A distributed medical image data storage method based on a cloud platform

Abstract

The present invention discloses a distributed medical image data storage method based on a cloud platform, which relates to the field of medical imaging technology. It includes that the original medical images of each medical site are encrypted and transmitted, double-verified, data-calibrated, and metadata-extracted to generate a standardized medical image sequence and bind it with structured metadata, and are divided into independent logical shards according to the site to form a standardized sub-site data stream; the compressed shard set is input into the initial segmentation model to calculate the loss gradient, and the directional and arbitrary gradient alignment is carried out in combination with historical replay data; based on the gradient alignment constraint of the joint loss function and the site identifier, synthetic medical images are generated through a diffusion model to form a structured replay data set; the cosine similarity of the gradient direction between the site-specific synthetic data generated by the diffusion model and the real data is improved, and the anatomical structure overlap rate is improved. Combining with the dynamic prompt vector optimization of reinforcement learning, the historical data replay efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular, to a distributed medical imaging data storage method based on a cloud platform. Background Art

[0002] The cross-site medical imaging joint training technology is a key path to break medical data islands and build a general medical AI model. The current mainstream methods are based on the federated learning framework, aggregating the model updates of each site through a parameter server, or synthesizing cross-site data with the help of a generative adversarial network to alleviate the distribution differences. Federated learning based on gradient sharing can achieve high detection accuracy in different site scenarios.

[0003] Existing technologies have tried to improve the model convergence by gradient clipping or dynamic weight adjustment, but have failed to solve the directional conflicts of the gradient vectors generated by the backpropagation of data at different sites in the parameter space. Although the gradient averaging strategy of traditional federated learning can reduce the update variance, it cannot eliminate the inconsistency of gradient directions. Summary of the Invention

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

[0005] Therefore, the present invention provides a distributed medical imaging data storage method based on a cloud platform to solve the problem of gradient direction conflicts caused by data distribution differences in cross-site medical imaging joint training.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a distributed medical imaging data storage method based on a cloud platform, which includes that the original medical images of each medical site are encrypted and transmitted, double-verified, data-calibrated, and metadata-extracted to generate a standardized medical image sequence and bind it with structured metadata, and are divided into independent logical shards by site to form a standardized sub-site data stream;

[0008] Parse the metadata of the sub-site data stream to identify the medical image modality category, set a dedicated compression path according to the medical image modality category and dynamically adjust the resource weights, calculate the activity entropy value, and divide the cold and hot storage layers according to the entropy value to form a compressed shard set;

[0009] Input the compressed shard set into an initial segmentation model to calculate the loss gradient, perform directional and arbitrary gradient alignment in combination with historical replay data, and form a joint loss function by combining the segmentation task loss;

[0010] Based on the gradient alignment constraint of the joint loss function and the site identifier, generate synthetic medical images through a diffusion model, and bind site parameters after screening by reinforcement learning to form a structured replay data set.

[0011] As a preferred solution of the distributed medical image data storage method based on the cloud platform according to the present invention, wherein: the generation of the standardized medical image sequence includes the following steps,

[0012] The original medical images of each medical site are uploaded to the cloud platform through an encrypted transmission channel, and double verification of the hash algorithm is performed;

[0013] Deeply analyze the DICOM file header, verify the transfer syntax identifier, and reject error files;

[0014] For the imaging differences of different devices, locate the key anatomical regions through a pre-trained anatomical structure detection model, dynamically adjust the weights of the bilinear interpolation kernel function, calibrate the medical image size, extract key metadata from the DICOM file header, and generate a standardized medical image sequence.

[0015] As a preferred solution of the distributed medical image data storage method based on the cloud platform according to the present invention, wherein: the formation of the standardized sub-site data stream refers to extracting device information, imaging protocol, and anonymized patient identification from the metadata, encoding the extracted metadata into a JSON-format structured tag, and binding it to the standardized medical image sequence.

[0016] As a preferred solution of the distributed medical image data storage method based on the cloud platform according to the present invention, wherein: the formation of the compressed shard set includes the following steps,

[0017] Parse the metadata of the standardized sub-site data stream and identify the modality categories of the medical images;

[0018] For CT and MRI shards, adopt a compression path combining a three-dimensional sparse convolutional layer and a deformable attention module, generate the spatial weight of the lesion area through the Grad-CAM heat map, and dynamically adjust the convolution kernel deformation offset to retain calcification points and vascular textures;

[0019] Apply discrete wavelet transform to decompose ultrasound shards into low-frequency approximation components and high-frequency detail components, and perform Huffman coding on the high-frequency subbands to compress blood flow noise;

[0020] Fuse the identified CT shards, MRI shards, and ultrasound shards, and output a compressed shard set.

[0021] As a preferred solution of the distributed medical image data storage method based on the cloud platform according to the present invention, wherein: the division of the cold and hot storage layers according to the entropy value includes the following steps,

[0022] Extract the gradient amplitude of the feature map during compression and dynamically allocate the modality weight coefficient;

[0023] Statistically calculate the activity entropy value of the feature map L1 norm from the compressed shard set, and set an entropy threshold to divide the compressed shards into hot and cold layers:

[0024] Compressed shards with entropy values higher than the entropy threshold are assigned to high-bandwidth storage nodes, and compressed shards with entropy values lower than the entropy threshold are secondarily compressed using the Zstandard algorithm and stored in low-performance storage nodes.

[0025] As a preferred solution of the distributed medical image data storage method based on the cloud platform described in the present invention, wherein: the generation of the joint loss function includes the following steps,

[0026] Input the compressed shard set with storage level labels into the initial segmentation model, calculate the loss value between the prediction result and the true label, and generate the original gradient vector;

[0027] When there is historical synthetic replay data, calculate the cosine similarity between the compressed shard set and the gradients of the historical data, and maximize the inner product to constrain the directional alignment;

[0028] Randomly split the current site data into a virtual training set and a test set, and enforce the same gradient direction for the training set and the test set for arbitrary alignment;

[0029] Fuse the directional alignment loss, the arbitrary alignment loss, and the original segmentation loss to generate the joint loss function.

[0030] As a preferred solution of the distributed medical image data storage method based on the cloud platform described in the present invention, wherein: the generation of the synthetic medical image includes the following steps,

[0031] Parse the site identifiers of the compressed shard set, and extract the gradient alignment information from the joint loss function as the generation constraint;

[0032] Load the pre-trained diffusion model, input the compressed shard feature map and the site identifier embedding vector, and control the generated data distribution through the learnable prompt vector;

[0033] In the noise addition stage, gradually superimpose Gaussian noise on the compressed shards to generate an intermediate noise sequence, and in the denoising stage, combine the site prompt vector to restore the original distribution to generate the synthetic medical image.

[0034] As a preferred solution of the distributed medical image data storage method based on the cloud platform described in the present invention, wherein: the formation of the structured replay data set means setting a reinforcement learning reward function, calculating the gradient direction cosine similarity, the feature distribution entropy, and the segmentation overlap rate between the synthetic data and the real data, screening the synthetic data with anatomical structure consistency, and binding the screened synthetic data with the site identifier, the noise step number, and the prompt vector version parameter.

[0035] In a second aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the distributed medical image data storage method based on a cloud platform as described in the first aspect of the present invention is implemented.

[0036] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of the distributed medical image data storage method based on a cloud platform as described in the first aspect of the present invention is implemented.

[0037] The beneficial effects of the present invention are as follows: The dynamic interpolation algorithm based on device metadata improves the sharpness of cross-device anatomical boundaries. DICOM metadata parsing and linear normalization eliminate outliers, providing a high-consistency data basis for subsequent training. The three-dimensional sparse convolution combined with the deformable convolution kernel guided by the Grad-CAM heatmap preserves texture features in the liver tumor region and saves storage space. The hot and cold stratification strategy based on activity entropy reduces the storage cost, and at the same time, the access latency of high-priority shards is shortened. The two-stage gradient alignment mechanism reduces the average gradient angle between heterogeneous sites, improves the Dice coefficient of lung nodule segmentation, and reduces the generalization error of new sites. The cosine similarity of the gradient direction between the site-specific synthetic data generated by the diffusion model and the real data is improved, and the anatomical structure overlap rate is increased. Combined with the dynamic prompt vector optimization of reinforcement learning, the historical data playback efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0039] Figure 1 It is a flowchart of the distributed medical image data storage method based on a cloud platform in Embodiment 1.

[0040] Figure 2 It is a flowchart of data standardization and sharding in Embodiment 1.

[0041] Figure 3 It is a schematic diagram of the data compression and storage stratification mechanism in Embodiment 1.

[0042] Figure 4 It is a schematic diagram of gradient alignment and synthetic data generation in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given in conjunction with the accompanying drawings of the specification.

[0044] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0045] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that can be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other from other embodiments.

[0046] Example 1, referring to Figures 1 to 4 , which is the first embodiment of the present invention. This embodiment provides a distributed medical image data storage method based on a cloud platform, including the following steps:

[0047] S1. The original medical images (such as MRI (Magnetic Resonance Imaging), CT (Computed Tomography)) from each medical site (such as Hospital A, Hospital B) are uploaded to the cloud platform through an encrypted transmission channel. A dual-verification mechanism is executed. Based on the SHA-3 hash algorithm, it detects possible data corruption during transmission and deeply analyzes the file header metadata to verify the compliance of core elements such as transmission syntax identifiers and required tag groups, rejecting non-medical images or files with incorrect formats. For the imaging differences of different devices (such as Siemens MRI, Philips CT), a pre-trained anatomical structure detection model is integrated, deployed after fine-tuning on the datasets of different devices, and key anatomical regions are located in real time (such as the edge of the corpus callosum in brain MRI, the bronchial bifurcation point in CT lung window). A spatial attention mask is generated according to the organ contour, dynamically adjusting the weight distribution of the bilinear interpolation kernel function, using a sharpening-guided weight matrix for the anatomical boundary region to enhance the sharpness of tissue edges; maintaining standard interpolation for homogeneous regions, and the calibrated medical images remove redundant blank areas to ensure that the anatomical structures are centered and displayed, outputting a medical image sequence with standardized dimensions. A double-layer bidirectional LSTM neural network is constructed, inputting device-specific parameters (including features such as the peak kilovoltage of CT devices, repetition time / echo time of MRI) extracted from the DICOM file header, and outputting the optimal truncation threshold range; performing dynamic range truncation according to the optimal truncation threshold, removing outliers outside the human tissue range (such as CT value < -1000 or > 3000); then linearly normalizing the maximum and minimum valid values of the current medical image; finally, applying histogram stretching to enhance the contrast to ensure the visual consistency of medical images from different sites.

[0048] Extract key metadata from the DICOM file header, including device information, imaging protocol, and patient identification.

[0049] Among them, the device information includes the manufacturer name and device model; the imaging protocol refers to the sequence type, slice thickness, and magnetic field strength; the patient identification includes the anonymized ID, removing sensitive information; encoding the extracted key metadata into a JSON-formatted structured tag, corresponding and binding it with the medical image to form an image-metadata paired dataset.

[0050] The paired dataset is divided into independent logical shards according to the site identifier, and each logical shard only contains the images and metadata of the current site; the logical shards enter the training queue in a preset order, and immediately trigger the deletion of the original medical image data and the release of memory after training is completed, forming a standardized site-based data stream.

[0051] S2. Analyze the device type and imaging protocol fields in the metadata of the standardized sub-station data stream, automatically identify the medical imaging modality category, and output a shard set classified by medical imaging modality. For example, if the device model field in the metadata is judged to be a high-resolution modality, and the dynamic blood flow imaging tag in the imaging protocol is recognized as a low-resolution time-series modality.

[0052] Design dedicated compression paths for different modalities as follows:

[0053] The CT / MRI compression path uses three-dimensional sparse convolutional layers and embeds a deformable attention module to dynamically adjust the geometric shape of the convolutional kernel; for lesion regions such as liver tumors, according to the spatial weights generated by the Grad-CAM heat map, the convolutional kernel produces the maximum pixel deformation offset at the tumor boundary to accurately capture heterogeneous texture features; capture multi-scale anatomical structures (such as organ boundaries, lesion regions) through multi-level dilated convolutions, retain high-frequency details (such as calcification points, vascular textures) during compression, and constrain the reconstruction quality through peak signal-to-noise ratio; the ultrasound compression path combines wavelet transform and entropy coding to decompose the image into low-frequency approximation components (retaining tissue contours) and high-frequency detail components (compressing blood flow noise).

[0054] Extract the gradient amplitude from the feature map of the compression model during the compression process, and quantify the contribution of different modalities to the update of the compression model parameters; design a reward function to dynamically adjust the modality weight coefficients according to the compression quality (PSNR / SSIM (structural similarity index)) and computational overhead (such as GPU memory occupancy); for example, if the PSNR (peak signal-to-noise ratio) of the CT shard drops by more than the compression threshold after compression, then reduce the weight of PSNR to reduce resource allocation.

[0055] Perform shard compression according to the dynamically adjusted modality weight coefficients to obtain the compressed shard set as follows:

[0056] The CT / MRI shards load a three-dimensional sparse convolutional model, are compressed in blocks and then spliced into complete ciphertext shards, and verify whether the PSNR meets the standard; the ultrasound shards apply discrete wavelet transform (DWT) to decompose the image, perform Huffman coding on the high-frequency subbands, and verify the SSIM and compression ratio.

[0057] Extract the activation features from the compressed shard set, calculate the L1 norm of the feature maps of each channel, calculate the probability distribution, and calculate the activity entropy value. The higher the entropy value, the stronger the feature diversity (such as the heterogeneous features in the tumor region), and the more frequent access is required.

[0058] Divide the storage strategy into cold layer and hot layer according to the activity entropy value; by analyzing the distribution of the activity entropy values of historical compressed shards, take the top 20% quantile as the activity entropy value threshold (e.g., entropy value ≥ 5.7), and allocate the shards with activity entropy values greater than the activity entropy value threshold (such as CT shards containing lesions) to high-bandwidth storage nodes to support low-latency access. For shards with activity entropy values less than or equal to the activity entropy value threshold (such as ultrasound shards of normal tissues), enable the Zstandard compression algorithm for further compression to save storage space. Integrate the activity entropy value list with the compressed shards to form a shard set with storage level labels.

[0059] S3. Input the shard set with storage level labels into the pre-trained initial segmentation model for forward inference, and calculate the loss value between the prediction result and the true label of each sample; based on the calculated loss value, backpropagate to generate the original gradient vector of the initial segmentation model parameters, and record the gradient distribution characteristics of the current site data shard set. At the same time, if there is historical synthetic replay data, synchronously calculate the gradient vector of the historical synthetic replay data for gradient feature extraction.

[0060] Calculate the gradient direction similarity between the current site data shard set and the historical replay data for directional gradient alignment, including initial stage processing and iterative stage processing.

[0061] Further explanation, the initial stage processing means that if the historical replay data is empty (such as when training site A for the first time), skip the directional alignment and directly enter the arbitrary alignment; the iterative stage processing means that the replay buffer loads synthetic data (such as the generated images of site A), extracts the gradient vector of the synthetic data, calculates the cosine similarity with the gradient of the current site shard set, and maximizes the inner product of the two to constrain the gradient directions to be consistent.

[0062] Randomly divide the current site data (such as the CT images of site B) into a virtual training set and a virtual test set, and simulate the cross-site data distribution difference for arbitrary gradient alignment, as follows:

[0063] First, randomly split the current site data into two subsets according to a safe preset ratio, and input them into the segmentation model to calculate the loss value and generate the corresponding gradient vectors; then force the gradient directions of the two subsets to be consistent, that is, maximize the loss gradient of the segmentation model parameters on the virtual training set and the loss gradient on the virtual test set, so that the segmentation model maintains a stable feature extraction ability on arbitrarily divided data subsets, thereby improving the generalization ability for unknown sites.

[0064] If there is historical replay data, add a directional alignment loss term and an arbitrary alignment loss; combine the original segmentation task loss with the task loss with the added directional alignment loss term and arbitrary alignment loss to form a joint loss function.

[0065] S4. Parse the site identifier for the compressed shard set and metadata label, and at the same time extract gradient alignment information from the joint loss function as a constraint condition for generating contract data; load the pre-trained diffusion model, the input channels include the compressed shard feature map and the embedded vector encoded by the site identifier. At the same time, initialize the learnable prompt vector, and each site corresponds to an independent prompt vector, which is used to control the distribution characteristics of the generated data; gradually add Gaussian noise to the compressed shard feature map to generate an intermediate noise image sequence. For example, for the CT shard of site B, a high-noise feature map is generated after 100 steps of noise accumulation; through the denoising network of the diffusion model, combined with the site prompt vector, gradually restore the original feature distribution. For example, input the noise feature map and the hospital prompt vector to generate a synthetic medical image that conforms to the imaging characteristics of site B.

[0066] Minimize the mean square error between the noise predicted by the denoising network and the real noise to ensure the fidelity of the generated medical image; limit the number of non-zero elements of the prompt vector through L1 regularization to reduce redundant information, and generate optimized diffusion model parameters and sparse prompt vectors; design a reinforcement learning reward function to dynamically adjust the prompt vector to balance the gradient difference and generation diversity between the synthetic data and the real data, specifically as follows:

[0067] Calculate the cosine similarity of the gradient direction between the synthetic data and the real data in the segmentation model. The higher the similarity, the greater the reward value; evaluate the generation diversity through the feature distribution entropy of the synthetic data, and impose a penalty when the entropy value is lower than the entropy threshold; calculate the overlap rate of the segmentation results of the synthetic data and the real data through the pre-trained segmentation model to evaluate the anatomical structure consistency.

[0068] Based on the gradient alignment result, calculate the cosine similarity between the synthetic data gradient and the historical replay data; set a similarity score threshold, and the data with a similarity score lower than the similarity threshold is eliminated, and the high-score data is screened out; bind the screened synthetic data with the corresponding site identifier and generation parameters (such as noise steps, prompt vector version) to form a structured replay data set.

[0069] This embodiment also provides a computer device, which is applicable to the case of a distributed medical image data storage method based on a cloud platform, including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to implement the distributed medical image data storage method based on the cloud platform as proposed in the above embodiment.

[0070] The computer device can be a terminal, which includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be achieved through WIFI, carrier network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0071] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the method for storing distributed medical image data based on a cloud platform as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0072] In summary, the dynamic interpolation algorithm based on device metadata in the present invention improves the sharpness of cross-device anatomical boundaries. The DICOM metadata parsing and linear normalization eliminate outliers, providing a data basis with high consistency for subsequent training. The three-dimensional sparse convolution combined with the deformation convolution kernel guided by the Grad-CAM heat map preserves texture features in the liver tumor area and saves storage space. The hot and cold stratification strategy based on activity entropy reduces storage costs, while shortening the access latency of high-priority shards. The two-stage gradient alignment mechanism reduces the average gradient angle between heterogeneous sites, improves the Dice coefficient of lung nodule segmentation, and reduces the generalization error of new sites. The gradient direction cosine similarity between the site-specific synthetic data generated by the diffusion model and the real data is improved, and the anatomical structure overlap rate is increased. Combined with the dynamic prompt vector optimization of reinforcement learning, the historical data replay efficiency is improved.

[0073] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A distributed medical image data storage method based on a cloud platform, characterized in that: including, The original medical images of each medical site are encrypted and transmitted, double-verified, data-calibrated, and metadata-extracted to generate a standardized medical image sequence that is bound to structured metadata, and is divided into independent logical shards by site to form a standardized sub-site data stream; Parse the metadata of the sub-site data stream to identify the medical image modality category, set a dedicated compression path according to the medical image modality category and dynamically adjust the resource weights, calculate the activity entropy value, and divide the hot and cold storage layers according to the entropy value to form a compressed shard set; Input the compressed shard set with storage layer labels into the initial segmentation model, calculate the loss value between the prediction result and the true label, and generate the original gradient vector; When there is historical synthetic playback data, calculate the cosine similarity between the compressed shard set and the historical data gradient, and maximize the inner product to constrain the directional alignment; Randomly split the current site data into a virtual training set and a test set, force the gradient directions of the training set and the test set to be the same, and perform arbitrary alignment; Fuse the directional alignment loss, the arbitrary alignment loss, and the original segmentation loss to generate a joint loss function; Based on the gradient alignment constraint of the joint loss function and the site identifier, generate synthetic medical images through a diffusion model, and bind site parameters after screening by reinforcement learning to form a structured playback data set.

2. The distributed medical image data storage method based on a cloud platform according to claim 1, wherein: The generation of the standardized medical image sequence includes the following steps, The original medical images of each medical site are uploaded to the cloud platform through an encrypted transmission channel, and double-verified by executing the hash algorithm; Deeply parse the DICOM file header, verify the transfer syntax identifier, and reject incorrect files; For imaging differences of different devices, locate key anatomical regions through a pre-trained anatomical structure detection model, dynamically adjust the weights of the bilinear interpolation kernel function, calibrate the size of the medical image, extract key metadata from the DICOM file header, and generate a standardized medical image sequence.

3. The distributed medical image data storage method based on a cloud platform according to claim 2, characterized in that: The formation of the standardized sub-site data stream means encoding the extracted key metadata into a JSON format structured label and binding it to the standardized medical image sequence.

4. The distributed medical image data storage method based on a cloud platform according to claim 3, characterized in that: The formation of the compressed shard set includes the following steps, Parse the metadata of the standardized sub-site data stream to identify the modality category of the medical image; For CT and MRI shards, adopt a compression path that combines a three-dimensional sparse convolutional layer and a deformable attention module, generate the spatial weight of the lesion area through a Grad-CAM heat map, and dynamically adjust the deformation offset of the convolutional kernel to retain calcification points and vascular textures; Apply discrete wavelet transform to decompose ultrasound shards into low-frequency approximation components and high-frequency detail components, and perform Huffman coding on the high-frequency subbands to compress blood flow noise; Fuse the identified CT shards, MRI shards, and ultrasound shards, and output a compressed shard set.

5. The distributed medical image data storage method based on a cloud platform according to claim 4, wherein: The division of the hot and cold storage layers according to the entropy value includes the following steps, Extract the gradient amplitude of the feature map during compression and dynamically allocate the modality weight coefficient; Statistically calculate the entropy value from the L1 norm of the feature map in the compressed shard set, and set an entropy threshold to divide the compressed shards into hot and cold layers: Compressed shards with entropy values higher than the entropy threshold are assigned to high-bandwidth storage nodes, and compressed shards with entropy values lower than the entropy threshold are secondarily compressed using the Zstandard algorithm and stored in low-performance storage nodes.

6. The distributed medical image data storage method based on a cloud platform according to claim 5, characterized in that: The generation of synthetic medical images includes the following steps: Parse the site identifiers of the compressed shard set, and extract gradient alignment information from the joint loss function as generation constraints; Load the pre-trained diffusion model, input the compressed shard feature map and the site identifier embedding vector, and control the generated data distribution through the learnable prompt vector; In the noise addition stage, gradually superimpose Gaussian noise on the compressed shard feature map to generate an intermediate noise sequence, and in the denoising stage, combine the site prompt vector to restore the original distribution to generate synthetic medical images.

7. The distributed medical image data storage method based on a cloud platform according to claim 6, wherein: The formation of the structured replay dataset means setting a reinforcement learning reward function, calculating the gradient direction cosine similarity, feature distribution entropy, and segmentation overlap rate between the synthetic data and the real data, screening the synthetic data with anatomical structure consistency, and binding the screened synthetic data with the site identifier, noise step number, and prompt vector version parameter.

8. The distributed medical image data storage method based on a cloud platform according to claim 2, wherein: The key metadata includes device information, imaging protocol, and patient identification.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the distributed medical image data storage method based on the cloud platform according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the distributed medical image data storage method based on the cloud platform according to any one of claims 1 to 8.

Images