Method and system for dynamic optimization of magnetic resonance image reconstruction task parameters combining metadata and reinforcement learning

Abstract

The application discloses a kind of magnetic resonance image reconstruction task parameter dynamic optimization method and system combined with metadata and reinforcement learning.Method includes: acquisition magnetic resonance equipment parameter and reconstruction task log constructs metadata association graph, carries out task initial allocation and constructs population;Population distribution is as reinforcement learning environment state, and association graph is combined to construct population balance factor;According to balance factor, the weight of elite selection and roulette selection is dynamically adjusted, and adaptive switching mutation strategy generates offspring;Cooling law is introduced to calculate metadata real-time access heat by neural network, and dynamically adjust knowledge migration probability;With the change value of cache hit rate as reward feedback, when it is lower than threshold value, migration probability is corrected and agent parameters are updated by combining Markov chain predicted future access sequence;Migration allocation is executed and node delay is monitored, and termination is satisfied when tolerance is met.The application realizes the adaptive scheduling and load balancing of magnetic resonance image reconstruction task.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method and system for dynamic optimization of parameters in magnetic resonance image reconstruction tasks that combines metadata and reinforcement learning. Background Technology

[0002] In magnetic resonance imaging (MRI), fat suppression techniques are crucial for improving lesion detection rates and eliminating chemical shift artifacts. Current mainstream techniques include frequency-selective saturation (FatSat / CHESS / SPIR), short-time inversion recovery (STIR), frequency-selective inversion pulse (SPAIR), and Dixon fat-water separation. Its inherent defects have been confirmed by multiple patent background technologies: CHESS / FatSat fat suppression uniformity is constrained by both B0 and B1 uniformity, and fat suppression is prone to failure in off-center high B0 offset areas such as the shoulder joint, neck, and distal limbs (see CN107369153B, US20070247153A1); STIR is not sensitive to magnetic field inhomogeneity but is not frequency selective, has poor specificity, and cannot be used for contrast-enhanced scanning (CN103202693A, US20070247153A1); SPAIR improves B1 robustness by adiabatic inversion but at the cost of extended scan time (CN120122042A, CN117872245A); Dixon requires multi-echo acquisition, and the scan time increases linearly with the number of echoes, and there are problems with water-fat exchange and motion sensitivity (US20070247153A1, Philips mDIXON XD Technology White Paper).

[0003] The aforementioned physical trade-offs result in the clinical fat-suppression sequence selection being highly dependent on anatomical location, field strength, and whether enhancement is present, leading to strong heterogeneity in reconstruction tasks within a shared computing cluster: the algorithmic complexity varies greatly among different fat-suppression sequences (T1WI-FS, T2WI-FS, STIR, Dixon, etc.); explicit references and temporal co-occurrence dependencies exist between multiple sequences and fat-suppression techniques for the same patient and the same location; and task arrival fluctuates non-linearly with the batch processing patterns of emergency / physical examinations. Existing patents have attempted to use reinforcement learning for cluster scheduling (such as CN119311407B based on deep reinforcement learning scheduling and CN202411347525.2 hierarchical RL allocation), or to use RL alone to optimize the distributed cache hit rate (see related applications for distributed cache optimization), but none of them have used the explicit + implicit associations of magnetic resonance metadata (sequence type / anatomical location / magnetic field homogeneity / fat-suppression label) as scheduling priors, nor have they used the change in cache hit rate as a closed-loop reward combined with future access sequence predictions to correct the migration probability.

[0004] Existing distributed task scheduling methods mainly suffer from the following shortcomings:

[0005] First, the task allocation strategy is static and rigid. Traditional methods mostly allocate reconstruction tasks based on polling or simple load balancing rules, which cannot perceive the algorithmic correlation between different fat-suppression sequences, the data dependencies between multiple sequences of the same patient, and the differentiated requirements of reconstruction parameters for different anatomical sites. When the clinical scanning mode changes, the static strategy is difficult to adjust dynamically, which can easily lead to some computing nodes being overloaded while others are idle.

[0006] Secondly, there is a lack of dynamic heat perception mechanisms. Existing caching and task scheduling methods generally use fixed thresholds or simple counting strategies to measure the priority of reconstruction tasks, which cannot capture the non-linear decay of reconstruction requests over time for different fat-suppressing sequences. In scenarios with rapidly changing access patterns, such as emergency rooms or batch physical examinations, this can easily lead to delays in high-priority reconstruction tasks or low-value tasks occupying computing resources for extended periods.

[0007] Third, migration decisions lack closed-loop feedback. Current task migrations largely rely on manually set rules or single performance indicators, without establishing a closed-loop mechanism for scheduling, execution, feedback, and optimization. The actual impact of migration on image reconstruction throughput and the timeliness of clinical reports is difficult to quantify and assess, and parameter adjustments lag behind changes in system state.

[0008] Fourth, the predictive capability is insufficient. Existing methods lack accurate prediction of the arrival trend of future reconstruction tasks, making it difficult to make forward-looking resource allocation decisions before sudden changes in load.

[0009] Therefore, there is an urgent need for a dynamic optimization method for task parameters that can integrate metadata association features of magnetic resonance image reconstruction tasks, dynamically sense the access popularity of different fat suppression sequences, and have closed-loop learning capabilities, in order to break through the performance bottleneck of current medical imaging computing clusters. Summary of the Invention

[0010] This invention addresses the problems in existing magnetic resonance image reconstruction task scheduling methods, such as static and rigid task allocation strategies, lack of dynamic heat perception mechanisms, lack of closed-loop feedback in transfer decisions, and insufficient prediction capabilities. It proposes a dynamic optimization method and system for magnetic resonance image reconstruction task parameters that combines metadata and reinforcement learning.

[0011] In a first aspect, embodiments of this application provide a method for dynamically optimizing parameters of a magnetic resonance image reconstruction task that combines metadata and reinforcement learning, the method comprising:

[0012] The hardware parameters of the magnetic resonance imaging (MRI) scanner and historical image reconstruction task logs are collected to construct a metadata association map. The metadata includes at least the scan sequence type, anatomical site identifier, magnetic field homogeneity index, and fat suppression effect label. Based on this, the metadata task is initially assigned and a population is constructed, and the knowledge transfer probability and skill factor are initialized.

[0013] The initially assigned population is used as the state of the reinforcement learning environment. Population distribution characteristics are periodically collected and combined with the association map to construct a population balance factor.

[0014] The weights of elite selection and roulette wheel selection are dynamically adjusted based on the population balance factor, and the mutation strategy is adaptively switched based on the metadata correlation to generate offspring individuals;

[0015] The cooling law is introduced to calculate the real-time access popularity of metadata through neural networks, and the knowledge transfer probability is dynamically adjusted in combination with the population balance factor to control the migration intensity of offspring individuals.

[0016] The change in cache hit rate of the image reconstruction task after migration is calculated as a reward feedback. When it is lower than the threshold, the knowledge transfer probability is corrected by combining the future access sequence predicted by the Markov chain and the agent parameters are updated.

[0017] The migration assignment is performed based on the corrected knowledge migration probability. The processing latency of each computing node is monitored. The process terminates when the tolerance is met; otherwise, the current distribution is used as the environment state for further iteration.

[0018] In a second aspect, embodiments of this application provide a system for dynamically optimizing parameters of a magnetic resonance image reconstruction task that combines metadata and reinforcement learning, applied to the method described in the first aspect, the system comprising:

[0019] The initial allocation module is used to collect hardware parameters of the magnetic resonance scanning equipment and historical image reconstruction task logs, and construct a metadata association map. The metadata includes at least scan sequence type, anatomical site identifier, magnetic field homogeneity index and fat suppression effect label. Based on this, the metadata task is initially allocated and a population is constructed, and the knowledge transfer probability and skill factor are initialized.

[0020] The environment perception module is used to take the initially assigned population as the state of the reinforcement learning environment, periodically collect population distribution characteristics, and construct a population balance factor in combination with the association map.

[0021] The strategy adjustment and mutation module is used to dynamically adjust the weights of elite selection and roulette wheel selection based on the population balance factor, and adaptively switch mutation strategies to generate offspring individuals based on metadata correlation.

[0022] The heat assessment and migration control module is used to introduce the cooling law to calculate the heat in real time through the neural network, and dynamically adjust the knowledge migration probability in combination with the population balance factor to control the migration intensity of the offspring individuals.

[0023] The reward feedback and parameter correction module is used to calculate the change in cache hit rate after migration as reward feedback. When it is lower than the threshold, it combines the future access sequence predicted by the Markov chain to correct the knowledge transfer probability and update the agent parameters.

[0024] The equilibrium determination and iteration control module is used to perform migration allocation based on the corrected knowledge migration probability, monitor the latency of each server node, terminate when the tolerance is met, otherwise use the current distribution as the environmental state to trigger the environment perception module to continue iterating.

[0025] Thirdly, embodiments of this application provide an electronic device, including:

[0026] processor;

[0027] Memory used to store processor-executable instructions;

[0028] The processor is configured to implement, when executing the instructions, the method for dynamic optimization of magnetic resonance image reconstruction task parameters combining metadata and reinforcement learning as described in the first aspect.

[0029] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the method for dynamically optimizing magnetic resonance image reconstruction task parameters by combining metadata and reinforcement learning as described in the first aspect.

[0030] This invention constructs a metadata association map that integrates explicit correlations between magnetic resonance imaging sequences (such as phase resolution dependence between Dixon three-echo sequences and same-plane correspondence between T1WI-FS and T2WI-FS) with implicit patterns of clinical access (such as continuous access to multiple sequences in emergency patients and historical comparison of the same site in follow-up cases). This incorporates the algorithmic coupling between different fat-suppression techniques such as Fat Sat, STIR, SPAIR, and Dixon, as well as the data dependence between multiple sequences of the same patient, into the initial allocation decision, thereby reducing cross-node data exchange overhead from the source and shortening the end-to-end reconstruction link.

[0031] Building upon this foundation, this invention introduces a population balance factor to quantify the system's imbalance risk and adaptively adjusts the weights of elite selection and roulette wheel selection: when the cluster load is skewed or a certain type of fat-suppressing task is piling up, it accelerates convergence to a better allocation, quickly releasing computational bottlenecks; when the system is running smoothly, it maintains population diversity to avoid getting trapped in local optima. This mechanism enables the scheduling strategy to maintain responsiveness and optimization stability in different clinical scenarios.

[0032] Furthermore, based on the population balance factor, an adaptive switching mutation strategy is employed: under imbalance conditions, highly correlated fat-suppressed sequence reconstruction tasks (such as T2WI-STIR and T1WI-Dixon aqueous phase from the same patient) are aggregated and migrated to the same computing node according to the metadata association map, significantly reducing cross-node data dependency overhead and reducing inter-sequence waiting latency; under stable conditions, a heterogeneous node allocation scheme is randomly explored to enhance the system's elastic capacity to handle sudden batch tasks.

[0033] To address the nonlinear decay characteristics of access requests for magnetic resonance reconstruction tasks, this invention combines the cooling law with neural networks to dynamically calculate the access heat of reconstruction tasks. This overcomes the shortcomings of traditional static evaluation methods that cannot capture the fluctuations in access patterns from "emergency peak - off-peak - nighttime trough," enabling migration decisions to focus on critical tasks in time-sensitive scenarios (such as DWI sequences for stroke patients and multi-phase dynamic scanning after tumor enhancement), ensuring that critical cases receive priority access to computing resources.

[0034] Finally, this invention uses the change in cache hit rate as a reward signal for reinforcement learning. When the hit rate is lower than a preset threshold, it combines a high-order weighted Markov chain to predict future access sequences and correct the migration probability, making the migration decision forward-looking. This mechanism significantly improves the overall reconstruction efficiency of magnetic resonance images in typical clinical scenarios involving multiple sequences and multiple fat-suppression techniques (such as complementary diagnosis of FatSat and STIR in combined shoulder joint scanning, and parallel reconstruction of abdominal mDIXON fat separation and conventional T2WI).

[0035] This invention significantly shortens the entire MRI examination process through intelligent scheduling, greatly reducing the waiting time for patients from the end of the scan to when radiologists can see usable images. This directly improves the timeliness of critical diagnostic and treatment scenarios such as emergency trauma, acute stroke, and tumor staging. At the same time, the improved stability of fat-suppressed sequence reconstruction quality reduces the need for repeated scans due to fat-suppressing failure, thus reducing the psychological burden on patients and the waste of medical resources caused by secondary exposure. In addition, the system's efficient support for concurrent fat-suppressing techniques allows clinicians to more freely combine different fat-suppressing strategies to obtain complementary diagnostic information, improving the detection rate and qualitative accuracy of small lesions (such as bone marrow infiltration and small soft tissue metastases), ultimately serving the goal of precision medicine. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a method for dynamically optimizing parameters of a magnetic resonance image reconstruction task by combining metadata and reinforcement learning, provided in an embodiment of this application.

[0037] Figure 2 This is a system architecture diagram for dynamic parameter optimization of magnetic resonance image reconstruction task by combining metadata and reinforcement learning, provided as an embodiment of this application.

[0038] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0039] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0040] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0041] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] Example 1

[0043] Figure 1 This is a schematic flowchart illustrating a method for dynamically optimizing parameters in a magnetic resonance image reconstruction task that combines metadata and reinforcement learning, as provided in an embodiment of this application. Figure 1 As shown, a method for dynamically optimizing parameters in magnetic resonance image reconstruction tasks, combining metadata and reinforcement learning, includes:

[0044] S1. Collect the hardware parameters of the magnetic resonance scanning equipment and the historical image reconstruction task log, and construct a metadata association map. The metadata includes at least the scan sequence type, anatomical site identifier, magnetic field homogeneity index and fat suppression effect label. Based on this, perform initial allocation of metadata tasks and construct a population, and initialize the knowledge transfer probability and skill factor.

[0045] The hardware parameters of an MRI scanner should include at least the number of CPU cores, memory capacity, GPU model and memory size, and network bandwidth of each computing node. Historical image reconstruction task logs should record at least the reception time of each reconstruction request, the type of the corresponding scan sequence (e.g., T1WI-FS, T2WI-FS, STIR, Dixon, etc.), the anatomical location (e.g., skull, shoulder joint, knee joint, abdomen, etc.), the magnetic field homogeneity index (B0 field offset), the fat suppression effect label (homogeneous / non-homogeneous / failed), and the actual computing resources consumed and processing time of the reconstruction task.

[0046] Specifically, in this embodiment, constructing the metadata association graph includes:

[0047] S1.1 Using unique metadata identifiers as nodes, the unique metadata identifiers include at least magnetic resonance scan sequence identifiers, anatomical site codes, and fat suppression technology type codes; constructing an initial topology graph using the reference relationships between metadata as directed edges, the reference relationships including multi-sequence scan associations of the same patient, control associations of different fat suppression technologies in the same site, and temporal associations of enhanced and non-enhanced scans.

[0048] In one specific embodiment, the system generates a globally unique metadata identifier (MetaID) for each reconstruction task generated by a magnetic resonance imaging (MRI) scan. This identifier encodes information in three dimensions: scan sequence type (SEQ_TYPE), anatomical location (ANAT_SITE), and fat suppression technique type (FS_TYPE). For example, MetaID = "T1WI-FS_KNEE_SPAIR" indicates that the task is a T1WI sequence image reconstruction using the SPAIR fat suppression technique for the knee joint; MetaID = "T2WI-FS_KNEE_SPAIR" indicates a T2WI sequence image reconstruction using the SPAIR fat suppression technique for the same patient's knee joint, and the two are associated through a multi-sequence scan because they belong to different fat suppression sequences of the same anatomical location; MetaID = "T1WI-FS_BREAST_Dixon" indicates a T1WI sequence image reconstruction using the Dixon fat suppression technique for the breast. The system analyzes the sequence dependencies recorded in the historical reconstruction task logs. Temporal correlations exist between plain and enhanced scans of the same patient (e.g., T1WI-FS plain and T1WI-FS enhanced scans), and control correlations exist between different fat-suppression techniques for the same anatomical site (e.g., SPAIR and Dixon). For each reference record found pointing from one metadata node to another, a directed edge is established between the two nodes. By traversing all reconstruction tasks and connecting all reference relationships, the system obtains a complete initial topology graph, which visually represents the dependency structure between different fat-suppression sequence reconstruction tasks.

[0049] S1.2. Based on the historical image reconstruction task log, count the co-occurrence frequency of any two metadata nodes in the same reconstruction session, and calculate the implicit association weight between nodes based on the co-occurrence frequency.

[0050] In one specific embodiment, the system reads historical image reconstruction task logs and divides the logs into several independent reconstruction sessions based on patient examinations and clinical request forms. A single session refers to a batch of consecutive image reconstruction requests submitted by a technician for the same patient during a single MRI examination. For each session, the system records all requested reconstruction metadata node combinations within that session; any two metadata nodes appearing in the same session are counted as one co-occurrence. After traversing all sessions, the system counts the total number of co-occurrences for each pair of nodes. A higher co-occurrence frequency indicates a clinical tendency to reconstruct these two sequences sequentially within the same examination procedure; for example, T1WI-FS and T2WI-FS sequences of the knee joint are often requested for reconstruction simultaneously in the same session. The system further normalizes the co-occurrence frequencies to obtain the implicit association weights between each pair of nodes.

[0051] S1.3. Combining the explicit topological structure formed by the reference relationship and the implicit association weight formed by the co-occurrence frequency, a metadata association map containing the hierarchical constraints of the magnetic field homogeneity region is generated; wherein, the initial allocation of the metadata task is specifically as follows: based on the hierarchical constraints of the magnetic field homogeneity region, metadata tasks with reference relationships or high co-occurrence frequencies are preferentially allocated to the same computing node or computing nodes with adjacent network topologies, so that different fat-suppression sequence image reconstruction tasks of the same anatomical site are processed by the same computing node, thereby reducing cross-node data exchange overhead.

[0052] In one specific embodiment, the system fuses the directed edges of the reference relationships obtained in S1.1 with the co-occurrence weights obtained in S1.2. For any two metadata nodes, the system comprehensively considers whether there is a reference edge between them, the frequency of co-occurrence, and whether they belong to the same anatomical location or the same magnetic field homogeneity partition, calculating a comprehensive association strength value for each pair of nodes. During fusion, reference relationships are given higher priority because they represent explicit dependencies between image reconstruction algorithms; co-occurrence relationships serve as a supplement, reflecting the collaborative reconstruction needs in actual clinical work; the magnetic field homogeneity region level is used to correct the association strength, with sequence reconstruction tasks located in the same magnetic field homogeneity partition (such as the central region, off-center region, or distal limb region) having stronger associations due to facing similar fat-suppression challenges. Finally, the system generates a metadata association map containing all nodes and the fused association edges. When performing the initial allocation, the system prioritizes assigning metadata reconstruction tasks with reference relationships or high co-occurrence frequency to the same computing node or computing nodes with adjacent network topologies. This ensures that different fat-suppressed sequence image reconstruction tasks (such as T1WI-FS and T2WI-FS) of the same anatomical site are processed by the same computing node, thereby reducing cross-node data exchange overhead in the subsequent reconstruction process.

[0053] S2. The initially assigned population is used as the state of the reinforcement learning environment. Population distribution characteristics are periodically collected and population balance factors are constructed by combining the association map.

[0054] In one specific embodiment, after completing the initial allocation, the system arranges the computing node numbers currently belonging to all metadata reconstruction tasks into a vector. Each element of this vector represents a computing node where a reconstruction task is currently located. For example, if there are N reconstruction tasks and M computing nodes, the population is a vector of length N, with each element ranging from 1 to M. All possible allocation schemes together constitute the decision space, and the population is a set of candidate allocation schemes generated within this decision space, with each scheme corresponding to a population individual.

[0055] Specifically, in this embodiment, the periodic collection of population distribution characteristics and the construction of a population balance factor in conjunction with the association map include:

[0056] S2.1. Using the metadata reconstruction tasks carried on each computing node as individuals in the population, and the computing node number as the decision variable dimension, construct the decision space.

[0057] S2.2 In the decision space, the mean vector and standard deviation vector of the location of each metadata reconstruction task are periodically collected to represent the central tendency and dispersion of the population, respectively.

[0058] In one specific embodiment, the system samples the current population at preset time intervals (e.g., every 30 seconds or after every 50 reconstruction tasks). For each dimension of the decision space (i.e., each reconstruction task), the system calculates the distribution ratio of the number of individuals assigned to each server across all individuals in the current population for that task, obtaining a position mean vector. This vector reflects the overall centroid distribution trend of the population. Simultaneously, the system calculates the dispersion of each individual relative to this mean within the same dimension, obtaining a position standard deviation vector. A larger standard deviation indicates a higher frequency of the reconstruction task being assigned to different computing nodes among different candidate schemes, suggesting that the population has not yet reached a consensus on the deployment location of the task, and the system is still in a significant exploratory phase.

[0059] S2.3. Traverse the metadata association graph, count the proportion of metadata association edges across computing nodes to the total number of association edges, and calculate the standard deviation entropy value in combination with the position standard deviation vector.

[0060] In one specific embodiment, the system reads the metadata association graph generated in S1 and traverses all association edges in the graph. For each edge, the system checks whether the two metadata reconstruction tasks it connects belong to the same computing node number in the current population's best individual; if they are different, the edge is counted as a cross-computing node association edge. After the statistics are completed, the system calculates the proportion of cross-computing node association edges to the total number of association edges. The higher this proportion, the more association sequence reconstruction tasks need to perform cross-node data transmission due to being assigned to different computing nodes, resulting in greater reconstruction latency. At the same time, the system calculates the entropy value of each dimension of the standard deviation vector obtained in S2.2. The more dispersed the standard deviation distribution, the higher the entropy value, indicating greater population diversity, but also meaning that the system has not yet converged to a stable state.

[0061] S2.4. Normalize and fuse the position mean vector, standard deviation entropy value, and cross-node associated edge ratio to generate the population balance factor. The value of the population balance factor increases with the increase of the cross-node associated edge ratio or the increase of the standard deviation entropy value, and is used to quantify the comprehensive level of system load imbalance risk and cross-node data access overhead.

[0062] In one specific embodiment, the system normalizes the position mean vector in S2.2, the standard deviation entropy value in S2.3, and the cross-node association ratio, eliminating differences in the dimensions and value ranges of each indicator. Subsequently, the system weights and fuses the three indicators according to preset weight ratios (e.g., mean vector 0.3, entropy value 0.3, cross-node ratio 0.4), obtaining a comprehensive value between 0 and 1, namely the population balance factor. A larger factor value indicates more cross-node association accesses and a more dispersed population distribution in the current system, comprehensively reflecting a high level of system load imbalance risk and data access overhead, requiring intervention and adjustment; a smaller value indicates a more balanced and stable current system state.

[0063] To accurately quantify the combined level of system load imbalance risk and data access overhead, this invention introduces the following weighted fusion formula to calculate the population balance factor:

[0064] ,

[0065] in, express The population balance factor at any given time has a range of [0,1]. The larger the value, the more unbalanced the system. It is the normalized norm of the position mean vector, representing the population centripetal tendency; The standard deviation entropy value is defined as:

[0066] , ,

[0067] The proportion of cross-node associated edges:

[0068] ,

[0069] in, For the edge set in the metadata association graph; For nodes and The correlation weight between them; These are preset weighting coefficients.

[0070] S3. Based on the population balance factor, dynamically adjust the weights of elite selection and roulette wheel selection, and generate offspring individuals by adaptively switching mutation strategies based on metadata correlation.

[0071] Specifically, in this embodiment, dynamically adjusting the weights of elite selection and roulette wheel selection based on the population balance factor includes:

[0072] S3.1 Set the population balance factor threshold.

[0073] In one specific embodiment, the system sets a threshold for the population balance factor based on historical operational data of the medical imaging computing cluster and clinical needs. For example, the initial baseline threshold can be set to 0.6. When the population balance factor calculated in real time exceeds this value, it indicates that the system has entered an imbalanced state requiring intervention; if it is below this value, the current allocation scheme is considered acceptable.

[0074] To achieve adaptive adjustment of the weights for elite selection and roulette wheel selection, this invention constructs the following nonlinear sigmoid mapping function:

[0075] ,

[0076] in, Weighting is assigned to the elite. Choose weights for roulette; express Temporal population balance factor; The preset balance factor threshold; , is a hyperparameter for adjusting the steepness. When , Enhance local search; when , To maintain population diversity.

[0077] S3.2 When the population balance factor is greater than the threshold, it is determined that the system load of the magnetic resonance image reconstruction task is in an unbalanced state. The weight of elite selection is increased to enhance the local search capability and accelerate convergence to the current better task allocation solution.

[0078] In one specific embodiment, the system detects a population balance factor of 0.75, exceeding the preset threshold of 0.6, and determines that the distribution of fat-suppressing sequence reconstruction tasks in the current system is unbalanced, with excessive cross-node data access and a high degree of load unevenness. At this point, the system increases the weight of the elite selection strategy in generating the next generation population. Elite selection prioritizes retaining the top few best-performing allocation schemes in each round of evolution to directly enter the next generation. Increasing this weight means that more excellent schemes are retained, and the algorithm will perform a fine-grained search around these known better solutions, accelerating convergence and finding an ideal deployment scheme that balances the load across computing nodes as quickly as possible.

[0079] S3.3 When the population balance factor is less than or equal to the threshold, the system is determined to be in a stable state. The weight of the roulette wheel selection is increased to maintain population diversity and prevent the system from falling into a local optimum task allocation scheme.

[0080] In one specific embodiment, the system detects a population balance factor of 0.42, which does not exceed the threshold of 0.6, indicating that the current system state is good and the distribution of each fat-suppressing sequence reconstruction task is relatively balanced. At this point, the system increases the weight of the roulette wheel selection strategy in individual selection. Roulette wheel selection allocates the probability of selection according to the fitness ratio of each candidate scheme, meaning that individuals with low fitness still have a certain probability of being selected for the next generation. Increasing this weight helps retain a certain proportion of exploratory individuals, maintains population diversity, and prevents the evolutionary process from prematurely converging to a locally optimal task allocation scheme.

[0081] Specifically, in this embodiment, the step of generating offspring individuals based on the adaptive switching mutation strategy according to metadata correlation includes:

[0082] S3.4. In the imbalanced state where the population balance factor is greater than the threshold, based on the metadata association map, retrieve the associated metadata nodes that have explicit references or high co-occurrence frequency relationships with the target metadata node, and execute the targeted prefetch mutation strategy: copy or migrate the reconstruction task individuals corresponding to the associated metadata nodes to the neighborhood of the computing node where the target metadata node is located, so that the lipid-suppressed sequence reconstruction tasks with high correlation are aggregated in the same computing node.

[0083] In one specific embodiment, the system initiates targeted prefetch mutation under imbalance conditions. For each target reconstruction task on a computationally overloaded node in the current allocation scheme, the system queries the metadata association map constructed by S1, retrieving highly associated metadata nodes that have a direct reference relationship with the task (such as the temporal association between plain and enhanced scans of the same patient) or rank in the top 10% of co-occurrence frequency. Subsequently, the system performs targeted prefetch mutation: the individual reconstruction tasks corresponding to these associated metadata nodes are copied or migrated to the neighborhood of the computational node where the target task is located, so that highly associated fat-suppressed sequence reconstruction tasks (such as T1WI-FS and T2WI-FS) are aggregated on the same computational node. This strategy can effectively reduce cross-node data transmission during subsequent reconstruction processes.

[0084] S3.5. Under the stable state where the population balance factor is less than or equal to the threshold, a random exploration mutation strategy is executed: within the decision space, metadata is randomly selected to reconstruct task individuals and redistributed to heterogeneous computing nodes in order to explore the potential optimal task allocation solution space.

[0085] In one specific embodiment, the system switches to random exploratory mutation in a stable state. At this point, the system no longer performs targeted migration based on association maps, but instead randomly selects several reconstruction task individuals from the decision space and reassigns them to heterogeneous computing nodes (such as computing nodes with different GPU models or memory configurations) that are not currently handling that task. Random exploratory mutation does not rely on any association information; its purpose is to proactively try new allocation combinations when the system is in good condition, discover potentially better deployment structures, and prepare candidate solutions in advance for changes in clinical scanning patterns.

[0086] S4. Introduce the cooling law to calculate the real-time access popularity of metadata through a neural network, and dynamically adjust the knowledge transfer probability in combination with the population balance factor to control the migration intensity of offspring individuals.

[0087] Specifically, in this embodiment, the introduction of the cooling law to calculate the real-time access heat of metadata through a neural network includes:

[0088] S4.1 Construct a dynamic heat assessment model for metadata and set the initial cooling coefficient and ambient temperature.

[0089] In one specific embodiment, the system constructs an independent heat assessment model for each metadata reconstruction task, based on the principle of Newton's law of cooling. The system first sets an initial cooling coefficient, which controls the rate of heat decay as inaccessibility increases. The initial value can be set based on historical reconstruction request frequency statistics of the medical image computing cluster; for example, a smaller value is set during high-frequency scanning periods to slow down heat decay, and a larger value is set during low-frequency periods to accelerate the elimination of cold data. Simultaneously, the system sets an ambient temperature parameter, representing the minimum baseline value for metadata heat decay within the system.

[0090] S4.2. Using the historical image reconstruction request timestamp sequence of metadata as input, the nonlinear decay characteristics of the access interval are fitted by a neural network, and the dynamically adjusted cooling coefficient is output.

[0091] In one specific embodiment, the system extracts the timestamp sequence of historical image reconstruction requests for each metadata reconstruction task over a past period and inputs it into a pre-trained lightweight neural network. The time window is set with an upper limit of 7 days and a lower limit of 24 hours. Specific rules are as follows: a storage limit is set, with each task retaining a maximum of 1000 timestamp records; records exceeding this limit are discarded in chronological order. During the initial acquisition, high-frequency tasks (daily reconstruction requests exceeding 100) prioritize records from the most recent 24 hours (if exceeding 1000, only the latest 1000 are retained), while low-frequency tasks prioritize records from the most recent 7 days (if exceeding 1000, only the latest 1000 are retained). All subsequent record maintenance uses 1000 records as the absolute upper limit. The network learns the personalized access rhythm of the metadata by analyzing the time interval distribution between adjacent reconstruction requests. For example, some compression sequences are frequently requested for reconstruction on weekday mornings but are almost never requested at night; the network will identify this periodic characteristic and adjust the cooling coefficient accordingly. The neural network is trained offline and inferred online, and outputs a dynamic cooling coefficient value for the current state of the metadata.

[0092] To overcome the shortcomings of traditional heat assessment methods in failing to capture nonlinear decay patterns, this invention proposes the following metadata heat assessment model that integrates a neural network dynamic cooling coefficient:

[0093] ,

[0094] Among them, dynamic cooling coefficient Real-time prediction by lightweight neural networks:

[0095] ,

[0096] in, Real-time access popularity for metadata task i; This is the time difference between the current moment and the most recent access. The dynamic cooling coefficient is the output of the neural network. To maintain ambient temperature and prevent the temperature from dropping to zero; For lightweight neural networks; This is a sequence of historical access time intervals; These are the parameters of the neural network.

[0097] S4.3. The time difference between the current time and the most recent reconstruction request time of the metadata is used as a cooling variable. Substitute it into the cooling law formula, and combine it with the dynamically adjusted cooling coefficient and the ambient temperature to calculate the real-time access heat value of each metadata reconstruction task. The real-time access heat value decreases exponentially with the increase of the time difference and is used to characterize the time priority of the corresponding fat-suppressed sequence image reconstruction task.

[0098] In one specific embodiment, the system obtains the current system time and reads the most recent reconstruction request record time of the metadata reconstruction task, calculating the time difference between the two. This time difference is used as a cooling variable, and together with the dynamic cooling coefficient output in S4.2 and the ambient temperature set in S4.1, it is substituted into the analytical expression of the cooling law for calculation. When the time difference is small, the heat value is close to the maximum value; as the time difference increases, the heat value gradually decreases exponentially and approaches the ambient temperature. In this way, the system generates a real-time access heat value between 0 and 1 for each metadata reconstruction task. This value is used to characterize the time priority of the corresponding fat-suppressed sequence image reconstruction task. The higher the heat value, the more urgent the current clinical need for this sequence type (such as the T1WI-FS sequence after emergency contrast-enhanced scanning), and the more priority it should receive computing resources.

[0099] Specifically, in this embodiment, the step of dynamically adjusting the migration probability in conjunction with the population balance factor to control the migration intensity of offspring individuals includes:

[0100] S4.4 Construct a nonlinear mapping function with the population balance factor as the independent variable and the knowledge transfer probability adjustment magnitude as the dependent variable. The mapping function is configured such that when the population balance factor increases, the absolute value of the knowledge transfer probability adjustment magnitude increases accordingly.

[0101] In one specific embodiment, the system constructs a nonlinear mapping function that takes the population balance factor generated by S2 as input and outputs a migration probability adjustment value. The function is designed such that when the population balance factor is small, the adjustment magnitude is very small or even close to zero; when the balance factor increases and exceeds a certain inflection point, the adjustment magnitude increases rapidly until it reaches its upper limit. This nonlinear design ensures that the system does not easily change its migration strategy when in a good state, but can respond quickly when the state imbalance worsens, significantly improving the initiative and flexibility of migration decisions.

[0102] S4.5. The real-time access popularity value and the knowledge transfer probability adjustment range are weighted and fused to obtain the target knowledge transfer probability.

[0103] In one specific embodiment, the system reads the heat value calculated in S4.3 and the adjustment range output in S4.4, and merges them according to a preset weight ratio. The heat value determines the urgency of the metadata reconstruction task at the current moment; the higher the heat value, the more urgently the reconstruction task of the compressed sequence needs to be deployed to a suitable location. The adjustment range reflects the intensity of the migration demand at the system-wide level. After weighted fusion, a target knowledge migration probability between 0 and 1 is obtained. This probability combines the individual task heat value and the overall system state, serving as the specific execution probability of this migration operation.

[0104] To integrate system state and data value, this invention constructs the following migration probability fusion function:

[0105] ,

[0106] The system state regulation function is as follows:

[0107] ,

[0108] in, For metadata tasks The migration probability; The Sigmoid function compresses the output to [0,1]. The weighted average of heat and system state; It is a nonlinear migration intensity modulator.

[0109] S4.6. Execute the migration of offspring individuals based on the target knowledge migration probability: When the real-time access popularity value of the metadata reconstruction task is higher than the preset popularity threshold, increase the migration intensity of the task to low-load computing nodes according to the target knowledge migration probability; when the real-time access popularity value of the metadata reconstruction task is lower than the preset popularity threshold, decrease the migration intensity of the task according to the target knowledge migration probability.

[0110] In one specific embodiment, the system performs migration decisions based on the target knowledge migration probability obtained in S4.5. For a certain metadata reconstruction task, if its real-time access popularity value is higher than a preset popularity threshold (e.g., 0.8) and the target knowledge migration probability reaches a specified level, the system migrates the task from its current computing node to a target computing node with a lower load, thereby shortening the queuing time for high-popularity fat-suppressed sequence reconstruction tasks. If the real-time access popularity value of the task is lower than another preset threshold (e.g., 0.3), it indicates that the task is unlikely to be requested for reconstruction in the near future. Even in the face of a high migration probability, the system does not actively initiate a migration operation, avoiding unnecessary cross-node data transmission and saving network bandwidth and system resources. The migration execution is further filtered by the popularity value to ensure that migration operations are concentrated on the most valuable hotspot reconstruction tasks (such as emergency contrast-enhanced scan sequences).

[0111] The system defines the overall load index as a weighted sum of CPU utilization, memory utilization, GPU utilization, and the current reconstruction task queue length, with weighting coefficients set to 0.25, 0.2, 0.3, and 0.25, respectively. Computing nodes with an overall load index more than 20% lower than the system average load index are considered low-load computing nodes. During migration, computing nodes with the lowest overall load index and no more than 3 network hops are prioritized.

[0112] S5. Calculate the change in cache hit rate of the image reconstruction task after migration as a reward feedback. When it is lower than the threshold, combine the future access sequence predicted by the Markov chain to correct the knowledge transfer probability and update the agent parameters.

[0113] Specifically, in this embodiment, the calculation of the cache hit rate change value of the image reconstruction task after migration as reward feedback includes:

[0114] S5.1. In the preset statistics window after the metadata reconstruction task migration and allocation is executed, the actual cache hit count and the total number of image reconstruction requests for each computing node are counted respectively, and the current cache hit rate is calculated.

[0115] In one specific embodiment, after completing a round of metadata reconstruction task migration and allocation, the system initiates a statistical window of a preset duration (e.g., running continuously for 5 minutes). During this window, the system records two types of data for each computing node in the cluster: first, the number of times the local cache directly hits when a client requests image reconstruction, i.e., the number of times the complete fat-suppressed sequence image reconstruction algorithm does not need to be re-executed; second, the total number of all image reconstruction requests received by that node. After the statistical window ends, the system divides the number of hits for each node by the corresponding total number of requests to obtain the current cache hit rate for each node. The system then performs a weighted average of the hit rates of each node according to the proportion of the total number of requests to the total number of requests, i.e., the overall cache hit rate of the system = Σ(hit rate of each node × proportion of the number of requests of each node).

[0116] S5.2. Subtract the current cache hit rate from the historical cache hit rate before migration to obtain the cache hit rate change value.

[0117] In one specific embodiment, the system reads the historical cache hit rate value from the storage module, based on the last statistical analysis performed before the current migration execution. The current hit rate calculated in step S5.1 is subtracted from this historical value, resulting in a difference that can be positive or negative—the change in cache hit rate. For example, if the system hit rate was 65% before migration and 72% after migration, the change is +7 percentage points; if it drops to 61% after migration, the change is -4 percentage points. This difference directly reflects the actual impact and extent of the migration operation on the system's cache utilization efficiency. The hit rate before migration is collected at the end of the most recent complete statistical window before the migration begins. The system automatically saves a snapshot of the cache hit rate at the end of each statistical window, and this snapshot is used as a baseline during migration execution.

[0118] S5.3 Map the cache hit rate change value to a preset reward range to generate an instant reward value for the reinforcement learning agent; wherein, when the cache hit rate change value is positive, it corresponds to a positive reward; when the cache hit rate change value is negative, it corresponds to a negative reward.

[0119] In one specific embodiment, the system pre-defines a reward mapping range, for example, limiting the reward value range to between -1.0 and +1.0. When the change in cache hit rate is positive, it is converted into a positive reward value according to a preset mapping coefficient; the larger the change, the higher the reward. When the change is negative, it is converted into a negative reward value; the greater the decrease, the heavier the penalty. Through this mapping method, the system quantifies the degree of improvement in cache performance into an immediate reward signal that the reinforcement learning agent can directly use.

[0120] Specifically, in this embodiment, the step of correcting the migration probability and updating the agent parameters based on the future visit sequence predicted by the Markov chain when the threshold is reached includes:

[0121] S5.4. A preset reward threshold is set. When the change in cache hit rate is lower than the reward threshold, it is determined that the current migration strategy does not provide sufficient acceleration benefits for the magnetic resonance image reconstruction task.

[0122] In one specific embodiment, the system pre-sets a reward threshold (e.g., -0.1 or 0) to determine whether the current migration strategy has brought about an actual performance improvement. When the instant reward value generated by S5.3 is lower than this threshold, the system determines that the current migration operation has failed to achieve the expected positive benefit, that is, the cache hit rate decreases or does not improve significantly after migration. This indicates that the current knowledge migration probability setting is unreasonable or the migration direction is incorrect, and a compensation mechanism needs to be triggered to correct the migration strategy.

[0123] S5.5 Extract the metadata access sequence from the historical reconstruction task log, construct a high-order weighted Markov chain model, and calculate the state transition probability matrix of each metadata reconstruction task at several future moments.

[0124] In one specific embodiment, after determining that the migration strategy is insufficiently beneficial, the system extracts metadata access sequences from the recent historical reconstruction task logs from storage, and organizes the reconstruction request records into a continuous access sequence according to time order. Based on this sequence, the system constructs a high-order weighted Markov chain model. "High-order" means that predicting the next state depends not only on the current state but also on combinations of multiple previous historical states, enabling the capture of longer access dependencies; "weighted" means that when calculating state transition probabilities, reconstruction request records more recent than the current time are given higher weights, making the model more focused on recent access trends. After the model is constructed, the system calculates the state transition probability matrix between each metadata reconstruction task at several future time points from the current time.

[0125] The system first extracts the k most recent historical reconstruction request records (k=3) from the access sequence and combines these k records into a higher-order state vector as a condition. Then, it statistically analyzes the subsequent transitions of all identical higher-order state combinations in the historical data and assigns exponentially increasing weight coefficients to samples that are closer to the current time.

[0126] S5.6 Based on the state transition probability matrix, predict the target metadata set that is highly likely to be accessed in the future access sequence, and the target metadata set corresponds to the expected high-priority fat-suppressing sequence reconstruction task.

[0127] In one specific embodiment, based on the state transition probability matrix calculated in S5.5, the system extrapolates multiple time steps backward, starting from the metadata tasks currently being requested for reconstruction on each computing node, to predict the sequence of metadata tasks most likely to be requested for reconstruction in the future. The system sets a probability threshold (e.g., 0.6) and filters out all metadata reconstruction tasks with predicted probabilities higher than this threshold, forming a high-probability target metadata set. This set corresponds to the expected high-priority fat-suppressed sequence reconstruction tasks (such as T1WI-FS sequences after emergency contrast-enhanced scans).

[0128] S5.7 Calculate the predicted access popularity weight of each metadata reconstruction task in the target metadata set, and add the weight to the real-time access popularity value to make a secondary correction to the knowledge transfer probability, so that the metadata reconstruction task with a high predicted access probability can obtain transfer resources first.

[0129] In one specific embodiment, the system calculates a predicted access popularity weight for each metadata reconstruction task in the target metadata set predicted in S5.6 and its corresponding predicted probability value; the higher the predicted probability, the greater the weight. Subsequently, the system adds this weight to the real-time access popularity value calculated in S4.3 to form a corrected comprehensive popularity value. Based on the corrected popularity value, the system performs a second correction on the knowledge transfer probability: increasing the transfer probability of tasks with high predicted probability to prioritize them for access to transfer resources; and decreasing the transfer probability of tasks with low predicted probability to avoid wasting system resources due to ineffective transfers.

[0130] To overcome the lag in short-term reward feedback, this invention introduces a high-order weighted Markov chain to predict future visit sequences and performs forward-looking corrections to the popularity value:

[0131] ,

[0132] Wherein, the nth-order state transition probability is:

[0133] ,

[0134] This is the revised predicted popularity. For metadata tasks In the future Step access probability; Weighted by historical step size, recent visits have a higher weight; To predict the weighting coefficients. For metadata tasks At any moment Whether it has been accessed.

[0135] S5.8. Using the instant reward value and the current state transition data, update the network parameters of the reinforcement learning agent using the temporal difference error algorithm.

[0136] In one specific embodiment, the system combines the state sequence, the sequence of actions performed, the immediate reward value generated in S5.3, and the state information of the next time step into a complete experience sample during the current transfer process. The system employs a temporal difference error algorithm, adding the actual immediate reward to the value estimate of the next state as the target value, and subtracting the value estimate of the current state to obtain an error signal. Based on this error signal, the system adjusts the network parameters of the reinforcement learning agent using gradient descent, gradually enhancing the agent's ability to distinguish between superior and inferior transfer strategies. After the parameter update, the agent will make more accurate transfer decisions in subsequent iterations.

[0137] The experience pool management mechanism is as follows: The system sets the upper limit of the experience pool capacity to 10,000 experience samples. When the experience pool reaches the capacity limit, a strategy combining priority sampling and random elimination is adopted to eliminate samples. The 20% of samples with the lowest reward value are eliminated first, and the remaining 80% are eliminated in first-in-first-out order according to time.

[0138] To continuously optimize the migration strategy, this invention employs a temporal difference error algorithm to update the agent parameters:

[0139] ,

[0140] The timing difference error is:

[0141] ,

[0142] These are the parameters of the agent network; For target network parameters; For TD error; Rewards mapped to changes in cache hit rate; This is the discount factor. for The environmental state at any given time; , Real-time access frequency of metadata The proportion of cross-node associated edges; This represents the average response latency across all servers. for Actions performed by the intelligent agent at any given moment. This represents the gradient of the Q-value with respect to the network parameters.

[0143] S6. Perform migration allocation based on the corrected knowledge migration probability, monitor the processing delay of each computing node, terminate when the tolerance is met, otherwise continue iterating with the current distribution as the environment state.

[0144] In one specific embodiment, the system uses the final knowledge transfer probability, corrected by S5, as the execution basis to perform cross-compute node migration allocation operations for each metadata reconstruction task. After the migration is completed, the system enters the monitoring phase, continuously collecting the average response latency of each compute node in the cluster for processing image reconstruction requests. The system pre-sets a tolerance range, for example, the deviation between the latency of each node and the average latency of the cluster does not exceed ±10%. The deviation is calculated as a relative deviation, i.e., deviation = |node latency - average cluster latency| / average cluster latency × 100%. When the relative deviation of all nodes does not exceed 10%, it is considered balanced. Simultaneously, a judgment frequency is added: a judgment is performed at the end of each iteration cycle. When two consecutive judgments meet the balance condition, the optimization is officially terminated to avoid premature termination due to short-term fluctuations. The system performs a judgment on the latency data at a preset period (e.g., every 30 seconds): if the latency deviation of all nodes is within the tolerance range, the system is considered to have reached a load-balanced state, the optimization process terminates, and the current allocation scheme is output as the final deployment result. If the latency deviation of any node exceeds the tolerance range, it indicates that the system has not yet converged to an equilibrium state. At this point, the system reorganizes the current metadata reconstruction task distribution state of each node into a new environmental state and returns to S2 to start the next iteration. Through this cyclical judgment mechanism, the system continuously adjusts and optimizes until the reconstruction task load of each computing node in the cluster tends to be balanced and the response latency stabilizes within an acceptable range.

[0145] The maximum number of iterations is limited to 100 rounds, and the maximum runtime is limited to 30 minutes. If the tolerance condition is not met after reaching any of the upper limits, the system will forcibly terminate the iteration and output the current optimal allocation scheme (i.e., the one with the best historical evaluation indicators) as the final deployment result. At the same time, the reasons for non-convergence will be recorded for analysis by the operation and maintenance personnel.

[0146] Example 2

[0147] like Figure 2 As shown, this application provides a system architecture diagram for dynamic parameter optimization of magnetic resonance image reconstruction tasks combining metadata and reinforcement learning, applied to the dynamic parameter optimization method for magnetic resonance image reconstruction tasks combining metadata and reinforcement learning as described in Embodiment 1, including:

[0148] The initial allocation module 210 is used to collect hardware parameters of the magnetic resonance scanning equipment and historical image reconstruction task logs, and construct a metadata association map. The metadata includes at least the scan sequence type, anatomical site identifier, magnetic field homogeneity index and fat suppression effect label. Based on this, the metadata task is initially allocated and a population is constructed, and the knowledge transfer probability and skill factor are initialized.

[0149] The environment perception module 220 is used to take the initially assigned population as the state of the reinforcement learning environment, periodically collect population distribution characteristics, and construct a population balance factor in combination with the association map.

[0150] The strategy adjustment and mutation module 230 is used to dynamically adjust the weights of elite selection and roulette wheel selection based on the population balance factor, and adaptively switch mutation strategies to generate offspring individuals based on metadata correlation.

[0151] The heat assessment and migration control module 240 is used to introduce the cooling law to calculate the real-time access heat through the neural network, and dynamically adjust the knowledge migration probability in combination with the population balance factor to control the migration intensity of the offspring individuals.

[0152] The reward feedback and parameter correction module 250 is used to calculate the change in cache hit rate after migration as reward feedback. When it is lower than the threshold, it combines the future access sequence predicted by the Markov chain to correct the knowledge transfer probability and update the agent parameters.

[0153] The equilibrium determination and iteration control module 260 is used to perform migration allocation based on the corrected knowledge migration probability, monitor the latency of each server node, terminate when the tolerance is met, otherwise use the current distribution as the environmental state to trigger the environment perception module to continue iterating.

[0154] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0155] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0156] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0157] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0158] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0159] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0160] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0161] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0162] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for dynamically optimizing parameters in magnetic resonance image reconstruction tasks by combining metadata and reinforcement learning, characterized in that, Includes the following steps: The hardware parameters of the magnetic resonance imaging (MRI) scanner and historical image reconstruction task logs are collected to construct a metadata association map. The metadata includes at least the scan sequence type, anatomical site identifier, magnetic field homogeneity index, and fat suppression effect label. Based on this, the initial allocation of metadata tasks is carried out and a population is constructed, and the knowledge transfer probability and skill factor are initialized. The initially assigned population is used as the state of the reinforcement learning environment. Population distribution characteristics are periodically collected and combined with the association map to construct a population balance factor. The weights of elite selection and roulette wheel selection are dynamically adjusted based on the population balance factor, and the mutation strategy is adaptively switched based on the metadata correlation to generate offspring individuals; The cooling law is introduced to calculate the real-time access popularity of metadata through neural networks, and the knowledge transfer probability is dynamically adjusted in combination with the population balance factor to control the migration intensity of offspring individuals. The change in cache hit rate after migration is calculated as a reward feedback. When it is lower than the threshold, the knowledge transfer probability is corrected by combining the future access sequence predicted by the Markov chain and the agent parameters are updated. The migration assignment is performed based on the corrected knowledge migration probability. The latency of each server node is monitored. The process terminates when the tolerance is met; otherwise, the current distribution is used as the environment state for further iteration.

2. The method according to claim 1, characterized in that, The construction of the metadata association graph includes: Using unique metadata identifiers as nodes, the unique metadata identifiers include at least magnetic resonance scan sequence identifiers, anatomical site codes, and fat suppression technology type codes; an initial topology graph is constructed using the reference relationships between metadata as directed edges, the reference relationships including multi-sequence scan associations of the same patient, control associations of different fat suppression technologies in the same site, and temporal associations of enhanced and non-enhanced scans; Based on the historical image reconstruction task log, the co-occurrence frequency of any two metadata nodes in the same reconstruction session is counted, and the implicit association weight between nodes is calculated based on the co-occurrence frequency. By combining the explicit topology formed by the reference relationships with the implicit association weights formed by the co-occurrence frequencies, a metadata association map containing the hierarchical constraints of the magnetic field homogeneity region is generated. Specifically, the initial allocation of metadata tasks involves prioritizing the allocation of metadata tasks with reference relationships or high co-occurrence frequencies to the same computing node or computing nodes with adjacent network topologies, based on the hierarchical constraints of the magnetic field homogeneity region. This ensures that different fat-suppression sequence image reconstruction tasks for the same anatomical site are processed by the same computing node, thereby reducing cross-node data exchange overhead.

3. The method according to claim 1, characterized in that, The periodic collection of population distribution characteristics and the construction of population balance factors in conjunction with the correlation map include: The decision space is constructed by taking the metadata reconstruction tasks carried on each computing node as individuals in the population and the computing node number as the decision variable dimension. In the decision space, the mean vector and standard deviation vector of the location of each metadata reconstruction task are periodically collected to represent the central tendency and dispersion of the population, respectively. Traverse the metadata association graph, count the proportion of metadata association edges across computing nodes to the total number of association edges, and calculate the standard deviation entropy value in combination with the position standard deviation vector; The mean vector of positions, the standard deviation entropy value, and the proportion of cross-node associated edges are normalized and fused to generate the population balance factor. The value of the population balance factor increases with the increase of the proportion of cross-node associated edges or the increase of the standard deviation entropy value, and is used to quantify the comprehensive level of system load imbalance risk and cross-node data access overhead.

4. The method according to claim 1, characterized in that, The dynamic adjustment of the weights of elite selection and roulette wheel selection based on the population balance factor includes: Set the population balance factor threshold; When the population balance factor is greater than the threshold, it is determined that the system load of the magnetic resonance image reconstruction task is in an unbalanced state. The weight of elite selection is increased to enhance the local search capability and accelerate convergence to the current better task allocation solution. When the population balance factor is less than or equal to the threshold, the system is determined to be in a stable state. The weight of the roulette wheel selection is increased to maintain population diversity and prevent the system from falling into a local optimum task allocation scheme.

5. The method according to claim 1, characterized in that, The adaptive switching mutation strategy based on metadata correlation to generate offspring individuals includes: In the imbalanced state where the population balance factor is greater than the threshold, the associated metadata nodes that have explicit references or high co-occurrence frequency with the target metadata node are retrieved according to the metadata association graph. The targeted prefetch mutation strategy is executed: the reconstruction task individuals corresponding to the associated metadata nodes are copied or migrated to the neighborhood of the computing node where the target metadata node is located, so that the lipid-suppressed sequence reconstruction tasks with high correlation are aggregated in the same computing node. In a stable state where the population balance factor is less than or equal to the threshold, a random exploration mutation strategy is executed: within the decision space, metadata is randomly selected to reconstruct task individuals and redistributed to heterogeneous computing nodes in order to explore the potential optimal task allocation solution space.

6. The method according to claim 1, characterized in that, The introduced cooling law calculates the real-time access heat of metadata through a neural network, including: Construct a dynamic heat assessment model for metadata and set the initial cooling coefficient and ambient temperature; Using the historical image reconstruction request timestamp sequence of metadata as input, the nonlinear decay characteristics of the access interval are fitted by a neural network, and the dynamically adjusted cooling coefficient is output. The time difference between the current time and the time of the most recent reconstruction request of the metadata is used as a cooling variable. Substituted into the cooling law formula, and combined with the dynamically adjusted cooling coefficient and the ambient temperature, the real-time access heat value of each metadata reconstruction task is calculated. The real-time access popularity value decreases exponentially with the increase of time difference, and is used to characterize the time priority of the corresponding fat-suppressed sequence image reconstruction task.

7. The method according to claim 6, characterized in that, The method of dynamically adjusting the migration probability in conjunction with the population balance factor to control the migration intensity of offspring individuals includes: A nonlinear mapping function is constructed with the population balance factor as the independent variable and the knowledge transfer probability adjustment magnitude as the dependent variable. The mapping function is configured such that when the population balance factor increases, the absolute value of the knowledge transfer probability adjustment magnitude increases accordingly. The target knowledge transfer probability is obtained by weighting and fusing the real-time access popularity value with the knowledge transfer probability adjustment range. Based on the target knowledge migration probability, the migration of offspring individuals is performed: when the real-time access popularity value of the metadata reconstruction task is higher than the preset popularity threshold, the migration intensity of the task to low-load computing nodes is increased according to the target knowledge migration probability; when the real-time access popularity value of the metadata reconstruction task is lower than the preset popularity threshold, the migration intensity of the task is reduced according to the target knowledge migration probability.

8. The method according to claim 1, characterized in that, The calculation of the change in cache hit rate after migration as a reward feedback includes: Within the preset statistics window after the metadata reconstruction task migration and allocation is performed, the actual cache hit count and the total number of image reconstruction requests for each computing node are counted to calculate the current cache hit rate. The difference between the current cache hit rate and the historical cache hit rate before migration is used to obtain the cache hit rate change value; The cache hit rate change value is mapped to a preset reward range to generate an instant reward value for the reinforcement learning agent; wherein, when the cache hit rate change value is positive, it corresponds to a positive reward; when the cache hit rate change value is negative, it corresponds to a negative reward.

9. The method according to claim 8, characterized in that, The step of adjusting the migration probability and updating the agent parameters based on the future visit sequence predicted by the Markov chain when the threshold is reached includes: A preset reward threshold is set. When the change in cache hit rate is lower than the reward threshold, it is determined that the current migration strategy does not provide sufficient acceleration benefits for the magnetic resonance image reconstruction task. Extract metadata access sequences from historical reconstruction task logs, construct a high-order weighted Markov chain model, and calculate the state transition probability matrix of each metadata reconstruction task at several future time points. Based on the state transition probability matrix, predict the target metadata set that is highly likely to be accessed in the future access sequence, and the target metadata set corresponds to the expected high-priority fat-suppressing sequence reconstruction task; Calculate the predicted access popularity weight of each metadata reconstruction task in the target metadata set, and add the weight to the real-time access popularity value to make a secondary correction to the knowledge transfer probability, so that the metadata reconstruction task with a high predicted access probability can obtain migration resources first. Using the instant reward value and the current state transition data, the network parameters of the reinforcement learning agent are updated using the temporal difference error algorithm.

10. A dynamic parameter optimization system for magnetic resonance image reconstruction tasks combining metadata and reinforcement learning, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The initial allocation module is used to collect hardware parameters of the magnetic resonance scanning equipment and historical image reconstruction task logs, and construct a metadata association map. The metadata includes at least scan sequence type, anatomical site identifier, magnetic field homogeneity index and fat suppression effect label. Based on this, the metadata task is initially allocated and a population is constructed, and the knowledge transfer probability and skill factor are initialized. The environment perception module is used to take the initially assigned population as the state of the reinforcement learning environment, periodically collect population distribution characteristics, and construct a population balance factor in combination with the association map. The strategy adjustment and mutation module is used to dynamically adjust the weights of elite selection and roulette wheel selection based on the population balance factor, and adaptively switch mutation strategies to generate offspring individuals based on metadata correlation. The heat assessment and migration control module is used to introduce the cooling law to calculate the heat in real time through the neural network, and dynamically adjust the knowledge migration probability in combination with the population balance factor to control the migration intensity of the offspring individuals. The reward feedback and parameter correction module is used to calculate the change in cache hit rate after migration as reward feedback. When it is lower than the threshold, it combines the future access sequence predicted by the Markov chain to correct the knowledge transfer probability and update the agent parameters. The equilibrium determination and iteration control module is used to perform migration allocation based on the corrected knowledge migration probability, monitor the latency of each server node, terminate when the tolerance is met, otherwise use the current distribution as the environmental state to trigger the environment perception module to continue iterating.

Images