Cloud-edge collaborative bone age detection system and method based on dynamic computing power scheduling

Abstract

The application relates to the technical field of cloud edge collaboration, and discloses a cloud edge collaboration bone age detection system and method based on dynamic computing power scheduling, which comprises an edge layer arranged at a user side or a local end of a medical institution and used for acquiring a bone image and selecting a running mode based on a software and hardware environment; the edge layer comprises an image acquisition module connected with an imaging device to acquire the bone image; an environment and computing power monitoring module acquires software and hardware environment indexes; a decision engine selects the running mode according to the acquired software and hardware environment indexes; an image preprocessing module performs image enhancement on the acquired bone image; a local lightweight inference module performs bone age detection based on an output image of the image preprocessing module; and a cloud server layer arranged at a remote central cloud computing center and provided with a deep bone development quantitative evaluation large model and used for performing bone age detection based on the output image of the image preprocessing module.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge collaborative technology, and in particular to a cloud-edge collaborative bone age detection system and method based on dynamic computing power scheduling. Background Technology

[0002] In recent years, AI-based bone age prediction technology based on deep learning models has gradually matured. To enable these complex AI models to serve a wider population, the technical architecture has begun to evolve towards networking and cloud-based approaches. How to utilize distributed computing architectures (such as cloud-edge collaboration and edge computing) to ensure the access of massive and diverse terminals while achieving optimal allocation of computing power and secure data flow is a crucial macro-level technical issue that urgently needs to be addressed in the field of distributed medical computing. Specifically focusing on the application scenario of hand bone age imaging detection, Chinese invention application CN121999991A discloses a remote AI detection method and system for digital X-ray images of bone age. This system adopts a "fully cloud-based remote transmission and centralized computing" architecture that relies entirely on a central cloud server. This solution involves static cloud-internal cluster scheduling in terms of computing power flow, lacks edge-end perception capabilities, and does not include flexible slicing and distribution mechanisms for cloud-edge collaboration or local disaster recovery and degradation mechanisms. Chinese invention application CN120451376A discloses a method for constructing a bone age detection model, a prediction method, and a device. It simultaneously inputs hand X-ray images into a global model and a specific local model based on K bone age segments. Global features and bone age-specific features are extracted and converted into first and second probability distributions, respectively. Finally, a fusion module outputs a third probability distribution to obtain the predicted bone age. However, this method does not involve cross-network topology computing power scheduling, nor does it employ edge-end environmental awareness or adaptive offline traffic distribution mechanisms. Therefore, the core problems of this existing technology include:

[0003] Computing power scheduling is entirely limited to server load balancing within the cloud. Its edge hardware is only used as a "transfer tool," and its remaining computing power is completely idle. It cannot dynamically switch computing pipelines according to the real-time hardware load at the edge.

[0004] It is extremely dependent on high-bandwidth networks. Once grassroots medical institutions face network congestion, high packet loss rate or sudden network outage, the cloud cannot receive images, which will completely paralyze the entire testing business and lack disaster recovery and degradation mechanisms.

[0005] Uploading the entire original X-ray image incurs huge bandwidth overhead, and the original image often carries the patient's external signs or unnecessary environmental labels; although existing technologies have related content anonymization at the transmission layer, they do not physically isolate the data source, and there is still a risk of privacy leakage during cloud-edge transmission. Summary of the Invention

[0006] The purpose of this invention is to solve the problems in the prior art.

[0007] The technical solution adopted by this invention to solve its technical problem is: to provide a cloud-edge collaborative bone age detection system based on dynamic computing power scheduling, comprising:

[0008] An edge layer, deployed on the user side or local end of a medical institution, is used to acquire skeletal images and select an operating mode based on the hardware and software environment. The edge layer includes an image acquisition module, which connects to imaging equipment to acquire skeletal images; an environment and computing power monitoring module, which acquires hardware and software environment indicators; a decision engine, which selects an operating mode based on the acquired hardware and software environment indicators; an image preprocessing module, which performs image enhancement on the acquired skeletal images; and a local lightweight inference module, which performs bone age detection based on the output image of the image preprocessing module.

[0009] The cloud server layer, deployed in a remote central cloud computing center, is equipped with a large-scale model for quantitative assessment of deep skeletal development, and performs bone age detection based on the output images from the image preprocessing module.

[0010] This invention also provides a cloud-edge collaborative bone age detection method based on dynamic computing power scheduling, which, based on the above system, includes the following steps:

[0011] The image acquisition module acquires raw hand X-ray image data, while the environment and computing power monitoring module starts polling the software and hardware environment indicators.

[0012] The decision engine determines the operating mode based on hardware and software environment indicators, and sends the control instructions containing the operating mode along with the raw hand X-ray image data to the image preprocessing module.

[0013] The image preprocessing module selects the preprocessing content according to the operating mode, and outputs the preprocessed image to the local lightweight inference module and / or cloud server layer;

[0014] Local lightweight inference modules and / or large-scale models for deep skeletal development quantitative assessment perform bone age detection based on preprocessed images.

[0015] Preferably, the software and hardware environment indicators include:

[0016] The local real-time remaining computing power value is the percentage of real-time available computing power load of the central processing unit and graphics processing unit of the current local computing node;

[0017] The system calculates the real-time network bandwidth value, sends probe messages to the remote central server, calculates the round-trip latency and network packet loss rate between the current terminal and the cloud server, and assesses the current available network bandwidth.

[0018] Preferably, the decision engine determines the operating mode based on software and hardware environment indicators, including two levels of decision-making:

[0019] First-level decision: If the real-time network bandwidth value is lower than the set hard threshold for network outage / extremely weak network, the network is determined to be unsuitable for flow and the full edge mode is selected; otherwise, proceed to the second-level decision.

[0020] Secondary decision-making: If the local real-time remaining computing power is lower than the device overload hard threshold, it is determined that the local computing power is severely insufficient and the full cloud mode is selected; otherwise, the cloud-edge collaboration mode is selected.

[0021] Preferably, the image preprocessing module selects the preprocessing content according to the operating mode, and outputs the preprocessed image to the local lightweight inference module and / or the cloud server layer, including the following steps:

[0022] Basic image processing is performed on the raw hand X-ray image data, including adjusting the global contrast of the image through histogram equalization algorithm to eliminate exposure differences; and removing noise introduced by the equipment by using median filtering or Gaussian noise reduction algorithm to improve the clarity of local texture features.

[0023] After the image enhancement process is completed, further processing is selected according to the operating mode indicated by the control command.

[0024] Preferably, the step of selecting further processing according to the operating mode indicated by the control command includes the following steps when the operating mode is full edge mode:

[0025] The image preprocessing module inputs the complete skeletal image after basic image enhancement into the local lightweight inference module inside the local edge layer to perform the entire process of in-situ offline inference.

[0026] Preferably, the step of selecting further processing according to the operating mode indicated by the control command includes the following steps when the operating mode is a full cloud mode:

[0027] The image preprocessing module performs lossless hard compression with a high compression rate on the raw hand X-ray image data after basic enhancement, and then uploads the entire image to the cloud server layer.

[0028] The large-scale quantitative assessment model for deep skeletal development at the cloud server layer performs bone age detection based on the complete image uploaded from the edge layer.

[0029] Preferably, the step of selecting further processing according to the operating mode indicated by the control command includes the following steps when the operating mode is cloud-edge collaboration mode:

[0030] The image preprocessing module calls the built-in lightweight object detection and spatial segmentation network to dynamically locate the core skeletal development assessment area in hand images, including phalanges, metacarpals and carpal bones;

[0031] Based on the positioning bounding box, the image is physically cropped and separated into several independent local joint image blocks in the spatial dimension, thereby removing large areas of blank background tissue and unnecessary patient identification labels;

[0032] The decision engine determines whether to use cloud inference or cloud-edge collaborative inference based on the current cloud load.

[0033] Preferably, the decision engine determines whether to use cloud inference or cloud-edge collaborative inference based on the current cloud load, including the following steps:

[0034] The decision engine refines the dispatch path based on the latest cloud load status, which includes the current concurrent queue length and server load coefficient in the cloud.

[0035] If the cloud load status shows queuing and congestion, the image preprocessing module will directly input the phalangeal image blocks with low computational complexity and small distortion directly to the local lightweight inference module for edge-end local inference. At the same time, the wrist bone image blocks with extremely high computational complexity and relying on deep large models will be packaged into a local joint feature stream and uploaded to the cloud server layer. Otherwise, the entire set of local joint image block sequences will be packaged into a local joint feature stream and uploaded to the cloud server layer.

[0036] The large-scale quantitative assessment model for deep skeletal development at the cloud server layer performs bone age detection based on local joint feature streams uploaded from the edge layer.

[0037] Preferably, the process of bone age detection at the cloud server layer includes the following steps:

[0038] The distributed task queue at the cloud server layer receives data packets uploaded from various edge layers across the entire network and uses a message queue mechanism to perform traffic shaping, deduplication, and queuing for high-concurrency requests.

[0039] After the distributed task queue completes task buffering, it distributes the tasks to the computing power cluster in an orderly manner.

[0040] For local joint feature streams, the computing power cluster performs multi-threaded parallel feature extraction on each independent joint image block of the input, and reconstructs the spatial and semantic dimensions through a multi-scale feature fusion layer. Finally, the comprehensive bone age value is calculated through a regression network. For the whole image, the computing power cluster completes the whole image spatial segmentation and large model deep inference in the cloud, and outputs the bone age result.

[0041] The present invention has the following beneficial effects:

[0042] (1) This invention breaks through the limitations of existing technologies that only perform load balancing within the central cloud server room, extending the reach of scheduling directly to the edge nodes of medical institutions. Through the edge's self-awareness of computing power and network, an elastic computing power grid composed of three dynamically switchable business pipelines—"full cloud, collaborative, and full edge"—is constructed, realizing the on-demand flow of computing power across boundaries between the "cloud" and the "edge." In the face of extreme situations such as cloud overload and congestion caused by large-scale screening or sudden network outages, the system can find the optimal solution by dynamically adjusting the computing ratio. Even in extreme network-free environments, primary hospitals can still complete the initial bone age screening by relying on the local degradation mechanism of Mode 3, maximizing the utilization of the idle value of edge hardware and ensuring the continuity and robustness of core medical services.

[0043] (2) Based on the hybrid inference collaborative path, this invention successfully realizes a flexible architecture that breaks down a hand X-ray into joints of different developmental sites and distributes them to local and remote computing sites simultaneously. The system can finely balance the computing load at both ends of the cloud and edge based on the instantaneous load of the central cloud and the local hardware status. For finger bone image blocks with low computational complexity and small distortion, local inference is retained at the edge, while wrist bone image blocks with extremely high computational complexity and dependence on deep large models are uploaded to the central cloud simultaneously. This achieves task decoupling and dynamic allocation at the single joint level, maximizing the combined efficiency of the distributed computing grid.

[0044] (3) This invention directly shreds the joint locally at the edge layer using an image preprocessing module, accurately removing over 60% of the useless blank background image area. When transmitting to the cloud, it can not only transmit the simplified local joint image blocks, but also further convert them into one-dimensional low-dimensional feature vectors (such as 512-dimensional embedding vectors) for pure data transfer. This not only completely frees the system from dependence on high bandwidth such as 5G private networks and greatly reduces the throughput overhead of public networks, but also completely erasures the patient's privacy information at the data source, making it extremely suitable for secure transfer in public network environments.

[0045] (4) This invention introduces a functionally integrated edge preprocessing mechanism. While the edge is finding joints (segmenting), it also performs image noise reduction and enhancement processing, ensuring that all data uploaded to the cloud-based distributed task queue is "high-value joint data" that has undergone refined preprocessing. Under this architecture, the remote central cloud server no longer needs to waste computing power on inefficient front-end processes such as basic "hand finding, joint finding, and image preprocessing." The cloud computing power cluster can be 100% used for the operation of the core high-precision bone age assessment and prediction model, which significantly improves the recognition accuracy, calculation speed, and overall concurrent throughput of the central cloud model.

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of a cloud-edge collaborative bone age detection system based on dynamic computing power scheduling, according to an embodiment of the present invention.

[0048] Figure 2 This is a flowchart illustrating a cloud-edge collaborative bone age detection method based on dynamic computing power scheduling, according to an embodiment of the present invention.

[0049] Figure 3 This is a flowchart illustrating step S1 of a cloud-edge collaborative bone age detection method based on dynamic computing power scheduling, according to an embodiment of the present invention.

[0050] Figure 4 This is a flowchart illustrating step S2 of a cloud-edge collaborative bone age detection method based on dynamic computing power scheduling, according to an embodiment of the present invention.

[0051] Figure 5 This is a flowchart of steps S3 to S5 of a cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0053] See Figure 1 As shown, the system of this embodiment is logically and physically divided into an edge layer and a cloud server layer. It relies on three dynamic control nodes and one closed-loop output node to achieve dynamic routing of the computing pipeline. The specific module division, control nodes, physical relationships, and core technical functions within each layer are as follows:

[0054] (1) Edge layer, deployed on the user side or local end of the medical institution, is responsible for the acquisition of raw data, hardware and software environment awareness, composite preprocessing and spatial segmentation of images, full-edge offline computing and path distribution. It integrates the following functional modules and control nodes:

[0055] The image acquisition and input module, the system's image data source receiver, is responsible for connecting to digital medical imaging equipment such as DR / CR to enable on-site reading, buffering, and digital signal input of raw hand X-ray image data. The output of this module is unidirectionally connected to the environmental and computing power monitoring module.

[0056] The environment and computing power monitoring module is a hardware unit for environmental awareness within the system. It is responsible for real-time polling and underlying hardware monitoring of the available computing power load and remaining memory of the local computing nodes' CPU and GPU. Simultaneously, it sends probe messages to detect the real-time available network bandwidth and packet loss rate between the terminal and the remote central server. This module receives trigger signals from the image acquisition input module at its input end, and its output end is unidirectionally connected to the decision engine.

[0057] The decision engine, the core collaborative scheduling and control hub of the system, is responsible for receiving real-time environmental indicators from the monitoring module and running its internal adaptive cost boundary function to dynamically generate traffic splitting control logic. Its inputs are connected to the environment and computing power monitoring modules, as well as the distributed task queue in the cloud (for receiving status feedback). Its output is unidirectionally connected to the image preprocessing module, and its control signals are connected to the full-edge computing nodes, full-cloud inference nodes, or cloud-edge collaborative inference nodes.

[0058] The image preprocessing module, a core component of edge-end composite image processing, possesses dual functions: composite image processing and spatial joint segmentation. On one hand, it performs image optimization actions such as histogram equalization, noise reduction, and enhancement on the raw X-ray images; on the other hand, it invokes a lightweight object detection and segmentation network to locate and physically crop out independent key joint regions such as phalanges, metacarpals, and carpal bones. The module's input connects to the control output of the decision engine, and its output has a bidirectional feedback loop connecting back to the decision engine. Furthermore, its data outputs are unidirectionally connected to the full-edge computing node, the full-cloud inference node, and the cloud-edge collaborative inference node.

[0059] The full-edge computing (policy transfer node) is an emergency degradation control node within the edge layer. Activated in response to the decision engine's hard switchover command for network outages / weak networks, it is responsible for forcibly severing all network communication connections. The node's input is connected to the image preprocessing module, and its output is unidirectionally connected to the local lightweight inference module.

[0060] The local lightweight inference module, the local computing core at the edge, internally deploys a lightweight compressed bone age regression prediction model. It is responsible for handling the entire local prediction process in a fully edge-computing environment, or for handling a portion of the lightweight joint region bone age inference tasks offloaded from the edge preprocessing module in a cloud-edge collaborative inference environment. The module's input is unidirectionally connected to the fully edge-computing node and the cloud-edge collaborative inference node, and its output is unidirectionally connected to the results output module.

[0061] (2) Cross-level dynamic control and closed-loop points, belonging to the flexible task distribution path connecting the edge layer and the cloud server layer and the final business closed-loop end, including:

[0062] The all-cloud inference (policy flow node) is a cross-layer computing power absolute offloading control node. Activated in response to edge overload instructions from the decision engine, it controls the edge layer to skip complex spatial joint segmentation actions, performing hard compression on the original hand image before transferring it over the network. The node's input connects to the decision engine and the image preprocessing module, while its data output crosses the network boundary, unidirectionally inputting to a distributed task queue in the cloud.

[0063] The cloud-edge collaborative inference (policy flow node) is a flexible, cross-layer task decoupling and distribution node. Activated in response to collaborative commands from the decision engine, it controls the data flow for refined splitting and distribution: it unidirectionally distributes low-complexity, low-distortion finger bone image blocks segmented by the image preprocessing module to the local lightweight inference module, while simultaneously uploading computationally complex wrist bone image blocks, which rely on deep, large models, via the network. The node's input connects to the image preprocessing module, and its outputs are unidirectionally connected to the local lightweight inference module and the distributed task queue in the cloud.

[0064] The results are output to the entire network's closed-loop business terminals, deployed in viewing rooms, printers, or user display terminals. Their inputs are connected in parallel to receive local preliminary screening results from the local lightweight inference module, and also connect unidirectionally across levels to receive high-precision inference reports from the cloud's high-performance computing cluster, enabling on-site visualization of the detection results and a closed-loop business process.

[0065] (3) Cloud server layer, deployed in a remote central cloud computing center, serves as a centralized high-precision computing pool for massive high-concurrency data across the entire network. It integrates the following two functional modules:

[0066] This distributed task queue is a cloud-based network throughput and traffic shaping buffer module. Its input receives compressed full graphs uploaded via cloud-wide inference nodes or local joint feature streams uploaded via cloud-edge collaborative inference nodes. It is responsible for queuing, deduplication, concurrency control, and asynchronous distribution of high-concurrency requests. This module has a bidirectional connection to the cloud-based high-performance computing cluster and a control signal feedback connection to the edge layer's decision engine.

[0067] The system utilizes a high-performance cloud computing cluster and a high-precision core inference engine. Internally, it deploys a large-scale model for quantitative assessment of deep skeletal development and interacts bidirectionally with a distributed task queue. It is responsible for performing multi-scale feature fusion and deep regression inference on individual joint slices (or performing full-process inference on the entire image), calculating the final quantitative comprehensive bone age value, and generating a standardized clinical bone age assessment report. Its output crosses network layer boundaries and is unidirectionally connected to the results output module.

[0068] (4) Explanation of the role of the hardware improvements in the system. In the system of this invention, the "environment and computing power monitoring module" involving hardware improvements plays a bridging and foundational role in the overall cloud-edge adaptive elastic scheduling. Traditional medical imaging AI systems lack direct perception of the edge hardware status and cannot cope with sudden network interruptions and local queuing overloads. By introducing this hardware monitoring unit, this invention enables the system to convert the physical CPU / GPU remaining load of local computing nodes and cross-network round-trip time (RTT) into quantifiable discrete decision-making criteria in real time. This provides the "decision engine" with the physical enablement conditions for dynamic hard switching between the three mutually exclusive hardware computing pipelines of "full cloud", "collaboration", and "full edge", ensuring the continuity and network robustness of medical bone age screening services from the hardware source.

[0069] See Figure 2 The diagram shown is a flowchart of a cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to an embodiment of the present invention, including the following steps:

[0070] S1, the image acquisition module acquires raw hand X-ray image data, while the environment and computing power monitoring module starts polling the software and hardware environment indicators;

[0071] S2, the decision engine determines the operating mode based on the software and hardware environment indicators, and sends the control instructions containing the operating mode along with the original hand X-ray image data to the image preprocessing module;

[0072] S3, the image preprocessing module selects the preprocessing content according to the running mode, and outputs the preprocessed image to the local lightweight inference module and / or cloud server layer;

[0073] S4, a local lightweight inference module and / or a large model for deep skeletal development quantitative assessment, performs bone age detection based on preprocessed images.

[0074] Specifically, the cloud-edge collaborative bone age detection method based on dynamic computing power scheduling in this embodiment of the invention further includes: S5, result cross-layer backhaul and business closure, that is, transmitting the bone age detection result to the local edge layer.

[0075] See Figure 3 As shown, step S1 performs image acquisition and environment awareness initialization, including the following steps:

[0076] S101: The "Image Acquisition Input Module" of the local edge layer reads the raw hand X-ray image data captured by imaging equipment such as DR / CR on-site through a standard medical digital imaging communication interface (DICOM) or data transmission pipeline, and caches it in local memory.

[0077] S102: After the "Image Acquisition Input Module" successfully receives the image data, it immediately sends a trigger signal to the "Environment and Computing Power Monitoring Module".

[0078] S103: The "Environment and Computing Power Monitoring Module" responds to this trigger signal by initiating parallel polling of software and hardware environment indicators through the underlying system's low-level interface and network detection mechanism.

[0079] Dynamically obtain the real-time available computing power load percentage of the central processing unit (CPU) and graphics processing unit (GPU) of the current local computing node, and record it as the real-time remaining computing power value P_edge;

[0080] Send probe messages to the remote central server, calculate the round-trip time (RTT) and network packet loss rate between the current terminal and the cloud server, and evaluate the current available network bandwidth, which is denoted as the real-time network bandwidth value B_net.

[0081] S104: The "Environment and Computing Power Monitoring Module" unidirectionally packages and inputs the collected parameter matrix (including raw image data, real-time remaining computing power value P_edge, and real-time network bandwidth value B_net) into the "Decision Engine".

[0082] See Figure 4 As shown, S2 executes multi-level traffic splitting and control scheduling logic, including the following steps:

[0083] S201: After receiving the parameter matrix, the "decision engine" synchronously retrieves the current concurrent queue length and server load coefficient in the cloud, which are fed back in real time by the "distributed task queue" in the cloud, and records them as the cloud load status L_cloud.

[0084] S202: The "Decision Engine" performs multi-level logical judgments on the P_edge, B_net, and L_cloud inputs into the preset adaptive cost boundary function to generate computation pipeline flow instructions.

[0085] Branch Decision A: If the real-time network bandwidth value B_net is lower than the set hard threshold B_min for network outage / extremely weak network, it is determined that the network does not meet the conditions for flow, and the "decision engine" issues a hard switch command to the next pipeline to activate "Mode 3: Full Edge Mode".

[0086] Branch decision B: If the real-time network bandwidth value B_net is higher than the normal network threshold, and the local real-time remaining computing power value P_edge is lower than the device overload hard threshold P_min (indicating that the local computing power is seriously insufficient), then the "decision engine" sends a control command to the next pipeline to activate "mode 2: full cloud mode".

[0087] Branch decision C: If the real-time network bandwidth value B_net is greater than the normal network threshold, and the local real-time remaining computing power value P_edge is sufficient (greater than P_min), then the "decision engine" sends a control command to the next pipeline to activate "Mode 1: Cloud-Edge Collaboration Mode".

[0088] S203: The "Decision Engine" will send the control command generated by the decision, along with the original hand X-ray image data, unidirectionally to the "Image Preprocessing Module".

[0089] See Figure 5 The diagram shown is a flowchart of steps S3 to S5.

[0090] Specifically, S3 performs composite image processing and spatial joint segmentation, including the following steps:

[0091] S310: The "Image Preprocessing Module" first performs basic image composite image processing on the raw hand X-ray image data according to the received control instructions:

[0092] The global contrast of the image is adjusted by a histogram equalization algorithm to eliminate exposure differences.

[0093] Median filtering or Gaussian noise reduction algorithms are used to remove noise introduced by the device and improve the clarity of local texture features.

[0094] S320: After image enhancement processing is completed, the "Image Preprocessing Module" further executes mutually exclusive spatial splitting and inference pipelines according to the operating mode indicated by the control instructions:

[0095] If the control command is Mode 1, Cloud-Edge Collaboration Mode, execute the following steps:

[0096] S321-A: The "Image Preprocessing Module" calls the built-in lightweight object detection and spatial segmentation network (such as a pruned and optimized YOLO or U-Net model) to dynamically locate the core skeletal development assessment areas such as phalanges, metacarpals, and carpal bones in hand images.

[0097] S321-B: Based on the positioning bounding box, the image is physically cropped and separated into several independent local joint image blocks in the spatial dimension, thereby removing large areas of blank background tissue and unnecessary patient identification labels.

[0098] S321-C: The image preprocessing module uses the local processing time and cropping completion status as closed-loop signals, and feeds them back to the decision engine to update the decision engine's strategy library.

[0099] S321-D: The decision engine refines the dispatch path based on the latest cloud load status L_cloud.

[0100] Path 1 (Pure edge processing + full cloud inference): The "image preprocessing module" packages the complete set of local joint image block sequences into a local joint feature stream, crosses hierarchical boundaries, and uploads it to the "distributed task queue" in the cloud through the communication network.

[0101] Path 2 (Hybrid Inference Collaboration): If the cloud load L_cloud indicates queuing congestion, the "Image Preprocessing Module" performs local decoupling and distribution: the finger bone image blocks with low computational complexity and small distortion are directly and unidirectionally input to the "Local Lightweight Inference Module" for edge-side local inference; at the same time, the wrist bone image blocks with extremely high computational complexity and dependence on deep large models are packaged and uploaded to the "Distributed Task Queue" of the remote cloud server layer via the network.

[0102] If the control command is in Mode 2, Full Cloud Mode, execute the following steps:

[0103] S322: After receiving the instruction, the "image preprocessing module" skips the complex spatial joint segmentation action locally and only performs lossless hard compression with a high compression rate on the original image that has undergone basic enhancement. Then, it crosses the layer boundary and directly uploads it to the "distributed task queue" of the cloud server layer, realizing the absolute offloading of computing power to the cloud.

[0104] If the control command is mode 3, full edge mode, perform the following steps:

[0105] S323: The "Decision Engine" issues a local closed-loop flow instruction to the system, logically suspending or closing the uplink communication interface facing the external network, realizing the offline elastic degradation of the computing pipeline to the local environment; the "Image Preprocessing Module" responds to this instruction, inputting the complete hand X-ray image data after basic image enhancement into the "Local Lightweight Inference Module" inside the local edge layer, completely detaching from the remote cloud server and shielding cross-network interaction, to perform the entire process of local offline inference.

[0106] Specifically, S4 performs cloud-based distributed buffering and parallel feature inference, including the following steps:

[0107] S401: The "distributed task queue" of the cloud server layer receives data packets uploaded from edge nodes across the entire network (local joint feature streams of mode 1, or whole graph data of mode 2), and uses the message queue mechanism to perform traffic shaping, deduplication, and queuing for high-concurrency requests.

[0108] S402: After the "distributed task queue" completes task buffering, it conducts bidirectional data flow with the "cloud high-performance computing cluster" to distribute tasks to the computing cluster in an orderly manner.

[0109] S403: The "Cloud-based High-Performance Computing Cluster" utilizes multi-card or distributed high-performance GPU computing resources to initiate high-precision inference for a large-scale model for deep skeletal development quantitative assessment.

[0110] If the received data is a joint feature stream, the computing cluster performs multi-threaded parallel feature extraction on each independent joint image block, and reconstructs the spatial and semantic dimensions through a multi-scale feature fusion layer. Finally, the comprehensive bone age value is calculated through a regression network.

[0111] If the received image is a compressed full image of mode two, the computing cluster will connect and complete the full image spatial segmentation and large model deep inference in the cloud, and output the bone age result.

[0112] Specifically, the cross-layer return of S5 execution results and business closure includes the following steps:

[0113] S501: For Mode 1 and Mode 2, after the "cloud-based high-performance computing cluster" completes high-precision inference and automatically generates a standardized clinical bone age detection report, its output end packages and encrypts the data, crosses hierarchical boundaries, and sends it back in a directional, one-way manner to the "local edge layer" that issued the request.

[0114] S502: For Mode 3, the "Local Lightweight Inference Module" completely handles the computational task locally and outputs the preliminary bone age screening results.

[0115] S503: The edge layer executes corresponding closed-loop actions based on the different operating modes mentioned above through the result output module:

[0116] For Mode 2 or Path 1 of Mode 1, the high-precision detection report returned by the remote cloud server is directly received and visualized; for Mode 3, the preliminary screening results output by the local lightweight inference module are read and visualized locally.

[0117] For path 2 (hybrid reasoning collaboration) in mode 1, the result output module synchronously listens to and receives the local calculation results and the high-precision local results returned from the remote end, performs multi-path asynchronous result alignment, timestamps are aligned and fused and stitched together the local bone age results with different spatial dimensions and different arrival times, and finally completes the visualization presentation and local closed loop.

[0118] Specifically, in the Mode 1 path of step S320, after the image preprocessing module completes the spatial dimension segmentation of key joints, its output is not limited to transmitting cropped joint image blocks. It can also directly call a lightweight feature extraction network locally to convert each segmented joint region image into a one-dimensional low-dimensional feature vector (such as a 512-dimensional embedding vector). The "distributed task queue" of the central cloud server layer directly receives and distributes the low-dimensional feature vector to the computing power cluster, and the "cloud-based high-performance computing power cluster" directly performs quantitative assessment and prediction of bone age based on the feature vector. This scheme reduces network bandwidth consumption to near zero by degenerating image transfer into data transfer.

[0119] Specifically, the distributed task flow and data interaction between the local edge layer and the cloud server layer of this invention are not limited to an active push streaming architecture using Webhook or standard HTTP / HTTPS. They can also be replaced by a "publish-subscribe" asynchronous topology architecture based on lightweight distributed MQTT, AMQP, or Kafka message queues. By introducing a distributed relay message broker at the cross-network boundary, with edge computing nodes acting as publishers and cloud-based distributed task queues acting as subscribers (or bidirectional subscriptions with role swapping), the system's resilience and asynchronous peak-shaving performance under high-concurrency bone age detection requests are enhanced. In this architecture, the relay message broker supports a long-connection bidirectional keep-alive heartbeat mechanism maintained by the client, as well as a one-way outbound connection handshake initiated by the edge node, to penetrate the complex internal firewall isolation zone of the medical institution's local area network. In addition, the publish-subscribe architecture supports dynamic degradation adjustment of service quality level (QoS), offline message persistence and re-transmission mechanism, and end-to-end asynchronous message acknowledgment (ACK) feedback loop, so as to ensure the lossless and deterministic flow of local joint feature streams and bone age assessment reports in heterogeneous network fluctuation environment.

[0120] Specifically, in the hybrid inference collaborative path of step S321-D, the inference task allocation ratio between the edge layer and the cloud server layer is not limited to a fixed joint type or skeletal part. It can also be replaced by an adaptive adjustment mechanism based on the real-time resource load and dynamic latency of the edge layer nodes: the decision engine obtains the system CPU remaining availability rate, system available memory remaining rate, GPU video memory remaining rate and the current local actual inference time of a single joint from the monitoring module in real time, and dynamically calculates the proportion of joints that undertake inference locally at the edge end through a preset edge resource attenuation formula.

[0121] Specifically, the physical deployment location and hardware form of the "edge layer" node described in this invention are not limited to independent local edge computing servers or integrated computing cabinets. They can also be replaced by a heterogeneous in-situ or localized deployment architecture that is infinitely close to the user terminal. Specifically, the local edge computing node can be directly embedded in image acquisition equipment (such as the image processing unit inside DR / CR medical imaging equipment, or the head chip of a hand X-ray machine) to achieve in-situ computing; or it can be deployed in a local area network node extremely close to the user terminal (such as a doctor's image reading workstation, a local PC industrial control computer, a microcomputing chip in a mobile handheld terminal, or a dedicated gateway within a hospital department) to achieve localized computing. This solution is designed to adapt to diverse and heterogeneous physical medical scenarios, further reducing transmission latency within the local area network.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cloud-edge collaborative bone age detection system based on dynamic computing power scheduling, characterized in that, include: An edge layer, deployed on the user side or local end of a medical institution, is used to acquire bone images and select the operating mode based on the software and hardware environment; the edge layer includes an image acquisition module, which connects to imaging equipment to acquire bone images; The environment and computing power monitoring module acquires software and hardware environment indicators; the decision engine selects the operating mode based on the acquired software and hardware environment indicators; and the image preprocessing module performs image enhancement on the acquired skeletal images. A local lightweight inference module performs bone age detection based on the output images of the image preprocessing module; The cloud server layer, deployed in a remote central cloud computing center, is equipped with a large-scale model for quantitative assessment of deep skeletal development, and performs bone age detection based on the output images from the image preprocessing module.

2. A cloud-edge collaborative bone age detection method based on dynamic computing power scheduling, based on the system described in claim 1, characterized in that, Includes the following steps: The image acquisition module acquires raw hand X-ray image data, while the environment and computing power monitoring module starts polling the software and hardware environment indicators. The decision engine determines the operating mode based on hardware and software environment indicators, and sends the control instructions containing the operating mode along with the raw hand X-ray image data to the image preprocessing module. The image preprocessing module selects the preprocessing content according to the operating mode, and outputs the preprocessed image to the local lightweight inference module and / or cloud server layer; Local lightweight inference modules and / or large-scale models for deep skeletal development quantitative assessment perform bone age detection based on preprocessed images.

3. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 2, characterized in that, The hardware and software environment indicators include: The local real-time remaining computing power value is the percentage of real-time available computing power load of the central processing unit and graphics processing unit of the current local computing node; The system calculates the real-time network bandwidth value, sends probe messages to the remote central server, calculates the round-trip latency and network packet loss rate between the current terminal and the cloud server, and assesses the current available network bandwidth.

4. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 3, characterized in that, The decision engine determines the operating mode based on software and hardware environment indicators, including two levels of decision-making: First-level decision: If the real-time network bandwidth value is lower than the set hard threshold for network outage / extremely weak network, the network is determined to be unsuitable for flow and the full edge mode is selected; otherwise, proceed to the second-level decision. Secondary decision-making: If the local real-time remaining computing power is lower than the device overload hard threshold, it is determined that the local computing power is severely insufficient and the full cloud mode is selected; otherwise, the cloud-edge collaboration mode is selected.

5. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 2, characterized in that, The image preprocessing module selects the preprocessing content according to the operating mode and outputs the preprocessed image to the local lightweight inference module and / or cloud server layer, including the following steps: Basic image processing is performed on the raw hand X-ray image data, including adjusting the global contrast of the image through histogram equalization algorithm to eliminate exposure differences; and removing noise introduced by the equipment by using median filtering or Gaussian noise reduction algorithm to improve the clarity of local texture features. After the image enhancement process is completed, further processing is selected according to the operating mode indicated by the control command.

6. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 5, characterized in that, The step of selecting further processing based on the operating mode indicated by the control command, when the operating mode is full edge mode, includes the following steps: The image preprocessing module inputs the complete skeletal image after basic image enhancement into the local lightweight inference module inside the local edge layer to perform the entire process of in-situ offline inference.

7. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 5, characterized in that, The step of selecting further processing based on the operating mode indicated by the control command, when the operating mode is full cloud mode, includes the following steps: The image preprocessing module performs lossless hard compression with a high compression rate on the raw hand X-ray image data that has undergone basic enhancement, and then uploads the entire image to the cloud server layer. The large-scale quantitative assessment model for deep skeletal development at the cloud server layer performs bone age detection based on the complete image uploaded from the edge layer.

8. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 5, characterized in that, The step of selecting further processing according to the operating mode indicated by the control command, when the operating mode is cloud-edge collaboration mode, includes the following steps: The image preprocessing module calls the built-in lightweight object detection and spatial segmentation network to dynamically locate the core skeletal development assessment area in hand images, including phalanges, metacarpals and carpal bones; Based on the positioning bounding box, the image is physically cropped and separated into several independent local joint image blocks in the spatial dimension, thereby removing large areas of blank background tissue and unnecessary patient identification labels; The decision engine determines whether to use cloud inference or cloud-edge collaborative inference based on the current cloud load.

9. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to claim 8, characterized in that, The decision engine determines whether to use cloud inference or cloud-edge collaborative inference based on the current cloud load, including the following steps: The decision engine refines the dispatch path based on the latest cloud load status, which includes the current concurrent queue length and server load coefficient in the cloud. If the cloud load status shows queuing and congestion, the image preprocessing module will directly input the phalangeal image blocks with low computational complexity and small distortion directly to the local lightweight inference module for edge-end local inference. At the same time, the wrist bone image blocks with extremely high computational complexity and relying on deep large models will be packaged into a local joint feature stream and uploaded to the cloud server layer. Otherwise, the entire set of local joint image block sequences will be packaged into a local joint feature stream and uploaded to the cloud server layer. The large-scale quantitative assessment model for deep skeletal development at the cloud server layer performs bone age detection based on local joint feature streams uploaded from the edge layer.

10. The cloud-edge collaborative bone age detection method based on dynamic computing power scheduling according to any one of claims 7 to 9, characterized in that, The process of bone age detection at the cloud server layer includes the following steps: The distributed task queue at the cloud server layer receives data packets uploaded from various edge layers across the entire network and uses a message queue mechanism to perform traffic shaping, deduplication, and queuing for high-concurrency requests. After the distributed task queue completes task buffering, it distributes the tasks to the computing power cluster in an orderly manner. For local joint feature streams, the computing power cluster performs multi-threaded parallel feature extraction on each independent joint image block of the input, and reconstructs the spatial and semantic dimensions through a multi-scale feature fusion layer. Finally, the comprehensive bone age value is calculated through a regression network. For the whole image, the computing power cluster completes the whole image spatial segmentation and large model deep inference in the cloud, and outputs the bone age result.

Images