An edge-computing-based medical image real-time analysis method, device and medium

Abstract

The application discloses a kind of based on edge computing's medical image real-time analysis method, equipment and medium, it is related to medical image processing technical field, including, establish communication channel and collect and transmit image data frame;Edge processing node carries out caching and pre-processing according to key frame management strategy, image data frame is divided into image sub-region;Lightweight feature transformation is carried out in image sub-region, and feature vector is extracted and local analysis confidence and analysis uncertainty are calculated;According to analysis uncertainty, processing time consumption and resource utilization, task unloading decision is made, and uplink data packet is transmitted to cloud analysis node;Cloud analysis node carries out fine analysis to data, generates cloud image analysis data and cloud analysis confidence, and is transmitted back to edge processing node;Edge processing node compares cloud and local analysis confidence and outputs image analysis final result.The stability of image processing, cloud edge collaborative efficiency and the improvement of diagnosis real-time are realized by key frame management and task unloading decision.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, device and medium for real-time analysis of medical images based on edge computing. Background Technology

[0002] With the improvement of resolution and sampling rate of medical imaging equipment, clinical diagnosis has placed higher demands on the real-time processing capability of image data. Conventional medical image analysis often relies on central servers for batch processing, using convolutional neural networks, feature extraction operators, and multi-layer fusion structures to achieve lesion identification and feature segmentation. Such methods are usually based on cloud computing platforms, using high-performance GPU clusters to complete deep learning inference and feature fusion analysis, thus playing an important role in image diagnosis assistance, quantitative measurement of lesions, and intelligent hierarchical assessment. In recent years, edge computing, as a distributed intelligent processing mode, has been introduced into the field of medical imaging to complete some feature calculations and decision inference near the image acquisition end, thereby reducing latency and improving real-time performance.

[0003] However, in the real-time analysis of medical images, conventional methods still have limitations in terms of the stability of data transmission links and the confidence fusion of multi-node collaborative processing. On the one hand, centralized cloud processing requires high bandwidth support, which can easily cause image frame loss and feedback delay when the network fluctuates or latency increases, affecting the timeliness of diagnosis. On the other hand, edge processing nodes lack an adaptive evaluation mechanism for analysis uncertainty when computing independently, making it difficult to achieve dynamic task allocation based on task reliability and resource occupancy status. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a real-time medical image analysis method based on edge computing to solve the problems of insufficient cloud-edge collaborative delay control and lack of confidence-adaptive fusion mechanism in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a real-time medical image analysis method based on edge computing, which includes establishing a communication channel with quality of service assurance between an image acquisition terminal and an edge processing node, and acquiring and transmitting image data frames.

[0008] Edge processing nodes manage the caching of each image data frame according to the keyframe management strategy;

[0009] The image data frames cached in the local circular buffer are preprocessed and divided into several image sub-regions;

[0010] Lightweight feature transformation is performed within the image sub-region to obtain the feature vector of the image sub-region, and the local analysis confidence and analysis uncertainty of the image data frame are calculated by the discriminant operator.

[0011] Based on the analysis uncertainty of the image data frame, the current total processing time, and the resource utilization rate, a task offloading decision is made, and an uplink data packet is generated and transmitted to the cloud analysis node.

[0012] The cloud analytics node receives and decodes uplink data packets, performs fine-grained analysis on the decoded image sub-regions, and generates cloud image analytics data and cloud analytics confidence scores.

[0013] Pack cloud image analysis data and cloud analysis confidence into downlink data packets and send them back to the edge processing node;

[0014] After receiving cloud image analysis data from downlink data packets, the edge processing node compares the cloud analysis confidence level with the local analysis confidence level and outputs the image analysis data with the higher confidence level.

[0015] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the communication channel with quality of service assurance includes establishing a communication transmission link through the TCP / IP protocol and marking the communication priority identifier using the IEEE 802.1Q protocol at the data link layer.

[0016] After the communication transmission link is established, the edge processing nodes are initialized and uniformly configured.

[0017] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the edge processing node performs cache management on each image data frame according to a keyframe management strategy, specifically as follows:

[0018] Image data frames are written to a local circular buffer, which is managed in a first-in-first-out manner using read and write pointers.

[0019] Edge processing nodes synchronize read and write pointers through a dual-thread mechanism. When transmission jitter is detected in the image data frames in the local circular buffer, the key frame management strategy is triggered.

[0020] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the specific steps of dividing the image into several sub-regions are as follows:

[0021] After the image data frame is preprocessed, the image data frame is divided into several non-overlapping sub-blocks, and the grayscale variance and edge intensity of each non-overlapping sub-block are calculated.

[0022] Calculate the region score for each non-overlapping sub-block based on grayscale variance and edge intensity;

[0023] Each non-overlapping sub-block is sorted from high to low according to its region score, and the top-scoring non-overlapping sub-blocks are selected as image sub-regions.

[0024] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the steps of performing a lightweight feature transformation operation within the image sub-region to obtain the feature vector of the image sub-region, and calculating the local analysis confidence and analysis uncertainty of the image data frame through a discriminant operator are as follows:

[0025] The image sub-region is input into a lightweight convolutional structure, which outputs the feature vector of the image sub-region through three convolutional layers.

[0026] Perform L2 normalization on the feature vectors of the image sub-regions to obtain the normalized feature vectors;

[0027] The weight coefficients of the image sub-regions are calculated based on their regional scores. The feature vectors of all image sub-regions are then weighted and fused based on their weight coefficients to obtain the global feature vector of the image data frame.

[0028] The global feature vectors are used to calculate the local analysis confidence and analysis uncertainty of the image data frame using a discriminant algorithm.

[0029] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the discriminant operator is composed of a linear transformation layer and a sigmoid activation function layer.

[0030] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the specific steps of determining task offloading based on the analysis uncertainty of the image data frame, the current total processing time, and resource utilization, and generating uplink data packets for transmission to the cloud analysis node are as follows:

[0031] Edge processing nodes acquire the analysis uncertainty of image data frames and calculate the current total processing time and resource utilization of image data frames in real time;

[0032] If the analysis uncertainty is greater than the uncertainty threshold, a reliability decision is triggered, and the image data frame is marked as an unreliable frame.

[0033] If the current total processing time exceeds the maximum allowable end-to-end delay, a delay decision is triggered, and the image data frame is marked as a delay-stressed frame.

[0034] If the resource utilization rate exceeds the maximum resource utilization rate limit, a resource decision is triggered, and the image data frame is marked as a high-load frame.

[0035] If the reliability decision, latency decision, and resource decision are not triggered, the local image analysis data is directly output at the edge processing node based on the image data frame.

[0036] If any decision is triggered, the current image data frame is sent to the candidate offload queue. The candidate offload queue compresses and packages the image data frame according to priority to generate an uplink data packet and uploads it to the cloud analysis node.

[0037] As a preferred embodiment of the edge computing-based real-time medical image analysis method of the present invention, the cloud analysis node receives and decodes uplink data packets, performs refined analysis on the decoded image sub-regions, and generates cloud image analysis data and cloud analysis confidence scores. The specific steps are as follows:

[0038] The cloud analytics node verifies the received uplink data packets;

[0039] The verified uplink data packets are decrypted using a shared key and decompressed using a decompression algorithm to obtain the image sub-regions and feature vectors of the image sub-regions in the uplink data packets.

[0040] The cloud analytics node performs multi-layer convolution and feature fusion on image sub-regions and their feature vectors using a high-precision medical image depth model to obtain cloud image analysis data and cloud analytics confidence scores.

[0041] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the edge computing-based real-time medical image analysis method as described in the first aspect of the present invention.

[0042] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the edge computing-based real-time medical image analysis method as described in the first aspect of the present invention.

[0043] The beneficial effects of this invention are as follows: By setting a keyframe management strategy and a local circular cache structure at the edge processing node, adaptive caching and jitter correction of continuous image frames are achieved, effectively avoiding data frame loss and temporal sequence disorder, thereby improving the stability of image data processing at the edge; by dividing the image into sub-regions and using preset discrimination operators, intelligent confidence assessment and uncertainty quantification at the edge are achieved, enabling adaptive task offloading decisions, thereby improving the overall cloud-edge collaboration efficiency; through task offloading decisions, the edge processing node achieves comprehensive adaptive scheduling of task reliability, latency constraints, and load status, thereby effectively controlling end-to-end processing latency while ensuring diagnostic accuracy. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a real-time medical image analysis method based on edge computing.

[0046] Figure 2 This is a flowchart of the keyframe management strategy.

[0047] Figure 3 A flowchart for calculating confidence levels in local analysis.

[0048] Figure 4 A flowchart for the task unloading decision. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Reference Figures 1-4 This is one embodiment of the present invention, which provides a real-time medical image analysis method based on edge computing, including the following steps:

[0053] S1. Establish a communication channel with quality of service guarantee between the image acquisition terminal and the edge processing node, and acquire and transmit image data frames.

[0054] Furthermore, the communication channel is transmitted through the hospital's internal local area network, and the communication protocol adopts a reliable transmission method based on TCP / IP. At the data link layer, the IEEE 802.1Q protocol is used to mark the communication priority identifier.

[0055] To ensure the real-time transmission of image data, edge processing nodes mark the image data frames that affect the analysis task as high-priority service categories, so that the image data frames are forwarded at the switch layer with priority over non-real-time data streams.

[0056] After the communication transmission link is established, the edge processing node initializes, reads the configuration file, and uniformly configures the communication and computing parameters.

[0057] The specific details of the unified configuration are as follows (A1-A4):

[0058] A1: Set the maximum allowable end-to-end latency for edge processing nodes, such as 200ms. End-to-end latency includes data acquisition latency, transmission latency, edge processing latency, and display latency.

[0059] It should be noted that the end-to-end latency will be written into the QoS control table during the analysis initiation phase.

[0060] A2: Edge processing nodes obtain real-time available bandwidth from the network interface and set an uplink bandwidth limit, usually 80% of the available bandwidth. They also limit the uplink data load of each frame of image data, usually to 256kbit. Data exceeding the limit will be compressed first.

[0061] A3: Set a maximum resource utilization limit, such as 80%, and set an uncertain threshold; when the edge processing node detects that the maximum resource utilization of either the CPU or GPU exceeds the limit, the number of concurrent tasks is reduced; when the uncertainty is greater than the uncertainty threshold, the task unloading mechanism is triggered.

[0062] A4: Every few seconds, such as 1 second, the edge processing node sends a link heartbeat packet to the cloud analysis node. The cloud analysis node returns an acknowledgment response. The edge processing node calculates the average link latency based on the round-trip time and records the moving average of the average link latency over the most recent few times. If the average link latency is detected to be greater than 50ms three times in a row, it is determined that the current link is in a congested state and a diversion operation is performed.

[0063] The diversion operations include reducing the uplink transmission frequency and lowering the priority of image data in the QoS control table.

[0064] It should be noted that the uncertainty threshold was obtained by randomly selecting 5,000 images from the multimodal medical image data sample, using a discriminant operator to calculate the confidence level of each medical image, recording the corresponding true labels, statistically analyzing the misclassification rate in different uncertainty intervals, plotting ROC curves, and selecting the inflection point that allows the misclassification rate to be <5% while maintaining the retention rate of >70% as the uncertainty threshold.

[0065] The maximum permissible end-to-end latency is determined based on the tolerance of doctors to perceived latency in real-time clinical diagnostic tasks. When the latency from image acquisition to display does not exceed 200ms, operation and visual feedback can be synchronized, meeting the real-time requirements of medical care.

[0066] The maximum resource utilization limit is determined by setting the maximum resource utilization of the edge processing node as the limit when the latency increases significantly and the stability decreases under load. It is usually set to 80%.

[0067] Furthermore, the image acquisition terminal transmits the acquired image data frames to the edge processing node in real time through a communication channel with guaranteed quality of service.

[0068] Among them, the image acquisition terminals include, but are not limited to, CT scanners, MRI systems, and digital ultrasound devices.

[0069] S2. Edge processing nodes perform cache management for each image data frame according to the keyframe management strategy.

[0070] Furthermore, when the edge processing node receives an image data frame, the communication thread parses the frame header information and verifies the continuity of the timestamp. If the frame sequence number of the image data frame is continuous and the CRC check passes, the current image data frame is written to the local circular buffer. If frame loss or delay is detected, the current image data frame is marked as an abnormal frame and a retransmission request is triggered.

[0071] The local circular buffer refers to a fixed-length circular data structure allocated in the main memory of the edge processing node, which can hold 30 frames of image data. Each buffer structure in the circular data structure includes the frame number, timestamp, image data pointer, and processing status identifier of the image data frame.

[0072] The processing status indicators include unprocessed, processing, and completed; 0, 1, and 2 are used to represent the processing status of the current image data frame, respectively.

[0073] The local circular buffer uses read and write pointers for first-in-first-out management. When the write pointer and read pointer coincide and the write operation continues, the oldest frame is automatically overwritten, thus achieving circular replacement.

[0074] Furthermore, to ensure the continuity and timing accuracy of the image data stream, the edge processing nodes employ a dual-thread mechanism for frame synchronization.

[0075] Dual-threading includes a communication thread and a processing thread.

[0076] The communication thread is responsible for receiving image data frames and updating the local circular buffer write pointer, while the processing thread is responsible for reading the latest unprocessed image data frames from the local circular buffer.

[0077] The two threads coordinate with each other using mutexes and semaphores to prevent data misalignment caused by concurrent writes.

[0078] When the timestamp interval between image data frames is detected to be greater than 1.5 times the average frame period, it indicates that there is transmission jitter, and the key frame management strategy is triggered.

[0079] The keyframe management strategy is as follows (B1-B2):

[0080] B1: If the number of frames remaining in the local circular buffer is ≥5, then the oldest frame is discarded and the read pointer is readjusted to make the current frame sequence number continuous.

[0081] B2: If the number of frames remaining in the local circular buffer is less than 5, then pause receiving new frames and wait for the buffer to be replenished.

[0082] S3. Preprocess the image data frames cached in the local circular buffer and divide them into several image sub-regions.

[0083] Furthermore, the edge processing node preprocesses the image data frames; the preprocessing includes grayscale normalization and spatial resampling.

[0084] Furthermore, the edge processing node performs grayscale normalization on the image data frames to eliminate the differences in imaging brightness between different image acquisition terminals. Specifically, the edge processing node calculates the moving average grayscale and grayscale standard deviation in the most recent frames, such as 16 frames of image data, and normalizes the pixel grayscale values ​​of the current image data frame. The grayscale distribution of the normalized image data frame is limited to the range of [-1, 1], so that images from different devices and under different exposure conditions can be analyzed at the same grayscale scale.

[0085] The edge processing node performs spatial resampling on the image data frames after grayscale normalization to unify the image data frames to a resolution of 512×512 pixels; the resampling operation is completed using a bilinear interpolation algorithm.

[0086] Furthermore, the edge processing node divides the preprocessed image data frame into 4×4 non-overlapping sub-blocks, each non-overlapping sub-block being 128×128 pixels in size, and calculates the grayscale variance and edge intensity of each non-overlapping sub-block.

[0087] The edge intensity is obtained by calculating the average horizontal and vertical gradient magnitudes using the Sobel operator.

[0088] The edge processing node calculates the region score for each non-overlapping sub-block according to the weighted formula, as follows:

[0089] ;

[0090] in, Indicates the first Region scores for non-overlapping sub-blocks Indicates the first The grayscale variance of non-overlapping sub-blocks Indicates the first The edge strength of each non-overlapping sub-block This represents the maximum value of the grayscale variance in the current image data frame. This represents the maximum edge intensity in the current image data frame. Indicates the weight of grayscale variance. This represents the edge strength weight.

[0091] It should be noted that in medical imaging, differences in tissue density directly correspond to changes in grayscale. The larger the grayscale variance, the more obvious the brightness change within the current area, often representing tissue interfaces and abnormal densities, such as lesion edges and bleeding areas. Therefore, grayscale variance is a key indicator for measuring local information content. Noise in medical images is mainly manifested as low-amplitude grayscale fluctuations. If the weight of grayscale variance is too large, noisy areas may be misjudged as highly significant areas. According to statistical analysis, the contribution of grayscale variance in the overall judgment is within a reasonable range of 50% to 70%. Therefore, the weight of grayscale variance is generally set to 0.6.

[0092] It should be noted that edge intensity is obtained by extracting the gradient magnitude using the Sobel operator, and is used to represent structural boundaries and texture clarity. In medical imaging, edge changes often correspond to organ contours, blood vessel orientation, or tumor boundaries, and are key indicators for judging lesion morphology. Edge intensity is sensitive to high-frequency noise, especially in ultrasound images with low signal-to-noise ratios where false edges may appear. Therefore, edge intensity should participate in the overall judgment, but should not dominate it. To balance noise suppression and structure preservation, the edge intensity weight is generally set to 0.4.

[0093] Furthermore, after the region scores of non-overlapping sub-blocks are calculated, the edge processing nodes are sorted from high to low according to the region scores, and the highest-scoring non-overlapping sub-blocks are selected as image sub-regions, such as selecting 4 non-overlapping sub-blocks.

[0094] S4. Perform a lightweight feature transformation operation within the image sub-region to obtain the feature vector of the image sub-region, and calculate the local analysis confidence and analysis uncertainty of the image data frame through a discriminant operator.

[0095] Furthermore, the edge processing node allocates an independent thread for feature extraction for each image sub-region.

[0096] The feature extraction algorithm is based on a lightweight convolutional structure; the lightweight convolutional structure consists of three layers, specifically C1-C3:

[0097] C1, First convolutional layer: The kernel size is 3×3, the stride is 1, and it extracts local gradient and texture change information. The number of output channels is usually set to 16.

[0098] C2, Second convolutional layer: The kernel size is 1×1, used for channel dimensionality reduction, compressing the number of output channels to 8.

[0099] C3, Global Average Pooling Layer: Transforms the two-dimensional feature map of an image sub-region into a single 8-dimensional feature vector, with each component representing the average response intensity of the image sub-region.

[0100] To ensure the comparability of features across different image sub-regions, L2 normalization is performed on each feature vector to unify the magnitude of all feature vectors to 1.

[0101] Furthermore, the edge processing node fuses the feature vectors of all image sub-regions in the same image data frame to obtain the global feature vector of the current image data frame.

[0102] The fusion method uses a weighted summation, expressed as:

[0103] ;

[0104] in, Indicates the first Global fusion feature vector of frame image data Indicates the first Weighting coefficients for each image sub-region Indicates the first Feature vectors extracted from each image sub-region This indicates the number of image sub-regions in the current frame.

[0105] The weight coefficients of the image sub-regions are obtained by normalizing the region scores, and are expressed as follows:

[0106] ;

[0107] in, Indicates the first The region score of each image sub-region Indicates the first The region score of each image sub-region This represents the sum of scores for all image sub-regions within the current frame.

[0108] It should be noted that, ≠ .

[0109] Furthermore, a set of pre-trained discriminant operators are set within the edge processing node; wherein, the discriminant operator is composed of a linear transformation layer and a sigmoid activation function layer, which is used to quickly estimate the local analysis confidence of the current image data frame.

[0110] The local analysis confidence score of the current image data frame is calculated using the discriminant operator and is expressed as follows:

[0111] ;

[0112] in, Indicates the first Local analysis confidence level of frame image data This represents the Sigmoid activation function. Represents the linear transformation weight vector. This represents the constant offset.

[0113] It should be noted that the pre-training of the discriminant operator is performed using an offline dataset, which contains multimodal medical image samples, such as CT, MRI, and ultrasound. Each medical image sample is accompanied by a doctor's diagnostic annotation, such as lesion area or healthy area. During the training phase, feature vectors of image sub-regions in each image data frame are calculated using the offline dataset, and lesion category labels are determined based on the annotation results. The lesion category labels include lesion label 0 and non-lesion label 1. The goal of pre-training is to minimize the error between the local analysis confidence value of the pre-training output and the actual diagnostic result by using a cross-entropy loss function. During offline training, a learning rate is set, such as 0.001, and the Adam optimizer is used for iterative training to obtain the finally converged linear transformation weight vector and constant offset.

[0114] Furthermore, based on the calculated local analysis confidence of the current image data frame, the analysis uncertainty is calculated simultaneously and expressed as:

[0115] ;

[0116] in, Indicates the first Uncertainty in the analysis of frame image data.

[0117] S5. Based on the analysis uncertainty of the image data frame, the current total processing time, and the resource utilization rate, make a task offloading decision and generate an uplink data packet to be transmitted to the cloud analysis node.

[0118] Furthermore, the edge processing node reads the local analysis confidence and analysis uncertainty of the current image data frame.

[0119] Based on the set maximum resource utilization limit, uncertain threshold, and maximum allowable end-to-end latency, the task unloading decision is executed.

[0120] Task unloading decisions include reliability decisions, latency decisions, and resource decisions.

[0121] The reliability decision refers to marking the current frame's image data frame as an "unreliable frame" if the analysis uncertainty is greater than the uncertainty threshold.

[0122] The delay decision refers to the edge processing node calculating the total processing time of the current image data frame in real time. If the total processing time is greater than the maximum allowable end-to-end delay, the current image data frame is marked as a "delay-stressed frame".

[0123] Resource determination refers to the process whereby, if the edge processing node obtains the average utilization rate of the CPU and GPU in real time and calculates the resource utilization ratio of the current image data frame, if the resource utilization ratio of the current image data frame is greater than the maximum resource utilization limit, it is determined to be a "high load frame".

[0124] If none of the three conditions—reliability judgment, latency judgment, and resource judgment—are met, the current image data frame will be directly used to generate local image analysis data at the edge processing node and then output.

[0125] The local image analysis data includes image data frame number, image data frame timestamp, global feature vector, local analysis confidence, analysis uncertainty, regional score of image sub-region, and lesion category label.

[0126] If any condition is triggered, the current image data frame is sent to the candidate offload queue. The candidate offload queue compresses and packages the image data frame in sequence according to the priority order of unreliable frames, time-stressed frames, and high-load frames, and then uploads it to the cloud analysis node. The compression uses the Zstd compression algorithm.

[0127] Furthermore, the edge processing nodes adopt a layered transmission mechanism, which includes a main transmission channel and a redundant control channel. The main transmission channel transmits uplink data packets via the UDP protocol, while the redundant control channel transmits acknowledgment signals and retransmission requests via the TCP protocol.

[0128] Before each transmission, the edge processing node adds a 12-byte header to the data packet. The header includes the packet type, length, and timestamp.

[0129] When a cloud analytics node receives an uplink data packet, it immediately returns an ACK confirmation message. If the edge processing node does not receive an ACK confirmation message within 200ms, it will retransmit the packet, with a maximum of 3 attempts.

[0130] Furthermore, to ensure the security of medical data during transmission, a two-layer encryption is used in the uplink stage. Specifically, AES-128 is used to symmetrically encrypt the feature stream and image stream at the uplink data packet layer, and an encrypted tunnel is established at the communication layer using the TLS1.3 protocol for end-to-end protection.

[0131] The uplink data includes the current image data frame number, the current image data frame acquisition timestamp, and the feature vector of the image sub-region.

[0132] S6. The cloud analytics node receives and decodes uplink data packets, performs refined analysis on the decoded image sub-regions, and generates cloud image analytics data and cloud analytics confidence scores.

[0133] Furthermore, the cloud analytics node continuously monitors the UDP main transmission channel and the TCP redundant control channel through the main receiving thread. When a new uplink data packet is detected, the cloud analytics node compares the packet type, length, and timestamp recorded in the header information of the uplink data packet. If the comparison is consistent, the uplink data packet is decrypted using the shared key using AES-128. After decryption, the Zstd decompression algorithm is used to restore the image and feature vector of the image sub-region.

[0134] Based on the acquired image and feature vectors of the image sub-regions, the cloud analysis node calls a high-precision medical image depth model for diagnostic analysis.

[0135] The high-precision medical image depth model is a 50-layer convolutional neural network structure, trained on a large number of labeled medical image samples.

[0136] The input layer of the high-precision medical imaging depth model receives the image sub-region image and feature vector uploaded by the edge processing node. The feature is stitched through the fusion layer. After stitching, it enters the fully connected layer and the Softmax classification layer, and outputs cloud image analysis data and cloud analysis confidence. Based on the output cloud analysis confidence, the maximum confidence is taken as the diagnostic result category of the cloud image analysis data, and the corresponding cloud analysis confidence is recorded.

[0137] S7. Package the cloud image analysis data and cloud analysis confidence into a downlink data packet and send it back to the edge processing node.

[0138] Furthermore, after the cloud analytics node completes the refined analysis and cloud confidence assessment of the uploaded data, it encapsulates the cloud image analysis data and cloud analytics confidence into downlink data packets and transmits the downlink data packets via the UDP protocol.

[0139] After receiving cloud image analysis data from downlink data packets, the edge processing node compares the cloud analysis confidence level with the local analysis confidence level and outputs the image analysis data with the higher confidence level.

[0140] This embodiment also provides a computer device applicable to the real-time medical image analysis method based on edge computing, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the real-time medical image analysis method based on edge computing as proposed in the above embodiment.

[0141] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0142] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the real-time medical image analysis method based on edge computing as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0143] In summary, this invention achieves adaptive caching and jitter correction for continuous image frames by setting keyframe management strategies and local circular caching structures at edge processing nodes, effectively avoiding data frame loss and temporal sequence disorder, thereby improving the stability of image data processing at the edge. By dividing the image into sub-regions and using preset discrimination operators, it achieves intelligent confidence assessment and uncertainty quantification at the edge, enabling adaptive task offloading decisions and improving overall cloud-edge collaboration efficiency. Through task offloading decisions, it achieves comprehensive adaptive scheduling of task reliability, latency constraints, and load status at edge processing nodes, thereby effectively controlling end-to-end processing latency while ensuring diagnostic accuracy.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time medical image analysis method based on edge computing, characterized in that: include, A communication channel with quality of service assurance is established between the image acquisition terminal and the edge processing node to acquire and transmit image data frames. Edge processing nodes manage the caching of each image data frame according to the keyframe management strategy; The image data frames cached in the local circular buffer are preprocessed and divided into several image sub-regions; Lightweight feature transformation is performed within the image sub-region to obtain the feature vector of the image sub-region, and the local analysis confidence and analysis uncertainty of the image data frame are calculated by the discriminant operator. Based on the analysis uncertainty of the image data frame, the current total processing time, and the resource utilization rate, a task offloading decision is made, and an uplink data packet is generated and transmitted to the cloud analysis node. The cloud analytics node receives and decodes uplink data packets, performs fine-grained analysis on the decoded image sub-regions, and generates cloud image analytics data and cloud analytics confidence scores. Pack cloud image analysis data and cloud analysis confidence into downlink data packets and send them back to the edge processing node; After receiving cloud image analysis data from downlink data packets, the edge processing node compares the cloud analysis confidence level with the local analysis confidence level and outputs the image analysis data with the higher confidence level.

2. The real-time medical image analysis method based on edge computing as described in claim 1, characterized in that: The communication channel with quality of service assurance includes establishing a communication transmission link through the TCP / IP protocol and marking the communication priority identifier using the IEEE 802.1Q protocol at the data link layer. After the communication transmission link is established, the edge processing nodes are initialized and uniformly configured.

3. The real-time medical image analysis method based on edge computing as described in claim 2, characterized in that: The edge processing node performs cache management for each image data frame according to the keyframe management strategy. The specific steps are as follows: Image data frames are written to a local circular buffer, which is managed in a first-in-first-out manner using read and write pointers. The edge processing node synchronizes read and write pointers through a dual-thread mechanism. When transmission jitter is detected in the image data frame in the local circular buffer, the key frame management strategy is triggered.

4. The real-time medical image analysis method based on edge computing as described in claim 3, characterized in that: The process of dividing the image into several sub-regions involves the following steps: After the image data frame is preprocessed, the image data frame is divided into several non-overlapping sub-blocks, and the grayscale variance and edge intensity of each non-overlapping sub-block are calculated. Calculate the region score for each non-overlapping sub-block based on grayscale variance and edge intensity; Each non-overlapping sub-block is sorted from high to low according to its region score, and the top-scoring non-overlapping sub-blocks are selected as image sub-regions.

5. The real-time medical image analysis method based on edge computing as described in claim 4, characterized in that: The steps involve performing a lightweight feature transformation operation within the image sub-region to obtain the feature vector of the image sub-region, and then calculating the local analysis confidence and analysis uncertainty of the image data frame using a discriminant operator. The image sub-region is input into a lightweight convolutional structure, which outputs the feature vector of the image sub-region through three convolutional layers. Perform L2 normalization on the feature vectors of the image sub-regions to obtain the normalized feature vectors; The weight coefficients of the image sub-regions are calculated based on their regional scores. The feature vectors of all image sub-regions are then weighted and fused based on these weight coefficients to obtain the global feature vector of the image data frame. The global feature vectors are used to calculate the local analysis confidence and analysis uncertainty of the image data frame using a discriminant algorithm.

6. The real-time medical image analysis method based on edge computing as described in claim 5, characterized in that: The discriminant operator consists of a linear transformation layer and a sigmoid activation function layer.

7. The real-time medical image analysis method based on edge computing as described in claim 6, characterized in that: The steps for determining task offloading based on the analysis uncertainty of image data frames, the current total processing time, and resource utilization, and generating uplink data packets for transmission to the cloud analysis node are as follows: Edge processing nodes acquire the analysis uncertainty of image data frames and calculate the current total processing time and resource utilization of image data frames in real time; If the analysis uncertainty is greater than the uncertainty threshold, a reliability decision is triggered, and the image data frame is marked as an unreliable frame. If the current total processing time exceeds the maximum allowable end-to-end delay, a delay decision is triggered, and the image data frame is marked as a delay-stressed frame. If the resource utilization rate exceeds the maximum resource utilization rate limit, a resource decision is triggered, and the image data frame is marked as a high-load frame. If the reliability decision, latency decision, and resource decision are not triggered, the local image analysis data is directly output at the edge processing node based on the image data frame. If any decision is triggered, the current image data frame is sent to the candidate offload queue. The candidate offload queue compresses and packages the image data frame according to priority to generate an uplink data packet and uploads it to the cloud analysis node.

8. The real-time medical image analysis method based on edge computing as described in claim 7, characterized in that: The cloud analytics node receives and decodes uplink data packets, performs refined analysis on the decoded image sub-regions, and generates cloud image analysis data and cloud analysis confidence scores. The specific steps are as follows: The cloud analytics node verifies the received uplink data packets; The verified uplink data packets are decrypted using a shared key and decompressed using a decompression algorithm to obtain the image sub-regions and feature vectors of the image sub-regions in the uplink data packets. The cloud analytics node performs multi-layer convolution and feature fusion on image sub-regions and their feature vectors using a high-precision medical image depth model to obtain cloud image analysis data and cloud analytics confidence scores.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the real-time medical image analysis method based on edge computing as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the real-time medical image analysis method based on edge computing as described in any one of claims 1 to 8.

Images