Artificial intelligence-based regional medical information sharing method and system

Abstract

The application discloses an artificial intelligence-based regional medical information sharing method, and relates to the field of information processing, which comprises the following steps: acquiring regional medical information; extracting text information in the regional medical information; dividing the text information into key text and regular text; completing text encryption of the key text to obtain text encryption information; performing information compression encryption on non-text information to obtain non-text encryption information; performing global encryption on the regular text, the text encryption information and the non-text encryption information to obtain medical encryption information; and uploading the medical encryption information to a medical information sharing terminal. The application can effectively realize information sharing of medical information in a state without information leakage.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of information processing, and in particular to a regional medical information sharing method and system based on artificial intelligence. BACKGROUND

[0002] With the increasing popularity of data informatization, medical information stored in each medical institution is also increasing. These medical information is an important basis for medical research, clinical diagnosis and treatment. If medical information sharing can be achieved between different medical institutions, the medical quality can be greatly improved, and the development of medicine can be promoted. However, there are many difficulties in the process of realizing medical information sharing. For example, since medical information is sensitive data, a sharing mode without proper encryption protection may lead to the leakage of patients' privacy, causing serious consequences.

[0003] The existing medical information sharing technology includes directly uploading and sharing after hiding the private information in the medical information through anonymization technology. Although this method encrypts the private information of the patient, it does not encrypt other non-private information, and it is still possible to infer the private information of the patient by combining the non-private information, resulting in the leakage of the patient's information. SUMMARY

[0004] Embodiments of the present application provide a regional medical information sharing method and system based on artificial intelligence, which is used to solve the problem that it is difficult to ensure that medical information is shared without information leakage in the prior art.

[0005] To achieve the above purpose, the embodiments of the present application adopt the following technical solutions:

[0006] In a first aspect, a regional medical information sharing method based on artificial intelligence is provided, which includes:

[0007] Obtaining regional medical information of a target region;

[0008] Positioning and extracting text information in the regional medical information based on a target detection algorithm;

[0009] Inputting the text information into a key information classification model, outputting an information classification result of the text information through the key information classification model, dividing the text information into key text and regular text according to the information classification result, and constructing the key information classification model based on a recurrent neural network model;

[0010] Completing text encryption of the key text based on a text encryption algorithm to obtain text encryption information;

[0011] Vectorizing non-text information in the regional medical information except the text information, and performing information compression encryption on the non-text information after the vectorization processing to obtain non-text encryption information;

[0012] The global encryption of the routine text, the text encrypted information and the non-text encrypted information in the regional medical information is completed based on a group encryption algorithm, and medical encrypted information is obtained;

[0013] The medical encrypted information is uploaded to a pre-constructed medical information sharing terminal through the cloud server.

[0014] Optionally, the key information classification model comprises a feature extraction module, an element marking module, a key degree module and an information classification module, and the information classification result of the text information output by the key information classification model comprises the following steps:

[0015] The text information is segmented to obtain a plurality of text information sequences;

[0016] For any text information sequence, the text information sequence is input into the feature extraction module, and the text information feature of the text information sequence is extracted by the feature extraction module;

[0017] The text information feature is input into the element marking module, the element type of the text information sequence is marked by the element marking module, and the text element sequence of the text information sequence is output according to the element type marking result;

[0018] The information weight of the text element sequence is calculated by the key degree module, and the key degree vector of the text element sequence is output according to the information weight;

[0019] All text information features and key degree vectors are sequentially input into the information classification module for information classification, and the information classification result of the text information is output by the information classification module.

[0020] Optionally, the text information feature of the text information sequence extracted by the feature extraction module comprises the following steps:

[0021] The text information sequence is converted into an information vector sequence by the feature extraction module;

[0022] The context semantic analysis of the information vector sequence is completed by the feature extraction module, and the text semantic features of all character vectors in the information vector sequence are extracted according to the context semantic analysis result;

[0023] The semantic correlation degrees between all character vectors are analyzed based on the feature extraction module, and the weights are assigned to all text semantic features according to the semantic correlation degrees;

[0024] The weighted fusion of all text semantic features with the weight assignment is completed by the feature extraction module, and the text information feature is output.

[0025] Optionally, the information weight of the text element sequence is calculated by the key degree module, and the key degree vector of the text element sequence is output according to the information weight comprises the following steps:

[0026] The sequence self-information of all text elements in the text element sequence is calculated by the key degree module and using the self-information formula, and the information entropy weight of all text elements is calculated based on the sequence self-information;

[0027] The sequence mutual information between all text elements is calculated by the key degree module and using the mutual information formula, and the mutual information weight of all text elements is calculated based on the sequence mutual information;

[0028] The information entropy weight and the mutual information weight are integrated to obtain the information weight of the text element sequence;

[0029] The text key degree matrix of the text element sequence is constructed based on the information weight;

[0030] The text key degree matrix is vectorized, and the key degree vector of the text element sequence is output by the key degree module.

[0031] Optionally, the non-text information in the medical information in the medical region except the text information is vectorized, and the non-text information after the vectorization is compressed and encrypted to obtain the non-text encrypted information, which comprises the following steps:

[0032] The non-text information is vectorized to obtain a non-text information vector;

[0033] The non-text information vector is decomposed based on an empirical mode decomposition algorithm to obtain a plurality of non-text sub-information;

[0034] For any non-text sub-information, lossless compression of the non-text sub-information is completed based on an encoding technology to obtain non-text compressed information;

[0035] The non-text compressed information is encrypted by using a chaotic system to obtain non-text encrypted sub-information;

[0036] All non-text encrypted sub-information is integrated to obtain the non-text encrypted information.

[0037] Optionally, the non-text information vector is decomposed based on the empirical mode decomposition algorithm to obtain a plurality of non-text sub-information, which comprises the following steps:

[0038] An extreme interval of the non-text information vector is extracted based on the empirical mode decomposition algorithm;

[0039] A local mean value of the non-text information vector is calculated according to the extreme interval and by using an integral average formula, and an information mean value curve of the non-text information vector is generated based on the local mean value;

[0040] Iteratively extracting all information components of the non-text information vector based on the information mean curve;

[0041] Obtaining a plurality of vector information matrices after reconstructing all information components;

[0042] Constructing a plurality of non-text sub-information based on all vector information matrices.

[0043] Optionally, the non-loss compression of non-text sub-information is completed based on the encoding technology to obtain non-text compressed information as follows:

[0044] For any information element in the non-text sub-information, a predicted value of the information element is generated according to a plurality of adjacent elements of the information element and through a linear prediction algorithm;

[0045] Calculating the absolute value of the difference between the element value and the predicted value of the information element to obtain a predicted element error;

[0046] Integrating all predicted element errors to obtain a predicted error non-text;

[0047] Decomposing the predicted error non-text based on a bit plane decomposition technology to obtain a plurality of error bit planes;

[0048] For any error bit plane, the error bit plane is divided into a plurality of error non-overlapping blocks;

[0049] Traversing all error non-overlapping blocks to extract elements, and synthesizing an element plane according to the element extraction result;

[0050] Completing double encoding of all element planes using an encoding technology, and completing element compression of all element planes based on the double encoding result to obtain non-text compressed information.

[0051] Optionally, the non-text compressed information is encrypted using a chaotic system to obtain non-text encrypted sub-information, including the following steps:

[0052] Driving the chaotic system to complete chaotic mapping of the non-text compressed information through a pre-set initial key, and generating a chaotic non-text sequence;

[0053] Quantizing the chaotic non-text sequence, and generating an integer index sequence according to the quantization result;

[0054] Constructing a Latin square based on the integer index sequence and using a dynamic padding rule;

[0055] Scrambling the non-text compressed information based on the Latin square to obtain scrambled non-text;

[0056] Performing modulo addition operation on the scrambled non-text and a diffusion key to obtain intermediate non-text, and the diffusion key is generated based on the Latin square;

[0057] Diffuse the intermediate non-text according to the initial key to obtain a diffused non-text;

[0058] Take the diffused non-text as new non-text compression information, and repeat the above shuffling and diffusing steps until a preset maximum number of encryptions is reached, and output non-text encryption sub-information.

[0059] Optionally, the method further comprises the following steps:

[0060] Receiving the medical encryption information, and extracting the pre-stored reference hash information in the medical encryption information;

[0061] Generating a hash path according to the medical encryption information, and calculating a root hash of the medical encryption information according to the hash path;

[0062] Integrating the hash path and the root hash to obtain verification hash information;

[0063] Checking whether the reference hash information and the verification hash information are consistent, if the reference hash information and the verification hash information are consistent, determining that the medical encryption information passes the verification;

[0064] If the reference hash information and the verification hash information are inconsistent, it is determined that the medical encryption information fails the verification.

[0065] In a second aspect, the present application provides a regional medical information sharing system based on artificial intelligence, characterized in that it comprises:

[0066] a memory configured to store instructions; and

[0067] a processor configured to call the instructions from the memory and capable of realizing the method of regional medical information sharing based on artificial intelligence according to any one of the first aspect when executing the instructions.

[0068] Through the above technical solution, the text information in the regional medical information is screened out, and the text information is divided into key text and regular text by extracting the text information features and the element types contained therein. The key text is separately subjected to local information encryption, which improves the confidentiality of the key information, prevents the leakage of key text in the process of medical information sharing, and effectively saves the computing power. Then, the non-text information in the regional medical information is subjected to information encryption. After the local encryption of the key text and the non-text information is completed, the regional medical information is subjected to global encryption. Through the above method, double encryption of the key text and the non-text information can be realized to protect the privacy and security thereof, and to ensure that there is no information leakage and information tampering in the process of medical information sharing.

[0069] Other features and advantages of the embodiments of the present application will be described in detail in the subsequent specific implementation manner part. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 A flowchart of a regional medical information sharing method based on artificial intelligence provided by an embodiment of the present application;

[0071] Figure 2 A structural diagram of a regional medical information sharing system provided by an embodiment of the present application;

[0072] Figure 3 A flowchart of compressing and encrypting non-text provided by an embodiment of the present application;

[0073] Figure 4 An example diagram of an error bit plane and an element plane provided by an embodiment of the present application;

[0074] Figure 5 A schematic diagram of an information compression process provided by an embodiment of the present application;

[0075] Figure 6 A flowchart of Latin square chaotic encryption based on an embodiment of the present application. DETAILED DESCRIPTION

[0076] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the specific embodiments described herein are only used to explain and illustrate the embodiments of the present application and should not be used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0077] It should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative positional relationship, movement condition, etc. between components in a certain specific posture (as shown in the drawings). If the specific posture changes, the directional indications also change accordingly.

[0078] In addition, if the description of "first", "second", etc. is involved in the embodiments of the present application, the description of "first", "second", etc. is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can be explicitly or implicitly included at least one of the features. In addition, the technical solutions of various embodiments can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor within the protection scope required by the present application.

[0079] Figure 1 The flowchart of the method for sharing regional medical information based on artificial intelligence according to the embodiments of the present application is schematically shown. As shown in Figure 1 The embodiments of the present application provide a method for sharing regional medical information based on artificial intelligence, which can include the following steps:

[0080] S101, acquiring regional medical information of a target region;

[0081] In the present embodiment, the regional medical information refers to various file information generated in the process of medical diagnosis of patients in medical institutions, including but not limited to electronic medical records, payment receipts, examination reports (blood routine report, urine routine report, etc.), imaging reports (X-ray film, CT imaging, etc.), registration cards, etc. Since the above-mentioned regional medical information contains information related to patient privacy such as patient identity (such as name, ID number), disease details (such as HIV test results) and biological feature data (such as gene sequence), in order to prevent the above-mentioned privacy information from being leaked in the process of information sharing, it is necessary to encrypt it before information sharing. In addition, when patients are treated in different medical institutions, relevant medical staff may refer to the regional medical information of the patient in other hospitals, and if the regional medical information is not encrypted in the process of cross-institution information sharing, the regional medical information of the patient may be at risk of being tampered with. Once the medical staff refers to the tampered regional medical information of the patient in the process of diagnosing and treating the disease of the patient, it may lead to misdiagnosis, thereby affecting the physical condition of the patient. Based on the above, in order to protect the regional medical information from being leaked and tampered with in the process of information sharing, it is necessary to encrypt the regional medical information before information sharing to protect the information security of the regional medical information.

[0082] S102, positioning and extracting text information in the regional medical information based on a target detection algorithm;

[0083] In the embodiment, the target detection algorithm is used to quickly frame the text information region in the regional medical information, and then text verification is performed on the text information region to check whether only text information is contained in the text information region. If the verification is passed, the segmented text information is output as the final result, that is, the extraction of the text information is completed. If the verification is not passed, the target detection algorithm is used to frame the text information region again until the verification is passed.

[0084] Specifically, the commonly used target detection algorithm includes YOLOv3, Faster R-CNN, etc. Taking YOLOv3 as an example, YOLOv3 uses Darknet-53 (a convolutional neural network model) as the backbone network, contains 53 convolutional layers, balances the calculation efficiency and feature extraction capability through residual connection, predicts the bounding box on 3 feature layers of different scales (13x13, 26x26, 52x52), respectively captures large, medium and small size text regions, introduces an anchor box mechanism, predefines 9 anchor boxes (3 for each scale) for the width-height ratio of text information (such as long bar-shaped test item name, square-shaped test result), improves the positioning accuracy of dense text, uses CIoU Loss (complete intersection over union loss) as its loss function, improves the matching degree of the overlap between the predicted box and the real box and its geometric relationship, and improves the text positioning accuracy. Finally, the binary cross entropy is used to determine whether each anchor box contains text information to avoid the interference of complex background (such as organ contour in medical images). The steps of using YOLOv3 to frame the text information in the regional medical information include: first, pre-processing the medical information such as denoising and geometric correction, and then unifying the size of all information in the regional medical information to facilitate subsequent text detection and positioning. The feature pyramid network (FPN) mechanism of YOLOv3 is used to perform multi-scale feature extraction and fusion on the regional medical information, that is, by capturing large-scale text (such as report title), medium-size text (such as test item name) and small-size text (such as drug dosage) in the regional medical information on 3 feature layers of different scales (13x13, 26x26, 52x52), respectively, then the text of different scales is spliced and merged, the framing and positioning of the text information is completed, and finally the boundary box coordinates output by YOLOv3 are used to segment the region containing the text information, and the text information region is output.

[0085] The text information region can be subjected to text verification using an OCR technique, the steps of which include: first, subjecting the text information region to grayscale, denoising (e.g., removing ink stains, creases, and other interference), and geometric correction (e.g., correcting the inclination angle of the text line), then segmenting the text information into individual characters, sending the segmented characters into an OCR engine, using pattern recognition and machine learning algorithms for recognition, and outputting the recognition results, e.g., using a CRNN model to output the confidence of each character, if the average confidence of each character in the text information region is greater than or equal to a preset confidence threshold, it is determined that the text information region verification is passed, otherwise, if the average confidence of each character in the text information region is less than the preset confidence threshold, it is determined that the text information region verification is not passed. The CRNN model (Convolutional Recurrent Neural Network) is a deep learning model that combines the advantages of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), containing convolutional layers, recurrent layers and transcription layers, capable of processing sequences of arbitrary length and not dependent on character segmentation or horizontal scale normalization, suitable for variable image text recognition tasks, in addition, CRNN can also be applied to other image sequence recognition tasks, such as music symbol recognition, showing good versatility.

[0086] Since the regional medical information contains text information, image information and other types of information, the encryption level and suitable encryption method of different types of information are not the same, therefore it is necessary to distinguish different types of text information. In addition, since the text information contains various types of data, these data contain patient identity card numbers, patient genetic test results and other high-sensitivity key texts, and also contain gender, general medical terms (such as upper respiratory tract infection, cold) and other low-sensitivity regular texts, therefore it is necessary to extract the text information in the regional medical information separately, classify the extracted text information, and encrypt the text information with different sensitivities to different degrees, in order to ensure the information security of the regional medical information in the process of medical information sharing.

[0087] S103, input the text information into a key information classification model, output the information classification result of the text information through the key information classification model, divide the text information into key text and regular text according to the information classification result, and the key information classification model is constructed based on a recurrent neural network model;

[0088] In this embodiment, the feature extraction module is constructed based on a bidirectional gated recurrent unit, which is a variant of recurrent neural network model, and the core of which is the gating mechanism, including update gate, keep gate and candidate state. The element labeling module is constructed based on a conditional random field (CRF) model, which is a discriminative probabilistic model commonly used for labeling or analyzing sequence data. CRF predicts by learning the conditional probability distribution between label sequences given the observed sequence, and the core idea is to maximize the conditional probability for model training. The key degree module is constructed based on a neural network model (which can be a fully connected neural network model), which is used to output the key probability of all text information sequences as key texts based on text information features and key degree vectors. The text information sequence with key probability greater than the preset probability threshold can be regarded as a key text, and the text information sequence with key probability less than or equal to the preset probability threshold can be regarded as a regular text.

[0089] In order to facilitate subsequent feature extraction of text information, it is necessary to standardize the text information in advance, and unify the non-standard abbreviations (such as "heart attack" and "myocardial infarction") and spelling variants (such as "fever" and "fever") in the text information into medical standard terms. In addition, due to the large volume of text information, it is not conducive to the processing of the key information classification model, so the text information is first divided into multiple text information sequences according to the preset length, and the text information sequences are sequentially input into the feature extraction module for feature extraction.

[0090] Specifically, the input text information first passes through the embedding layer of the feature extraction module, and the text information sequence is first processed by word segmentation, and then each word is converted into a fixed-dimensional vector using word embedding technology to form a character vector. Integrating the character vector will obtain an information vector sequence. The information vector sequence is processed from left to right by the forward GRU of the encoding layer in the gated recurrent unit, capturing the historical information of the information vector sequence, and the information vector sequence is processed from right to left by the reverse GRU, capturing the future information of the information vector sequence. The captured historical information and future information are spliced to obtain a hidden state sequence containing bidirectional context information, i.e. text semantic features.

[0091] Since the information vector sequence can be long, global feature extraction of the text information sequence is needed. By introducing an attention mechanism, the feature extraction module can better capture the influence between character vectors that are far apart. The semantic correlation between all character vectors can be calculated by a similarity formula, and the higher the semantic correlation between character vectors, the higher the weight allocated. Finally, all text semantic features with completed weight allocation are weighted and fused to obtain text information features, which are taken as the output of the feature extraction module.

[0092] The text information features are taken as the input of the element labeling module, which is constructed based on a CRF model. By modeling the relationship between the observation features and the text elements and the transition probability between the text elements, the globally optimal text element sequence is finally output. The text information features are first mapped to a low-dimensional feature space with a dimension equal to the number of entity categories, to obtain low-dimensional features. The number of entity categories refers to the number of information element types in the text information, and the information element type refers to the type of text elements contained in the text information, including identity information, disease name, disease type, etc. The text elements can be used as labels of the text information. The scores between the low-dimensional features and different text element sequences are calculated using the pre-set inference algorithm (such as the Viterbi algorithm) in the CRF model, and the text element sequence with the highest score is taken as the output.

[0093] The number of occurrences of each type of text element is counted according to the text element sequence, and the occurrence probability of each type of text element in the entire text element sequence is calculated according to the number of occurrences of the text elements. Then, the sequence self-information of all text elements in each text element sequence of the text information sequence is calculated using the self-information formula, and the calculated sequence self-information is sequentially input into the calculation formula of the information entropy weight to obtain the information entropy weight of all text elements. The sequence mutual information between all text elements is calculated using the mutual information formula, and the calculated sequence mutual information is sequentially input into the calculation formula of the mutual information weight to obtain the mutual information weight of all text elements. The information entropy weight and the mutual information weight are integrated to obtain the information weight of the text element sequence, and a text key degree matrix of the text element sequence is constructed based on the information weight. The text key degree matrix is vectorized and output by the key degree module to obtain the key degree vector of the text element sequence. The text key degree matrix can be converted into a key degree vector using the flattening method, i.e., the matrix is sequentially spliced into a vector in the order of rows or columns, directly retaining all original data.

[0094] Finally, all text information features and key degree vectors are sequentially input into the information classification module for information classification, and the information classification result of the text information is output by the information classification module. The key degree module is constructed based on a neural network model (which can be a fully connected neural network model).

[0095] Since the text information contains various types of data, the data contains high-sensitivity key text and low-sensitivity regular text, and the amount of information of the text information is large, in order to save computing power and improve encryption efficiency, the regional medical information text information is extracted separately, and the extracted text information is classified, the text information is divided into key text and regular text according to the sensitivity, and different levels of encryption are performed on the two, which not only ensures the information security of the regional medical information, but also saves the computing power and reduces the encryption cost of the regional medical information.

[0096] S104, text encryption of the key text is completed based on the text encryption algorithm, and text encryption information is obtained;

[0097] In the embodiment, the commonly used text encryption algorithm includes attribute-based encryption (ABE), homomorphic encryption and lattice-based encryption, etc. Taking the lattice-based encryption algorithm as an example, the text encryption of the key text is completed by using the lattice-based encryption algorithm, and the text encryption information includes the following steps:

[0098] 1.1, preset the encryption parameters of lattice encryption, the encryption parameters include polynomial dimension N, modulus q and modulus p, define the polynomial ring, define the value range of the polynomial coefficient (such as -1, 0, 1), define the polynomial ring, the modulus q must be much larger than the modulus p, and the modulus q and the modulus p are relatively prime, for example, N=443, p=3, q=677;

[0099] 1.2, generate a private key (f, g), randomly generate two polynomials, i.e. polynomial f and polynomial g, whose coefficients are randomly distributed in -1, 0, 1, and polynomial f has inverse elements under modulus q and modulus p, and polynomial f and polynomial g form a private key (f, g), wherein the inverse element refers to in the modulus operation, if there is a number b, such that a x b = 1 (mod c), then b is called the inverse element of a under modulus c;

[0100] 1.3, generate a public key h, calculate the public key polynomial: , is the inverse element of the polynomial f under the modulus q, is the modulus operation;

[0101] 1.4, divide the key text into multiple key subtexts according to the fixed length, and encode all the key subtexts according to the preset encoding rule to obtain multiple data information sequences, the encoding rule can be ASCII or Unicode encoding rule;

[0102] 1.5, mapping all data information sequences into polynomial coefficient space, i.e. converting all data information sequences into a plurality of information polynomials respectively, for example, the data information sequence is [10, 20, 30], assuming p = 3 and N = 5, then each number in the data information sequence is modulo 3 to get [1, 2, 0], and since N = 5, two zeros need to be filled in, to get the polynomial 1 + 2x + 0x² + 0x³ + 0x 4 , and the integration result is 1 + 2x, i.e. the data information sequence [10, 20, 30] is converted into the polynomial 1 + 2x;

[0103] 1.6, randomly generating a noise polynomial r, whose coefficient value range is (-1, 0, 1), for ensuring that different ciphertexts are generated in the process of encrypting text information, for improving the security of text encryption information and increasing the difficulty of cracking text encryption information;

[0104] 1.7, performing encryption operation on the information polynomial to encrypt it, and the encryption formula is: s = r x h + t mod q, wherein t is the information polynomial, s is the encrypted information polynomial, and the text encryption information is obtained by integrating and splicing all the completed encrypted information polynomials.

[0105] Since the lattice-based encryption algorithm has high complexity and can resist traditional computing power attacks (such as brute force cracking and side channel attacks), it can effectively ensure the information security of the key text. In addition, the key text often needs to be stored for a long time due to its importance to the patient, and the lattice-based encryption algorithm can resist attacks by future quantum computers, avoiding the risk of text encryption information leakage due to the breakthrough of quantum computing technology, so the lattice-based encryption algorithm can ensure the long-term security of the key text information. In addition, due to the high complexity of the lattice-based encryption algorithm, in order to reduce the encryption computing power burden and encryption time, the key text in the text information needs to be selected, and only the key text is encrypted based on the lattice, which can not only reduce the encryption computing power and encryption time, but also adapt to the access needs of different medical staff, simplify the key distribution process, for example, medical staff with lower access rights are only provided with the key of medical encryption information, and the key of medical encryption information can only access regular text, and only medical staff with higher access rights can provide the key of the key text. Through the above method, the leakage risk of the key text can be further reduced, and the access process of regional medical information can be standardized.

[0106] S105, vectorizing the non-text information in the regional medical information except the text information, and performing information compression encryption on the non-text information after the vectorization to obtain non-text encryption information;

[0107] In the present embodiment, the non-text information mainly refers to image information, such as X-ray film, CT image, etc. The image information is usually stored in the form of two-dimensional matrix, such as single-channel matrix of gray-scale image or RGB three-channel matrix of color image, etc. Therefore, the image matrix of non-text information can be unfolded into a one-dimensional vector in a certain order (such as row priority), and the non-text information vector is obtained.

[0108] Then, the extreme value points of the non-text information vector are extracted, and the extreme value intervals of the non-text information vector are divided according to the extreme value points. The integral average value of the pixel value in each extreme value interval is calculated by using the integral average formula, and the integral average values of adjacent extreme value intervals are connected by using straight lines to form a continuous curve, i.e. the information mean curve. The information candidate component is obtained by subtracting the information mean curve from the non-text information vector, and the process is repeated until the information candidate component meets the IMF condition, and the first IMF (information component) is obtained. The remaining non-text information vector is taken as a new non-text information vector, and the EMD decomposition step is repeated until no more information components can be extracted, and finally a plurality of information components and corresponding residuals are obtained. Each information component and corresponding residual is rearranged into a two-dimensional matrix according to the size (such as MxN) of the non-text information, i.e. the vector information matrix, and the vector information matrix is normalized and denoised to generate a sub-image, i.e. the non-text sub-information.

[0109] For any non-text sub-information, the pixel values of a plurality of adjacent elements of any information element are used for linear prediction to obtain the predicted value of the information element. Then, the absolute value of the difference between the element value and the predicted value is calculated to obtain the predicted element error, i.e. the error value of prediction. After calculating the predicted element errors of all information elements, the predicted element errors are taken as the pixel values of the image, and all pixel values are integrated to obtain the prediction error non-text, which can also be called the prediction error image.

[0110] The prediction error image is decomposed into a plurality of error bit planes. For any error bit plane, the error bit plane is first divided into a plurality of error non-overlapping blocks according to a preset fixed size. Then, the binary pixel values in each error non-overlapping block are combined into a bit stream in raster scan order, the element extraction is completed, and the bit stream is converted into a plurality of decimal pixel values, and all decimal pixel values are combined into an element plane. After all error bit planes are converted into element planes, each pixel plane is double-encoded by using two different encoding techniques, which can be Huffman encoding and arithmetic encoding, the compression of the element plane is completed, and the non-text compressed information is obtained.

[0111] The initial key is converted into control parameters and initial values of the chaotic system by a key expansion algorithm, and the control parameters and initial values are used to drive the chaotic system to generate a high-entropy chaotic sequence, i.e., a chaotic non-text sequence. The floating-point numbers in the chaotic non-text sequence are projected into the integer interval to obtain an integer index sequence. The integer index sequence is used as an index to construct a Latin square, and a plurality of different Latin squares are obtained by using the same method. The non-text compression information is divided into sub-blocks matched with the order of the Latin square, and the different sub-blocks are permuted by using different Latin squares. The row position and column position of the data in the sub-blocks are permuted according to the element values of the Latin square. The permuted sub-blocks are spliced according to the arrangement positions before the division to obtain a permuted non-text. Diffusion parameters are extracted from the Latin square, and a diffusion key is generated by a nonlinear function. Then, the permuted non-text and the diffusion key are subjected to modulo addition operation to obtain intermediate non-text. Finally, the intermediate non-text and the initial key are subjected to modulo addition operation to complete the diffusion operation and obtain diffusion non-text. The diffusion non-text is used as the information to be encrypted, and the steps of permutation and diffusion are repeated until a preset maximum encryption number (which can be three times) is reached, to obtain non-text encryption sub-information. The non-text encryption sub-information is spliced to obtain non-text encryption information.

[0112] S106, complete global encryption of the regular text, text encryption information and non-text encryption information in the regional medical information based on a group encryption algorithm, to obtain medical encryption information;

[0113] In this embodiment, commonly used encryption algorithms include AES (Advanced Encryption Standard), SM4 (Chinese national cryptographic algorithm), and RSA. To ensure the security of regular text, encrypted text information, and non-encrypted text information, AES and SM4 can be combined as group encryption algorithms to encrypt these types of information. First, the AES algorithm is used to encrypt the information to be encrypted, which includes regular text, encrypted text information, and non-encrypted text information. Specifically, first, select the AES key length (128-bit, 192-bit, or 256-bit), then select a suitable encryption mode (such as ECB, CBC, GCM, etc.); use a secure random number generator to generate the AES key, then uniformly encode the information to be encrypted to obtain encoded information. The encoded information can be a byte array or a string. In addition, if the length of the encoded information is not a multiple of 16 bytes, a padding algorithm (such as PKCS#7) is used to pad the data to meet the AES block size requirement; arrange the encoded information that meets the AES block size requirement into a 4x4 matrix to obtain the encoded information matrix; arrange the AES key into a 4x4 matrix to obtain the key matrix; then perform an XOR operation on the encoded information matrix and the key matrix to complete the initial round of key encryption, obtaining the first-level encryption matrix; perform multiple rounds (up to 8 rounds) of compound transformation operations such as byte substitution, row shifting, column obfuscation, and key encryption on the first-level encryption matrix to obtain the medical encrypted information;

[0114] The encryption process involves several steps: Byte substitution (using an S-box non-linear substitution table to map each byte (matrix element) in the first-level encryption matrix to a new value, breaking the statistical characteristics of the data); Row shifting (performing a cyclic shift operation on each row of the first-level encryption matrix—the first row remains unchanged, the second shifts one position to the left, the third shifts two positions to the left, and the fourth shifts three positions to the left—to increase the diffusion effect; Column obfuscation (multiplying each column of the first-level encryption matrix to ensure encryption strength); and Key encryption (XORing the first-level encryption matrix after byte substitution, row shifting, and column obfuscation with an extended key matrix). The extended key matrix is ​​obtained by transforming the AES key into a round key generated by a key expansion algorithm. The key expansion algorithm expands a shorter key into a longer, more complex, and more secure key by hashing the shorter key using an encryption function.

[0115] After encrypting the information to be encrypted using the AES algorithm, the AES key can be encrypted using the SM4 algorithm. The encryption steps of the SM4 algorithm include: first, selecting an elliptic curve and a base point G; second, selecting a random private key d, the length of which can be 256 bits; and third, using the private key d to calculate the public key P=d. G, the public key P is used for encryption; a random number k is selected, which has the same length as the order n of the elliptic curve; the point Q on the elliptic curve is calculated, and the calculation formula is Q=k G; calculate the shared secret S=H(M||Q), where H is a hash function, M is the message to be encrypted, i.e. the AES key, and || is the OR operation; calculate the temporary public key T=S G, the public key P is used for encryption; a random number k is selected, which has the same length as the order n of the elliptic curve; the point Q on the elliptic curve is calculated, and the calculation formula is Q=k

[0116] The AES encryption algorithm has the advantages of fast encryption speed and high security. Especially for regional medical information with large data volume, using the AES encryption algorithm can ensure the security of regional medical information while increasing the encryption speed, reducing the algorithmic occupation of servers or terminal devices, and reducing the overall operation cost of the medical information system. However, as a symmetric encryption algorithm, AES requires that the information sending end and the information receiving end share the key. However, directly transmitting the AES key poses a risk of interception. Text encryption information and non-text encryption information have already been encrypted once, so even if the AES key is intercepted, the risk of leakage is still relatively small. However, for regular text, once the AES key is maliciously intercepted, it is likely to cause the leakage or tampering of regular text. Although the information contained in the regular text is not critical, if it is tampered with, it will still affect the medical decision-making of medical staff. For example, tampering with regular diagnostic data (such as disease classification codes) may distort the results of epidemiological statistics, leading to the formulation of incorrect prevention and control strategies by relevant departments. Therefore, it is necessary to ensure that regular text does not appear in large-scale leakage events or tampering events. The SM2 encryption algorithm protects the transmission of the AES key through asymmetric encryption, ensuring that only the recipient with the private key can decrypt the AES key. This mechanism makes up for the key distribution defect of symmetric encryption. In addition, since only short data (256-bit AES key) needs to be asymmetrically encrypted, the algorithmic power is greatly saved, reducing the overall operation cost of the medical information system. In addition, by encrypting information in multiple different ways and different rounds, the information security of regional medical information can be ensured while adapting to the access needs of different medical roles, simplifying the key distribution process. For example, for medical staff with low access rights, only the key of medical encryption information is provided, and only medical staff with high access rights can provide the key of critical text. Through the above method, the leakage risk of critical text can be further reduced, and the access process of regional medical information can be standardized.

[0117] S107, upload the medical encryption information to the pre-constructed medical information sharing terminal through the cloud server.

[0118] In this embodiment, the reference hash path is generated according to the medical encryption information, and the reference root hash is calculated according to the reference hash path. The reference hash path and the reference root hash are integrated into the reference hash information. Specifically, the medical encryption information to be uploaded is divided into multiple data blocks according to a preset fixed size (such as 1 MB), each data block is independently verifiable, and a hash algorithm (such as SHA-256 or SM3) is applied to each data block to generate a unique hash value as a leaf node of a Merkle tree. The hash values of the adjacent two leaf nodes are spliced to obtain a spliced hash value. The hash value of the spliced hash value is calculated again using the hash algorithm, and the hash value of the spliced hash value is used as the parent node hash. The step of generating the parent node hash is repeated until a unique root hash is generated. The unique root hash is used as the reference root hash for subsequent verification of the integrity of the uploaded medical encryption information. The reference hash path refers to the hash path used as the verification reference for subsequent information integrity verification. The hash path includes all parent node hashes and spliced hash values.

[0119] With reference to Figure 2 The client uploads the medical encryption information to the medical information sharing terminal through the cloud server by connecting to the cloud server through the SSL / TLS protocol or the VPN. When the medical encryption information is uploaded, the corresponding reference hash information and metadata are attached for subsequent verification by the terminal. The metadata includes the upload time of the medical encryption information, the encryption method of the medical encryption information, the key index of the medical encryption information, the identity of the uploader, and other information. The key index points to the location of a specific key stored in the key management system (KMS). The key only exists in the KMS and the memory and is not saved or transmitted to avoid being intercepted. Even if the metadata is leaked, attackers cannot obtain the key based on the index alone (additional KMS permission control must be broken), which greatly ensures the security of the key.

[0120] When the target user (usually a medical staff) needs to query regional medical information, the medical staff needs to log in to the medical information sharing terminal using personal identity credentials (such as an employee ID + dynamic password, biometric identification, or digital certificate), verify the identity information of the medical staff through the medical information sharing terminal, determine the query authority of the medical staff, and call the corresponding key index according to the query authority. For example, if the authority of the medical staff is low, the medical information sharing terminal will only call the key index of the medical encrypted information. After the medical information sharing terminal sends the key index of the medical encrypted information to the key management system (KMS), the KMS will only call the key of the medical encrypted information. The medical staff can only decrypt the medical encrypted information and view the regular text, but cannot view other key texts and non-text information. Only by obtaining the key of the key text and the key of the non-text information can other key texts and non-text information be viewed, and the key of the key text and the key of the non-text information require higher query authority. Based on this, the regional medical information is encrypted by using a hierarchical encryption method, which is beneficial to the management of the key and the allocation of the authority of the medical staff while ensuring the security of the regional medical information. Different keys are allocated to medical staffs with different authorities to ensure the security of the regional medical information and standardize the allocation of the authority of the regional medical information and the access process.

[0121] In one embodiment, the key information classification model includes a feature extraction module, an element labeling module, a key degree module, and an information classification module. The information classification result of the text information output by the key information classification model includes the following steps:

[0122] Segmenting the text information to obtain a plurality of text information sequences;

[0123] For any text information sequence, input the text information sequence into the feature extraction module, and extract the text information features of the text information sequence through the feature extraction module;

[0124] Input the text information features into the element labeling module, label the element types of the text information sequence through the element labeling module, and output the text element sequence of the text information sequence according to the element type labeling result;

[0125] Calculate the information weight of the text element sequence through the key degree module, and output the key degree vector of the text element sequence according to the information weight;

[0126] Input all the text information features and key degree vectors into the information classification module in sequence for information classification, and output the information classification result of the text information through the information classification module.

[0127] In the embodiment, the feature extraction module is constructed based on a bidirectional gated recurrent unit, which is a variant of recurrent neural network model, and the core of which is a gating mechanism including an update gate, a keep gate and a candidate state. The element labeling module is constructed based on a conditional random field (CRF) model, which is a discriminative probabilistic model commonly used for labeling or analyzing sequence data. The CRF model predicts by learning the conditional probability distribution between label sequences given an observed sequence, and the core idea is to maximize the conditional probability for model training. The key degree module is constructed based on a neural network model (which can be a fully connected neural network model) and is used to output the key probability of all text information sequences as key texts based on the text information features and the key degree vector. The text information sequence with a key probability greater than a preset probability threshold can be regarded as a key text, and the text information sequence with a key probability less than or equal to the preset probability threshold can be regarded as a regular text.

[0128] In order to facilitate subsequent feature extraction of the text information, it is necessary to pre-process the text information to standardize the text information. Non-standard abbreviations (such as "heart attack" and "myocardial infarction") and spelling variants (such as "fever" and "fever") in the text information are unified as medical standard terms. The UMLS (Unified Medical Language System) can be used to standardize the text information. In addition, due to the large volume of text information, it is not conducive to the processing of the key information classification model, so the text information is divided into multiple text information sequences according to a preset length, and the text information sequences are sequentially input into the feature extraction module for feature extraction.

[0129] Specifically, the input text information is first processed by the embedding layer of the feature extraction module. The text information sequence is first segmented, and then each word is converted into a fixed-dimensional vector using word embedding technology to form a character vector. Integrating the character vectors will obtain an information vector sequence. The information vector sequence is processed from left to right by the forward GRU of the encoding layer in the gated recurrent unit to capture the historical information of the information vector sequence, i.e., to capture the relationship between the current character vector and the previous text. The reverse GRU processes the information vector sequence from right to left to capture the future information of the information vector sequence, i.e., to capture the relationship between the current character vector and the next text. The captured historical information and future information are spliced to obtain a hidden state sequence containing bidirectional context information, i.e., text semantic features.

[0130] Since the information vector sequence can be long, global feature extraction is needed for the text information sequence. By introducing an attention mechanism, the feature extraction module can better capture the influence between character vectors that are far apart. The semantic correlation between all character vectors can be calculated using a similarity formula. Common similarity formulas include the cosine similarity formula and the Euclidean distance formula. The similarity between character vectors is calculated as the semantic correlation by inputting the character vectors into the similarity formula. The higher the semantic correlation between character vectors, the higher the weight assigned. This is to determine the key words in the text information and give them more attention, effectively filtering out irrelevant information, so that the final text information features can better reflect the true features of the text information sequence. Finally, all text semantic features with assigned weights are fused to obtain text information features, which are used as the output of the feature extraction module.

[0131] The text information features are used as the input of the element labeling module, which is based on the CRF model. By modeling the relationship between observation features and text elements and the transition probability between text elements, the globally optimal text element sequence is output. The text information features are first mapped to a low-dimensional feature space with a dimension equal to the number of entity categories, resulting in low-dimensional features. The number of entity categories refers to the number of information element types in the text information. Information element types refer to the types of text elements contained in the text information, including identity information, disease name, disease type, etc. Text elements can be used as labels for text information. The scoring between low-dimensional features and different text element sequences is calculated using the inference algorithm (such as the Viterbi algorithm) pre-set in the CRF model. The text element sequence with the highest score is output. Specifically, the emission probability between low-dimensional features and different text elements is calculated, which represents the likelihood of generating a low-dimensional feature given a text element. The state transition probability between text elements is calculated, which represents the likelihood of transitioning from one text element to another. Finally, the scoring between low-dimensional features and different text element sequences is calculated based on the emission probability and state transition probability, and the text element sequence with the highest score is selected as the output. Labeling the text elements in the text information is to facilitate the subsequent calculation of information entropy weights and mutual information weights for different text information in the text information, so that the information entropy weights and mutual information weights are used as one of the core factors for screening key texts.

[0132] According to the text element sequence, the number of occurrences of each type of text element is counted, and the occurrence probability of each type of text element in the entire text element sequence is calculated according to the number of occurrences of the text element. Then, the sequence self-information of all text elements in each text element sequence of the text information sequence is calculated using the self-information formula. The calculated sequence self-information is sequentially input into the calculation formula of the information entropy weight to calculate the information entropy weight of all text elements. The sequence mutual information between all text elements is calculated using the mutual information formula, and the calculated sequence mutual information is sequentially input into the calculation formula of the mutual information weight to calculate the mutual information weight of all text elements. The information entropy weight and the mutual information weight are integrated to obtain the information weight of the text element sequence, and a text key degree matrix of the text element sequence is constructed based on the information weight. The text key degree matrix is vectorized and output by the key degree module to obtain the key degree vector of the text element sequence. The text key degree matrix can be converted into a key degree vector by using the flattening method, that is, the matrix is sequentially spliced into a vector in the order of rows or columns, and all original data is directly retained.

[0133] Finally, all text information features and key degree vectors are sequentially input into the information classification module for information classification, and the information classification result of the text information is output by the information classification module. The key degree module is constructed based on a neural network model (which can be a fully connected neural network model). Taking a fully connected layer neural network model as an example, the key degree module is introduced through the fully connected layer and uses an activation function (such as ReLU, Sigmoid, Tanh) to introduce nonlinear features. After feature fusion of the text information features and the key degree vector, a high-order comprehensive feature is obtained, and finally a Sigmoid activation function is used to output the text information as a key text probability value (0~1). The key text probability value is used as the information classification result.

[0134] In addition, the key information classification model has completed model training in advance. The training set is pre-constructed and trained. The training set contains historical medical text information of different text lengths and different information densities. The training set is input into the untrained key information classification model, and the prediction value is output by the untrained key information classification model. The prediction value is compared with the true label, the loss value (such as cross-entropy loss) is calculated, and the untrained key information classification model is continuously optimized according to the loss value until the model index (accuracy, precision, recall, etc.) of the untrained key information classification model reaches the preset index threshold or the training times reaches the preset maximum training times, and the training is completed.

[0135] Since the text information contains various types of data, these data contain patient ID numbers, patient genetic test results, and other sensitive key texts, as well as gender, general medical terms (such as upper respiratory tract infection, cold), and other less sensitive regular texts. Due to the large amount of information in the text information, in order to save computing power and improve encryption efficiency, the regional medical information text information is extracted separately, and the extracted text information is classified according to the sensitivity of the text information. Different levels of encryption are performed on the key texts and regular texts to ensure the security of the regional medical information, while saving computing power and reducing the encryption cost of the regional medical information.

[0136] In one embodiment, the text information features of the text information sequence are extracted by the feature extraction module, including the following steps:

[0137] The text information sequence is converted into an information vector sequence by the feature extraction module.

[0138] The context semantic analysis of the information vector sequence is completed by the feature extraction module, and the text semantic features of all character vectors in the information vector sequence are extracted according to the context semantic analysis results.

[0139] The semantic correlation between all character vectors is analyzed based on the feature extraction module, and the weights of all text semantic features are assigned according to the semantic correlation.

[0140] The weighted fusion of all text semantic features with weight distribution is completed by the feature extraction module, and the text information features are output.

[0141] In this embodiment, the feature extraction module is constructed based on a bidirectional gated recurrent unit and an attention mechanism. The bidirectional gated recurrent unit can learn the features of time series data through two directions, forward and backward, to obtain forward and backward hidden state representations. Then, the gated recurrent unit integrates and filters information based on the two hidden states to obtain the final representation result. This structure can better capture the long-term dependencies in time series data and improve the performance and generalization ability of the model. The input text information is first processed by the embedding layer of the feature extraction module. The text information sequence is first segmented, and then each word is converted into a fixed-dimensional vector using word embedding technology to form a character vector. The integration of the character vector will obtain an information vector sequence. Word embedding is a method of mapping words or phrases to vectors in the real number domain. Each word or phrase has a corresponding vector representation. This representation method not only reduces the dimensionality of the data, but also captures the semantic relationships between words.

[0142] The forward GRU encodes in the gated recurrent unit processes the sequence of information vectors from left to right, capturing the history information of the sequence of information vectors, i.e., capturing the relationship of the current character vector with the previous context, and the backward GRU processes the sequence of information vectors from right to left, capturing the future information of the sequence of information vectors, i.e., capturing the relationship of the current character vector with the next context, and the captured history information and future information are spliced to obtain a sequence of hidden states containing bidirectional context information, i.e., text semantic features. The text semantic features contain not only the local semantics of each position (such as specific symptoms or terms), but also the global context dependence.

[0143] Since the sequence of information vectors can be long, global feature extraction is needed for the sequence of text information, and by introducing an attention mechanism, the feature extraction module can better obtain the influence between character vectors that are far apart. The semantic correlation between all character vectors can be calculated by a similarity formula, and common similarity formulas include the cosine similarity formula and the Euclidean distance formula. The similarity between character vectors is calculated as the semantic correlation by inputting the character vectors into the similarity formula, and the principle is that the character vectors with higher correlation strength are generally closer in position in the medical text, for example, the words "diabetes" and "insulin", so their context environment is similar when training the word embedding model, resulting in their vectors being close in space. At this time, whether using cosine similarity or Euclidean distance, the proximity can be detected, thereby reflecting the semantic correlation between them. The higher the semantic correlation between character vectors, the higher the weight assigned, which is to determine the key words in the text information and give more attention to the key words, effectively filtering out irrelevant information, so that the final text information features can better reflect the true features of the sequence of text information. Finally, all text semantic features with completed weight assignment are weighted and fused to obtain text information features, which are taken as the output of the feature extraction module.

[0144] Identifying the semantic features of the text information is also to more accurately filter out the key text, because the text information usually contains a large number of complex terms, redundant descriptions and implicit logic. Directly filtering the key text based on surface features (such as keyword frequency) is easy to miss semantic associations or introduce noise, and extracting the semantic features of the text information captures the deep associations between words, phrases and paragraphs by mapping the text to a high-dimensional semantic space, thereby accurately locating the key text.

[0145] In one embodiment, the information weight of the sequence of text elements is calculated by the key degree module, and the key degree vector of the sequence of text elements is output according to the information weight, including the following steps:

[0146] The sequence self-information of all text elements in the text element sequence is calculated by the key degree module and using the self-information formula, and the information entropy weight of all text elements is calculated based on the sequence self-information;

[0147] The sequence mutual information between all text elements is calculated by the key degree module and using the mutual information formula, and the mutual information weight of all text elements is calculated based on the sequence mutual information;

[0148] The information weight of the text element sequence is obtained by integrating all the information entropy weights and mutual information weights;

[0149] The text key degree matrix of the text element sequence is constructed based on the information weight;

[0150] The text key degree matrix is vectorized and the key degree vector of the text element sequence is output by the key degree module.

[0151] In the embodiment, the text element sequence refers to the text elements contained in the text information sequence, the text element refers to the information type with a higher frequency in the historical regional medical information, for example, identity information, disease name, disease type, drug information, etc. are all text elements, and different types of text elements are numbered, for example, the identity information is numbered as A1, the disease name is numbered as A2, and the text element sequence is a sequence containing the text element number, for example, a subtext in the text information sequence contains identity information A1, mild disease A4, blood information A6, and drug information A9, and the text element sequence K=(A1, A4, A6, A9) of the subtext.

[0152] The number of occurrences of each type of text element is counted according to the text element sequence, and the occurrence probability of each type of text element in all text element sequences is calculated according to the number of occurrences of the text element, and then the sequence self-information of all text elements in each text element sequence of the text information sequence is calculated using the self-information formula, and the self-information formula is as follows:

[0153] ,

[0154] wherein, is the occurrence probability of the i-th type of text element in all text element sequences, and n is the number of types of text elements.

[0155] The calculated all sequence self-information is sequentially input into the calculation formula of the information entropy weight to calculate the information entropy weight of all text elements, and the calculation formula of the information entropy weight is as follows:

[0156] ,

[0157] The sequence mutual information between all text elements is calculated using the mutual information formula, and the mutual information formula is as follows:

[0158] ,

[0159] in, This represents the probability that text elements of type i and type j appear in the same text element sequence in the entire text element sequence. This represents the probability of the j-th type of text element appearing in the entire sequence of text elements.

[0160] The calculated mutual information of all sequences is sequentially input into the formula for calculating mutual information weights to obtain the mutual information weights of all text elements. The formula for calculating mutual information weights is as follows:

[0161] ,

[0162] The information weights of the text element sequence are obtained by integrating all information entropy weights and mutual information weights, and a text keyity matrix of the text element sequence is constructed based on the information weights, for example. For each element in the text key matrix, when a equals b The information entropy weight of the text element, when a is not equal to b, The mutual information weights between text elements.

[0163] The text key matrix is ​​vectorized, and the key vector of the text element sequence is output through the key module. The text key matrix can be transformed into a key vector by using the flattening method, that is, the matrix is ​​concatenated into a vector in the order of rows or columns, and all the original data is directly retained.

[0164] Calculating sequence self-information reflects the independent information content of text elements. Higher sequence self-information indicates a more dispersed distribution of the text element, a larger amount of information it contains, a higher risk of privacy leakage, and higher criticality. Calculating the self-information of text elements can reveal their potential importance. Sequence mutual information represents the strength of the correlation between two text elements. Higher mutual information indicates a strong correlation between text elements, a surge in the risk of combined leakage, and therefore indicates higher criticality, requiring strict encryption. Information entropy weight and mutual information weight represent the criticality of a single text element among all text elements. Therefore, the criticality vector of each text information sequence can measure its importance compared to other text information sequences, making the criticality vector one of the factors for selecting key text.

[0165] In one embodiment, reference is made to Figure 3 The process involves vectorizing non-text information (excluding text information) in medical information within a medical area, and then compressing and encrypting the vectorized non-text information to obtain encrypted non-text information. This includes the following steps:

[0166] vectorizing the non-text information to obtain a non-text information vector;

[0167] decomposing the non-text information vector based on an empirical mode decomposition algorithm to obtain a plurality of non-text sub-information;

[0168] for any non-text sub-information, completing lossless compression of the non-text sub-information based on an encoding technique to obtain non-text compressed information;

[0169] encrypting the non-text compressed information using a chaotic system to obtain non-text encrypted sub-information;

[0170] integrating all the non-text encrypted sub-information to obtain non-text encrypted information.

[0171] In this embodiment, the non-text information mainly refers to image information, such as X-ray films, CT images, etc. The image information is usually stored in the form of a two-dimensional matrix, such as a single-channel matrix of a grayscale image or a RGB three-channel matrix of a color image, etc. Therefore, the image matrix of the non-text information can be unfolded into a one-dimensional vector in a specific order (such as row priority) to obtain a non-text information vector. For example, image information with a size of MxN can be converted into a vector with a length of MxN, and each element corresponds to the pixel value of a pixel. This process preserves the spatial relationship of the pixels, facilitating the generation of non-text sub-information after the EMD decomposition of the non-text information vector.

[0172] First, the pixel values of each data point in the non-text information vector are checked from left to right one by one to extract the extreme value points of the non-text information vector, and the extreme value intervals of the non-text information vector are divided according to the extreme value points. The integral mean value of the pixel values in each extreme value interval is calculated using the integral average formula, and the integral mean value is the local average value of the non-text information vector. After calculating the integral mean values of all the extreme value intervals, the integral mean values of adjacent extreme value intervals are connected by straight lines to form a continuous curve, i.e. an information mean curve. The information candidate component is obtained by subtracting the information mean curve from the non-text information vector, and this process is repeated until the information candidate component meets the IMF condition to obtain the first IMF (information component). The remaining non-text information vector is taken as a new non-text information vector to repeat the EMD decomposition step until no more information components can be extracted, and finally a plurality of information components and corresponding residuals are obtained. Each information component and corresponding residual is rearranged into a two-dimensional matrix according to the size (such as MxN) of the non-text information, i.e. a vector information matrix, and the vector information matrix is normalized and denoised to generate a sub-image, i.e. non-text sub-information.

[0173] For any non-text sub-information, the pixel values of a plurality of adjacent elements of any information element thereof are linearly predicted to obtain a predicted value of the information element. Then, the absolute value of the difference between the element value and the predicted value is calculated to obtain a predicted element error, i.e., an error value of the prediction. After the predicted element errors of all information elements are calculated, the predicted element errors are taken as pixel values of the image, and all pixel values are integrated to obtain a predicted error non-text, which can also be referred to as a predicted error image.

[0174] The predicted error image is decomposed into a plurality of error bit planes, which can be divided into high bit planes and low bit planes. The pixel values (gray values or pixel brightness) of the predicted error image are binarized to obtain binary pixel values, which are then arranged in order from high bits to low bits to form error bit planes. For any error bit plane, the error bit plane is divided into a plurality of error non-overlapping blocks in a non-overlapping manner according to a preset fixed size in the order from left to right and from top to bottom. Then, the binary pixel values in each error non-overlapping block are combined into a bit stream in a raster scan order, element extraction is completed, and the bit stream is converted into a plurality of decimal pixel values, which are combined into an element plane. After all error bit planes are converted into element planes, each pixel plane is double-encoded using two different encoding techniques, which can be Huffman encoding and arithmetic encoding, compression of the element plane is completed, and non-text compressed information is obtained.

[0175] The initial key is preset (such as a 128-bit binary string), the initial key is converted into control parameters and initial values of a chaotic system through a key expansion algorithm, the control parameters and the initial values are used to drive the chaotic system to generate a high-entropy chaotic sequence, i.e., a chaotic non-text sequence. The floating-point numbers in the chaotic non-text sequence are projected into an integer interval to obtain an integer index sequence. The integer index sequence is taken as an index for constructing a Latin square, i.e., the integer index sequence is used to construct a Latin square. Specifically, an empty initial matrix is first initialized, and then the initial matrix is filled row by row. For example, the integer index sequence is {Q1, Q2, Q3, …, Qn}, and the value of each element in the integer index sequence belongs to {0, 1, 2, …, n-1}. The first row of the initial matrix is filled according to the first row of the integer index sequence to obtain (0, 1, 2, …, n-1). The arrangement of the i-th row (i is greater than or equal to 2) of the initial matrix is that the i-1-th row is cyclically right-shifted by Qi times.

[0176] After obtaining multiple different Latin squares by using the same method, the non-text compression information is divided into sub-blocks matching the order of the Latin square, different Latin squares are used to perform permutation operation on different sub-blocks, and the row position and column position of the data in the sub-blocks are permuted according to the element values of the Latin square. The permuted sub-blocks are spliced according to the arrangement position before division to obtain the permuted non-text. The diffusion parameters (such as row sum and column XOR value) are extracted from the Latin square, the diffusion key is generated through a nonlinear function, then the permuted non-text and the diffusion key are subjected to modulo addition operation to obtain intermediate non-text, and bit-level diffusion is realized. Finally, the intermediate non-text and the initial key are subjected to modulo addition operation to complete the diffusion operation, and the diffusion non-text is obtained. The diffusion non-text is used as the information to be encrypted, and the permutation and diffusion steps are repeatedly executed until the preset maximum encryption number (which can be three times) is reached, and the non-text encryption sub-information is obtained. The non-text encryption sub-information is spliced to obtain the non-text encryption information.

[0177] In one embodiment, the step of decomposing the non-text information vector based on the empirical mode decomposition algorithm to obtain multiple non-text sub-information includes the following steps:

[0178] The extreme value interval of the non-text information vector is extracted based on the empirical mode decomposition algorithm;

[0179] The local mean value of the non-text information vector is calculated according to the extreme value interval and through the integral average formula, and the information mean value curve of the non-text information vector is generated based on the local mean value;

[0180] All information components of the non-text information vector are iteratively extracted based on the information mean value curve;

[0181] After reconstructing all information components, multiple vector information matrices are obtained;

[0182] Multiple non-text sub-information is constructed based on all vector information matrices.

[0183] In this embodiment, since the Empirical Mode Decomposition (EMD) algorithm is based on the time-scale characteristics of the signal itself and does not require pre-setting basis functions, it has stronger adaptability than Fourier analysis and wavelet analysis. Theoretically, it can be applied to any type of signal decomposition. The basic principle of the EMD algorithm is to decompose the real-valued signal into a series of intrinsic mode functions by separating the local mean. In this application, a chaotic encryption algorithm is used to encrypt non-textual information. The chaotic system used for encryption includes low-dimensional chaotic information and high-dimensional chaotic systems. However, chaotic sequences generated by low-dimensional chaotic systems have disadvantages such as short cryptographic periods, low precision, and small key spaces. While chaotic sequences generated by high-dimensional chaotic systems can overcome the above shortcomings, the increase in the dimension of the chaotic system will inevitably increase the computational complexity, affect the computational precision, and make the encryption algorithm very complex, which may lead to difficulties in decryption later. Therefore, in order to overcome the above problems, it is necessary to use the Empirical Mode Decomposition (EMD) algorithm to decompose the non-textual information into several non-textual sub-information, and then compress and encrypt the non-textual sub-information.

[0184] First, examine the pixel value of each data point in the non-text information vector from left to right. If the pixel value of a data point is greater than the pixel values ​​of its left and right adjacent data points, then the data point is marked as a maximum point; conversely, if the pixel value of a data point is less than the pixel values ​​of its left and right adjacent data points, then the data point is marked as a minimum point. Furthermore, for the start and end points of the non-text information vector, only the pixel values ​​of the adjacent data points on one side need to be compared. Arranging the maximum and minimum points in the order of their markings divides the non-text information vector into multiple extreme value intervals, which are the regions between adjacent maximum and minimum points.

[0185] The integral average of pixel values ​​within each extreme value interval is calculated using the integral average formula. The integral average is a local average of the non-text information vector. The integral average formula is as follows:

[0186] ,

[0187] in, For the first The pixel values ​​corresponding to each maximum or minimum point. For the first The pixel values ​​corresponding to each maximum or minimum point. It is a non-textual information vector.

[0188] After calculating the integral average of all extreme value intervals, the integral average values of adjacent extreme value intervals are connected by straight lines to form a continuous curve, i.e. the information average curve. The traditional envelope method calculates the average value through the upper and lower envelope lines, while the integral average method directly divides the intervals based on the extreme points, uses the integral to reflect the overall energy of the pixel values in the interval, avoids the overshoot or undershoot errors that may be introduced by envelope interpolation (such as the oscillation at the endpoints of cubic spline interpolation), and can more intuitively reflect the statistical characteristics of the pixel values in the extreme value interval. Subtract the information average curve from the non-text information vector to obtain the information candidate component, and repeat this process until the information candidate component meets the IMF condition, which means that the number of extreme points (maximum or minimum points) is close to the number of zero-crossing points, and the envelope average is zero, obtaining the first IMF (information component). Repeat the EMD decomposition steps with the remaining non-text information vector as the new non-text information vector until no more information components can be extracted, and finally obtain several information components and corresponding residuals.

[0189] Each information component and corresponding residual is rearranged into a two-dimensional matrix according to the size of the non-text information (such as MxN), i.e. the vector information matrix. After normalization and denoising of the vector information matrix, a sub-image, i.e. non-text sub-information, is generated. This is because the amplitude range of each information component may be different, and normalization can ensure that they have consistent contrast when visualized. In addition, high-frequency information components may contain noise, so it is necessary to use Gaussian filtering, median filtering and other methods for denoising to make the final non-text sub-information have higher quality.

[0190] In one embodiment, lossless compression of non-text sub-information is completed based on encoding technology to obtain non-text compressed information as follows:

[0191] For any information element in the non-text sub-information, a predicted value of the information element is generated according to a plurality of adjacent elements of the information element and by a linear prediction algorithm;

[0192] The absolute value of the difference between the element value and the predicted value of the information element is calculated to obtain a predicted element error;

[0193] All predicted element errors are integrated to obtain a predicted error non-text;

[0194] The predicted error non-text is decomposed based on bit plane decomposition technology to obtain a plurality of error bit planes;

[0195] For any error bit plane, the error bit plane is divided into a plurality of error non-overlapping blocks;

[0196] Element extraction is performed on all error non-overlapping blocks, and an element plane is synthesized according to the element extraction result;

[0197] The double encoding of the whole element plane is completed by using the encoding technology, and the element compression of the whole element plane is completed based on the double encoding result to obtain the non-text compression information.

[0198] In the embodiment, the non-text information is image information, wherein a large amount of pixel information is contained, and correlation exists between adjacent pixels, and the pixel values of the adjacent pixels are close in most cases. Based on the characteristic, the predicted value of the information element is obtained by linear prediction of the pixel values of the plurality of adjacent elements of the information element. Since the non-text information is image information, the non-text sub-information is a sub-image of the image information, the information element and the adjacent element are pixels of the sub-image, the predicted value is a predicted pixel value, the element value of the information element refers to the original pixel value of the pixel in the sub-image, the pixel value refers to color or brightness information contained by each pixel in the image, for example, the pixel value of a digital image is usually represented by an integer from 0 to 255, 0 represents black, 255 represents white, and the intermediate value represents different gray scales. The linear prediction algorithm is a method for predicting future values by linear regression analysis, for example, the mean value of the pixel values between all adjacent elements of a certain information element is calculated as the predicted element of the information element. Then, the absolute value of the difference between the element value and the predicted value is calculated to obtain the predicted element error, that is, the error value of the prediction. The histogram of the predicted element error calculated by the method usually presents a Laplace-like distribution, and a large amount of redundant space can be used for compression. After the predicted element error of all information elements is calculated, the predicted element error is taken as the pixel value of the image, and all pixel values are integrated to obtain the predicted error non-text, which can also be called the predicted error image.

[0199] The prediction error image is decomposed into a plurality of error bit planes, and the error bit planes can be divided into high bit planes and low bit planes. By performing binary representation on the pixel values (grayscale values or pixel brightness) of the prediction error image, binary pixel values are obtained, and then the binary pixel values are arranged from high bits to low bits to form error bit planes. Each error bit plane contains bit information of the same position of all pixels in the prediction error image, and the importance of these bit planes gradually decreases because the information of high bits is usually more important than the information of low bits. For example, each pixel value of an 8-bit grayscale image can be decomposed into 8 binary bits to form 8 bit planes from the lowest bit to the highest bit, and each bit plane only contains 0 or 1 values of the corresponding bit. Since the high bit planes obtained by decomposing the prediction error image still have many 0s, and the low bit planes present irregular and chaotic conditions of 0s and 1s, a bit plane block synthesis pixel method is designed to aggregate the information of each bit plane, to compress the 0 values of the high bit planes and the chaotic data of the low bit planes as much as possible, so as to free up more embeddable space. More embeddable space allows non-text encryption information to be dispersed in multiple dimensions (such as pixel values, frequency domain coefficients, bit planes, etc.), avoiding attackers from obtaining the complete key through local cracking, and better supporting subsequent more complex chaotic encryption operations, and improving the security of non-text encryption information.

[0200] Reference Figure 4 For any error bit plane, the error bit plane is first divided into a plurality of error non-overlapping blocks in a non-overlapping block manner according to a preset fixed size in a left-to-right and top-to-bottom order. Non-overlapping block division refers to dividing a large-size image into a plurality of mutually non-overlapping sub-blocks in image processing. This division manner makes each sub-block independent in space and does not interfere with each other, thereby simplifying the processing process. Then, the binary pixel values in each error non-overlapping block are combined into a bit stream in a raster scan order, element extraction is completed, and the bit stream is converted into a plurality of decimal pixel values, and all decimal pixel values are combined into an element plane. The bit stream refers to a data stream composed of a string of binary numbers, and the raster scan order refers to an order of scanning from left to right, from top to bottom, scanning a row first, and then moving to the starting position of the next row to continue scanning. The element plane can also be referred to as a pixel plane, that is, a plane image containing pixel values.

[0201] Reference Figure 5After converting all error bit planes into element planes, each pixel plane is double-encoded using two different encoding techniques: Huffman coding and arithmetic coding. Specifically, the frequency of pixel values ​​within each element plane is statistically analyzed to obtain pixel frequencies. A corresponding Huffman dictionary is generated based on these frequencies, assigning short codes to high-frequency pixel values ​​and long codes to low-frequency values. This method ensures that the compressed data has a smaller file size while maintaining the integrity of the original data. Since the element planes corresponding to higher bit planes have more consecutively repeated pixel values, Huffman coding can significantly shorten the code length of high-frequency symbols (such as long strings of 0s or 1s), significantly reducing the number of pixels. The element planes corresponding to lower bit planes may have fewer consecutively repeated pixel values, but local repetitions may exist. Huffman coding can compress these limited repetitive pixel values. Because the element planes corresponding to higher bit planes after Huffman coding may still have consecutive identical pixel values ​​(such as long strings of 0s or 1s), arithmetic coding is used to further eliminate repetitive pixel values ​​in the element planes corresponding to higher bit planes, completing further compression of the element planes and obtaining non-text compressed information.

[0202] Huffman coding is an entropy coding algorithm used for lossless data compression. Its basic principle is to construct a Huffman tree (also called an optimal binary tree) and allocate different code lengths based on the frequency of character occurrences, thereby achieving data compression. Arithmetic coding is also a lossless data compression method, belonging to entropy coding. Its core idea is to encode the entire input message into a decimal between 0 and 1 based on the probability of different symbols appearing in the input message. A representative decimal within the final mapping interval is selected as the encoded output, and this decimal is represented in binary form as the final encoded result.

[0203] The above method enables compression of each non-text sub-information, resulting in compressed non-text information with a significantly reduced data volume compared to the original non-text sub-information. This effectively improves the efficiency of subsequent chaotic encryption. Simultaneously, compression eliminates statistical redundancy between pixels (such as consecutive identical pixel values), allowing chaotic encryption to focus more on effective information and preventing redundant data from diluting the diffusion of the chaotic encryption algorithm. Furthermore, compression disrupts the spatial correlation between pixels, and the diffusion characteristics of chaotic encryption can further scramble the block structure. With these dual measures, it becomes more difficult to deduce the plaintext pattern from the ciphertext, thus improving the security of non-text encrypted information.

[0204] In one embodiment, using a chaotic system to chaotically encrypt non-text compressed information to obtain non-text encrypted sub-information includes the following steps:

[0205] The chaotic system is driven by a pre-set initial key to complete the chaotic mapping of non-textual compressed information and generate chaotic non-textual sequences.

[0206] quantizing the chaotic non-text sequence and generating an integer index sequence according to the quantization result;

[0207] constructing a Latin square based on the integer index sequence and using a dynamic padding rule;

[0208] scrambling the non-text compression information based on the Latin square to obtain scrambled non-text;

[0209] performing modulo addition operation on the scrambled non-text and the diffusion key to obtain intermediate non-text, the diffusion key being generated based on the Latin square;

[0210] performing diffusion operation on the intermediate non-text according to the initial key to obtain diffusion non-text;

[0211] taking the diffusion non-text as new non-text compression information and repeatedly performing the scrambling and diffusion steps until a preset maximum number of encryptions is reached, and outputting the non-text encryption sub-information.

[0212] In this embodiment, the principle of chaotic encryption is based on the nonlinearity and sensitivity to initial conditions of chaotic systems. By generating complex random sequences, data confusion and diffusion are achieved, thereby realizing encryption. Chaotic systems have inherent randomness and extreme sensitivity to initial conditions, which makes chaotic sequences exhibit good randomness and complexity in encryption, making them difficult to predict and reconstruct.

[0213] Reference Figure 6 The initial key is pre-set (such as a 128-bit binary string). The initial key is converted into control parameters and initial values of a chaotic system (such as the μ value of the Logistic map and the σ value of the Lorenz system) through a key expansion algorithm. The control parameters and initial values are used to drive the chaotic system to generate high-entropy chaotic sequences, i.e., chaotic non-text sequences. Since the chaotic non-text sequence is a sequence of floating-point numbers, which are realized by the combination of mantissa and exponent, the floating-point numbers in the chaotic non-text sequence are linearly quantized or nonlinearly quantized to obtain an integer index sequence. This process is a process of mapping the floating-point numbers in the chaotic non-text sequence to the integer interval. The integer index sequence is used as an index to construct a Latin square. Specifically, an empty initial matrix is first initialized, and then the initial matrix is filled row by row. For example, the integer index sequence is {Q1, Q2, Q3, …, Qn}, and each element in the integer index sequence belongs to {0, 1, 2, …, n-1}. The first row of the initial matrix is filled according to the integer index sequence, obtaining (0, 1, 2, …, n-1). The arrangement of the i-th row (i is greater than or equal to 2) of the initial matrix is i-1 row cyclically right shifted by Qi times, until the number of rows and the number of columns are the same. For example, the final Latin square can be:

[0214] ,

[0215] After obtaining multiple different Latin squares by the same method, the non-text compression information is divided into sub-blocks matching the order of the Latin square, different Latin squares are used to perform permutation operations on different sub-blocks, and the row and column positions of the data in the sub-blocks are permuted according to the element values of the Latin square, that is, the arrangement order of the i-th row data in the sub-block is rearranged according to the arrangement order of the i-th row element of the Latin square, and the arrangement order of the i-th column data in the sub-block is rearranged according to the arrangement order of the i-th column element of the Latin square, for example, if the 3rd row of the Latin square is [3, 0, 2, 1], the original row arrangement order of the sub-block is permuted from [a, b, c, d] to [b, d, a, c]. The permuted sub-blocks are spliced according to the arrangement positions before division to obtain the permuted non-text. The diffusion parameters (such as row sum and column XOR value) are extracted from the Latin square, and the diffusion key is generated by a nonlinear function. The nonlinear function can be an S-box or a Feistel network. Taking the S-box as an example, the diffusion parameter block is divided into S-box input units (such as 4 bits or 8 bits), and the S-box is replaced by table lookup and dynamic generation to obtain the diffusion key. Then, the permuted non-text and the diffusion key are subjected to modulo addition operation to obtain the intermediate non-text, that is, the image obtained in the middle of encryption, to realize bit-level diffusion. Finally, the intermediate non-text and the initial key are subjected to modulo addition operation to complete the diffusion operation and obtain the diffusion non-text. The diffusion non-text is used as the encrypted information to repeat the permutation and diffusion steps until the maximum number of encryption (which can be three times) is reached, and the non-text encryption sub-information is obtained. The modulo addition operation refers to d = e + f (mod n), where d is the result of the modulo addition operation, and mod represents the modulo operation, that is, the remainder operation.

[0216] The multi-level permutation and modulo addition diffusion are realized by the dynamically generated Latin square, the correlation between adjacent pixels is destroyed, and the statistical attack resistance is improved. At the same time, according to the high redundancy characteristics of the non-text information, the same color pixel continuity is quickly broken by chaotic permutation, and the low efficiency problem of traditional encryption algorithm is avoided. In summary, the chaotic mapping and Latin square permutation operation have low complexity, are suitable for large data encryption of image information (non-text information) in regional medical information, guarantee the security of non-text information, and also have fast encryption speed.

[0217] In one embodiment, the method further comprises the following steps:

[0218] Receiving the medical encryption information, and extracting the reference hash information pre-stored in the medical encryption information;

[0219] Generating a hash path according to the medical encryption information, and calculating a root hash of the medical encryption information according to the hash path;

[0220] Integrating the hash path and the root hash to obtain verification hash information;

[0221] The reference hash information is compared with the verification hash information to determine whether the medical encryption information passes the verification.

[0222] If the reference hash information is inconsistent with the verification hash information, it is determined that the medical encryption information fails the verification.

[0223] In the embodiment, the reference hash information is pre-calculated, which includes a reference hash path and a reference root hash. The calculation method of the hash path and the root hash is exactly the same as the reference hash information. First, a hash path is generated according to the medical encryption information, and a root hash is calculated according to the hash path, and the hash path and the root hash are integrated into the reference hash information. Specifically, the medical encryption information that has completed uploading is divided into multiple data blocks according to a preset fixed size (such as 1 MB), ensuring that each data block is independently verifiable, and a hash algorithm (such as SHA-256 or national secret SM3) is applied to each data block to generate a unique hash value as a leaf node of a Merkle tree. The hash values of the adjacent two leaf nodes are spliced to obtain a spliced hash value, and the hash algorithm is used again on the spliced hash value to calculate the hash value of the spliced hash value, and the hash value of the spliced hash value is taken as the parent node hash. Repeat the step of generating the parent node hash until a unique root hash is generated. The hash path is the collection of all parent node hashes and spliced hash values. Compare whether the root hash is consistent with the reference hash, and verify whether the hash path is consistent with the reference hash path. If both are consistent, it is determined that the medical encryption information passes the verification. If any one of the root hash and the hash path is inconsistent with the reference hash or the reference hash path in the reference hash information, it is determined that the medical encryption information fails the integrity verification and may be tampered with, and the uploader of the medical encryption information needs to re-encrypt and upload the regional medical information.

[0224] The application also discloses a regional medical information sharing system based on artificial intelligence.

[0225] The memory is configured to store instructions; and

[0226] The processor is configured to call the instructions from the memory and can implement the method for sharing regional medical information based on artificial intelligence according to any one of the above when executing the instructions.

[0227] The processor can be a central processing unit (CPU), of course, according to actual use, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), ready-to-program gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. can also be used, and the general-purpose processor can be a microprocessor or any conventional processor, etc. The application does not limit this.

[0228] The memory can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device, or an external storage device of the computer device, for example, a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD), or a flash card (FC) provided on the computer device, or a combination of the internal storage unit and the external storage device of the computer device. The memory is configured to store a computer program and other programs and data required by the computer device, and can be configured to temporarily store data that has been output or is to be output. The present application does not limit this.

[0229] The embodiment of the present application further provides a machine readable storage medium, which stores instructions for causing a machine to execute the method for sharing regional medical information based on artificial intelligence.

[0230] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0231] The present application is described with reference to the flowcharts and / or block diagrams of the method, device (system), computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be realized by computer program instructions, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The device that implements the function specified in one or more flows and / or blocks. Figure 1 The device that implements the function specified in one or more flows and / or blocks.

[0232] These computer program instructions can also be stored in a computer readable storage medium capable of directing the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a product including instruction devices, which implement the flowcharts and / or block diagrams. Figure 1 The device that implements the function specified in one or more flows and / or blocks. Figure 1 The device that implements the function specified in one or more flows and / or blocks.

[0233] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0234] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0235] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the processor can execute instructions. The memory can also include non-volatile memory, such as read only memory (ROM), electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), or flash memory, for example. Memory is an example of computer readable media.

[0236] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0237] It should also be noted that the terms "comprising," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0238] ​​The above merely provides an example of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of claims of the present application.

Claims

1. A method for sharing regional medical information based on artificial intelligence, characterized by The method comprises the following steps: Obtain regional medical information of a target area; Locate and extract text information in the regional medical information based on a target detection algorithm; Input the text information into a key information classification model, segment the text information, and obtain a plurality of text information sequences, wherein the key information classification model is constructed based on a recurrent neural network model, and the key information classification model comprises a feature extraction module, an element labeling module, a key degree module, and an information classification module; For any text information sequence, input the text information sequence into the feature extraction module, and extract text information features of the text information sequence through the feature extraction module; Input the text information features into the element labeling module, label the element types of the text information sequence through the element labeling module, and output a text element sequence of the text information sequence according to the element type labeling result; Calculate the sequence self-information of all text elements in the text element sequence through the key degree module and by using a self-information formula, and calculate the information entropy weight of all text elements based on the sequence self-information; Calculate the sequence mutual information between all text elements through the key degree module and by using a mutual information formula, and calculate the mutual information weight of all text elements based on the sequence mutual information; Integrate all the information entropy weights and mutual information weights to obtain the information weight of the text element sequence; Construct a text key degree matrix of the text element sequence based on the information weight; Vectorize the text key degree matrix, and output a key degree vector of the text element sequence through the key degree module; Input all the text information features and the key degree vector into the information classification module in sequence for information classification, and output an information classification result of the text information through the information classification module; Divide the text information into key text and regular text according to the information classification result; Complete text encryption of the key text based on a text encryption algorithm to obtain text encrypted information; Vectorize non-text information in the regional medical information other than the text information, and perform information compression encryption on the vectorized non-text information to obtain non-text encrypted information; Complete global encryption of the regular text, the text encrypted information, and the non-text encrypted information in the regional medical information based on a group encryption algorithm to obtain medical encrypted information; Upload the medical encrypted information to a pre-constructed medical information sharing terminal through a cloud server.

2. The method according to claim 1, characterized in that, The extraction of the text information features of the text information sequence through the feature extraction module comprises the following steps: Convert the text information sequence into an information vector sequence through the feature extraction module; Complete context semantic analysis of the information vector sequence through the feature extraction module, and extract text semantic features of all character vectors in the information vector sequence according to the context semantic analysis result; Analyze the semantic correlation between all character vectors based on the feature extraction module, and assign weights to all text semantic features according to the semantic correlation; Weighted fusion of all text semantic features with assigned weights is completed through the feature extraction module, and the text information features are output.

3. The method according to claim 1, characterized in that, The method comprises the following steps of: vectorizing non-text information in medical area medical information to obtain non-text information vector; decomposing the non-text information vector based on an empirical mode decomposition algorithm to obtain a plurality of non-text sub-information; performing lossless compression on the non-text sub-information based on an encoding technique to obtain non-text compressed information; encrypting the non-text compressed information by using a chaotic system to obtain non-text encrypted sub-information; and integrating all the non-text encrypted sub-information to obtain non-text encrypted information.

4. The method according to claim 3, characterized in that, The method comprises the following steps of: extracting extreme value intervals of the non-text information vector based on the empirical mode decomposition algorithm; calculating local mean values of the non-text information vector according to the extreme value intervals and through an integral average formula, and generating an information mean curve of the non-text information vector based on the local mean values; iteratively extracting all information components of the non-text information vector based on the information mean curve; reconstructing the all information components to obtain a plurality of vector information matrices; and constructing the plurality of non-text sub-information based on the all vector information matrices.

5. The method according to claim 3, characterized in that, The method comprises the following steps of: for any information element in the non-text sub-information, generating a predicted value of the information element according to a plurality of adjacent elements of the information element and through a linear prediction algorithm; calculating a difference absolute value between an element value of the information element and the predicted value to obtain a predicted element error; integrating all the predicted element errors to obtain predicted error non-text; decomposing the predicted error non-text based on a bit plane decomposition technique to obtain a plurality of error bit planes; dividing any error bit plane into a plurality of error non-overlapping blocks; traversing all the error non-overlapping blocks to extract elements, and synthesizing element planes according to the element extraction results; performing double encoding on all the element planes by using the encoding technique, and performing element compression on all the element planes based on the double encoding results to obtain the non-text compressed information.

6. The method according to claim 3, characterized in that, The method comprises the following steps of: driving the chaotic system to complete chaotic mapping of the non-text compressed information by using a pre-set initial key, and generating a chaotic non-text sequence; quantizing the chaotic non-text sequence, and generating an integer index sequence according to the quantization result; constructing a Latin square based on the integer index sequence and by using a dynamic padding rule; scrambling the non-text compressed information based on the Latin square to obtain scrambled non-text; performing modulo addition operation on the scrambled non-text and a diffusion key to obtain intermediate non-text, the diffusion key being generated based on the Latin square; performing diffusion operation on the intermediate non-text according to the initial key to obtain diffusion non-text; taking the diffusion non-text as new non-text compressed information, and repeatedly performing the above scrambling and diffusion steps until a preset maximum encryption number is reached, and outputting the non-text encrypted sub-information.

7. The method of claim 1, characterized in that, The method further comprises the following steps of: receive the medical encryption information, and extract reference hash information pre-stored in the medical encryption information; generate a hash path according to the medical encryption information, and calculate a root hash of the medical encryption information according to the hash path; integrate the hash path and the root hash to obtain verification hash information; check whether the reference hash information is consistent with the verification hash information, and if the reference hash information is consistent with the verification hash information, determine that the medical encryption information passes the check; if the reference hash information is not consistent with the verification hash information, determine that the medical encryption information fails the check. 8.A regional medical information sharing system based on artificial intelligence, characterized by, comprise: a memory configured to store instructions; and a processor configured to call the instructions from the memory and capable of implementing the method according to any one of claims 1 to 7 for sharing regional medical information based on artificial intelligence when executing the instructions.

Images