sequence based on the word vector of the target known non-standard sequence; inputting the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model, and outputting standard sequences corresponding to all non-standard sequences; comparing the output standard sequences and standard sequences inherent in the set of magnetic resonance pulse sequences with a standard sequence list required for the current inspection item, judging whether there is a missing scan sequence according to the comparison result, and warning when a missing scan is found.

[0012] According to another aspect of an embodiment of the present disclosure, a storage medium is also provided, which includes a stored program, wherein the program is executed by a processor when the program is running.

[0013] According to another aspect of the embodiments of the present disclosure, a specification quality control device for magnetic resonance imaging sequences is also provided, comprising: a sequence set acquisition module configured to acquire a set of magnetic resonance pulse sequences to be checked for compliance, and divide all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences, and unknown non-standard sequences; a known non-standard sequence word vector acquisition module configured to, for each known non-standard sequence, acquire a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores mapping relationships between all current known non-standard sequences and corresponding standard sequences, and word vectors of each known non-standard sequence; an unknown non-standard sequence word vector determination module configured to, for each unknown non-standard sequence, determine target known non-standard sequences that co-occur with all known non-standard sequences in historical inspection records of all known non-standard sequences in the set of magnetic resonance pulse sequences, and determine a word vector of the unknown non-standard sequence based on the word vectors of the target known non-standard sequences; a standard sequence output module configured to input the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model, and output standard sequences corresponding to all non-standard sequences; and a missed scan sequence inspection module configured to compare the output standard sequences and standard sequences inherent to the set of magnetic resonance pulse sequences with a list of standard sequences required for a current inspection item, determine whether there is a missed scan sequence according to a comparison result, and give a warning when a missed scan is found.

[0014] According to another aspect of the embodiments of the present disclosure, a specification quality control device for magnetic resonance imaging sequences is also provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions to process the following processing steps: acquire a set of magnetic resonance pulse sequences to be checked for compliance, and divide all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences, and unknown non-standard sequences; for each known non-standard sequence, acquire a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores mapping relationships between all current known non-standard sequences and corresponding standard sequences, and word vectors of each known non-standard sequence; for each unknown non-standard sequence, determine target known non-standard sequences that co-occur with all known non-standard sequences in historical inspection records of all known non-standard sequences in the set of magnetic resonance pulse sequences, and determine a word vector of the unknown non-standard sequence based on the word vectors of the target known non-standard sequences; input the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model, and output standard sequences corresponding to all non-standard sequences; compare the output standard sequences and standard sequences inherent to the set of magnetic resonance pulse sequences with a list of standard sequences required for a current inspection item, determine whether there is a missed scan sequence according to a comparison result, and give a warning when a missed scan is found.

[0015] The application first acquires and classifies the original magnetic pulse sequence names to be checked, clearly distinguifies the magnetic pulse sequence into standard sequences, known non-standard sequences and unknown non-standard sequences, thereby directly processing the heterogeneous and non-standard sequence text data, rather than processing the image data generated afterwards, laying the foundation for real-time checking. Then, the known non-standard sequences are accurately mapped, and their word vectors are directly obtained by querying the pre-built dictionary, solving the problem of not being able to directly understand the sequence meaning due to non-standard naming. Next, the first-appeared unknown non-standard sequence is intelligently inferred, the other known non-standard sequences co-occurring with it are found out by analyzing the historical inspection records containing the same batch of known non-standard sequences, and the semantic representation of the unknown non-standard sequence is derived based on the word vectors of these co-occurring known non-standard sequences, thereby realizing the reasonable speculation of the word vector of the first-appeared unknown non-standard sequence, realizing the instant understanding ability of the newly named sequence, and breaking through the limitation of the coverage range of the dictionary. Then, the prediction model is used to complete the standardized output of all non-standard sequences, and the sequences with different names are uniformly mapped to the standard sequence list, completing the automatic conversion from the original device output to the standardized understanding. Finally, the actual sequence list after standardization is compared with the preset scanning list in real time, and a warning is given immediately when a missing sequence is found, so that the checking action can be completed when the scanning is just finished and the patient has not left. Therefore, by abandoning the traditional path of transmitting and identifying a large amount of image data afterwards, and instead directly performing instant and light intelligent analysis and standardization on the sequence name text generated by the device, the technical effect of completing sequence compliance checking and missing scan warning immediately after scanning is completed is realized. Further, the technical problem that the existing MR compliance quality control scheme in the prior art can only be reviewed afterwards due to the dependence on image data, the limitation of data transmission and processing delay, and cannot identify sequence compliance and warn potential missing scan risk immediately during scanning or after scanning is solved. BRIEF DESCRIPTION OF DRAWINGS

[0016] The drawings described herein are used to provide further understanding of the present disclosure, and form a part of the present application. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure, and do not constitute an improper limitation on the present disclosure. In the drawings:

[0017] Figure 1 is a hardware structure block diagram of a computing device for implementing the method according to Embodiment 1 of the present disclosure;

[0018] Figure 2 is a flowchart of the specification quality control method of the magnetic resonance imaging sequence according to Embodiment 1 of the present disclosure;

[0019] Figure 3 is an Embedding network training diagram of the specification quality control method of the magnetic resonance imaging sequence according to Embodiment 1 of the present disclosure;

[0020] Figure 4 is a schematic diagram of a prediction model of a normative quality control method of a magnetic resonance imaging sequence according to the embodiment 1 of the present disclosure;

[0021] Figure 5 is a schematic diagram of a normative quality control device of a magnetic resonance imaging sequence according to the embodiment 2 of the present disclosure;

[0022] Figure 6 is a schematic diagram of a normative quality control device of a magnetic resonance imaging sequence according to the embodiment 3 of the present disclosure. DETAILED DESCRIPTION

[0023] In order to make the technical personnel in the art better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, not all. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative labor should be within the scope of protection of the present disclosure.

[0024] It should be noted that the terms "first", "second", and the like in the specification and claims of the present disclosure and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] Embodiment 1

[0026] According to the present embodiment, a normative quality control method embodiment of a magnetic resonance imaging sequence is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0027] The method embodiments provided by the present embodiment can be executed in a mobile terminal, a computer terminal, a server or a similar computing device. Figure 1A hardware structure block diagram of a computing device for implementing a specification quality control of a magnetic resonance imaging sequence is shown. As shown Figure 1 , the computing device can include one or more processors (the processor can include but not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory for storing data, a transmission device for communication function, and an input / output interface. The memory, the transmission device, and the input / output interface are connected with the processor through a bus. In addition, a display, a keyboard, and a cursor control device connected with the input / output interface can be included. Those skilled in the art can understand that Figure 1 , the structure shown is only schematic, and does not limit the structure of the above-mentioned electronic device. For example, the computing device can include more or less components than those shown in Figure 1 , or have a different configuration from that shown in Figure 1 .

[0028] It should be noted that the one or more processors and / or other data processing circuits described above can be referred to as "data processing circuits" herein. The data processing circuit can be embodied in whole or in part as software, hardware, firmware, or any combination thereof. In addition, the data processing circuit can be a single independent processing module, or all or part of any one of the other elements combined into the computing device. As referred to in the embodiments of the present disclosure, the data processing circuit serves as a processor to control, for example, the selection of the variable resistance terminal path connected with the interface.

[0029] The memory can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the specification quality control method of the magnetic resonance imaging sequence in the embodiments of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, i.e. implements the specification quality control method of the magnetic resonance imaging sequence of the above-mentioned application program. The memory can include a high-speed random access memory, and can also include a non-volatile memory such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory can further include a memory remotely disposed relative to the processor, which can be connected to the computing device through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0030] The transmission device is used to receive or send data via a network. The network can include a wireless network provided by a communication provider of the computing device. In one example, the transmission device includes a network interface controller (NIC) that can connect to other network devices through a base station to communicate with the Internet. In one example, the transmission device can be a radio frequency (RF) module that is used to communicate with the Internet in a wireless manner.

[0031] The display can be a liquid crystal display (LCD) that is touch screen type, for example, which can enable a user to interact with a user interface of the computing device.

[0032] It should be noted that, in some optional embodiments, the above-mentioned Figure 1 The computing device shown can include hardware elements (including circuitry), software elements (including computer code stored on a computer readable medium), or a combination of both hardware and software elements. It should be noted that Figure 1 is only one example of a particular implementation and is intended to demonstrate a type of components that can be present in the computing device described above.

[0033] Terminology:

[0034] 1. Magnetic resonance imaging sequence: This term focuses on the physical underlying pulse timing combination that generates the magnetic resonance signal, which is the core technical basis for MR imaging.

[0035] 2. Magnetic resonance pulse sequence: This term is based on the clinical application level of the magnetic resonance pulse sequence, emphasizing the imaging results and clinical diagnostic value corresponding to the pulse sequence.

[0036] In this application, the terms "magnetic resonance pulse sequence" and "magnetic resonance imaging sequence" refer to the same technical concept in essence, and can be used interchangeably. Specifically, "magnetic resonance imaging sequence" refers to the overall technical solution designed to obtain a magnetic resonance image with specific contrast; its technical essence and implementation core lies in a specific "magnetic resonance pulse sequence", i.e. a combination of a series of radio frequency pulses and gradient pulses arranged accurately in time. The "magnetic resonance pulse sequence" described in this application refers to the complete pulse sequence for implementing magnetic resonance imaging.

[0037] Under the above operating environment, according to a first aspect of the embodiment, a specification quality control method of a magnetic resonance imaging sequence is provided, which is implemented by Figure 1 the processor shown in Figure 2 The flowchart of the method is shown, referring to Figure 2 The method includes:

[0038] S202: Obtain a set of magnetic resonance pulse sequences to be checked for compliance, and divide all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences, and unknown non-standard sequences;

[0039] S204: For each known non-standard sequence, obtain a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores the mapping relationship between all current known non-standard sequences and corresponding standard sequences, and the word vector of each known non-standard sequence;

[0040] S206: For each unknown non-standard sequence, determine target known non-standard sequences co-occurring with all known non-standard sequences in the historical inspection records containing all known non-standard sequences in the set of magnetic resonance pulse sequences, and determine the word vector of the unknown non-standard sequence based on the word vector of the target known non-standard sequence;

[0041] S208: Input the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model, and output the standard sequences corresponding to all non-standard sequences;

[0042] S210: Compare the output standard sequences and the standard sequences inherent to the set of magnetic resonance pulse sequences with the list of standard sequences required for the current inspection item, and determine whether there are missed scanning sequences according to the comparison result, and give a warning when a missed scanning sequence is found.

[0043] Specifically, in a clinical magnetic resonance examination process, the compliance checking refers to all magnetic resonance pulse sequence data generated from a single MR examination to be checked for compliance (for example, after a patient completes a head scan examination, all pulse sequences involved in the examination). Unlike the prior art, which relies on massive image data for post-analysis, the present application directly starts from the sequence name text output by the device, the data volume is extremely light, which provides the premise for real-time processing. Based on this, the present application pre-constructs a structured standard sequence knowledge base according to authoritative medical guidelines (such as relevant specifications of the Chinese Medical Association Radiology Branch or the Chinese Medical Association) and expert consensus. Among them, the knowledge base defines the minimum sequence set necessary for different examination sites and clinical purposes. Therefore, according to the scan sequence quality control standards of the authoritative medical guidelines, all sequences of the magnetic resonance pulse sequence set to be checked for compliance can be divided into standard sequences, known non-standard sequences and unknown non-standard sequences (corresponding to step S202). Among them, the standard sequence refers to a sequence that completely meets the scan sequence quality control standards of the authoritative medical guidelines (including sequence naming, sequence list for each site examination and imaging direction requirements, etc.). For example, in a head scan, the standard sequences include T2WI_Axial, T2WI_Axial_Flair (or T2_FLAIR_Axial), T1WI_Axial (or T1WI_Axial_Flair), and T1WI_Sagittal (or T1WI_Sagittal_Flair / T2WI_Sagittal).

[0044] The known non-standard sequence refers to a sequence that does not meet the standard sequence specification requirements in the authoritative medical guidelines, but the characteristics (such as self-defined naming, parameter range, device model belonging to, and clinical use, etc.) of the sequence have been pre-included and labeled by the structured standard sequence knowledge base. For example: a certain brand of MR device, due to manufacturer settings, represents the T1WI_Sag standard sequence of the Chinese Medical Association standard naming as eT1W_TSE_Sag, and represents the PDWI_Ax_FS standard naming as eePDW_TSE_SPAIR_TRA standard sequence. Although it does not conform to the naming specification, it has been pre-verified against the standard of the authoritative medical guidelines, and the correspondence between the self-defined naming and the standard sequence has been recorded, and it is determined to be a known non-standard sequence.

[0045] The unknown non-standard sequence refers to a sequence that neither meets the standard sequence specification of the authoritative medical guidelines, nor has its characteristics (such as a never-before-seen sequence name, an abnormal parameter combination, a special sequence generated by a new device, or a sequence generated by a technician's mistake operation) pre-recorded in the structured standard sequence knowledge base, so that its attributes (purpose, whether it is a key sequence, whether it meets clinical needs, etc.) cannot be directly identified. For example, a small brand MR device newly introduced by a primary hospital generates a sequence named T2-FLAIR-Plus. By comparing the standard sequence list of the Chinese Medical Association and the pre-recorded feature library, neither the name nor the corresponding parameter information is found, so the sequence belongs to an unknown non-standard sequence.

[0046] Then, for each known non-standard sequence, the corresponding word vector is obtained by querying the pre-constructed dictionary (corresponding to step S204). Among them, the dictionary pre-stores the correspondence between the known non-standard sequence and the standard sequence and the word vector of each known non-standard sequence, for example, the self-defined name eT1W_TSE_Sag of a certain brand MR device is recorded in the dictionary as: corresponding to the standard sequence T1WI_Sag of the Chinese Medical Association. It should be noted that the construction of the dictionary needs to be based on the learning of the historical examination data of all hospitals in a region, and in the application process, the dictionary can be queried in real time for a single examination. When the processor processes the real-time magnetic resonance pulse sequence data of a single examination, it can immediately identify the corresponding standard sequence of the known non-standard sequence and query the corresponding word vector, which not only ensures the accuracy of compliance judgment, but also avoids the minute-level or even hour-level delay caused by image transmission and processing in the prior art.

[0047] It should be noted that the word vector of each known non-standard sequence is a string of calculable digital codes (such as [0.35, 1.28, 0.76,..., 0.41]) converted by computer algorithms (such as natural language processing or feature encoding technology) from the key information of the known non-standard sequence (such as sequence name, core parameters TR / TE / flip angle, device model, and clinical use label, etc.). Compared with the complex process of extracting features from image pixels in the prior art, the conversion process from text to vector has very low computational overhead and extremely fast speed. Combined with the real-time process of quality control, when the processor receives a known non-standard sequence in a certain examination, it automatically starts the dictionary query operation. Specifically, first, the key identifier (such as the sequence name or the device model) of the known non-standard sequence is extracted to match the known non-standard sequence list in the pre-constructed dictionary. If the matching is successful, the corresponding standard sequence of the known non-standard sequence and its corresponding word vector are obtained synchronously.

[0048] Meanwhile, for the determination of the unknown non-standard sequence word vector, the processor first filters out each single historical examination record containing all known non-standard sequences in the current magnetic pulse sequence set from the accumulated historical examination records according to the sequence co-occurrence rule in the same MR examination. Then, the target known non-standard sequence that coexists with all known non-standard sequences in the current magnetic pulse sequence set in the same historical examination record is further extracted. Further, based on the word vector of the target known non-standard sequence, the semantic representation of the unknown non-standard sequence is derived to determine the word vector (corresponding to step S206) of the unknown non-standard sequence, so as to realize reasonable speculation of the word vector of the first-appeared unknown non-standard sequence and realize instant understanding of the newly-named sequence, thereby breaking through the limitation of the coverage range of the dictionary.

[0049] It should be noted that the current magnetic pulse sequence set refers to the single examination sequence group to be checked for compliance this time. Meanwhile, containing all known non-standard sequences means that all known non-standard sequences in the current magnetic pulse sequence set need to be completely covered in the single examination corresponding to the historical examination record. For example, if the examination to be quality controlled this time is a head scan, the sequence set thereof contains known non-standard sequence A (corresponding to the standard sequence DWI B1000 of the Chinese Medical Association) and known non-standard sequence B (corresponding to the standard sequence T1WI axial of the Chinese Medical Association), and there is an unknown non-standard sequence C. After screening, 120 historical head scan examination records containing A and B are obtained, and in the 120 records, it is found that other known non-standard sequences also frequently coexist, i.e., 95 records contain known non-standard sequence D (corresponding to the standard sequence T2WI-FS of the Chinese Medical Association), 15 records contain known non-standard sequence E (corresponding to the standard sequence FLAIR of the Chinese Medical Association), and 10 records contain known non-standard sequence F (corresponding to the standard sequence ADC map of the Chinese Medical Association). These known non-standard sequences (D, E, and F) are the target known non-standard sequences coexisting with known non-standard sequences A and B. Subsequently, the word vector of unknown non-standard sequence C can be generated based on the word vector of known non-standard sequences D, E, and F and the co-occurrence frequency, so as to realize reasonable semantic inference of the first-appeared sequence.

[0050] Further, the word vectors of all known non-standard sequences and unknown non-standard sequences are input into the pre-trained prediction model together, and the corresponding standard sequences of all non-standard sequences (known non-standard sequences and unknown non-standard sequences) are output (corresponding to step S208), completing the automatic conversion from the original device output to the standardized understanding. Among them, the prediction model is a mapping model of "sequence word vector features-standard sequence" based on a deep learning algorithm (such as a convolutional neural network model). This model processes lightweight word vectors, not huge image matrices, so its reasoning process can be completed in milliseconds, which is in sharp contrast to the need to run complex convolutional neural networks (CNN) to process high-dimensional image data in the prior art. Therefore, the prediction model can match the features of each input non-standard sequence word vector according to the learned mapping rules, and finally quickly and accurately output the unique "Chinese Medical Association" standard sequence corresponding to each non-standard sequence. For example, the prediction model can learn that the word vector [0.35, 1.28, 0.76,...] corresponds to the standard sequence DWI B1000, and the word vector [0.335, 1.295, 0.77,...] corresponds to the standard sequence FLAIR.

[0051] Finally, the processor compares the standard sequences inherent in the magnetic resonance pulse sequence set (i.e. sequences that fully comply with the "Chinese Medical Association" standard and do not need to be mapped, such as a T2WI_Axial sequence directly named by a hospital according to the specification), the standard sequences corresponding to the non-standard sequences output by the prediction model, and the standard sequence list that should exist for the current examination item (a set of mandatory scanning standard sequences defined in advance for a specific examination item, for example, for a head scan examination, the following four sequence categories are required: T2WI_Axial, T2WI_Axial_Flair (or T2_FLAIR_Axial), T1WI_Axial (or T1WI_Axial_Flair), T1WI_Sagittal (or T1WI_Sagittal_Flair / T2WI_Sagittal)), and checks whether each mandatory scanning standard sequence in the standard sequence list exists in the actual magnetic resonance pulse sequence set. If the comparison result is that there is a standard sequence in the standard sequence list that does not exist in the actual set, it is determined that there is a missing scanning sequence, and a real-time warning is triggered immediately (such as pushing a pop-up window to the examination technician's operation terminal to prompt "suspected missing FLAIR axial sequence, please confirm", and marking the missing scanning risk of the examination on the regional quality control platform) (corresponding to step S210).

[0052] As the content of the foregoing background art, the leakage detection method by periodically reviewing the image data has an essential defect: the discovery of the leakage scanning problem has a serious time lag. Since the patient has left the examination room and cannot return in time for a supplementary scan, it directly leads to the failure of this examination, and the clinical diagnosis information is incomplete. This not only prolongs the diagnosis cycle and increases the burden of the patient's second visit to the doctor, but also seriously reduces the diagnosis and treatment efficiency and patient satisfaction.

[0053] Therefore, the present application first acquires and classifies the original magnetic resonance pulse sequence names to be checked, clearly distinguishes the magnetic resonance pulse sequence into standard sequences, known non-standard sequences and unknown non-standard sequences, thereby directly processing the heterogeneous and non-standard sequence text data, rather than processing the image data generated afterwards, laying the foundation for real-time checking. Then, the known non-standard sequence is accurately mapped, and its word vector is directly obtained by querying the pre-built dictionary, solving the problem of not being able to directly understand the meaning of the sequence due to non-standard naming. Next, the first-appeared unknown non-standard sequence is intelligently inferred, by analyzing the historical examination records containing the same batch of known non-standard sequences, finding out other known non-standard sequences co-occurring with them, and based on the word vectors of these co-occurring known non-standard sequences, deducing the semantic representation of the unknown non-standard sequence, thereby realizing the reasonable speculation of the first-appeared unknown non-standard sequence, realizing the instant understanding ability of the newly named sequence, breaking through the limitation of the coverage of the dictionary. Then, the prediction model is used to complete the standardized output of all non-standard sequences, and the sequences with different names are uniformly mapped to the standard sequence list, completing the automatic conversion from the original device output to the standardized understanding. Finally, the actual sequence list after standardization is compared with the preset scanning list in real time, and a warning is given immediately when a missing sequence is found, so that the checking action can be completed as soon as the scanning is finished and the patient has not left. Therefore, by abandoning the traditional path of transmitting and identifying a large amount of image data after the event, the present application directly analyzes and standardizes the sequence name text generated by the device in real time and light, realizes the technical effect of completing the sequence compliance checking and leakage scanning warning as soon as the scanning is finished. Further solve the technical problems in the prior art that the existing MR compliance quality control scheme can only be reviewed after the event due to the dependence on image data and the delay of data transmission and processing, and cannot identify the sequence compliance and warn the potential leakage scanning risk as soon as the scanning is finished.

[0054] Optionally, the dictionary is constructed by collecting magnetic resonance pulse sequence data in the examination records of all hospitals in the region, and constructing a paired sample composed of known non-standard sequences and standard sequences; training the Embedding network using the paired sample to learn the mapping relationship from the non-standard sequence description to the standard sequence name; using the trained Embedding network to generate a corresponding word vector for each known non-standard sequence; and constructing the dictionary based on the paired sample and the generated word vector; wherein each record in the dictionary includes: a known non-standard sequence, a corresponding standard sequence, and a word vector of the known non-standard sequence.

[0055] Specifically, constructing the dictionary first requires batch collection of all hospitals participating in compliance review in the region, and accumulation of historical magnetic resonance pulse sequence data (including various sequence descriptions such as naming and parameter information generated by different device models and different technicians) in MR examinations. Then, in the historical magnetic resonance pulse sequence data, all sequences that are known to be non-standard according to the standard of the Chinese Medical Association but can be explicitly matched to the corresponding standard sequence are screened to form a one-to-one paired sample of “known non-standard sequence→standard sequence”.

[0056] Further, the paired sample is used to train the Embedding network to learn how to map the non-standard sequence description to the corresponding standard sequence name to generate a dedicated word vector for the known non-standard sequence. The Embedding network is a deep learning model specifically used to learn data feature mapping. Specifically, the trained Embedding network is used to process each known non-standard sequence in turn, and the core features (such as name keywords, parameter ranges, and device model association information) of the sequence are converted into a string of calculable numerical codes (i.e., word vectors, such as [0.35, 1.28, 0.76,..., 0.41]);

[0057] Finally, the paired sample and the generated word vector are stored in a structured manner to construct a queryable dictionary. The specific storage form can be referred to in Table 1.

[0058] Table 1

[0059]

[0060] In the above manner, a query database covering all known non-standard sequences in the region is constructed, and when subsequent non-standard sequences are processed, the corresponding standard sequence and word vector can be quickly obtained by querying the dictionary, thereby providing core support for real-time quality control.

[0061] Optionally, the embedding is trained as follows: construct an embedding network and a first classification network; input the one-hot encoding of known non-standard sequences in the paired samples into the embedding network to determine the corresponding initial word vectors; input the initial word vectors into the first classification network to predict the standard sequences corresponding to the initial word vectors; optimize the embedding layer parameters and the first classification network parameters by minimizing the loss between the predicted standard sequences and the standard sequences in the paired samples until a preset termination condition is reached.

[0062] Specifically, an embedding network is trained to learn the semantic mapping relationship from "non-standard sequences" to "standard sequences" and encode this semantic information into dense word vectors. The training data for this embedding network consists of collected "known non-standard sequence - standard sequence" paired samples. The input is the one-hot encoding of the known non-standard sequence, and the supervision label is the one-hot encoding or class label of its corresponding standard sequence.

[0063] Specifically, refer to Figure 3 As shown, an embedding network (which can be viewed as a trainable lookup matrix) and a subsequent first-class classification network (such as a fully connected layer) are first constructed. During the training of the embedding network, the one-hot encodings C1~C1 of the known non-standard sequences are used. m (A binary number matrix, for example, using "1" and "0" to mark keywords, parameter ranges, and other core information in the sequence name) is input into the embedding network. This embedding network extracts and compresses features from the one-hot encoding, outputting initial word vectors. Subsequently, the generated initial word vectors are input into the neural network and the first classification network. The first classification network, based on its own initial parameter settings and feature matching rules, performs feature parsing and category matching on the input initial word vectors, outputting the standard sequence prediction results i1~i corresponding to the initial word vectors. m (For example, predicting that "diffusion-weighted imaging-1000" corresponds to the standard sequence "DWI B1000").

[0064] Then, a preset loss function (such as a cross-entropy loss function) is used to calculate the deviation (i.e., loss value) between the "predicted standard sequence" output by the first classification network and the "true standard sequence" in the paired sample. The greater the loss value, the greater the deviation between the prediction result and the true situation, and vice versa. Further, based on the loss value, the feature encoding parameters (affecting the accuracy of word vector generation) of the Embedding network and the class matching parameters (affecting the prediction accuracy) of the first classification network are adjusted in reverse through a back propagation algorithm to iteratively reduce the loss value. The prediction process of "one-hot encoding input → initial word vector generation → standard sequence prediction → loss calculation → parameter optimization" is repeatedly executed until the preset termination condition (such as the number of training iterations reaching a set threshold or the loss value being stably lower than a preset precision threshold) is met. When the training is completed, the trained Embedding network parameters are fixed.

[0065] In the above manner, the trained Embedding network can generate accurate word vectors for each known non-standard sequence and be stored in the aforementioned dictionary for direct calling in subsequent processes.

[0066] Optionally, for each unknown non-standard sequence, determining the target known non-standard sequence co-occurring with all known non-standard sequences in the historical examination records containing all known non-standard sequences in the set of magnetic resonance pulse sequences, includes: determining all known non-standard sequences in the set of magnetic resonance pulse sequences as the context known sequence set; searching all examination records containing each known non-standard sequence in the context known sequence set in the historical examination database; and counting other known non-standard sequences co-occurring with the context known sequence set members in the found examination records as target known non-standard sequences.

[0067] Specifically, for a newly emerged unknown non-standard sequence, its corresponding word vector representation cannot be directly obtained due to the lack of historical records. To this end, the present scheme proposes an inference method based on known context co-occurrence patterns. The core of this method is that the semantic function of an unknown sequence can be indirectly represented by the typical historical co-occurrence patterns associated with the known sequences co-occurring in the current examination. The specific implementation steps are as follows:

[0068] 1. Determine the context known sequence set: in the set of magnetic resonance pulse sequences to be checked for compliance, all known non-standard sequences are defined as the context known sequence set. For example, a certain examination contains an unknown non-standard sequence SN unknown and a plurality of known non-standard sequences {SN1, SN2, SN3, SN5}, then {SN1, SN2, SN3, SN5} constitutes the context known sequence set.

[0069] 2. Retrieve historical co-occurrence records: In the historical examination database, retrieve all historical examination records that contain {SN1, SN2, SN3, SN5} simultaneously.

[0070] 3. Extract target co-occurrence sequences: In the above retrieved historical records, count other known non-standard sequences that co-occur in addition to {SN1, SN2, SN3, SN5} as target known non-standard sequences, denoted as SN x , which is the typical co-occurrence sequence of the context known sequence set, as the target known non-standard sequence. x

[0071] By mining the fixed co-occurrence rules of known sequences, a deducible semantic embedding is provided for the first occurrence of unknown sequences, thereby supporting their classification and mapping in compliance verification.

[0072] Optionally, based on the word vector of the target known non-standard sequence, the operation of determining the word vector of the unknown non-standard sequence includes: counting the co-occurrence frequency of each target known non-standard sequence in the found examination records; determining the weight distribution corresponding to each target known non-standard sequence based on the co-occurrence frequency; and performing weighted operation on the word vector of each target known non-standard sequence obtained from the dictionary according to the weight distribution, and determining the operation result as the word vector of the unknown non-standard sequence.

[0073] Specifically, continuing the above example, in all historical examination records retrieved that contain {SN1, SN2, SN3, SN5} simultaneously, count the co-occurrence frequency of each target known non-standard sequence SN x in addition to the members of the context known sequence set. For example, if it is a head scan examination, 1000 historical head scan records containing known non-standard sequences {SN1, SN2, SN3, SN5} simultaneously are screened out, of which 650 records show that the target known non-standard sequence SN6 co-occurs with {SN1, SN2, SN3, SN5}, and 350 records show that the target known non-standard sequence SN7 co-occurs with {SN1, SN2, SN3, SN5}. The co-occurrence frequency of SN6 is 650, and the co-occurrence frequency of SN7 is 350.

[0074] Then, according to the co-occurrence frequency, a co-occurrence weight vector is constructed to determine the weight distribution corresponding to each target known non-standard sequence, so as to quantify each SN x ​The significance of the set of context-known sequences as historical partners. That is, the higher the co-occurrence frequency of the target known non-standard sequence, the greater the weight, and the stronger the influence of its word vector in the generation of the unknown sequence word vector. Specifically, by normalizing (such as applying the Softmax function) the co-occurrence frequency of each target known non-standard sequence into a weight value w x , and the sum of the weights of all target known non-standard sequences is 1. Among them, the weight value w x represents the SN x The significance of the set of {SN1, SN2, SN3, SN5} historical partners. For example, in the above example, the co-occurrence frequency of SN6 is 650, the co-occurrence frequency of SN7 is 350, and the total frequency is 1000, then:

[0075] The weight value of SN6 = 650 / 1000 = 0.65;

[0076] The weight value of SN7 = 350 / 1000 = 0.35;

[0077] The weight distribution is clear, and the contribution of the word vector of SN6 to the unknown non-standard sequence word vector is higher than that of SN7.

[0078] Further, according to the weight distribution, the word vector of the unknown non-standard sequence SN unknown is weighted and summed, and the operation result is determined as the word vector of the unknown non-standard sequence. Specifically, the corresponding word vector Embedding(SN x ) of each SN x is extracted from the pre-constructed dictionary, and according to the weight w x , the weighted sum is calculated according to the following formula:

[0079] (1)

[0080] For example, taking a 4-dimensional word vector as an example:

[0081] The word vector of the target known non-standard sequence SN6: [0.328, 0.651, 0.489, 0.712], the weight value is 0.65;

[0082] The word vector of the target known non-standard sequence SN6: [0.473, 0.689, 0.605, 0.516], the weight value is 0.35;

[0083] The word vector of SN6 and the word vector of SN7 are weighted according to their respective weight values (the word vector is calculated by multiplying the weight value, and then the sum is calculated) to obtain:

[0084] Dimension 1: (0.328*0.65) + (0.473*0.35) = 0.2132 + 0.16555 = 0.37835≈0.378;

[0085] Dimension 2: (0.651*0.65) + (0.689*0.35) = 0.42315 + 0.24115 = 0.6643≈0.664;

[0086] Dimension 3: (0.489*0.65) + (0.605*0.35) = 0.31785 + 0.21175 = 0.5296≈0.530;

[0087] Dimension 4: (0.712*0.65) + (0.516*0.35) = 0.4628 + 0.1806 = 0.6434≈0.643;

[0088] Finally, the operation result is determined as the unknown non-standard sequence SN unknown word vector: [0.394, 0.669, 0.543, 0.623].

[0089] In the above manner, the reasonable inference of the unknown non-standard sequence feature is realized, the word vector thereof can accurately match the input requirements of the subsequent prediction model, and support is provided for the standardization mapping of the non-standard sequence.

[0090] Optionally, the prediction model includes a BERT network and a second classification network; and the word vectors of all known non-standard sequences and unknown non-standard sequences are input into the pre-trained prediction model together to output the standard sequences corresponding to all non-standard sequences, including: inputting the word vectors of all known non-standard sequences and unknown non-standard sequences into the BERT network together to determine the context vectors of each non-standard sequence; inputting all context vectors output by the BERT network into the second classification network together to output the standard sequences corresponding to all non-standard sequences.

[0091] Specifically, the prediction model adopts a collaborative architecture of “BERT network + second classification network” to predict the standard sequences corresponding to all non-standard sequences. Since the non-standard sequence set to be quality controlled (including known non-standard sequences and unknown non-standard sequences) is for the same inspection, there is a natural clinical correlation between the sequences (for example, the DWI, T1WI and FLAIR sequences of head scan belong to a fixed combination), therefore the BERT network can capture the context correlation features between multiple sequences, make up for the limitation that a single word vector can only represent the features of the sequence itself, and make the subsequent classification more accurate.

[0092] Reference Figure 4As shown, the word vectors of all known non-standard sequences to be quality controlled and the word vectors of the unknown non-standard sequence calculated are taken as a whole sequence set x1~x m , which is input into the BERT network. Then, the BERT network analyzes the association relationship (such as imaging logical association and examination scene consistency) of each sequence in the set with other sequences through an internal self-attention mechanism, and performs feature enhancement on the original word vectors based on the association relationship, and finally outputs a context vector for each non-standard sequence.

[0093] Subsequently, the second classification network takes the output of the BERT network, and based on the enhanced context vector, batch completes the category mapping of the non-standard sequence to the standard sequence, and supports multi-sequence synchronous processing, and adapts to the real-time requirement of regional quality control. Specifically, all context vectors output by the BERT network are taken as batch input, which is simultaneously transmitted into the second classification network. The second classification network is based on the mapping rule of "context vector feature to standard sequence category" learned in the pre-training stage, and performs category matching on each input context vector. That is, by calculating the similarity of the context vector and the feature of each standard sequence category, the standard sequence with the highest similarity is selected as the output result. Then, the classification network synchronously outputs the standard sequences y1~y m corresponding to all non-standard sequences, realizing batch mapping.

[0094] In summary, the BERT network excavates the context association between sequences, so that the feature representation of the non-standard sequence is more consistent with the clinical examination scene. The second classification network realizes accurate category matching based on the enhanced features, and the two work together to solve the mapping error problem caused by the lack of scene association of a single word vector, and through batch processing, the efficiency requirement of real-time quality control of regional multi-hospital is met, and finally it is ensured that all non-standard sequences can be accurately mapped to the corresponding standard sequence of the Chinese Medical Association, laying a foundation for subsequent missed scan comparison.

[0095] Optionally, the prediction model is optimized and corrected in the following manner: samples are extracted from examinations that have completed compliance verification, and original image data of the sampled examinations is obtained; a pre-independently trained medical image classifier is used to classify each sequence image in the original image data to obtain a standard sequence type determined based on image content as a correction label; the determination result of the medical image classifier and the recognition result of the prediction model for the same examination are compared, and samples with differences therebetween are collected to construct a difference data set; wherein each record in the difference data set includes: a non-standard sequence as input of the prediction model, a standard sequence output by the prediction model, a standard sequence determined by the medical image classifier, and a magnetic resonance pulse sequence set corresponding to the sampled examination; the prediction model is supervised fine-tuned using the difference data set, wherein the supervision signal used in the fine-tuning is the correction label.

[0096] Specifically, in one preferred embodiment, the present scheme introduces a convolutional neural network (CNN) based medical image classifier as an "arbitrator" independent of the text recognition process, for continuous optimization and correction of the text recognition model (BERT network and subsequent classification network), to improve the robustness and generalization ability of the system in real scenarios. This process does not involve fine-tuning of the Embedding network, only optimizing the sequence understanding and mapping part. The specific steps include:

[0097] 1. Regular sampling and image recognition

[0098] 1) The system regularly (e.g., daily or weekly) randomly samples a certain proportion (e.g., 5%) of samples from the examinations processed by the text process.

[0099] 2) For each sampled examination, obtain its original DICOM image data. The trained CNN image classifier classifies these data images sequentially, identifying their actual standard sequence types (e.g., T1WI_Ax, T2WI_Sag, DWI_B1000, etc.), as correction labels.

[0100] 3) The image classifier is independently trained using a high-quality, expert-labeled image-sequence label dataset to ensure its recognition results have high reliability and can be used as a "gold standard" reference.

[0101] 2. Result comparison and difference data collection

[0102] 1) Compare the recognition results of the CNN image classifier with the sequence recognition results of the text recognition model (BERT + second classification network) for the same examination item by item.

[0103] 2) Record the samples that do not agree, and build a "text-image difference dataset". Each record in the difference dataset contains: the original non-standard sequence description text, the standard sequence predicted by the text recognition model, the standard sequence determined by the CNN image arbitrator (considered as the more reliable label for this examination), and the context information of all sequences in the examination.

[0104] 3. Model optimization (fine-tuning) based on difference data

[0105] 1) Use the collected difference dataset to supervise the fine-tuning of the BERT network and the second classification network in the text recognition process.

[0106] Wherein, the operation of fine-tuning input is: after the non-standard sequence text and its context in the difference data are converted into word vectors by the Embedding layer (fixed parameters, not updated), they are input into the BERT network.

[0107] The operation of fine-tuning the label is: using the standard sequence label provided by the CNN arbitrator as the supervision signal of the fine-tuning stage, wherein the supervision signal is the correction label.

[0108] The training target is: by minimizing the difference between the prediction result of the text recognition model and the image arbitrator label (such as cross-entropy loss), propagating the error back and updating the weight parameters of the BERT network and the sequence labeling network.

[0109] It should be noted that the extraction of known non-standard sequences needs to extract unstructured examination information and sequence names from the hospital PACS / RIS, and use the fine-tuned text recognition model to map the heterogeneous sequence names (such as Philips' eT1W_TSE_Sag and Siemens' tse_t1_sag) defined by each manufacturer and device to the sequence categories in the structured standard sequence knowledge base. This process realizes millisecond-level real-time compliance screening, which can immediately determine potential omissions based on the sequence name during scanning and issue a preliminary warning to the technician.

[0110] At the same time, in order to overcome the errors that may occur when the LLM analyzes highly unconventional sequence names, a CNN-based image classifier is introduced. The workflow is: when the confidence of the LLM's judgment is lower than a certain threshold, or the system requires final verification, the CNN module will analyze the DICOM image generated by the scan itself, and directly classify it into a standard sequence according to the image contrast, anatomical structure and weight features. This method does not rely on sequence text description, and fundamentally solves the problem of naming ambiguity.

[0111] In addition, the key innovation of the present application is to build a closed-loop learning system. All cases confirmed by the CNN arbitrator and failed by the LLM for the first time (including misjudgment and omission) will be automatically collected and form an "ambiguous sequence-standard mapping" incremental database. The incremental database will be used for incremental training or prompt word optimization of the LLM periodically. When the same or similar ambiguous sequence name is encountered next time, the LLM can make accurate judgments with the help of historical experience, thereby gradually reducing the dependence on CNN image-level verification, and continuously improving the accuracy and efficiency of rapid screening in the long run.

[0112] Through this process, the text recognition model can learn from the feedback provided by the image arbitrator, which is closer to the real scan content, and gradually correct its judgment on easily confused, rare or new non-standard sequence descriptions, achieving closed-loop self-optimization.

[0113] In addition, with reference to Figure 1 It is shown that according to the second aspect of the present embodiment, a storage medium is provided. The storage medium includes a stored program, wherein the program is executed by a processor when the program is run. The above-mentioned method is executed by the processor.

[0114] Thus, according to the present embodiment, the present application achieves the technical effect of completing sequence compliance checking and missed scan early warning immediately after scanning by abandoning the traditional path of post-transmission and identification of massive image data, and instead directly performing instant and lightweight intelligent analysis and standardization on the sequence name text generated by the device. Further, the technical problem of the prior art that the existing MR compliance quality control scheme can only be reviewed after the fact due to its dependence on image data and its being subject to data transmission and processing delays, and cannot identify sequence compliance and warn of potential missed scan risks during scanning or immediately after scanning is solved.

[0115] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily essential to the present application.

[0116] From the above description of the embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and a general hardware platform as necessary, and of course it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product in essence or in the part that contributes to the prior art, and the computer software product is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions to make a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) execute the methods described in the various embodiments of the present application.

[0117] Embodiment 2

[0118] Figure 5 A normative quality control device for a magnetic resonance imaging sequence according to the present embodiment is shown, which corresponds to the method according to embodiment 1. Reference is made to Figure 5As shown, the device comprises: a sequence set acquisition module 510, configured to acquire a magnetic resonance pulse sequence set to be checked for compliance, and divide all sequences in the magnetic resonance pulse sequence set into standard sequences, known non-standard sequences, and unknown non-standard sequences; a known non-standard sequence word vector acquisition module 520, configured to, for each known non-standard sequence, acquire a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores mapping relationships between all current known non-standard sequences and corresponding standard sequences, and word vectors of each known non-standard sequence; an unknown non-standard sequence word vector determination module 530, configured to, for each unknown non-standard sequence, determine target known non-standard sequences co-occurring with all known non-standard sequences in historical inspection records of all known non-standard sequences in the magnetic resonance pulse sequence set, and determine a word vector of the unknown non-standard sequence based on a word vector of the target known non-standard sequence; a standard sequence output module 540, configured to input the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model, and output standard sequences corresponding to all non-standard sequences; and a missing scan sequence inspection module 550, configured to compare the output standard sequences and standard sequences inherent to the magnetic resonance pulse sequence set with a list of standard sequences required for a current inspection item, determine whether there is a missing scan sequence according to a comparison result, and give a warning when a missing scan is found.

[0119] It should be noted that the magnetic resonance imaging sequence specification quality control device provided in this embodiment can realize all functions and steps in the method embodiment, solve the same technical problem, and achieve the same technical effect. Therefore, no further description is given to the same parts.

[0120] According to this embodiment, the application directly performs instant and lightweight intelligent analysis and standardization on the sequence name text generated by the device, instead of the traditional path of performing post-transmission and identification on massive image data, thereby achieving the technical effect of completing sequence compliance checking and missing scan warning immediately after scanning is completed. Thus, the technical problem that the existing MR compliance quality control scheme in the prior art can only be reviewed after scanning due to its dependence on image data and its being subject to data transmission and processing delay, and cannot identify sequence compliance and warn potential missing scan risks during scanning or immediately after scanning is completed is solved.

[0121] Embodiment 3

[0122] Figure 6 The magnetic resonance imaging sequence specification quality control device according to the embodiment is shown. The device corresponds to the method according to embodiment 1. For details, refer to Figure 6As shown, the device comprises: a processor 610; and a memory 620 connected with the processor 610, used to provide the processor 610 with instructions to process the following processing steps: acquiring a set of magnetic resonance pulse sequences to be checked for compliance, and dividing all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences and unknown non-standard sequences; for each known non-standard sequence, acquiring a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores a mapping relationship between all known non-standard sequences and corresponding standard sequences at present, and a word vector of each known non-standard sequence; for each unknown non-standard sequence, determining a target known non-standard sequence co-occurring with all known non-standard sequences in a historical inspection record containing all known non-standard sequences in the set of magnetic resonance pulse sequences at the same time, and determining a word vector of the unknown non-standard sequence based on the word vector of the target known non-standard sequence; inputting the word vectors of all known non-standard sequences and unknown non-standard sequences into a pre-trained prediction model together, and outputting standard sequences corresponding to all non-standard sequences; comparing the output standard sequences and standard sequences inherent in the set of magnetic resonance pulse sequences with a list of standard sequences required for the current inspection item, and judging whether there is a missed scanning sequence according to the comparison result, and warning when a missed scanning sequence is found.

[0123] It should be noted that the magnetic resonance imaging sequence compliance quality control device provided in the embodiment can realize all the functions and steps in the method embodiments, solve the same technical problem, and achieve the same technical effect. The same parts will not be described again.

[0124] According to the embodiment, the application directly performs instant and lightweight intelligent analysis and standardization on the sequence name text generated by the device, instead of the traditional path of post-transmission and identification of massive image data, thereby achieving the technical effect of completing sequence compliance checking and missed scanning warning immediately after scanning is completed. Further, the technical problem that the existing MR compliance quality control scheme in the prior art can only be reviewed after scanning due to its dependence on image data and its being subject to data transmission and processing delay, and cannot identify sequence compliance and warn potential missed scanning risks immediately during scanning or after scanning is completed is solved.

[0125] The above-mentioned sequence numbers of the embodiments of the application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0126] In the above-described embodiments of the application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0127] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented by other ways. Among them, the above-described device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0128] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0129] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0130] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0131] The above is only the preferred embodiment of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method of normative quality control of a magnetic resonance imaging sequence, characterized in that, The method comprises the following steps: acquiring a set of magnetic resonance pulse sequences to be checked for compliance, and dividing all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences, and unknown non-standard sequences; for each known non-standard sequence, acquiring a corresponding word vector by querying a pre-constructed dictionary, wherein the dictionary stores the mapping relationship between all known non-standard sequences and corresponding standard sequences, and the word vector of each known non-standard sequence; for each unknown non-standard sequence, determining a target known non-standard sequence that co-occurs with all known non-standard sequences in the history inspection records containing all known non-standard sequences in the set of magnetic resonance pulse sequences, and determining the word vector of the unknown non-standard sequence based on the word vector of the target known non-standard sequence; inputting the word vectors of all known non-standard sequences and the unknown non-standard sequence into a pre-trained prediction model, and outputting the standard sequences corresponding to all non-standard sequences; comparing the output standard sequences and the inherent standard sequences of the set of magnetic resonance pulse sequences with a list of standard sequences required for the current inspection item, judging whether there is a missed scanning sequence according to the comparison result, and warning when a missed scanning sequence is found, and wherein for each unknown non-standard sequence, the operation of determining a target known non-standard sequence that co-occurs with all known non-standard sequences in the history inspection records containing all known non-standard sequences in the set of magnetic resonance pulse sequences comprises: determining all known non-standard sequences in the set of magnetic resonance pulse sequences as a context known sequence set; finding all inspection records containing each known non-standard sequence in the context known sequence set in the history inspection database; counting other known non-standard sequences that co-occur in addition to the members of the context known sequence set in the found inspection records as the target known non-standard sequence, and wherein the operation of determining the word vector of the unknown non-standard sequence based on the word vector of the target known non-standard sequence comprises: counting the co-occurrence frequency of each target known non-standard sequence in the found inspection records; determining the weight distribution corresponding to each target known non-standard sequence based on the co-occurrence frequency; performing weighted operation on the word vectors of each target known non-standard sequence obtained from the dictionary according to the weight distribution, and determining the operation result as the word vector of the unknown non-standard sequence.

2. The method of claim 1, wherein, The dictionary is constructed in the following way: collecting magnetic resonance pulse sequence data in the inspection records of all hospitals in the region, and constructing a paired sample composed of known non-standard sequences and standard sequences; training an Embedding network using the paired sample to learn the mapping relationship from non-standard sequence description to standard sequence name; using the trained Embedding network to generate a corresponding word vector for each known non-standard sequence; constructing the dictionary based on the paired samples and the generated word vectors; wherein each record in the dictionary comprises a known non-standard sequence, a corresponding standard sequence, and a word vector of the known non-standard sequence.

3. The method of claim 2, wherein, The Embedding network is trained in the following manner: constructing an Embedding network and a first classification network; inputting the one-hot encoding of the known non-standard sequence in the paired sample into the Embedding network to determine an initial word vector corresponding thereto; inputting the initial word vector into the first classification network to predict a standard sequence corresponding to the initial word vector; optimizing the Embedding layer parameters and the first classification network parameters by minimizing the loss between the predicted standard sequence and the standard sequence in the paired sample until a preset termination condition is reached.

4. The method of claim 1, wherein, The prediction model comprises a BERT network and a second classification network; and the word vectors of all known non-standard sequences and unknown non-standard sequences are inputted into the pre-trained prediction model to output the standard sequences corresponding to all non-standard sequences, comprising: inputting the word vectors of all known non-standard sequences and unknown non-standard sequences into the BERT network to determine a context vector of each non-standard sequence; inputting all context vectors outputted by the BERT network into the second classification network to output the standard sequences corresponding to all non-standard sequences.

5. The method of claim 1, wherein, The prediction model is optimized and corrected in the following manner: sampling from the completed compliance verification checks and obtaining original image data of the sampled checks; classifying each sequence image in the original image data using a pre-independently trained medical image classifier to obtain a standard sequence type determined based on image content as a correction label; comparing the determination result of the medical image classifier with the recognition result of the prediction model for the same check, collecting samples having differences therebetween to construct a difference data set; wherein each record in the difference data set comprises a non-standard sequence inputted into the prediction model, a standard sequence outputted by the prediction model, a standard sequence determined by the medical image classifier, and a set of magnetic resonance pulse sequences corresponding to the sampled check; using the difference data set to supervise fine-tuning of the prediction model, wherein the supervision signal used for fine-tuning is the correction label.

6. A storage medium, characterized by The storage medium comprises a stored program, wherein the program is executed by a processor when the program is run to perform the method of any one of claims 1 to 5.

7. A specification quality control device for a magnetic resonance imaging sequence, characterized in that comprising: a sequence set acquisition module configured to acquire a set of magnetic resonance pulse sequences to be subjected to compliance verification, and divide all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences, and unknown non-standard sequences; a known non-standard sequence word vector acquisition module configured to, for each known non-standard sequence, acquire a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores a mapping relationship between all current known non-standard sequences and corresponding standard sequences, and a word vector of each known non-standard sequence; An unknown non-standard sequence word vector determination module is configured to determine, for each unknown non-standard sequence, a target known non-standard sequence co-occurring with all known non-standard sequences in historical examination records containing all known non-standard sequences in the set of magnetic resonance pulse sequences, and determine a word vector of the unknown non-standard sequence based on a word vector of the target known non-standard sequence; A standard sequence output module is configured to input the word vectors of all known non-standard sequences and the unknown non-standard sequence into a pre-trained prediction model, and output a standard sequence corresponding to each non-standard sequence; A missed scanning sequence examination module is configured to compare the output standard sequence with a standard sequence inherent to the set of magnetic resonance pulse sequences and a list of standard sequences required for a current examination item, determine whether there is a missed scanning sequence based on a comparison result, and give a warning when a missed scanning sequence is found, and wherein For each unknown non-standard sequence, the operation of determining a target known non-standard sequence co-occurring with all known non-standard sequences in historical examination records containing all known non-standard sequences in the set of magnetic resonance pulse sequences comprises: determining all known non-standard sequences in the set of magnetic resonance pulse sequences as a context known sequence set; finding, in a historical examination database, all examination records containing each known non-standard sequence in the context known sequence set; counting, in the found examination records, other known non-standard sequences co-occurring with the context known sequence set except for members of the context known sequence set as the target known non-standard sequence, and wherein The operation of determining a word vector of the unknown non-standard sequence based on a word vector of the target known non-standard sequence comprises: counting a co-occurrence frequency of each target known non-standard sequence in the found examination records; determining a weight distribution corresponding to each target known non-standard sequence based on the co-occurrence frequency; performing a weighted operation on the word vector of each target known non-standard sequence obtained from the dictionary according to the weight distribution, and determining an operation result as the word vector of the unknown non-standard sequence.

8. A specification quality control device for a magnetic resonance imaging sequence, characterized in that comprise: a processor; and a memory connected with the processor and configured to provide the processor with instructions for processing the following processing steps: obtaining a set of magnetic resonance pulse sequences to be checked for compliance, and dividing all sequences in the set of magnetic resonance pulse sequences into standard sequences, known non-standard sequences, and unknown non-standard sequences; for each known non-standard sequence, obtaining a corresponding word vector by querying a pre-constructed dictionary; wherein the dictionary stores a mapping relationship between all current known non-standard sequences and corresponding standard sequences, and a word vector of each known non-standard sequence; for each unknown non-standard sequence, determining a target known non-standard sequence co-occurring with all known non-standard sequences in historical examination records containing all known non-standard sequences in the set of magnetic resonance pulse sequences, and determining a word vector of the unknown non-standard sequence based on a word vector of the target known non-standard sequence; ​ Inputting the word vectors of the all known non-standard sequences and the unknown non-standard sequence into a pre-trained prediction model together, outputting the standard sequence corresponding to all non-standard sequences; Comparing the output standard sequence and the standard sequence inherent to the set of magnetic resonance pulse sequences with the list of standard sequences that should be possessed by the current examination item, judging whether there is a missed scanning sequence according to the comparison result, and warning when a missed scanning sequence is found, and wherein For each of the unknown non-standard sequence, the operation of determining the target known non-standard sequence co-occurring with the all known non-standard sequences in the historical examination records containing all the known non-standard sequences in the set of magnetic resonance pulse sequences simultaneously, comprises: Determining all the known non-standard sequences in the set of magnetic resonance pulse sequences as a context known sequence set; In the historical examination database, searching for all the examination records containing each known non-standard sequence in the context known sequence set simultaneously; In the searched examination records, counting other known non-standard sequences co-occurring with the context known sequence set members as the target known non-standard sequence, and wherein Based on the word vector of the target known non-standard sequence, the operation of determining the word vector of the unknown non-standard sequence comprises: Counting the co-occurrence frequency of each of the target known non-standard sequence in the searched examination records; Based on the co-occurrence frequency, determining the weight distribution corresponding to each of the target known non-standard sequence; according to the weight distribution, performing a weighted operation on the word vector of each of the target known non-standard sequence obtained from the dictionary, and determining the operation result as the word vector of the unknown non-standard sequence.

Images