Medical image data processing method and device and storage medium

Abstract

The invention relates to the technical field of medical image data processing scheme design, in particular to a medical image data processing method and device and a storage medium. The method comprises the following steps: constructing a cross-modal feature space through multi-modal self-supervised learning, eliminating cross-center difference by utilizing dynamic domain adaptation driven by equipment parameters, optimizing three-dimensional processing efficiency by adopting a layered sparse convolutional network, and introducing topological constraints to enhance the interpretability of a result. On the premise of ensuring data privacy, the optimized federated learning framework realizes high efficiency and safety of cross-mechanism collaborative training through collaborative design of gradient compression and noise injection. According to the scheme, the technical bottlenecks of annotation dependence, domain offset, calculation delay, untrusted results and the like are broken through, the accuracy, generalization and clinical acceptability of medical image analysis are remarkably improved, and a systematic solution is provided for intelligent medical treatment. The innovations jointly promote the leap-forward development of the medical image analysis technology to precision, real-time and credibility.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image data processing solution design, and in particular to a medical image data processing method, device and storage medium. Background Art

[0002] Although medical image processing technology has made significant progress in recent years, it still faces multi-dimensional technical bottlenecks. Traditional methods are highly dependent on data annotated by professional physicians, but the annotation cost is high and there are subjective differences. Especially for rare diseases, the lack of data makes model training difficult. Existing deep learning models have significant defects in multimodal data fusion. The differences in physical properties of different imaging devices (preferably including CT, MRI, and PET in the present invention) make it difficult to align the feature space, and the cross-modal information utilization efficiency is low. In addition, the domain offset problem caused by the parameter differences of medical equipment (preferably including the kilovolt peak of CT and the magnetic field strength of MRI in the present invention) seriously restricts the generalization ability of the model. The performance of the same algorithm in different medical institutions fluctuates significantly, which hinders clinical large-scale application.

[0003] There is a common contradiction between computational efficiency and accuracy in existing three-dimensional medical image processing. The computational delay of full-resolution three-dimensional convolutional networks cannot meet the real-time requirements of emergency scenarios, while reducing the resolution or simplifying the model will lead to an increase in the missed detection rate of subtle lesions. In terms of interpretability, there are cognitive differences between traditional visualization methods (the present invention preferably includes heat map generation) and clinical diagnostic logic, and doctors do not have enough trust in AI decision-making. The contradiction between privacy protection and data collaboration is particularly prominent, the phenomenon of medical data islands is common, and it is difficult to balance the communication overhead of traditional federated learning with the strength of privacy protection. In addition, compatibility issues between medical image analysis systems and existing hospital equipment (the present invention preferably includes PACS systems) occur frequently, and differences in data formats from different manufacturers lead to interruptions in clinical workflows and low efficiency in human-computer collaboration. These systemic defects jointly limit the clinical transformation and popularization of medical AI technology.

[0004] Therefore, the existing technology needs to be further developed. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above technical deficiencies and provide a method, device and storage medium for processing medical image data to solve the problems existing in the prior art.

[0006] To achieve the above technical objectives, according to a first aspect of the present invention, a method for processing medical image data is provided, comprising:

[0007] S1. Pre-training unlabeled medical images using a multimodal self-supervised learning framework to construct a cross-modal feature mapping matrix containing a shared feature space of CT, MRI, and PET modalities.

[0008] S2. Using a dynamic domain adaptation module to normalize the device parameters of the input image, the device parameters include at least the kVp value of CT, the magnetic field strength and slice thickness parameters of MRI;

[0009] S3. Applying Hierarchical Sparse Convolutional Networks to Process 3D Volume Data;

[0010] S4. Introducing differentiable topological constraints during the feature fusion phase, the Betti number distribution of the feature map is calculated using a persistent homology algorithm, and similarity matching is performed with a pre-set anatomical topological template.

[0011] S5. Distributed model updates based on the federated learning framework, using differential privacy mechanisms to inject gradient noise.

[0012] S6. Deploy a real-time inference acceleration engine to dynamically adjust image resolution based on GPU memory capacity.

[0013] S7. Generate a multi-granularity interpretability report, including pixel-level heat maps, lesion area geometric parameters, and a visualization chain of clinical decision pathways.

[0014] Specifically, the construction of the cross-modal feature mapping matrix in step S1 specifically includes:

[0015] A deformable convolution layer is set in the CT to MRI mapping path, and the deformation amplitude of the convolution kernel is constrained by the Hounsfield unit value;

[0016] A channel attention mechanism is used in the PET to CT mapping path, and a weight coefficient of 0.7-1.2 is assigned to high metabolic areas with SUVmax values ​​> 2.5.

[0017] Specifically, the device parameter normalization process in step S2 adopts:

[0018] An attenuation curve correction model is constructed for CT images based on kVp values, automatically compensating 15%-20% of soft tissue contrast when kVp < 100 is detected;

[0019] The Bloch equation is used to simulate the MRI image to generate a B1 field correction map to compensate for the geometric distortion caused by gradient nonlinearity.

[0020] Specifically, the layered sparse convolutional network in step S3 includes:

[0021] The coronal convolution layer is followed by a dilated convolution with a dilation rate of 2, and the receptive field is expanded to 23×23 pixels;

[0022] A switchable normalization layer is embedded in the sagittal convolution layer, and its statistics are dynamically calculated from the three adjacent slices.

[0023] Specifically, step S3 includes:

[0024] A 5×5 convolution kernel was used in the coronal plane, a 3×3 convolution kernel was used in the sagittal plane, and a 1×1 convolution kernel was used in the axial plane.

[0025] Specifically, the topology constraint in step S4 is implemented as follows:

[0026] Perform α-shape surface reconstruction on the feature map and calculate its 0-dimensional Betty number b0 and 1-dimensional Betty number b1;

[0027] When b0>3 or b1>5 is detected, the feature recalibration mechanism is triggered to adjust the channel weights through the graph convolutional network.

[0028] Specifically, the communication optimization of federated learning in step S5 includes:

[0029] The Krum algorithm is used to filter abnormal gradient updates, and the model parameter distillation mechanism is activated when the client data distribution difference is greater than 0.4;

[0030] Gradient compression uses a ternary quantization strategy to map the 32-bit floating-point gradient to three discrete values ​​{-1, 0, +1}.

[0031] Specifically, step S5 includes:

[0032] The noise standard deviation σ is controlled between 0.1 and 0.3.

[0033] According to a second aspect of the present invention, there is provided a medical image data processing device, comprising:

[0034] A multimodal adaptation module is used to pre-train unlabeled medical images using a multimodal self-supervised learning framework to construct a cross-modal feature mapping matrix that contains a shared feature space for CT, MRI, and PET modalities; and to use a dynamic domain adaptation module to normalize device parameters on the input image, where the device parameters include at least the kVp value for CT and the magnetic field strength and slice thickness parameters for MRI.

[0035] The control module is used to apply a hierarchical sparse convolutional network to process three-dimensional volume data; to introduce differentiable topological constraints in the feature fusion stage, calculate the Betti number distribution of the feature map through a continuous homology algorithm, and perform similarity matching with a preset anatomical structure topology template; to update the distributed model based on the federated learning framework, and use a differential privacy mechanism to inject gradient noise; to deploy a real-time inference acceleration engine, and dynamically adjust the image resolution according to the GPU memory capacity; and to generate multi-granularity interpretable reports, including pixel-level heat maps, lesion area geometric parameters, and a visual chain of clinical decision paths.

[0036] According to a third aspect of the present invention, there is provided a storage medium comprising: a memory; and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the above-mentioned method for processing medical image data is implemented.

[0037] Beneficial effects:

[0038] This invention achieves multi-dimensional breakthroughs in medical image processing through an innovative technical architecture. First, a pre-training framework based on multimodal self-supervised learning significantly reduces reliance on labeled data. A cross-modal feature sharing mechanism unlocks the potential value of unlabeled data, improving model robustness in small sample size scenarios. The dynamic domain adaptation module, through a device parameter-driven physical correction model, effectively eliminates image characteristic differences across centers and devices, ensuring stable diagnostic performance across different medical institutions. A hierarchical sparse convolutional network optimizes computational paths based on anatomical prior knowledge, reducing 3D processing latency to a clinically acceptable range while maintaining accuracy. This approach balances the real-time nature of emergency scenarios with the detection of subtle lesions. A topologically constrained interpretability generation mechanism combines mathematical morphological features with clinical cognition, providing a decision-making basis consistent with medical logic and significantly enhancing physicians' trust in AI results. The optimized federated learning framework, while ensuring data privacy, achieves efficient and secure cross-institutional collaborative training through the collaborative design of gradient compression and noise injection. Furthermore, a dynamic resource scheduling strategy automatically optimizes the balance between processing accuracy and speed based on hardware conditions, broadening the scope of device applicability for this technology. These innovations have jointly promoted the leapfrog development of medical image analysis technology towards precision, real-time and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flowchart of a method for processing medical image data provided in a specific embodiment of the present invention;

[0040] Figure 2 It is a schematic diagram of the system composition of the medical image data processing device provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is clearly and completely described below in conjunction with the drawings of the present invention. Based on the embodiments in this application, other similar embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of this application. In addition, the directional words mentioned in the following embodiments, such as "up", "down", "left", and "right", are only reference to the directions of the drawings. Therefore, the directional words used are used to illustrate rather than limit the invention.

[0042] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0043] See also Figure 1 The present invention provides a method for processing medical image data, comprising:

[0044] S1. Pre-training unlabeled medical images through a multimodal self-supervised learning framework to construct a cross-modal feature mapping matrix, which contains the shared feature space of the three modalities of CT, MRI, and PET.

[0045] Specifically, the construction of the cross-modal feature mapping matrix in step S1 specifically includes:

[0046] A deformable convolution layer is set in the CT to MRI mapping path, and the deformation amplitude of the convolution kernel is constrained by the Hounsfield unit value;

[0047] A channel attention mechanism is used in the PET to CT mapping path, and a weight coefficient of 0.7-1.2 is assigned to high metabolic areas with SUVmax values ​​> 2.5.

[0048] It should be further explained that, regarding step S1, the solution designed in the preferred embodiment of the present invention includes:

[0049] Design contrast loss function:

[0050]

[0051] in:

[0052] s CT,MRI : The cosine similarity between CT and MRI modality feature vectors is calculated as Reflects the consistency of anatomical structures between the two modalities;

[0053] τ = 0.07: Temperature coefficient, which adjusts the steepness of the probability distribution. Experiments show that when τ < 0.1, it can better distinguish similar sample pairs (preferably including tumors and normal tissues in this invention);

[0054] N = 256: The number of negative samples, determined by optimizing GPU memory capacity on Tesla V100.

[0055] Cross-modal projection head:

[0056] The 3-layer MLP dimensions are 256, 128, and 64, and the experimental results show that:

[0057] When the hidden layer is 128-dimensional, the feature separability index (Silhouette Score) reaches 0.68, which is better than other configurations;

[0058] The output dimension is 64, which satisfies the minimum orthogonal basis requirement for the fusion of three modal features.

[0059] S2. Use the dynamic domain adaptation module to normalize the device parameters of the input image, where the device parameters include at least the kVp value of CT, the magnetic field strength and layer thickness parameters of MRI.

[0060] Specifically, the device parameter normalization process in step S2 adopts:

[0061] An attenuation curve correction model is constructed for CT images based on kVp values, automatically compensating 15%-20% of soft tissue contrast when kVp < 100 is detected;

[0062] The Bloch equation is used to simulate the MRI image to generate a B1 field correction map to compensate for the geometric distortion caused by gradient nonlinearity.

[0063] It should be further explained that, regarding step S2, the solution designed in the preferred embodiment of the present invention includes:

[0064] Design CTkVp compensation formula:

[0065]

[0066] in:

[0067] I raw : Original CT value (Hounsfield Units);

[0068] kVp: X-ray tube peak voltage (range 80-140 kV);

[0069] Coefficient 0.18: obtained by fitting the water phantom experiment, which reduces the standard deviation of the CT value of soft tissue from 25HU to 18HU at 80kV. MRI geometric correction model:

[0070]

[0071] Parameter Description:

[0072] B0: actual magnetic field strength (unit T), B ref =3.0T is the standard reference value;

[0073] a n : Polynomial coefficients, calibrated by spherical harmonic analysis, to compensate for the displacement caused by gradient nonlinearity;

[0074] Experimental data: When B0=1.5T, the maximum geometric distortion is corrected from 5.3mm to 1.2mm.

[0075] S3. Applying hierarchical sparse convolutional networks to process 3D volume data.

[0076] Specifically, the layered sparse convolutional network in step S3 includes:

[0077] The coronal convolution layer is followed by a dilated convolution with a dilation rate of 2, and the receptive field is expanded to 23×23 pixels;

[0078] A switchable normalization layer is embedded in the sagittal convolution layer, and its statistics are dynamically calculated from the three adjacent slices.

[0079] Specifically, step S3 includes:

[0080] A 5×5 convolution kernel was used in the coronal plane, a 3×3 convolution kernel was used in the sagittal plane, and a 1×1 convolution kernel was used in the axial plane.

[0081] It should be further explained that, regarding step S3, the solution designed in the preferred embodiment of the present invention includes:

[0082] Coronal convolution kernel design:

[0083] The basis for selecting the 5×5 kernel is that the average width of the human coronal anatomical structure (preferably including the lung lobe boundary in the present invention) is ≈20 mm, which corresponds to the effective sensing range of the 5×5 kernel at a resolution of 1 mm.

[0084] Dynamic statistics calculation:

[0085]

[0086] in, The nth feature map of the adjacent i-th slice (i∈{-1,0,1}); 3N: normalization factor to ensure comparability of statistics across slices.

[0087] Effect: In the liver segmentation task, the Dice coefficient increased from 0.83 to 0.89.

[0088] S4. Differentiable topological constraints are introduced in the feature fusion stage. The Betti number distribution of the feature map is calculated through the continuous homology algorithm, and similarity matching is performed with the preset anatomical structure topology template.

[0089] Specifically, the topology constraint in step S4 is implemented as follows:

[0090] Perform α-shape surface reconstruction on the feature map and calculate its 0-dimensional Betty number b0 and 1-dimensional Betty number b1;

[0091] When b0>3 or b1>5 is detected, the feature recalibration mechanism is triggered to adjust the channel weights through the graph convolutional network.

[0092] It should be further explained that, regarding step S4, the solution designed in the preferred embodiment of the present invention includes:

[0093] α-shape parameterization:

[0094] α=0.7: empirical value, which satisfies the present invention preferably includes the following formula:

[0095]

[0096] Here, ||e|| is the longest edge in the Delaunay triangulation, ensuring that small structures such as blood vessels are preserved.

[0097] Betti number threshold setting:

[0098] b0>3: indicates abnormal separation of organs (preferably including interlobar effusion in the present invention);

[0099] b1>5: reflects that the complexity of the vascular network is too high (preferably including tumor neovascularization in the present invention);

[0100] Design recalibration formula:

[0101]

[0102] in:

[0103] Hc: eigenvector of channel c;

[0104] bk: is the kth Betti number, encoding the degree of topological anomaly;

[0105] tanh: Limit the adjustment range to [-1, 1] to avoid gradient explosion.

[0106] S5. Perform distributed model updates based on the federated learning framework and use differential privacy mechanisms to inject gradient noise.

[0107] Specifically, the communication optimization of federated learning in step S5 includes:

[0108] The Krum algorithm is used to filter abnormal gradient updates, and the model parameter distillation mechanism is activated when the client data distribution difference is greater than 0.4;

[0109] Gradient compression uses a ternary quantization strategy to map the 32-bit floating-point gradient to three discrete values ​​{-1, 0, +1}.

[0110] Specifically, step S5 includes:

[0111] The noise standard deviation σ is controlled between 0.1 and 0.3.

[0112] It should be further explained that, regarding step S5, the solution designed in the preferred embodiment of the present invention includes:

[0113] Differential privacy noise injection:

[0114]

[0115] Δf: gradient sensitivity, controlled by clipping to L2 norm ≤ 1.5;

[0116] ε=1: Privacy budget, which satisfies the present invention preferably includes the following formula:

[0117]

[0118] Among them, T = 100 is the number of training rounds, δ = 10 -5 is the failure probability.

[0119] Design a ternary quantization function:

[0120]

[0121] in:

[0122] Threshold 0.5σ: verified on the MNIST and BraTS datasets, making the quantization error rate < 8%;

[0123] Communication compression ratio: 32-bit floating point is converted to 2-bit ternary value, with a compression rate of 93.75%.

[0124] S6. Deploy a real-time inference acceleration engine to dynamically adjust image resolution based on GPU memory capacity.

[0125] Specifically, the resolution adjustment strategy in step S6 is:

[0126] Create a lookup table for resolution and video memory usage, and automatically switch to low-resolution mode when the remaining video memory is less than 2GB;

[0127] A lightweight U-Net network is deployed in the bypass branch, and its encoder depth is reduced to 1 / 3 of the original model.

[0128] It should be further explained that, regarding step S6, the solution designed in the preferred embodiment of the present invention includes:

[0129] The video memory-resolution strategy of the present invention preferably includes the following as shown in Table 1:

[0130] Table 1 Video memory-resolution strategy

[0131] Remaining video memory Downsampling rate Technical basis ≥6GB 1× 5.2GB of video memory required for full resolution processing 4-6GB 4× Bilinear interpolation downsampling, PSNR>40dB <4GB 8× Enable lightweight U-Net, with 0.9M parameters

[0132] Lightweight U-Net structure:

[0133] Number of encoder layers: 2 layers (originally 4 layers), with the number of channels per layer being 64, 32, and 16, respectively.

[0134] Skip connection: Only retained in the second layer, reducing memory usage by 50%.

[0135] Effect: The Dice coefficient dropped to <0.03, and the inference speed increased by 4 times.

[0136] It is understandable that:

[0137] 1X (full resolution):

[0138] The original image resolution is kept unchanged, and all spatial dimensions (length, width, and number of layers) are processed according to the original size.

[0139] Example:

[0140] If the original CT image is 512×512×32 voxels (i.e., 512×512 pixels per layer, 32 layers in total), 1X downsampling maintains this resolution.

[0141] 4X (quarter resolution):

[0142] Each spatial dimension is reduced to 1 / 2 of its original size (area is reduced to 1 / 4, volume is reduced to 1 / 8).

[0143] Mathematical expression:

[0144] The input size H×W×D is transformed into the output size (H / 2)×(W / 2)×(D / 2).

[0145] Example:

[0146] 512×512×32 is converted to 256×256×16 (the total number of pixels is reduced to 1 / 8 of the original).

[0147] 8X (one-eighth resolution):

[0148] Each spatial dimension is reduced to 1 / 4 of its original size (area is reduced to 1 / 16, volume is reduced to 1 / 64).

[0149] Example:

[0150] 512×512×32 is converted to 128×128×8 (the total number of pixels is reduced to 1 / 64 of the original).

[0151] S7. Generate a multi-granularity interpretability report, including pixel-level heat maps, lesion area geometric parameters, and a visualization chain of clinical decision pathways.

[0152] It should be further explained that, regarding step S7, the solution designed in the preferred embodiment of the present invention includes:

[0153] Improved Grad-CAM++:

[0154]

[0155] in:

[0156] The gradient sensitivity of category c to the feature map position (i, j);

[0157] Sobel(A ij ): Edge strength, when λ = 0.3, 82% doctor agreement was achieved in lung nodule detection;

[0158] Experimental verification: When λ = 0.5, the false positive heat map is reduced by 23%. Geometric parameter calculation:

[0159] Surface area error control:

[0160]

[0161] The step size is guaranteed to be ≤ 0.5 mm by the Marching Cubes algorithm.

[0162] Specifically, the present invention has been verified for its technical effects, and the verification results show that the present invention preferably includes the following as shown in Table 2:

[0163] Table 2 Technical effect verification results

[0164]

[0165] As can be understood, this invention achieves multi-dimensional breakthroughs in medical image processing through its innovative technical architecture. First, a pre-training framework based on multimodal self-supervised learning significantly reduces reliance on labeled data. A cross-modal feature sharing mechanism exploits the potential value of unlabeled data, improving model robustness in small sample size scenarios. The dynamic domain adaptation module, through a device parameter-driven physical correction model, effectively eliminates image characteristic differences across centers and devices, ensuring stable diagnostic performance across different medical institutions. A hierarchical sparse convolutional network optimizes computational paths based on anatomical prior knowledge, reducing 3D processing latency to a clinically acceptable range while maintaining accuracy, balancing the real-time requirements of emergency scenarios with the detection of subtle lesions. A topologically constrained interpretability generation mechanism combines mathematical morphological features with clinical cognition, providing a decision-making basis consistent with medical logic and significantly enhancing physicians' trust in AI results. The optimized federated learning framework, while ensuring data privacy, achieves efficient and secure cross-institutional collaborative training through the collaborative design of gradient compression and noise injection. Furthermore, a dynamic resource scheduling strategy automatically optimizes the balance between processing accuracy and speed based on hardware conditions, broadening the scope of device applicability for this technology. These innovations have jointly promoted the leapfrog development of medical image analysis technology towards precision, real-time and reliability.

[0166] See also Figure 2The present invention provides another embodiment, which provides a medical image data processing device, the medical image data processing device comprising:

[0167] The multimodal adaptation module 100 is used to pre-train unlabeled medical images using a multimodal self-supervised learning framework to construct a cross-modal feature mapping matrix containing a shared feature space of the three modalities of CT, MRI, and PET; and to use a dynamic domain adaptation module to normalize device parameters on the input image, wherein the device parameters include at least the kVp value of CT and the magnetic field strength and slice thickness parameters of MRI;

[0168] Control module 200 is used to apply a hierarchical sparse convolutional network to process three-dimensional volume data; to introduce differentiable topological constraints in the feature fusion stage, calculate the Betti number distribution of the feature map through a continuous homology algorithm, and perform similarity matching with a preset anatomical structure topology template; to update the distributed model based on the federated learning framework, and inject gradient noise using a differential privacy mechanism; to deploy a real-time inference acceleration engine, and dynamically adjust the image resolution according to the GPU memory capacity; and to generate a multi-granularity interpretable report, including a pixel-level heat map, lesion area geometric parameters, and a visualization chain of clinical decision paths.

[0169] It should be noted that this invention achieves multi-dimensional breakthroughs in medical image processing through its innovative technical architecture. First, a pre-training framework based on multimodal self-supervised learning significantly reduces reliance on labeled data. A cross-modal feature sharing mechanism exploits the potential value of unlabeled data, improving model robustness in small sample size scenarios. The dynamic domain adaptation module, through a device parameter-driven physical correction model, effectively eliminates image characteristic differences across centers and devices, ensuring stable diagnostic performance across different medical institutions. A hierarchical sparse convolutional network optimizes computational paths based on anatomical prior knowledge, reducing 3D processing latency to a clinically acceptable range while maintaining accuracy, balancing the real-time requirements of emergency scenarios with the detection of subtle lesions. A topologically constrained interpretability generation mechanism combines mathematical morphological features with clinical cognition, providing a decision-making basis consistent with medical logic and significantly enhancing physicians' trust in AI results. The optimized federated learning framework, while ensuring data privacy, achieves efficient and secure cross-institutional collaborative training through the collaborative design of gradient compression and noise injection. Furthermore, a dynamic resource scheduling strategy automatically optimizes the balance between processing accuracy and speed based on hardware conditions, broadening the scope of device applicability for this technology. These innovations have jointly promoted the leapfrog development of medical image analysis technology towards precision, real-time and reliability.

[0170] In a preferred embodiment, the present application further provides a storage medium, comprising:

[0171] A memory; and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method for processing medical image data is implemented. The computer device can be broadly defined as a server, a terminal, or any other storage medium with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, a memory, a network interface, a communication interface, etc. connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and an internal memory. An operating system, a computer program, etc. may be stored in or on the non-volatile storage medium. The internal memory can provide an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface and the communication interface of the computer device can be used to connect to and communicate with external devices via a network. When the computer program is executed by the processor, the steps of the method of the present invention are performed.

[0172] The present invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the method of an embodiment of the present invention to be performed. In one embodiment, the computer program is distributed on a plurality of computer devices or processors coupled to a network so that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, can be performed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations can be performed by one or more computer devices or processors, and one or more other method steps / operations can be performed by one or more other computer devices or processors. One or more computer devices or processors can perform a single method step / operation, or perform two or more method steps / operations.

[0173] It will be understood by those skilled in the art that the method steps of the present invention can be instructing the relevant hardware through a computer program. The present invention preferably includes a computer device or a processor to perform, and the computer program can be stored in a non-transitory computer-readable storage medium, which causes the steps of the present invention to be performed when the computer program is executed. Depending on the circumstances, any reference to memory, storage, database or other media herein may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0174] As can be understood, this invention achieves multi-dimensional breakthroughs in medical image processing through its innovative technical architecture. First, a pre-training framework based on multimodal self-supervised learning significantly reduces reliance on labeled data. A cross-modal feature sharing mechanism exploits the potential value of unlabeled data, improving model robustness in small sample size scenarios. The dynamic domain adaptation module, through a device parameter-driven physical correction model, effectively eliminates image characteristic differences across centers and devices, ensuring stable diagnostic performance across different medical institutions. A hierarchical sparse convolutional network optimizes computational paths based on anatomical prior knowledge, reducing 3D processing latency to a clinically acceptable range while maintaining accuracy, balancing the real-time requirements of emergency scenarios with the detection of subtle lesions. A topologically constrained interpretability generation mechanism combines mathematical morphological features with clinical cognition, providing a decision-making basis consistent with medical logic and significantly enhancing physicians' trust in AI results. The optimized federated learning framework, while ensuring data privacy, achieves efficient and secure cross-institutional collaborative training through the collaborative design of gradient compression and noise injection. Furthermore, a dynamic resource scheduling strategy automatically optimizes the balance between processing accuracy and speed based on hardware conditions, broadening the scope of device applicability for this technology. These innovations have jointly promoted the leapfrog development of medical image analysis technology towards precision, real-time and reliability.

[0175] The various technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification as long as such combination does not conflict.

[0176] The specific embodiments of the present invention described above do not limit the scope of protection of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for processing medical image data, characterized in that: The method comprises: S1. Pre-training unlabeled medical images using a multimodal self-supervised learning framework to construct a cross-modal feature mapping matrix containing a shared feature space of CT, MRI, and PET modalities. S2. Using a dynamic domain adaptation module to normalize the device parameters of the input image, the device parameters include at least the kVp value of CT, the magnetic field strength and slice thickness parameters of MRI; S3. Applying Hierarchical Sparse Convolutional Networks to Process 3D Volume Data; S4. Introducing differentiable topological constraints during the feature fusion phase, the Betti number distribution of the feature map is calculated using a persistent homology algorithm, and similarity matching is performed with a pre-set anatomical topological template. S5. Distributed model updates based on the federated learning framework, using differential privacy mechanisms to inject gradient noise. S6. Deploy a real-time inference acceleration engine to dynamically adjust image resolution based on GPU memory capacity. S7. Generate a multi-granularity interpretability report, including pixel-level heat maps, lesion area geometric parameters, and a visualization chain of clinical decision pathways.

2. The method for processing medical image data according to claim 1, characterized in that: The construction of the cross-modal feature mapping matrix in step S1 specifically includes: A deformable convolution layer is set in the CT to MRI mapping path, and the deformation amplitude of the convolution kernel is constrained by the Hounsfield unit value; A channel attention mechanism is used in the PET to CT mapping path, and a weight coefficient of 0.7-1.2 is assigned to high metabolic areas with SUVmax values ​​> 2.

5.

3. The method for processing medical image data according to claim 1, wherein: The device parameter normalization process in step S2 adopts: An attenuation curve correction model is constructed for CT images based on kVp values, automatically compensating 15%-20% of soft tissue contrast when kVp < 100 is detected; The Bloch equation is used to simulate the MRI image to generate a B1 field correction map to compensate for the geometric distortion caused by gradient nonlinearity.

4. The method for processing medical image data according to claim 1, wherein: The layered sparse convolutional network in step S3 includes: The coronal convolution layer is followed by a dilated convolution with a dilation rate of 2, and the receptive field is expanded to 23×23 pixels; A switchable normalization layer is embedded in the sagittal convolution layer, and its statistics are dynamically calculated from the three adjacent slices.

5. The method for processing medical image data according to claim 4, characterized in that: The step S3 comprises: A 5×5 convolution kernel was used in the coronal plane, a 3×3 convolution kernel was used in the sagittal plane, and a 1×1 convolution kernel was used in the axial plane.

6. The method for processing medical image data according to claim 1, wherein: The topology constraint in step S4 is implemented as follows: Perform α-shape surface reconstruction on the feature map and calculate its 0-dimensional Betty number b0 and 1-dimensional Betty number b1; When b0>3 or b1>5 is detected, the feature recalibration mechanism is triggered to adjust the channel weights through the graph convolutional network.

7. The method for processing medical image data according to claim 4, characterized in that: The communication optimization of the federated learning in step S5 includes: The Krum algorithm is used to filter abnormal gradient updates, and the model parameter distillation mechanism is activated when the client data distribution difference is greater than 0.4; Gradient compression uses a ternary quantization strategy to map the 32-bit floating-point gradient to three discrete values ​​{-1, 0, +1}.

8. The method for processing medical image data according to claim 7, characterized in that: The step S5 comprises: The noise standard deviation σ is controlled between 0.1 and 0.

3.

9. A medical image data processing device, characterized in that: include: A multimodal adaptation module is used to pre-train unlabeled medical images using a multimodal self-supervised learning framework to construct a cross-modal feature mapping matrix that contains a shared feature space for CT, MRI, and PET modalities; and to use a dynamic domain adaptation module to normalize device parameters on the input image, where the device parameters include at least the kVp value for CT and the magnetic field strength and slice thickness parameters for MRI. The control module is used to apply a hierarchical sparse convolutional network to process 3D volume data; it is used to introduce differentiable topological constraints during the feature fusion stage, calculate the Betti number distribution of the feature map through a continuous homology algorithm, and perform similarity matching with a preset anatomical structure topology template; it is used to update the distributed model based on the federated learning framework and use a differential privacy mechanism to inject gradient noise; it is used to deploy a real-time inference acceleration engine and dynamically adjust the image resolution based on the GPU memory capacity; Used to generate multi-granularity interpretable reports, including pixel-level heat maps, lesion area geometric parameters, and visualization chains of clinical decision pathways.

10. A storage medium, characterized in that: include: Memory; and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method for processing medical image data according to any one of claims 1 to 8 is implemented.

Images