Digestive tract disease data encryption acquisition method and risk prediction system

Abstract

The invention relates to the technical field of electronic data digital processing, and discloses a digestive tract disease data encryption acquisition method and a risk prediction system. The method comprises the following steps: carrying out real-time semantic discrimination on an endoscope video stream, and distinguishing high / low risk data frames; homomorphic encryption based on lattice cryptography is adopted for high-risk frames, efficient symmetric encryption is adopted for low-risk frames, and mixed encrypted data streams are formed; and constructing a structured data packet containing the risk label and the encrypted data, and performing secure transmission. The system comprises a front-end differential encryption subsystem and a rear-end ciphertext domain joint prediction server, and the rear-end ciphertext domain joint prediction server realizes accurate risk prediction while protecting privacy through two-way feature extraction and security fusion. According to the method, the encryption calculation overhead is remarkably reduced, and the real-time performance, the accuracy and the end-to-end data security are considered.

Description

Technical Field

[0001] This invention belongs to the field of electronic data digital processing technology, specifically relating to a method for encrypted acquisition of digestive tract disease data and a risk prediction system. Background Technology

[0002] With the deep integration of artificial intelligence and medical big data technologies, gastrointestinal disease risk prediction systems based on high-dimensional time-series data such as endoscopic videos and physiological signals have shown great potential in clinical auxiliary diagnosis. These systems rely on the real-time collection, transmission, and analysis of massive amounts of patient data to achieve accurate identification and risk grading of early lesions (such as polyps, ulcers, or cancerous areas). However, the high sensitivity of medical data requires it to meet strict privacy protection standards throughout its entire lifecycle. Traditional practices generally employ cryptographic methods such as homomorphic encryption to perform end-to-end encryption of the original data to ensure confidentiality during transmission and computation.

[0003] In scenarios involving the prediction of gastrointestinal diseases, the core data source typically consists of continuous endoscopic video streams and their associated temporal annotations, exhibiting significant high dimensionality, redundancy, and spatiotemporal sparsity. Existing encryption schemes often ignore the inherent structural characteristics of such data, applying uniform strength full encryption to the entire video or all frame sequences. While this "one-size-fits-all" encryption strategy ensures security, the extremely high computational complexity of homomorphic operations leads to severe computational latency and resource consumption when processing high-resolution, high-frame-rate video streams. This results in model inference response times far exceeding the millisecond thresholds allowed for emergency or intraoperative decision-making, making it difficult to support real-time risk warning requirements.

[0004] Existing technologies present a fundamental contradiction in balancing data security and predictive timeliness: on the one hand, high-strength encryption is a necessary condition for meeting the Personal Information Protection Law and medical data compliance requirements; on the other hand, the computational overhead of full encryption directly restricts the real-time performance of risk prediction systems. Especially in emergency endoscopy, doctors need to know the risk level of suspicious areas within seconds to decide whether to perform a biopsy or intervention. However, existing encryption mechanisms cannot dynamically adjust security strategies based on the clinical value of the data content, resulting in both excessive encryption of non-lesion areas leading to wasted computational resources and a failure to implement differentiated protection for highly sensitive segments such as key lesion coordinates. Summary of the Invention

[0005] To address the aforementioned technical challenges, this invention provides a method for encrypted acquisition of digestive tract disease data and a risk prediction system. The core technical concept lies in establishing a differential encryption and ciphertext domain joint prediction mechanism based on clinical semantic perception. This mechanism first uses a lightweight real-time discrimination model at the data acquisition front end to analyze the original digestive tract endoscopy video stream frame by frame, automatically identifying and marking "high-risk" data segments containing suspected lesions and background "low-risk" data segments. Next, different encryption strengths are applied to data segments with different risk levels; specifically, a computationally intensive fully homomorphic encryption scheme is applied only to high-risk segments carrying core diagnostic information, while a computationally efficient symmetric encryption scheme is used for low-risk segments. Finally, a dual-path parallel ciphertext domain prediction model is constructed on the backend server. This model processes homomorphically encrypted data and symmetrically encrypted data separately, and performs secure fusion at the feature layer. Ultimately, while protecting patient privacy, it achieves rapid and accurate prediction of digestive tract disease risks.

[0006] This invention provides a method for encrypting and acquiring digestive tract disease data, which includes the following steps: Acquire raw digestive tract medical data stream, which is video sequence data acquired in real time by endoscopic equipment; The video sequence data is subjected to real-time semantic discrimination processing to generate risk semantic labels associated with each data frame. The real-time semantic discrimination processing is implemented by a pre-trained lightweight convolutional neural network model, which is used to identify suspected lesion areas in video frames and to determine each data frame as a high-risk data frame or a low-risk data frame according to a preset lesion confidence threshold. Based on the risk semantic tags, differentiated encryption operations are performed on the video sequence data to generate a hybrid encrypted data stream. Specifically, the differentiated encryption operations include: when a data frame is determined to be a high-risk data frame, a lattice-based homomorphic encryption module is invoked to encrypt the pixel matrix of the data frame using the public key of the medical institution, generating a homomorphically encrypted data block; when a data frame is determined to be a low-risk data frame, a symmetric encryption module is invoked to encrypt the pixel matrix of the data frame using a randomly generated session key, generating a symmetric encrypted data block, and simultaneously, the session key is encrypted using the public key of the medical institution to generate an encrypted session key. Construct a structured encrypted data packet, which includes a data packet header, a metadata segment, and a data payload segment. The data packet header records the unique identifier and timestamp of the data packet. The metadata segment records the index of each data frame, its corresponding risk semantic label, and the encrypted session key. The data payload segment stores the homomorphic encrypted data block and the symmetric encrypted data block in chronological order. The structured encrypted data packet is sent to the remote risk prediction server via a secure transmission protocol.

[0007] Furthermore, the lightweight convolutional neural network model is a depthwise separable convolutional structure with no more than 10 layers and fewer than 1 million parameters. It undergoes model quantization to ensure a single-frame inference latency of less than 10 milliseconds on the embedded processor of the endoscope device. The lesion confidence threshold is a preset floating-point value between 0.75 and 0.95, used to balance the recall rate of lesion detection with the computational load of data processing.

[0008] Furthermore, the lattice-based homomorphic encryption module employs a rescaled fully homomorphic encryption scheme with Chebyshev polynomial approximation, and its encryption parameters are set as follows: the polynomial modulus degree is 2. 15 The ciphertext modulus is 1228 bits, and the noise standard deviation introduced in the encryption operation is 3.2 to support deep neural network inference that performs at least 10 multiplication operations and 100 addition operations in the ciphertext field, while maintaining a cryptographic security level of no less than 128 bits.

[0009] Furthermore, the symmetric encryption module employs the Advanced Encryption Standard (AES) algorithm, with a key length set to 256 bits, and operates in Galois counter mode to provide confidentiality and integrity verification for encrypted data. The session key is generated by a cryptographically secure pseudo-random number generator at the start of each independent endoscopic examination session, ensuring one-time pad encryption.

[0010] This invention provides a gastrointestinal disease risk prediction system, comprising: A front-end data processing subsystem, deployed on a gastrointestinal endoscope or its connected edge computing unit, integrates the following: A data acquisition interface is used to acquire raw video sequence data from the endoscope's image sensor; The real-time semantic discrimination unit is equipped with a lightweight convolutional neural network model, which is used to receive the original video sequence data and output risk semantic labels indicating the presence of suspected lesions frame by frame. The differentiated encryption execution unit includes a homomorphic encryption module and a symmetric encryption module. The differentiated encryption execution unit selectively calls the homomorphic encryption module or the symmetric encryption module to encrypt the corresponding data frame based on the risk semantic tags output by the real-time semantic discrimination unit. The data packet construction and transmission unit is used to encapsulate the encrypted data frame, risk semantic tags and related metadata into a structured encrypted data packet and send it through the network interface.

[0011] The gastrointestinal disease risk prediction system also includes: A backend risk prediction server, deployed in the cloud or a medical data center, integrates the following: The data packet receiving and parsing unit is used to receive and parse the structured encrypted data packets from the front-end data processing subsystem, and separate the homomorphic encrypted data blocks, symmetric encrypted data blocks, and encrypted session keys. The key management and decryption unit uses the pre-stored private key of the medical institution to decrypt the encrypted session key, obtains the session key used for symmetric encryption, and uses the session key to decrypt the symmetric encrypted data block to obtain the plaintext data of the low-risk data frame. The ciphertext domain joint prediction unit includes a homomorphic feature extraction network and a plaintext feature extraction network. The homomorphic feature extraction network directly takes the homomorphically encrypted data block as input and performs convolution and activation function approximation operations in the ciphertext domain to extract deep encrypted features of high-risk lesion regions. The plaintext feature extraction network takes the decrypted low-risk data frame as input and extracts plaintext features of the background environment and tissue texture. The feature fusion and risk judgment unit receives the deep encryption features and the plaintext features, and generates the final encryption risk score through feature splicing and fully connected layer calculation under the control of the secure multi-party computation protocol. The result decryption and presentation unit sends the encrypted final risk score back to the authorized medical staff terminal, where the terminal device uses its locally stored private key to decrypt the score and presents the final risk level and confidence value to the user.

[0012] Furthermore, the activation function used in the homomorphic feature extraction network is a third- or fifth-order polynomial function to approximately correct the behavior of linear units within the ciphertext domain. Its pooling operation is replaced by a fixed-stride convolution operation to avoid performing a non-linear maximum selection operation on the homomorphically encrypted data. The network weight parameters are quantized to a fixed point after training to match the integer arithmetic characteristics of the homomorphic encryption scheme.

[0013] Furthermore, the secure multi-party computation protocol is an addition and multiplication protocol based on secret sharing, ensuring that feature vectors generated by different encrypted paths can be aggregated without revealing their original information during feature fusion. The weight matrix of the fully connected layer is held by the server and participates in the calculation in encrypted form, guaranteeing the intellectual property security of the prediction model itself.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces a real-time discrimination mechanism based on clinical semantics to achieve differentiated encryption processing of medical data streams. It applies computationally expensive homomorphic encryption only to high-risk data frames containing key diagnostic information, while using efficient symmetric encryption for a large amount of background data. This reduces the overall encryption computational overhead by at least one order of magnitude without sacrificing the security level of key information, fundamentally solving the performance bottleneck caused by full homomorphic encryption.

[0015] 2. The dual-path parallel ciphertext domain prediction model constructed in this invention can collaboratively process the high-risk features of homomorphic encryption and the low-risk context features of plaintext decryption, achieving complete information utilization. This architecture ensures that the accuracy of prediction depends not only on the lesion itself, but also comprehensively considers the surrounding environmental information, improving the robustness and accuracy of the prediction model. At the same time, the high-risk data is always kept in ciphertext state on the server side throughout the entire prediction process, achieving end-to-end privacy protection.

[0016] 3. The systematic design of this invention moves data identification and encryption to the front end of data acquisition, reducing the risk of exposure of raw sensitive data during transmission in the network. By combining structured encrypted data packets with backend secure decryption and ciphertext calculation, a complete closed-loop solution for predicting the risk of digestive tract diseases is formed, which takes into account real-time performance, accuracy and security, and has extremely high clinical application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the encrypted acquisition method and risk prediction system for digestive tract disease data proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the differential encryption and ciphertext domain joint prediction mechanism based on clinical semantic perception in this invention; Figure 3 This is a logical flow diagram of the real-time semantic identification and differentiated encryption execution within the front-end data processing subsystem of this invention; Figure 4 This is a logical flowchart of the joint prediction and feature fusion of the encrypted domain within the backend risk prediction server in this invention. Figure 5 This is a schematic diagram illustrating the composition structure and metadata organization relationship of the structured encrypted data packet in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the front-end and back-end in this invention. Detailed Implementation

[0018] Please refer to Figures 1 to 6This invention provides a method for encrypting and acquiring digestive tract disease data and a risk prediction system. Its core lies in achieving differentiated encryption of endoscopic video streams through lightweight semantic discrimination at the front end, and constructing a joint prediction architecture in the back end that coordinates the ciphertext and plaintext domains. This significantly reduces computational overhead and meets the real-time requirements of emergency scenarios while ensuring patient privacy and security. The following will elaborate on the specific implementation of this technical solution.

[0019] The method for encrypting and acquiring digestive tract disease data includes the following steps: First, acquiring raw digestive tract medical data streams, which are video sequence data collected in real time by endoscopic equipment. Each frame of the image consists of a red, green, and blue three-channel pixel matrix with a resolution of not less than 720×576 and a frame rate of not less than 25 frames per second; Second, performing real-time semantic discrimination processing on the video sequence data to generate risk semantic tags associated with each data frame; Next, performing differentiated encryption operations on the video sequence data according to the risk semantic tags to generate a hybrid encrypted data stream; Subsequently, constructing a structured encrypted data packet, which includes a data packet header, metadata segments, and a data payload segment; Finally, sending the structured encrypted data packet to a remote risk prediction server through a secure transmission protocol.

[0020] In the above method, the step of acquiring the raw gastrointestinal medical data stream is performed by the image sensor of the endoscope. This sensor outputs an uncompressed raw pixel data stream through a standard video interface. The data format is an unencoded bitmap format, the color space is the ITU-T 601 standard, and the sampling precision is 8 bits per channel. This raw data stream is directly input to the front-end data processing subsystem deployed inside the endoscope or in an edge computing unit directly connected to it, ensuring that the data enters a controlled processing flow before leaving the acquisition device, thus preventing the exposure of raw sensitive information on the network.

[0021] The step of performing real-time semantic discrimination processing on video sequence data is implemented through a pre-trained lightweight convolutional neural network model. This model adopts a depthwise separable convolutional structure with a total of nine layers, including four deep convolutional layers and five pointwise convolutional layers, with a total of 923,456 parameters. All weight parameters are processed by 8-bit pointwise quantization, and the activation function is uniformly a modified linear unit.

[0022] During the training phase, the model utilizes a large dataset of annotated gastrointestinal endoscopic images with lesion regions for supervised learning. The lesion types include polyps, ulcers, bleeding points, and early-stage cancerous areas. The final converged model achieved an average lesion detection recall of 13.7% and a precision of 89.2% on the test set. During the inference phase, the model is deployed on a front-end embedded processor with a clock speed of at least 1.2 GHz and a memory bandwidth of at least 16 Gigabytes per second. The single-frame inference latency is controlled within 8 milliseconds, meeting real-time processing requirements.

[0023] For each input video frame, the model outputs a lesion confidence value between 0 and 1. If the value is greater than a preset lesion confidence threshold, the frame is considered a high-risk data frame; otherwise, it is considered a low-risk data frame. The lesion confidence threshold is set to 0.85, which was determined through cross-validation. This value effectively suppresses the false alarm rate while ensuring high recall, thus balancing computational load and diagnostic sensitivity.

[0024] The step of performing differentiated encryption based on risk semantic tags is completed by a differentiated encryption execution unit. This unit integrates two independent encryption modules: a homomorphic encryption module based on lattice ciphers and a symmetric encryption module. When a data frame is determined to be a high-risk data frame, the system calls the homomorphic encryption module to encrypt the pixel matrix of the frame using a public key pre-distributed by the medical institution. This homomorphic encryption module employs a rescaled fully homomorphic encryption scheme approximated by Chebyshev polynomials, with the following core parameter configuration: the polynomial modulus degree is 32768, i.e., 2^32. 15 The ciphertext modulus is 1228 bits; the noise introduced during encryption follows a discrete Gaussian distribution with a mean of 0 and a standard deviation of 3.2. This parameter combination ensures that the encryption scheme maintains a security strength of at least 128 bits while supporting at least 10 ciphertext multiplications and 100 ciphertext additions, sufficient to support subsequent deep neural network inference within the ciphertext domain. The encrypted data block is stored as an integer vector, with each pixel value mapped to a ciphertext element of length 1228 bits.

[0025] When a data frame is determined to be a low-risk data frame, the system invokes the symmetric encryption module. This module employs the Advanced Encryption Standard (AES) algorithm, operates in Galois counter mode, and uses a 256-bit key. The session key used for encryption is dynamically generated by a cryptographically secure pseudo-random number generator at the start of each endoscopic examination session, ensuring one-time pad encryption.

[0026] The session key was then asymmetrically encrypted using the aforementioned medical institution's public key to generate an encrypted session key, which was used to securely recover the original session key on the server side later. The symmetric encryption operation performed streaming encryption on the entire frame pixel matrix, and the output data blocks maintained the same size and structure as the original frame, but the content was unreadable ciphertext data.

[0027] The step of constructing the structured encrypted data packet is performed by the data packet construction and transmission unit. This data packet uses a fixed header and a variable-length payload structure. The packet header contains a 64-bit globally unique identifier, generated by concatenating the device serial number and a timestamp hash, and a 32-bit timestamp recording the precise moment the data packet was generated, with microsecond precision. The metadata segment is organized in key-value pairs, sequentially recording the index number of each data frame, the corresponding risk semantic tag (represented by a single-byte boolean value), and the encryption session key (filled only when a low-risk frame exists).

[0028] The data payload segment arranges all encrypted data blocks sequentially according to the original temporal order of the video frames. Homomorphic encrypted data blocks corresponding to high-risk frames are interleaved with symmetric encrypted data blocks corresponding to low-risk frames, with the order strictly corresponding to the index numbers in the metadata segment. The entire data packet uses compact binary encoding without redundant padding to minimize transmission bandwidth usage.

[0029] The step of sending the structured encrypted data packets to the remote risk prediction server is completed through a transport layer security protocol, version 1.3 or higher. An elliptic curve Diffie-Hellman key exchange mechanism is used to establish a secure channel, ensuring the confidentiality and integrity of the data during transmission. The data packet sending frequency is synchronized with the video capture frame rate. The system has traffic shaping capabilities and automatically activates local caching during network congestion to prevent data loss.

[0030] On the backend risk prediction server side, upon receiving a structured encrypted data packet, the packet receiving and parsing unit first performs decapsulation. This unit deduplicates and sorts the packets based on the identifier and timestamp in the packet header, then separates the metadata segment and the data payload segment. Based on the risk semantic tags in the metadata segment, the system classifies the encrypted data blocks in the data payload segment into a set of homomorphic encrypted data blocks and a set of symmetric encrypted data blocks.

[0031] The key management and decryption unit uses the medical institution's private key pre-stored in the hardware security module to decrypt the encrypted session key in the metadata segment, recovering the original 256-bit session key. Subsequently, this session key is used to decrypt all symmetric encrypted data blocks, obtaining the plaintext pixel matrix of the low-risk data frame. This plaintext data is then fed into the plaintext feature extraction network for further processing.

[0032] The ciphertext domain joint prediction unit contains two parallel feature extraction paths. The homomorphic feature extraction network directly takes the homomorphically encrypted data block as input and performs convolution operations within the ciphertext domain. The network structure consists of five stacked convolutional layers, each with a 3×3 kernel size, a stride of 1, and uniform padding. Since homomorphic encryption does not support non-linear activation functions, the modified linear unit in the network is replaced with a third-order polynomial function, the expression of which is: ; Where the coefficient This polynomial provides a high-fidelity approximation of the corrected linear unit within the interval -10 to +10. Pooling operations are replaced by convolutional layers with a stride of 2 and a fixed kernel weight of 0.25, achieving mean pooling and avoiding the nonlinear incomputability issues associated with max pooling. After training, the network weights are converted to 16-bit specific-point integers to match the integer arithmetic characteristics of homomorphic encryption schemes.

[0033] The plaintext feature extraction network takes the decrypted low-risk frames as input, adopts a standard residual network structure, contains 18 residual blocks, and outputs a 512-dimensional feature vector to represent contextual information such as background texture, cavity morphology, and lighting conditions.

[0034] The feature fusion and risk determination unit receives feature vectors from two paths. The high-risk feature vector is in encrypted form with a dimension of 256; the low-risk feature vector is in plaintext form with a dimension of 512. The two are fused using a secure multi-party computation protocol. This protocol, based on an additive secret-sharing mechanism, divides the plaintext feature vector into two random shares, which interact with the encrypted feature vector on different computation nodes. The feature concatenation operation is completed within the secret-sharing domain, forming a 768-dimensional fused feature vector. This vector is then input into a three-layer fully connected network. The network weights are held by the server and participate in the computation in encrypted form, ultimately outputting an encrypted scalar risk score. This score ranges from 0 to 100, representing the overall risk level of gastrointestinal lesions.

[0035] The result decryption and presentation unit transmits the encrypted risk score back to the authorized medical staff's terminal through a secure channel. The terminal device has a built-in user private key to decrypt the score and generate a visual report based on preset risk level classification standards (e.g., 0 to 30 is low risk, 31 to 70 is medium risk, and 71 to 100 is high risk). The report also displays the confidence level value and keyframe index for doctors' decision-making reference.

[0036] Throughout the entire process, an exception handling mechanism is implemented. If the front-end semantic discrimination model times out of inference, the system automatically marks the current frame as high-risk and initiates homomorphic encryption to ensure security as a priority. If network transmission is interrupted, the local cache retains a maximum of the most recent 100 frames of data, which will be retransmitted after the connection is restored. If the server fails to decrypt, the system records an error log and triggers a manual review process, while suspending data reception for that device to prevent invalid data from contaminating the prediction model.

[0037] Through the above implementation methods, this invention achieves efficient and accurate prediction of gastrointestinal disease risks while protecting patient privacy. The differentiated encryption strategy strictly limits the application of homomorphic encryption to high-value data segments, reducing overall encryption latency to less than 1 / 10 of traditional full-data homomorphic encryption schemes. Simultaneously, the dual-path prediction architecture ensures that diagnostic accuracy is unaffected by the encryption strategy, meeting the stringent requirements of real-time performance and security, especially in clinical and emergency settings.

Claims

1. A method for encrypted acquisition of digestive tract disease data, characterized in that, include: Acquire raw digestive tract medical data stream, which is video sequence data acquired in real time by endoscopic equipment; The video sequence data is subjected to real-time semantic discrimination processing to generate risk semantic labels associated with each data frame. The real-time semantic discrimination processing is implemented by a pre-trained lightweight convolutional neural network model, which is used to identify suspected lesion areas in video frames and to determine each data frame as a high-risk data frame or a low-risk data frame according to a preset lesion confidence threshold. Based on the risk semantic tags, differentiated encryption operations are performed on the video sequence data to generate a hybrid encrypted data stream. Specifically, the differentiated encryption operations include: when a data frame is determined to be a high-risk data frame, a lattice-based homomorphic encryption module is invoked to encrypt the pixel matrix of the data frame using the public key of the medical institution, generating a homomorphically encrypted data block; when a data frame is determined to be a low-risk data frame, a symmetric encryption module is invoked to encrypt the pixel matrix of the data frame using a randomly generated session key, generating a symmetric encrypted data block, and simultaneously, the session key is encrypted using the public key of the medical institution to generate an encrypted session key. Construct a structured encrypted data packet, which includes a data packet header, a metadata segment, and a data payload segment. The data packet header records the unique identifier and timestamp of the data packet. The metadata segment records the index of each data frame, its corresponding risk semantic label, and the encrypted session key. The data payload segment stores the homomorphic encrypted data block and the symmetric encrypted data block in chronological order. The structured encrypted data packet is sent to the remote risk prediction server via a secure transmission protocol.

2. The method for encrypting and acquiring digestive tract disease data according to claim 1, characterized in that, The lightweight convolutional neural network model is a depthwise separable convolutional structure with no more than 10 layers, fewer than 1 million parameters, and 8-bit point quantization. The lightweight convolutional neural network model is configured to have a single-frame inference latency of less than 10 milliseconds. The lesion confidence threshold is a preset floating-point value between 0.75 and 0.

95.

3. The method for encrypting and acquiring digestive tract disease data according to claim 2, characterized in that, The lattice-based homomorphic encryption module employs a rescaled fully homomorphic encryption scheme with Chebyshev polynomial approximation, and its encryption parameters are set as follows: the polynomial modulus degree is 2. 15 The password text modulus is 1228 bits, the noise standard deviation introduced in the encryption operation is 3.2, and the encryption parameters are configured to support at least 10 multiplication operations and 100 addition operations in the ciphertext field while maintaining a password security level of not less than 128 bits.

4. The method for encrypting and acquiring digestive tract disease data according to claim 3, characterized in that, The symmetric encryption module uses the Advanced Encryption Standard algorithm, with a key length of 256 bits, and operates in Galois counter mode; the session key is generated by a cryptographically secure pseudo-random number generator at the start of each independent endoscopic examination session.

5. The method for encrypting and acquiring digestive tract disease data according to claim 4, characterized in that, In the step of constructing the structured encrypted data packet, the header of the data packet contains a 64-bit globally unique identifier and a 32-bit timestamp; The metadata segment is organized in key-value pairs, recording the index number, risk semantic tag, and encryption session key of each data frame; the data payload segment stores homomorphic encrypted data blocks and symmetric encrypted data blocks in the original video frame time sequence.

6. The method for encrypting and acquiring digestive tract disease data according to claim 5, characterized in that, The step of sending the structured encrypted data packet to the remote risk prediction server is completed through a transport layer security protocol, the version of which is no lower than 1.3, and a secure channel is established using an elliptic curve Diffie-Hellman key exchange mechanism.

7. The method for encrypting and acquiring digestive tract disease data according to claim 6, characterized in that, After receiving the structured encrypted data packet, the remote risk prediction server parses out the homomorphic encrypted data block, the symmetric encrypted data block, and the encrypted session key. The encrypted session key is decrypted using the medical institution's private key. After obtaining the session key, the symmetric encrypted data block is decrypted to obtain the plaintext pixel matrix of the low-risk data frame.

8. The method for encrypting and acquiring digestive tract disease data according to claim 7, characterized in that, The remote risk prediction server constructs a ciphertext domain joint prediction unit, which includes a homomorphic feature extraction network and a plaintext feature extraction network. The homomorphic feature extraction network takes a homomorphically encrypted data block as input and performs convolution operations and polynomial activation function approximation in the ciphertext domain to extract deep encrypted features of high-risk lesion regions. The plaintext feature extraction network takes decrypted low-risk data frames as input and extracts plaintext features of the background environment and tissue texture.

9. The method for encrypting and acquiring digestive tract disease data according to claim 8, characterized in that, The activation function in the homomorphic feature extraction network is a third-order polynomial function, expressed as follows: Pooling operations are replaced by convolutional layers with a stride of 2 and a fixed weight of 0.25; network weight parameters are quantized at specific points by 16-bit and then used in the ciphertext field calculation.

10. A risk prediction system for digestive tract diseases, characterized in that, include: The front-end data processing subsystem, deployed on the gastrointestinal endoscope or its connected edge computing unit, integrates the following: A data acquisition interface is used to acquire raw video sequence data from the endoscope's image sensor; The real-time semantic discrimination unit is equipped with a lightweight convolutional neural network model, which is used to receive the original video sequence data and output risk semantic labels indicating the presence of suspected lesions frame by frame. The differentiated encryption execution unit includes a homomorphic encryption module and a symmetric encryption module. The differentiated encryption execution unit selectively calls the homomorphic encryption module or the symmetric encryption module to encrypt the corresponding data frame based on the risk semantic tags output by the real-time semantic discrimination unit. The data packet construction and sending unit is used to encapsulate the encrypted data frame, risk semantic tags and related metadata into a structured encrypted data packet and send it through the network interface; The backend risk prediction server, deployed in the cloud or medical data center, integrates the following: The data packet receiving and parsing unit is used to receive and parse the structured encrypted data packets from the front-end data processing subsystem, and separate the homomorphic encrypted data blocks, symmetric encrypted data blocks, and encrypted session keys. The key management and decryption unit uses the pre-stored private key of the medical institution to decrypt the encrypted session key, obtains the session key used for symmetric encryption, and uses the session key to decrypt the symmetric encrypted data block to obtain the plaintext data of the low-risk data frame. The ciphertext domain joint prediction unit includes a homomorphic feature extraction network and a plaintext feature extraction network. The homomorphic feature extraction network directly takes the homomorphically encrypted data block as input and performs convolution and activation function approximation operations in the ciphertext domain to extract deep encrypted features of high-risk lesion regions. The plaintext feature extraction network takes the decrypted low-risk data frame as input and extracts plaintext features of the background environment and tissue texture. The feature fusion and risk judgment unit receives the deep encryption features and the plaintext features, and generates the final encryption risk score through feature splicing and fully connected layer calculation under the control of the secure multi-party computation protocol. The result decryption and presentation unit sends the encrypted final risk score back to the authorized medical staff terminal, where the terminal device uses its locally stored private key to decrypt the score and presents the final risk level and confidence value to the user.

Images