A multi-modal tumor recognition method and system based on riemannian manifold fusion

Abstract

The application discloses a kind of multi-modal tumor identification method and system based on riemannian manifold fusion, belong to tumor identification technical field, to solve the problem that existing multi-modal tumor identification scheme exists spatial and temporal heterogeneous deviation, weak anti-interference ability, image and impedance data deep physical coupling is insufficient, identification conflict lacks effective arbitration;The application first obtains the image and impedance multi-modal original data of space-time synchronization, is completed standardization preprocessing by hardware level clock alignment and cross-modal space coordinate mapping;Again, based on dynamic rhythmization feature distribution, image recognition is completed, impedance recognition is completed based on fluid physical convection and diffusion mechanism, and the corresponding identification result is respectively output;Finally, through the conflict arbitration mechanism of riemannian manifold space mapping and curvature guidance, the dual-mode identification result is fused, and the final tumor identification result is output.The application effectively eliminates the spatial and temporal heterogeneous deviation of multi-modal data, improves the robustness and anti-interference ability of feature extraction in complex clinical environment, realizes the deep physical cooperation of dual-mode data, improves the accuracy and decision fault tolerance of early tumor identification.

Description

Technical Field

[0001] This invention relates to the field of tumor identification, and in particular to a multimodal tumor identification method and system based on Riemannian manifold fusion. Background Technology

[0002] Improving the clinical treatment effect of tumors and prolonging patient survival depends heavily on the early detection and identification of tumors. Currently, clinical diagnosis mainly relies on medical imaging examinations and electrophysiological tests. Among existing technologies, identification schemes based on a single modality have obvious limitations. Medical imaging can intuitively show the macroscopic anatomical morphology of tumors, but it is easily affected by imaging noise and uneven illumination, and it is not sensitive enough to microscopic tissue changes caused by tumor stroma pressure. While electrical impedance testing has high sensitivity to functional changes and small electrophysiological fluctuations in tissues, its data is mainly presented as a one-dimensional time series, which is weak in spatial resolution and morphological localization, and is easily affected by power frequency interference and baseline drift. To overcome the limitations of single-modal detection, several multimodal fusion schemes have emerged in recent years, attempting to combine image and impedance data to improve detection accuracy. However, these existing multimodal schemes still face the following difficulties in practical applications: Existing fusion technologies often focus on simple cascading of independent features from each modality or simple classification of fully connected layers, lacking in-depth physical mapping analysis between the macroscopic morphology of images and the microscopic differences in impedance. Due to the significant spatiotemporal heterogeneity between medical image data and impedance time-series signals in terms of physical dimensions, acquisition frequency, and spatial distribution characteristics, conventional feature extraction and fusion methods struggle to achieve alignment and feature overlap of heterogeneous data in the same physical coordinate system. Furthermore, in the face of complex physiological environments such as visual occlusion, tissue deformation, or impedance electrode detachment, existing multimodal fusion lacks dynamic anti-interference and fault-tolerance mechanisms. When transient biological noise or recognition conflicts occur in a certain modality, the model is prone to fusion weight imbalance, leading to the collapse of the overall recognition result. Existing AI recognition models also struggle to extract stable features from underlying physical laws when processing multi-source heterogeneous data, resulting in poor robustness of diagnostic recommendations under varying clinical pressures. In summary, existing technologies still suffer from problems such as large spatiotemporal heterogeneity deviations, weak anti-interference capabilities, insufficient deep physical coupling, and lack of effective arbitration in dealing with identification conflicts when processing deep fusion of medical images and electrical impedance signals. There is an urgent need for a more robust tumor identification method that can achieve multimodal deep physical collaboration and dynamic fault tolerance. Summary of the Invention

[0003] The main objective of this invention is to provide a multimodal tumor identification method and system based on Riemannian manifold fusion, which solves the problems of large spatiotemporal heterogeneity deviation, weak anti-interference ability, shallow physical coupling, and lack of effective arbitration to deal with identification conflicts.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a multimodal tumor identification method based on Riemannian manifold fusion, the method comprising the following steps: S1. Multimodal data acquisition and preprocessing to obtain spatiotemporally synchronized raw image datasets and impedance datasets; S2. Based on the rhythmic feature-driven structure, image recognition is completed, and the image recognition result is output. ; S3. Impedance identification is completed based on the fluid-like feature deconstruction and re-aggregation mechanism, and the impedance identification results are output. ; S4. Fusion through evidence emergence mechanism and The final tumor identification result was obtained based on Riemannian manifold evolution. .

[0005] In the preferred embodiment, step S1 specifically includes the following steps: S11. A multimodal acquisition device integrating multi-level penetration imaging units and dynamic impedance sensing units is used to image the suspected tumor target area. First, the FPGA main control unit generates frequency-synchronized microsecond-level TTL trigger pulses to drive the medical imaging sensor and the high-density electrode array respectively. Subsequently, two-dimensional structural images and three-dimensional volume data covering different depths of surface, middle, and deep tissues were generated. Simultaneously, gradient-based weak electrical stimulation with varying current and frequency was applied, and the impedance change sequences of grid nodes divided by step size within the target area at different depths and time periods were recorded to obtain the original image dataset. Compared with the original impedance dataset ; To ensure hardware-level signal synchronization, the FPGA-level physical clock alignment model is defined as follows: (1); in, This is a microsecond-level TTL pulse trigger signal vector; For synchronous sampling period; For rectangular pulse window functions; The pulse width; This represents the convolution operation; Hardware clock locking eliminates multimodal spatiotemporal deviation errors caused by system timeline jitter; S12. Perform denoising, enhancement, and registration extraction operations sequentially. A dynamic threshold denoising algorithm combined with neighborhood similarity filtering is used to remove imaging noise. An adaptive grayscale calibration algorithm is used to enhance tissue edge contrast and a light and shadow equalization adjustment mechanism is used to correct uneven illumination. Based on the feature anchor point alignment mechanism, images of different depths and modalities are aligned to the same spatial coordinate system. The joint calculation formula for dynamic threshold noise reduction and adaptive grayscale calibration is as follows: (2); in, To enhance the denoised spatial image data; and These represent the Fast Fourier Transform and its inverse transform, respectively. It is a non-local mean filter weight kernel; This is a cutoff frequency set dynamically based on the region. The order of the high-pass gain; It is an adaptive correction function; This is the Laplacian operator, used to extract the gradient at the edge of spatial organization; Subsequently, a coaxial mapping mechanism supplemented by a homography matrix is ​​introduced to align the coordinate system: (3); in, To calibrate the coordinates of the image pixel system; For the physical coordinates of the multimodal sensing system; This is the depth scale scaling factor; This is the homography transformation matrix for cross-modal space alignment; Obtaining the homography transformation matrix aligned across modal spaces Then, spatial resampling is used to eliminate the enhanced image data. Due to the physical parallax of multi-source sensors, it will be based on the physical coordinate system. The acquired discrete two-dimensional structural images and three-dimensional volume data are mapped to a unified pixel coordinate system. Perform projection reconstruction to generate a continuous voxel matrix with isotropic resolution, and output a normalized 3D image dataset. This variable will then be used in subsequent image recognition steps. S13. A fluctuation interference filtering algorithm is used to remove power frequency interference from impedance data, and a trend fitting correction mechanism is used to eliminate baseline drift. The formulas for impedance value normalization, impedance trend fitting, and baseline drift elimination are as follows: (4); in, The standard impedance time series after baseline drift elimination and normalization; For the first Legendre polynomials are used to fit low-frequency baseline drift profiles. The coefficient matrix is ​​the polynomial fitting matrix. To fit the highest order; This represents the average impedance within a single acquisition cycle. To prevent tiny constants with a denominator of zero; Based on cross-modal alignment matrix By anchoring the physical coordinates of impedance sensing grid nodes to the image pixel grid system and performing spatial overlap between physical electrical features and visual anatomical features, a complete standard spatiotemporal synchronous impedance dataset with consistent dimensions is obtained. This variable is then used for impedance identification.

[0006] In the preferred embodiment, step S2 specifically includes the following steps: S21. Calculate the image grayscale gradient, divide the feature regions, and generate dynamically rhythmic segment groups. The gradient operator identifies abrupt edge regions and smooth texture regions in the image. Regions with gray-level gradients greater than the upper threshold are classified as abrupt edge regions, while regions with gray-level gradients less than the lower threshold are classified as smooth texture regions. An adaptive rhythm allocation function based on multi-peak distribution is used to assign observation rhythms to each region. A low rhythm coefficient is assigned to abrupt edge regions, and each segment is scanned more times. A high rhythm coefficient is assigned to smooth texture regions, and each segment is scanned less times. Texture abrupt regions discovered during the scanning process are re-segmented, and continuous smooth regions are merged to generate dynamically rhythmic segment groups. In rhythm assignment calculation, a fractional-order KAN network dynamics model is used to capture the nonlocal characteristics of tumor growth. Its mathematical evolution function is defined as: (5); in, It is a Caputo-type fractional calculus operator that has a long-term memory of the tumor margin infiltration history. The calculated adaptive rhythm coefficient matrix; For mapping constants of hidden layer dimensions of child nodes within the KAN network; The number of feature channels in the input image; This is a parameterized B-spline activation function used for smoothing nonlinear feature responses; This is the outer nonlinear aggregation function; The stochastic dynamic perturbation term introduced for the nonlinear oscillator; The pixel-level adaptive rhythm coefficient matrix is ​​obtained through a fractional-order KAN network dynamic model. Then, spatial feature aggregation and nonlinear parameter mapping are performed on each dynamically rhythmic segment. First, spatial integration is performed on the pixel rhythm within the segment to calculate the first... Regional rhythm expectation value of each dynamic rhythmic segment Subsequently, the scalar expectation value is mapped using a pre-defined rhythm and motion mapping function. Transformed into a multi-dimensional parameter configuration vector containing the number of scans, switching frequency, and deformation boundaries. The calculation formula is: (6); in, This represents the total number of observation scans allocated to this segment; The time interval for dynamic viewpoint switching; This is a boundary limit for the generation of the viewpoint in this section; and For hardware response scaling factor, ensure Mapping to the desired and ; It is the Sigmoid smoothing function; This is the edge detection constant; The maximum physical transformation threshold set for the system; Derive the scan parameter configuration vector for this segment. Finally, the configuration vectors of all segments are aggregated to form a global configuration set. ; S22. Generate multi-view wave sequences and perform rhythm synchronization adjustment. Apply sequentially to each rhythmic segment Local grayscale fluctuations, horizontal / vertical texture flipping, Directional offset and pixel-level contour scaling are used to generate fluctuation sequences from multiple perspectives. Under dynamic switching frequencies, local grayscale fluctuations and texture flipping are applied to each segment to simulate the dynamic visual scanning process. A rhythm synchronization adjustment mechanism is used to match the perspective switching frequency with the segment observation rhythm. Fast rhythm segments have short perspective switching intervals, while slow rhythm segments have long switching intervals, forming a dynamic perspective feature data stream. ; S23. Use the first four layers of a deep feature extraction network to extract feature information, generate viewpoint feature sequences, and concatenate them to form a feature package. Extract feature information from each segment from each viewpoint to generate a viewpoint feature sequence; calculate the attention weight of the feature sequence in the time dimension, strengthen the dominant rhythm features that account for a larger proportion, and retain the secondary rhythm features that account for a smaller proportion; process the feature sequence... , , Multi-scale feature aggregation and concatenation form a 2048-dimensional segmental multi-scale rhythmic feature package; the formula for calculating the attention weight of the feature sequence time dimension and the multi-scale aggregation mechanism are expressed as follows: (7); (8); in, For time step The following is a matrix showing the distribution of attention; A learnable query and key mapping weight matrix; Dimension factor for feature scaling; Used for weight normalization to extract primary and secondary rhythm features; This is a concatenation operation along the feature channel dimension; For two-dimensional convolution operations of the receptive field at the corresponding scale; For Hadamard product operations; The generated 2048-dimensional segmental multi-scale rhythmic feature packet tensor ; S24. Establish rhythm correlations, merge global rhythm flows, and perform rhythm stability verification. Establish rhythm associations between different segments using inter-segment rhythm association matrices, and package local rhythm features. Merged into an 8192-dimensional global rhythmic flow The rhythm stream was reduced to 1024 dimensions using a feature dimension compression algorithm and input into a fully connected classification layer to obtain preliminary recognition results. Rhythm stability was verified by repeating feature extraction under various rhythmic stresses. If the results are consistent, the image recognition result is output. Otherwise, select the category with the highest percentage as... ; including global rhythm flow With recognition results The mapping network model is expressed as: (9); (10); in, This is the inter-segment rhythmic association adjacency matrix learned through a graph convolutional network; This refers to the global dimension reduction projection matrix and bias terms; For the applied first A series of rhythmic pressure perturbation vectors were used to simulate the visual characteristic response changes of tumor interstitial pressure. It is a multilayer perceptron classifier; Finally, after stability testing, the global rhythm flow vector is output. With recognition results ; In the preferred embodiment, step S3 specifically includes the following steps: S31. Convert the preprocessed impedance time series into a three-dimensional virtual fluid particle point array and generate an impedance diffusion band covering the spatiotemporal domain. Particle coordinates correspond to spatial positions With time The particle property value is the rate of change of impedance at that spacetime point. Furthermore, the particle density is positively correlated with the magnitude of impedance change; the spatiotemporal synchronous impedance dataset is analyzed using fluid physics convection and diffusion equations. Dynamically transformed into a particle density field The calculation equation is: (11); in, This represents the virtual particle density field distribution of charge carriers in a porous medium. for The velocity vector of biological tissue fluid diffusing into the flow field; for The anisotropic diffusion coefficient tensor; Here, is the divergence operator, representing the local net convective outflow. The Laplace operator dominates the concentration gradient diffusion caused by Brownian motion. By numerically solving this partial differential equation using the alternating direction implicit difference method, a three-dimensional impedance diffusion band tensor continuously covering the spatiotemporal domain is generated. ; S32. Apply aggregate rhythm, discrete rhythm, and oscillatory rhythm sequentially for multi-level analysis. Convergent rhythm through the central gravitational action function Iterative aggregation of particles toward the local mean point; discrete rhythms exert random dispersion forces. Disrupt highly consistent regions; The oscillating rhythm is applied alternately in multiple rounds with both aggregation and dispersion rhythms, affecting the particle distribution after oscillation. Perform three-level parsing: Macroscopic structural layer features are extracted through eigenvalue analysis of structural tensors. The micro-texture layer extracts features based on the local information entropy of the evolution field. The micro-signal sensing layer employs a weak signal capture device based on Lorentz resonance to capture... Level continuous impedance fluctuation signal, i.e., micro-variable sensing layer characteristics ; S33. Aggregate three-level features to generate impedance behavior units and cluster them to output impedance identification results. Based on the interpolation algorithm, the physical electrode data is reconstructed to logical resolution. Three-level features are then logically aggregated according to their weights: large for the macroscopic structure layer, small for the microscopic texture layer and the micro-variation perception layer, generating corresponding... Impedance behavior units in spatial regions; clustering behavior units with similarity coefficients greater than a threshold based on feature similarity coefficients, extracting time-domain and frequency-domain features of each group of behavior units, and splicing them together to form a 1024-dimensional impedance core feature. The input classification layer yields preliminary results, which are then verified through dynamic consistency testing (where the feature change rate is less than the minimum value for several consecutive frames). The output impedance identification result is then determined. .

[0007] In the preferred embodiment, step S4 specifically includes the following steps: S41. Construction of Evidence Space and Riemannian Manifold Mapping Will and Treating both sources of evidence as independent sources, we define category support, category repulsion, fluctuation sensitivity, and rhythmic stress response as quantitative dimensions of the evidentiary events; and we consider the high-dimensional features of both sources of evidence. And the preliminary identification results are mapped to a 256-dimensional high-dimensional evidence Riemannian manifold space. In the context of mapping, the mathematical topological transformation is defined as follows: (12); in, For constructing a locally metrically flat non-Euclidean Riemannian manifold space; and Let be the kernel functions for the nonlinear homeomorphic mapping from the image and the impedance domain to the evidence space, respectively; For the spatial tensor direct sum operation under Lie groups; Every point on the manifold The geometric deformation is subject to the Riemannian metric tensor Strict constraints.

[0008] S42. Based on the attention allocation mechanism and curvature-guided conflict arbitration interaction, the final result is output. First, the excitation weight of image evidence to impedance evidence is calculated based on the attention allocation mechanism, and the clustering threshold of impedance behavior units is adjusted to recalculate the category tendency. At the same time, the feature extraction weight of the corresponding region in the image is strengthened according to the location of the impedance anomaly region, and the attention distribution of the adjustment law feature package is fine-tuned. Subsequently, the shortest path along the geodesic is sought in the manifold space. If the two modes point in the same direction, the evolution along the manifold geodesic is used to find the energy endpoint. If the two modes conflict in their identification, the mode with the smoother curvature is automatically given higher dominance, and the curvature singularity caused by the Ricci flow smoothing transient biological noise is introduced. The calculation formula is as follows: (13); in, For the time-evolving manifold metric tensor; Let Ricci curvature tensor be the constant. Ricci flow enables isotropic heat conduction on the multimodal feature fusion surface, eliminating sudden interference; After each round of excitation and evolution, the similarity matching degree between the evidence is calculated. If the similarity matching degree is greater than the threshold, excitation stops. If the two pieces of evidence converge to the same category, that category is directly output. If the convergence categories differ but the trends are close to the same category, the curvature-guided weighted probability sum of the image and impedance is calculated. The final decision function is: (14); in, and These represent candidate tumor categories with influence and impedance, respectively. The posterior probability distribution under a single modality, which is derived from the global rhythm stream output in step S24. The core impedance characteristics of the output in step S33 The original confidence of a single-modal network for a specific class is obtained independently after normalization by the Softmax function of each sub-network. Let Riemann scalar curvature response function be defined in the constructed high-dimensional evidence manifold space. In the middle, local curvature The uncertainty and information entropy of the feature distribution in this region are characterized. If the feature points are mapped to folded regions with high curvature, it indicates that the local data of this mode has high noise, blurred boundaries, or severe microscopic heterogeneity perturbation; if they are mapped to regions with gentle curvature, it indicates that the evidence has stability and certainty. When visual imaging is disturbed, causing image curvature During a surge, weaken The dominant role is shifted smoothly to impedance evidence with lower curvature. ;vice versa; Finally, calculate all candidate categories. The joint probability weighted sum under curvature weight modulation, and through Operator optimization outputs the class with the highest global joint confidence as the final tumor identification result. .

[0009] To achieve the above method, the present invention also provides a multimodal tumor recognition system based on Riemannian manifold fusion. This system specifically includes: a multimodal data acquisition module, an image rhythm recognition module, an impedance particle flow analysis module, and a Riemannian manifold fusion arbitration module, wherein: The multimodal data acquisition module is used to acquire multimodal raw data, including spatiotemporally synchronized image datasets and impedance datasets. This module generates frequency-synchronized microsecond-level TTL trigger pulses through the FPGA main control unit to drive the medical imaging sensor and high-density electrode array respectively. It eliminates multimodal spatiotemporal deviation errors caused by system time axis jitter based on a hardware-level physical clock alignment model. At the same time, this module introduces a coaxial mapping mechanism supplemented by homography matrix to align the coordinate system, projecting and reconstructing discrete two-dimensional structural images and three-dimensional volume data onto a unified pixel coordinate system, and anchoring the impedance physical coordinates to the image pixel grid system, thereby generating a standardized three-dimensional image dataset and a standard spatiotemporally synchronized impedance dataset. The image rhythm recognition module is used to complete image recognition based on dynamic rhythmic feature allocation and output the image recognition results. Internally, the module first calculates the image grayscale gradient and divides the image into edge abrupt change regions and texture smooth regions. Then, it runs a fractional-order KAN network dynamic model and uses a calculus operator and a nonlinear smooth activation function to represent the long-term memory of the tumor edge invasion history to calculate the pixel-level adaptive rhythm coefficient matrix, and then maps it to generate a scan parameter configuration vector containing the number of scans and the switching frequency. Based on the above parameters, the module generates a dynamic perspective feature data stream, extracts multi-scale rhythm feature packages and fuses them into a global rhythm stream. Finally, it performs stability verification under various rhythmic pressure perturbations simulating tumor stroma pressure and outputs the image recognition results. The impedance particle flow analysis module is used to deconstruct and extract impedance based on fluid physics mechanisms to complete impedance identification and output the impedance identification results. This module performs cross-dimensional data physics mapping, converting the spatial location and time of the preprocessed impedance time series into virtual fluid particle coordinates and the impedance change rate into virtual particle attribute values. Then, using the fluid physics convection and diffusion equations containing divergence and Laplace operators, the impedance dynamics are transformed into a continuous three-dimensional impedance diffusion band tensor. This module alternately applies aggregation rhythms and discrete rhythms to the diffusion band to form an oscillating rhythm, and sequentially analyzes and extracts macroscopic structural layer features, microscopic texture layer features, and micro-variable sensing layer features that capture microvolt-level continuous impedance fluctuation signals. These features are then aggregated into impedance behavior units according to preset weights, and the impedance identification results are output after clustering and dynamic consistency verification. The Riemannian manifold fusion arbitration module is used to fuse the image recognition results and impedance recognition results based on manifold space mapping and curvature guidance mechanism to output the final tumor recognition result. This module uses the bimodal features and their preliminary recognition results as independent evidence sources, and maps them to a locally metrically flat high-dimensional evidence Riemannian manifold space through nonlinear homeomorphic mapping kernel function and direct sum operation. During the geodesic evolution process, this module introduces Ricci flow to perform isotropic heat conduction, smoothing the curvature singularities caused by transient biological noise. When the two modes have recognition conflicts, it performs conflict arbitration interaction excitation based on curvature guidance. If visual imaging is disturbed and causes a surge in local curvature of the image, this module automatically weakens the posterior probability dominance of the image evidence and smoothly shifts the decision focus to the impedance evidence with lower curvature. Finally, it calculates the joint probability weighted sum of all candidate categories under curvature weight modulation, and optimizes the output of the category with the highest global joint confidence as the final tumor recognition result.

[0010] This invention provides a multimodal tumor recognition method and system based on Riemannian manifold fusion. It integrates multi-level penetration imaging units and dynamic impedance sensing units to synchronously acquire data, effectively eliminating multimodal spatiotemporal deviation errors caused by system time axis jitter using an FPGA-level physical clock alignment model. A coaxial mapping mechanism using homography matrices is employed to project and reconstruct discrete two-dimensional structural images and three-dimensional volume data onto a unified pixel coordinate system, achieving spatial overlap between physical electrical features and visual anatomical features, thus solving the spatial misalignment problem caused by physical parallax of multi-source sensors. Based on a dynamic rhythmic feature allocation image recognition process, a fractional-order KAN network dynamics model is used to accurately capture non-local features of tumor growth, generating a multi-view dynamic feature data stream through adaptive rhythmic allocation. Simultaneously, rhythmic pressure perturbations simulating changes in the visual feature response to tumor stroma pressure are introduced for stability verification, improving the robustness and reliability of image feature extraction under complex physiological conditions. An innovative impedance analysis mechanism uses fluid physics convection and diffusion equations to convert impedance time sequence... The system transforms cross-dimensional structures into a three-dimensional virtual fluid particle density field, generating an impedance diffusion band that continuously covers the spatiotemporal domain. Combining alternating aggregated and discrete oscillatory rhythms, it achieves multi-level analysis of the macroscopic structural layer, micro-texture layer, and micro-variation sensing layer, capturing continuous impedance fluctuation signals at the microvolt level and overcoming the spatial resolution limitations of traditional single-dimensional impedance detection. In the fusion arbitration stage, dual-modal features and their preliminary identification results are mapped to a high-dimensional evidence Riemannian manifold space. Ricci flow is introduced to perform isotropic heat conduction, smoothing curvature singularities caused by transient biological noise and eliminating sudden interference. Simultaneously, an innovative curvature-guided conflict arbitration interaction excitation mechanism automatically weakens the posterior probability dominance of image evidence when two independent evidence sources conflict or when visual imaging is interfered with, causing a surge in local image curvature. It smoothly shifts the decision focus to impedance evidence with lower curvature, ultimately outputting the category with the highest global joint confidence. This significantly improves the system's dynamic fault tolerance in dealing with complex clinical interference and the accuracy of multimodal deep collaborative diagnosis. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a multimodal tumor identification method based on Riemannian manifold fusion according to the present invention; Figure 2 This is a block diagram of the module structure of a multimodal tumor recognition system based on Riemannian manifold fusion according to the present invention; Detailed Implementation Example 1 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. In specific engineering practice, this embodiment takes the combined screening and clinical diagnosis of early breast cancer using ultrasound imaging and electrical impedance tomography (EIT) as an application scenario, and provides an engineering-specific and parameter-specific description of the multimodal tumor identification method and system based on Riemannian manifold fusion proposed in this invention. This invention aims to solve the problems of information offset and anti-interference in the process of acquisition, feature extraction, and fusion decision-making of complex heterogeneous physiological data by establishing a rigorous mathematical logic and physical space mapping model.

[0012] like Figure 1-2 As shown, a multimodal tumor identification method and system based on Riemannian manifold fusion is presented, wherein: In the preferred embodiment, the Riemannian manifold fusion-based multimodal tumor recognition system is a physical hardware and software joint platform for high-precision identification of complex and heterogeneous breast tumors. Specifically, it includes a multimodal data acquisition module, an image rhythm recognition module, an impedance particle flow analysis module, and a Riemannian manifold fusion arbitration module. At the hardware architecture configuration level, the underlying timing scheduling center of the multimodal data acquisition module uses a Xilinx Zynq UltraScale+ MPSoC series field-programmable gate array with nanosecond-level instruction cycles as the data processing center to meet the high-throughput data synchronous acquisition requirements of multiple concurrent sensors. The medical image sensing unit uses a wideband linear array piezoelectric ultrasound probe with a central operating frequency range of 7.5MHz to 14MHz to acquire soft tissue morphological echo signals with high spatial resolution. The impedance sensing array is configured in a ring or matrix arrangement. Channel flexibility The high-density surface microelectrode array has a front-end signal conditioning circuit composed of an analog front-end (AFE) chip with a high common-mode rejection ratio, which is responsible for the constant current injection of microampere-level alternating current and the synchronous measurement of microvolt-level boundary voltage. The upper-level computing load of the system, including the image rhythm recognition module, the impedance particle flow analysis module, and the Riemann manifold fusion arbitration module, is deployed in an edge computing workstation equipped with dual NVIDIA RTX6000 Ada Generation professional-grade GPUs. Data transmission between physical devices relies on the PCIe 4.0 high-speed serial computer expansion bus, establishing a complete closed-loop data flow path from the acquisition of underlying physical signals to the operation of high-level geometric topology tensors.

[0013] In the preferred scheme, in step S1: the multimodal data acquisition and preprocessing stage obtains the original image dataset and impedance dataset with strict spatiotemporal synchronization characteristics by establishing a low-level physical clock alignment mechanism and a three-dimensional spatial coordinate system mapping. First, a multimodal acquisition device integrating multi-level penetration imaging units and dynamic impedance sensing units performs parallel scanning of the suspected lesion target area. In the clock synchronization mechanism, a high-precision digital phase-locked loop inside the FPGA main control unit generates a reference clock signal, which in turn generates frequency-synchronized microsecond-level TTL trigger pulses to synchronously drive the multi-frequency current source injection channels of the ultrasound imaging sensor's transmitting and receiving arrays and the high-density electrode array. The system then acquires three-dimensional ultrasound volume data covering the superficial, middle, and deep layers of breast tissue. Simultaneously, the impedance module... to Within the characteristic frequency band, a logarithmic law is set. A discrete detection frequency, with an applied amplitude of to Furthermore, by using gradient-type weak alternating current stimulation that meets human safety standards, the impedance magnitude and phase changes of each grid node divided according to a predetermined physical step size are recorded over time to obtain the original image dataset. Compared with the original impedance dataset To eliminate sampling time difference at the hardware source, the system operates as defined by the FPGA-level physical clock alignment model in formula (1). In the specific parameter calibration, the microsecond-level TTL pulse trigger signal vector... Synchronous sampling period Set as Microseconds, covering rapidly varying impedance signals caused by physiological metabolism, rectangular pulse window function pulse width Set as The microsecond time ensures that the time window for the ultrasonic longitudinal wave to complete the round-trip acoustic propagation in the tissue is fully covered without any truncation effect; the hardware clock locking mechanism reduces the multimodal spatiotemporal deviation error caused by crystal oscillator drift within the system. After the above synchronous acquisition steps are completed, the system sequentially performs denoising, enhancement, and registration extraction operations; for the inherent speckle noise in ultrasound imaging, the system uses a dynamic threshold denoising algorithm combined with nonlocal mean filtering for spatial smoothing, and its joint calculation formula (2) is used. In specific implementation, the spatial image data after denoising is enhanced. The calculation process relies on Fourier transform. and its inverse transformation Frequency domain modulation, nonlocal mean filter weight kernel The search window size is set to Pixels, similarity block size set to Pixel; Cutoff frequency dynamically set based on region Values High-pass gain order Set as Adaptive correction function The parameters are affected by the Laplace operator. Extracted spatial tissue edge gradient control aims to smooth uniformly distributed fat areas while sharpening the edge contours of microcalcifications. Subsequently, the system introduces a coaxial mapping mechanism supplemented by a homography matrix to align the coordinate systems of the multiple sensors. The geometric mapping relationship is determined by formula (3), in which the image pixel system calibration coordinates are used. Physical coordinates of multimodal sensing system The transformation between them is subject to the cross-modal space alignment homography transformation matrix. Constraints: Before leaving the factory or during periodic calibration, the system is calibrated using a standard physical phantom containing known acoustic scatterers and conductivity anomalies, and the specific coefficient values ​​of the matrix are obtained by solving for them, including the depth scale scaling factor. Based on the degree of tissue deformation under pressure to Adaptive adjustment between; obtain Subsequently, the system performs a spatial resampling operation based on cubic spline interpolation to eliminate physical parallax caused by the spatial position difference between the ultrasonic probe and the electrode array. The discretely acquired two-dimensional structural images and three-dimensional volume data are then projected and reconstructed onto a continuous voxel matrix with isotropic resolution, outputting a standardized three-dimensional image dataset. ; In impedance preprocessing, the system uses a bandpass finite impulse response filter to remove... Power frequency interference and high-frequency thermal noise are eliminated by trend fitting correction mechanism to remove the slow baseline drift caused by the patient's respiratory rhythm or heart position micro-movement; the process of impedance value normalization, trend fitting and elimination of baseline drift is realized by formula (4). In the specific execution of the algorithm, the system calculates the standard impedance time series after eliminating baseline drift and normalization. When using orthogonal... Legendre polynomial Approaching the low-frequency profile, because the normal human respiratory rate is distributed in to Between, the highest order of fitting Set as This setting, after extensive verification, effectively isolates periodic physiological displacement interference without losing low-frequency components related to tumor metabolism; the formula includes the average impedance within a single acquisition cycle. and a small constant set to prevent the denominator from being zero. This is used to ensure the stability of numerical calculations; finally, the system is based on the cross-modal space alignment matrix obtained above. This method maps and anchors the physical coordinates of discrete impedance sensing grid nodes to a standardized image voxel grid system, achieving precise spatial overlap between physical electrical conduction features and visual anatomical features. The output is a complete standard spatiotemporal synchronous impedance dataset that meets the dimensionality consistency requirements of subsequent tensor operations. .

[0014] In the preferred embodiment, in step S2: the image feature recognition process abandons the traditional global static convolution mode and instead constructs a network structure based on dynamic rhythmic feature allocation and nonlinear dynamic feature response to improve the sensitivity to the boundaries of local heterogeneity of tissues. The system first performs gradient calculation based on 3D Gaussian smoothing to obtain the gradient magnitude of image grayscale changes, and then divides the image into regions with distinct feature attributes based on this. In this operation, the system traverses the 3D matrix using a defined gradient operator, setting a normalized upper limit threshold. The lower threshold is The grayscale gradient is greater than The region typically includes the tumor capsule, microcalcifications, or vascular sections, and is systematically and clearly defined as a marginal mutation region; the grayscale gradient is less than [missing information]. The regions are typically composed of homogeneous fat or normal glands and are classified as areas with smooth textures. For these two types of regions, the system sets differentiated observation rhythms by running an adaptive rhythm allocation function that adapts to a multi-peak distribution. For edge mutation regions containing rich pathological structural information, the system assigns a range of values. to The low rhythm coefficient instructs subsequent network structures to perform high-sampling-rate, high-frequency dense scanning extraction within a unit space of that segment; while for regions with smooth textures, a range of values ​​is assigned. to The high rhythm coefficient reduces redundant consumption of computing resources. At the same time, if the system detects atypical texture abrupt changes in the originally planned smooth area during the iterative scanning process, it triggers a spatial resegmentation operation and merges adjacent continuous smooth sub-regions into a matrix to dynamically generate rhythmic segment groups with adaptive boundaries. In the calculation of specific rhythm parameter allocation for each segment, the system loads a pre-trained fractional-order KAN network dynamic model to deeply characterize the nonlocal spatial diffusion characteristics of tumor growth. Its core mathematical evolution function is formula (5), and the calculation process involves Caputo-type fractional-order calculus operators. Its order Set as Compared to traditional integer derivatives, this fractional derivative allows the model to incorporate a long-term historical memory term that decays over a longer spatial distance, more accurately reflecting the mathematical law of nonlinear invasive growth of tumor cells along the stroma; the computational objective is to obtain a pixel-level adaptive rhythm coefficient matrix. The model internally sets a constant for the mapping of hidden layer dimensions. for Number of input image feature channels Based on multi-scale input settings The parameterized B-spline activation function used in the formula Select B-spline bases possess excellent local control and smoothness, and when combined with outer nonlinear aggregation functions... And the introduction of minimum variance to prevent the model from getting trapped in local optima Nonlinear oscillator with stochastic dynamic perturbation term Together, they complete the calculation of highly nonlinear characteristic responses; Obtain the pixel-level adaptive rhythm coefficient matrix Then, the system performs a spatial surface integral on the rhythm coefficient matrix elements within each dynamically segmented rhythmic region, using the following formula: ; Normalization yields the first Regional rhythm expectation value of each segment Next, the system uses a pre-constructed rhythm-action mapping function, i.e., formula (6), to calculate the expected value of this scalar. Convert to a structured multidimensional parameter configuration vector The configuration vector specifically includes the total number of observation scans allocated to this segment. Time interval for dynamic perspective switching and the physical transformation boundary limiting during the viewpoint generation process In parameter design, the hardware response scaling factor and They are respectively labeled as and (This needs to be fine-tuned based on the specific GPU computing power) so that the calculated expected number of scans is limited to... to Within the integer range of the order of 1; smoothing function Using standard logistic functions and edge detection constants Set as experience points The system's maximum physical transformation threshold Constrained within the maximum safe range allowed before any abrupt changes in the organization's topology; the system traverses and derives the configuration vectors of all segments, assembling them to form a global scanning strategy configuration set. ; Following the above configuration set, the system then generates multi-view fluctuation sequences for image samples and performs rhythmic synchronization adjustment; for each rhythmic segment entering the processing queue, the system limits its amplitude. Multiple affine and nonlinear transformations are applied sequentially within the inner layer, including but not limited to introducing variances not exceeding a certain value. The local grayscale Gaussian fluctuation distribution, the execution of spatial horizontal and vertical texture mirror flipping, and the application of random polar angles in to The system uses coordinate axis rotation and offset within a certain range, along with pixel-level contour scaling based on a mesh deformation algorithm, to artificially synthesize a variety of wave sequence sample groups with different viewpoints. The system controls the timing of viewpoint frame loading, ensuring that the dynamic switching frequency of the viewpoint remains synchronized with the previously calculated observation rhythm. For abrupt change segments determined to have a fast rhythm, the viewpoint data loading time interval is allocated as follows: / frames of brief intervals, while slow rhythm segments are allocated as follows The longer intervals between frames generate a dynamic viewpoint feature data stream that conforms to the clinical temporal saccade pattern. ; In the subsequent feature extraction and aggregation encoding stage, the system inputs the generated data stream into a deep residual network with pre-trained weighted reusable deep residual network, such as a custom variant with a structure similar to the first four main blocks of the ResNet-50 architecture, and outputs a high-dimensional perspective feature sequence. The system introduces a time-dimensional self-attention mechanism in this cascade to calculate the attention weight of each feature frame, so as to enhance the main rhythm frame features with the most significant pathological manifestations, while retaining the secondary rhythm frame features containing complementary contextual information. The specific derivation and calculation are realized by formula (7), where the time step is... The following attention distribution matrix The computation depends on the learnable query projection matrix. Bond projection matrix The dimensions of both are set to a feature scaling dimension factor of 64. Adjustments are made to prevent the dot product value from becoming too large, which could cause the Softmax function to enter the gradient saturation region; the system synchronously adjusts the receptive field scale of the feature sequence according to the obtained attention weight distribution. , and The two-dimensional parallel convolution aggregation operation is performed, and the operation equation is referred to formula (8); after the multi-scale feature tensor is spliced ​​in the channel dimension, the Hadamard product operation is performed. Multiplying by the weight matrix, the final fused output has Segmental multi-scale rhythmic feature bag tensor of high-dimensional semantic representation ; To achieve spatial and semantic coherence between segments, the system introduces a graph topology to establish rhythmic associations and performs global feature dimensionality reduction and rigorous physical stability verification logic. Through inter-layer information transfer learning in a graph convolutional network, the system obtains an inter-segment rhythmic association adjacency matrix describing the spatial proximity and feature similarity between different dynamically rhythmic segments. The locally extracted feature package Fusion into images containing complete semantic meaning Global rhythm flow The mathematical expression of its feature mapping network structure is shown in formula (9), where the global dimension reduction projection matrix is... and its bias terms Automatic optimization is achieved through backpropagation of the network, and LeakyReLU is used for nonlinear activation to prevent negative gradient vanishing; To address occasional misjudgments caused by image noise interference, the system enforces a rhythm stability check procedure before final output: the system constructs a series of rhythm pressure perturbation vectors that characterize the fluctuations in visual response caused by simulated tumor stroma pressure. , from Get the set total number of tests The system appends these multiple sets of independent and identically distributed perturbation vectors to the global rhythm stream and feeds them into a multilayer perceptron classifier. The image is then mapped to the posterior probability distribution of the candidate categories using the Softmax function, resulting in the final image recognition result. Formula (10) is used to apply the maximum likelihood criterion. Find out; the verification rules require that, in repeated operations... In each inference loop with independent perturbation parameters, the number of times the system's judgment results are consistent must be greater than or equal to [the number of times]. Only after this process can the verification be completed and the output be generated. If significant discrepancies prevent the threshold from being reached, the algorithm will automatically regress and output the label with the highest frequency for each category as the corrected version. Along with dimensional reduction to dimensional stable global rhythmic flow vector Submit them together to the downstream fusion module.

[0015] In the preferred embodiment, in step S3: for impedance time series data with multimodal data concentrated in the time domain and arranged in a one-dimensional array, the impedance particle flow analysis module of the present invention adopts an analysis approach based on the cross-fusion of computational fluid dynamics and micro-electromagnetism, avoiding the drawback of traditional time-domain series data that rely on networks for both long and short periods and are prone to losing information on minute spatial deformations. First, the system executes a cross-dimensional data-physical mapping process, converting the preprocessed and coordinate-anchored impedance time series into a set of flow particle points within a three-dimensional virtual fluid space; the mapping rule specifies the spatial Cartesian coordinates of the virtual particle. and time scale Associated with the spatiotemporal points acquired by the sensor, the fundamental property values ​​of the particles are quantitatively assigned as the first-order rate of change of impedance over time at that specific spatiotemporal point. Furthermore, constraints are constructed to determine the virtual particle distribution density within the local space. The impedance change is numerically proportional to the magnitude of the impedance change, and its physical mapping and the rigorous mathematical logic of field initialization are as follows: First, calculate the scalar field of the rate of change of impedance at each voxel node in three-dimensional space in the time dimension. The calculation formula is: ; in, The rate of change of impedance at the voxel level; The impedance sampling time resolution of the system hardware; Subsequently, based on the physical assumption of a positive correlation between impedance change amplitude and particle density, a nonlinear density mapping functional constrained by tissue fluid osmotic dynamics is constructed to initialize the initial state of the particle density field at the evolution initiation point. : ; in, This represents the initial spatial virtual particle density distribution field. The baseline particle density constants for healthy surface / middle / deep tissues set for the system; The limiting amplification coefficient characterizing the abnormal increase in local ion concentration caused by tumor cell metabolism; It is a hyperbolic tangent function, used to limit the divergence of density field singularities caused by extreme anomalous electrical noise points, and to ensure the stability of physical evolution; For adjusting hyperparameters for dynamic sensitivity; Next, in order to establish a continuous field that conforms to the smooth transition law of biological tissue conductivity among discrete data, the calculated data is... As initial boundary conditions, i.e., when the evolution time hour, After substituting the fluid convection and diffusion equations, i.e., formula (11), within the set virtual simulation time window... Within this framework, the implicit difference method with alternating directions is used to numerically solve the partial differential equations, yielding the results as the virtual dynamic time... Evolutionary transient density field set In the parameter configuration of this physical model, the percolation characteristics of charge carriers in porous biological tissue media are taken into account, and the diffusion velocity vector of biological tissue fluid on the flow field is considered. The modulus length is limited to to Flow within the low-speed range of pixels / frames; the anisotropic diffusion coefficient tensor that governs the intensity of microscopic Brownian motion concentration diffusion. The main diagonal element is labeled in to Within the range; in terms of differential operators, the divergence operator is adopted. The net outflow flux within this infinitesimal volume element is continuously solved using the Laplace operator. The system describes the spatial diffusion process driven by the concentration gradient; to extract the spatiotemporal dynamic characteristics covering the entire fluid evolution cycle, the system analyzes the solved transient density field. Performing multiple functional integrations along the virtual evolution time axis, the output is a high-dimensional impedance diffusion band feature tensor covering the spatiotemporal domain of interest. ; The system then executes a multi-level deep analytical procedure based on the generated impedance diffusion band tensor, sequentially applying specific external force rhythms; the system first applies a cohesive rhythm simulating cohesive centripetal force, and the calculation model defines the central gravitational force function as... The elastic coefficient Determined by the eigenvalue of tissue capacitance, the system forces free particles within the region to iteratively converge toward their local statistical mean point, thus highlighting macroscopic tissue plaques with highly consistent electrophysiological characteristics. Conversely, the system alternately applies discrete rhythms with randomly oriented forces, as a function... Random, dispersed perturbations are injected in the form of [a method to break up] false high-concentration clusters that may be caused by poor contact in local areas; the aggregation and dispersion processes alternate at a fixed periodic frequency to form an oscillating rhythm. Under this oscillating excitation state, the system's three-level concurrent feature analyzer analyzes the particle distribution field. Perform operations synchronously: For the macroscopic structural layer, the algorithm constructs a three-dimensional Hessian structure tensor of the local flow field and solves for its eigenvalues. Based on the proportional relationship between the eigenvalues, it extracts macroscopic structural layer features that describe the overall three-dimensional boundary electrical properties of the tumor lesion. For the micro-texture layer, the algorithm defines a sliding statistical window and calculates the local spatial Shannon information entropy distribution map of the particle evolution field to characterize the micro-texture layer features caused by the increased disorder of cell arrangement due to different cell proliferation activities within the tumor. Furthermore, for the micro-variation sensing layer, the system is configured with a narrowband weak signal digital filter based on the Lorentz resonance principle. This filter dynamically locks the resonant frequency band to the frequency range related to cancer cell membrane rupture and abnormal opening and closing of ion channels. It captures and extracts high-frequency fluctuation signals of microvolt-level continuous impedance caused by changes in the lesion's microenvironment with high fidelity, forming the micro-variation sensing layer characteristics. ; In the impedance aggregation and output stage, the system first runs a cubic spline spatial interpolation algorithm to smoothly reconstruct the discrete data array acquired and mapped by a finite number of physical electrodes into a high-density logical resolution voxel matrix that matches medical images. Subsequently, it performs a spatial tensor stitching operation on three layers of features. During this process, the system applies unequal weights to the three layers of features: given the guiding role of macroscopic contour features in spatial localization, the macroscopic structure layer features are given the largest normalized weight of 0.6, while the micro-texture layer providing detail and the easily disturbed micro-variation sensing layer are given relatively smaller and balanced weights of 0.2 respectively. This logical aggregation produces a data matrix with a physical scale accuracy of [missing information]. Impedance behavior units within a spatial voxel grid region; the system calculates the feature similarity coefficients between adjacent behavior units based on multidimensional vector inner products. For adjacent units with coefficient values ​​greater than a preset clustering merging threshold of 0.75, a density-based spatial clustering algorithm with noise is applied to group them into clusters. The central time-domain statistics and frequency-domain amplitude and phase response parameters of each effective pathological cluster are extracted and concatenated to form a structured 1024-dimensional impedance core feature tensor. ; Finally, this tensor is fed into a fully connected classification layer of a feedforward neural network to obtain preliminary pathological classification results, and the rate of change of classification features of impedance data detected by the system for several consecutive frames is required to be lower than a preset extremely low safety threshold. Subsequently, the output impedance identification results were stabilized through dynamic consistency verification. .

[0016] In the preferred scheme, in step S4: the final fusion of multimodal recognition results does not adopt a simple hard voting or fixed weight splicing method, but is instead based on the arbitration mechanism guided by Riemannian manifold space mapping and Ricci flow curvature. The system first obtains the image recognition results. and its characteristic tensor , and impedance identification results and its characteristic tensor At the mathematical level, they are independently considered as two mutually orthogonal probabilistic evidence sources. To achieve lossless expression of evidence in non-Euclidean space, the system defines a quantitative dimension model of evidence events with four attributes: the initial classification confidence of network inference is defined as class support, the degree of rejection of a certain class by different modalities is defined as class rejection, and the fluctuation sensitivity and responsiveness to specific rhythmic perturbation pressure tested in the previous steps are used. According to the mapping topology transformation rules explained in formula (12), the system constructs a non-Euclidean high-dimensional evidence Riemannian manifold space with flat local metrics. In this process, the system specifies the kernel functions from the image domain to the manifold evidence space. and kernel function mapped from impedance domain By performing strict spatial tensor direct sum operations under specific Lie group algebraic structures The global rhythmic flow of the image with dimensions located in different feature spaces With impedance core characteristics Unified mappings increase dimensionality and coexist in evidence of Riemannian manifold space with a constant dimensionality of 256. On the point cloud manifold surface, the core feature of this manifold space is that the points Any tiny geometric deformation of the tangent space or the evolution of the geodesic path within its neighborhood no longer follows the Euclidean distance, but is instead governed by its second-order covariant symmetric Riemannian metric tensor. The strict computational constraints of the field result in excellent mathematical tolerance for heterogeneous data alignment. After the manifold mapping is completed, the system enters the conflict arbitration and interaction stimulation process based on the attention allocation mechanism and curvature guidance to determine the final output result. During the initial arbitration, the system performs cross-modal adaptive stimulation probing: based on the aforementioned quantitative indicators of each dimension, the system calculates the unidirectional stimulation weight scalar of image evidence to impedance evidence, and adjusts the previous clustering threshold of the impedance behavior unit accordingly, triggering the network to recalculate the class tendency probability to verify the robustness of the result. At the same time, based on the three-dimensional coordinate space position of the anomaly center with extremely low impedance characteristics located by impedance detection, the system instructs the image feature extraction network to increase the filter weight coefficients in the feature extraction process in the target region locally, and refines and fine-tunes the local attention distribution of the previously proposed segment rhythm feature package. In the subsequent optimization solution path, the system uses the variational method to find the shortest metric path of the geodesic connecting the two modal characterization feature points on the manifold space surface; if the multimodal judgment conclusions are highly consistent at this time, that is, the geodesic length is extremely short and the gradient convergence is consistent, the system directly evolves along the manifold geodesic without hindrance, and outputs the result at the energy minimum endpoint; when significant two-modal identification and classification conflicts occur, such as ultrasound morphology indicating benign cysts, but the impedance mode strongly supports the conclusion of malignant necrosis due to the detection of a sharp increase in conductivity caused by internal liquefaction, the system gives the dominant power through an automated mechanism. The system innovatively introduces the pure geometric equation Ricci flow mechanism to smooth out the manifold space extreme value distortion caused by patient body movement, heartbeat or external environmental electromagnetic transient biological noise. Its core mathematical evolution is formula (13); during the calculation execution, the time variable Iteratively evolving manifold metric tensor The partial derivatives are determined by the Ricci curvature tensor obtained from the trace of the tangential curvature of space. Negative regulation; this partial differential evolution process continuously simulates isotropic thermodynamic heat conduction effects on the high-dimensional feature fusion interface of multimodal modes, smoothing out the protrusions formed on the feature surface due to sudden physical disturbances; After undergoing excitation and manifold evolution adjustments within the set maximum iteration limit for each round, the system calculates the cosine similarity matching coefficient between two orthogonal evidence point vectors in the tangent space. If this similarity matching coefficient exceeds the system's pre-defined safe stopping threshold of 0.9, the subsequent redundant excitation calculation loop is immediately terminated. If the medical diagnosis categories pointed to by the two convergence paths are completely consistent, the system will directly output this stable medical category determination; however, if the specific pathological categories converged by the two verification paths have divergence errors, but their statistical confidence distribution evolution trends are gradually approaching the same dominant category, the system's weighted determination function logic system defined by formula (14) will be used; in this determination function, the normalized output representing the candidate specific tumor category is calculated independently through the Softmax activation function of each modal subnetwork. The posterior probability distribution of the single-item probability vector and Each of these will be multiplied by a dynamic scalar weighting coefficient calculated based on the Riemann scalar curvature response function. and In the non-Euclidean evidence manifold space constructed based on topology Scalar local curvature at a specific point within The direct inverse proportional characterization represents the confidence level of the data acquired by the corresponding modal sensor at this physical location and its inherent implicit information entropy. If the feature point is mapped to a Gaussian curvature space region with extremely high curvature due to the presence of a large amount of unfiltered noise or the severe perturbation caused by the microscopic heterogeneity of malignant tumors, the system determines that the reliability of the single-mode data is on the verge of collapse; conversely, if it is located in a region with a very smooth tangent plane and a small positive definite curvature gradient, it confirms that the evidence source acquired by the mode has high purity, certainty, and physical stability. Therefore, through this adaptive geometric evaluation mechanism, when the system is affected by adverse visual imaging interference conditions such as inadequate application of probe coupling agent, the resulting local information chaos in the image will lead to a decrease in the corresponding geometric curvature of the image. The values ​​surged or even diverged; according to the formula, the system immediately exhibited an exponential automatic decay that weakened the representation of image confidence. The proportion of a variable in the total weighted sum of all fused decisions, i.e., the dominance, smoothly and completely transfers the decision focus and reliance on trust to the resistance evidence, which has smaller curvature and is not similarly contaminated. Branches ensure the system continues to run and outputs accurate results; the processing flow under reverse physical interference follows the same mirror inversion replacement rule; In the final stage of the computation cycle, the system processor obtains all candidate medical test categories through parallel computing. The joint conditional probability weighted sum matrix column vector is subjected to strict bidirectional modulation by manifold curvature weights, and the extreme value is obtained by calling the extremum. The optimization operator selects and compares specific item categories with the highest global statistical joint confidence parameters within the matrix. This category is then locked and output as the system's final tumor identification result in a hybrid format of numerical and textual reports. .

[0017] In order to provide clinicians with diagnostic evidence that complies with medical ethics and has traceability and interpretation, the multimodal tumor identification system, in addition to providing qualitative conclusions, is also equipped with a dedicated result visualization mapping interface software module. When rendering and outputting results at the clinical diagnostic terminal, this interface simultaneously parses the impedance and the optimal response feature tensor retained in the image modality. The system reconstructs the three-dimensional impedance fluid diffusion zone data into an isosurface cloud map with color gradient and displays it on the grayscale ultrasound two-dimensional tomographic image. It also displays the dynamic fusion weight allocation details calculated by formula (14), thereby retrospectively reflecting the degree of dependence of the algorithm decision on different pathological features, meeting the clinical diagnostic requirement for the traceability of the algorithm decision path, and is different from the single scalar classification probability distribution result output by the conventional pure black box fully connected neural network scheme that does not provide the physical meaning of intermediate steps.

[0018] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A multimodal tumor identification method based on Riemannian manifold fusion, characterized in that, Includes the following steps: S1. Acquire multimodal raw data, including spatiotemporally synchronized image datasets and impedance datasets; S2. Image recognition is completed based on dynamic rhythmic feature allocation, and the image recognition result is output. S3. Based on fluid physics mechanisms, impedance identification is deconstructed and extracted to complete impedance identification, and the impedance identification results are output. S4. Based on the manifold space mapping and curvature guidance mechanism, the image recognition results and impedance recognition results are fused to output the final tumor recognition result.

2. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 1, characterized in that, Step S1 specifically includes the following steps: Two-dimensional structural images, three-dimensional volume data, and impedance change sequences at different depths were acquired using multimodal acquisition equipment. The spatiotemporal bias is eliminated by using a hardware-level physical clock alignment model, and discrete image data is projected and reconstructed to a unified pixel coordinate system using a cross-modal space alignment matrix, generating a standardized 3D image dataset and a standard spatiotemporal synchronization impedance dataset.

3. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 1, characterized in that, Step S2 specifically includes the following steps: Calculate the gradient of gray-level changes in the image and divide the image into regions with abrupt edge changes and regions with smooth textures; A nonlinear dynamic model is constructed. By introducing a calculus operator and a nonlinear smooth activation function to characterize the long-term memory of tumor edge invasion history, an adaptive rhythm coefficient matrix is ​​calculated and mapped to generate a scan parameter configuration vector containing the number of scans and the switching frequency. Based on the scanning parameter configuration vector, a dynamic viewpoint feature data stream is generated, multi-scale rhythm feature packages are extracted, inter-segment rhythm correlations are established, and the image recognition results are output with dimensionality reduction.

4. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 3, characterized in that, The process of dividing the image into regions with abrupt edge changes and regions with smooth textures, and generating a dynamic viewpoint feature data stream, specifically includes: Regions with grayscale gradients greater than the upper limit threshold are classified as edge abrupt regions and assigned low rhythm coefficients, while regions with grayscale gradients less than the lower limit threshold are classified as texture smooth regions and assigned high rhythm coefficients. During the scanning process, regions with abrupt texture changes are re-segmented and continuous smooth regions are merged to generate dynamic rhythmic segment groups. Local grayscale fluctuations, texture flips, orientation shifts, and pixel-level contour scaling are applied to each of the dynamically rhythmic segments to generate fluctuation sequences from multiple perspectives. The perspective switching frequency is matched with the observation rhythm of the dynamic rhythmic segment by a rhythm synchronization adjustment mechanism. Short perspective switching intervals are set for fast rhythmic segments and long perspective switching intervals are set for slow rhythmic segments, forming a dynamic perspective feature data stream.

5. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 3, characterized in that, The process of extracting multi-scale rhythmic feature packages, establishing inter-segment rhythmic associations, and reducing the dimensionality of the output image recognition results specifically includes: Calculate the attention weights of the feature sequence in the time dimension, and strengthen the main rhythm features while preserving the secondary rhythm features through a learnable query and key mapping weight matrix; Two-dimensional convolution operations with receptive fields of different scales are used to perform multi-scale feature aggregation on feature sequences and splice them to form a multi-scale rhythmic feature package; The inter-segment rhythm association adjacency matrix is ​​obtained by learning through graph convolutional networks, and multi-scale rhythm feature packets are fused into a global rhythm stream. Feature extraction was repeatedly performed under various rhythmic pressure perturbations to implement rhythmic stability verification. The rhythmic pressure perturbations were used to simulate the visual feature response changes of tumor stroma pressure, and the final image recognition result was output when the verification results reached a consistency threshold.

6. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 1, characterized in that, Step S3 includes a data transformation step: The spatial location and time of the preprocessed impedance time series are mapped to virtual particle coordinates, and the impedance change rate is mapped to virtual particle attribute values. By utilizing the fluid convection and diffusion equations, which include a divergence operator characterizing the net outflow of local convection and a Laplace operator characterizing the diffusion of concentration gradient, the impedance dynamics are transformed into a particle density field, generating a continuous three-dimensional impedance diffusion band tensor.

7. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 6, characterized in that, It also includes the feature parsing step: An oscillating rhythm is formed by alternately applying an aggregation rhythm for iterative aggregation and a discrete rhythm for breaking up highly consistent regions to the three-dimensional impedance diffusion band tensor. Extract macroscopic structural layer features, microscopic texture layer features, and micro-variable sensing layer features to capture microvolt-level continuous impedance fluctuation signals; The impedance behavior unit is generated by aggregating three-level features according to preset weights. After clustering and dynamic consistency verification, the impedance identification result is output.

8. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 1, characterized in that, Step S4 includes a topology mapping step: The image recognition results, impedance recognition results and their corresponding high-dimensional features are regarded as independent sources of evidence. Through the nonlinear homeomorphism mapping kernel function and the direct sum operation, a locally metrically flat non-Euclidean Riemannian manifold space is constructed. An evolutionary equation incorporating the Ricci curvature tensor and the metric tensor is introduced to perform isotropic heat conduction, smoothing curvature singularities caused by transient biological noise.

9. The multimodal tumor identification method based on Riemannian manifold fusion according to claim 8, characterized in that, It also includes conflict arbitration procedures: In the Riemannian manifold space, the shortest path of geodesics is sought. If two independent sources of evidence conflict in their identification, the mode with the gentler curvature is given higher dominance. The joint probability weighted sum of all candidate categories under curvature weight modulation is calculated. If visual imaging is disturbed, causing a surge in local curvature of the image, the dominance of the posterior probability of the image evidence is weakened, and the decision focus is shifted to impedance evidence with lower curvature. The category with the highest global joint confidence is output as the final tumor identification result.

10. A multimodal tumor identification system, characterized in that, include: The multimodal data acquisition module is used to acquire raw multimodal data, including spatiotemporally synchronized image datasets and impedance datasets; The image rhythm recognition module is used to complete image recognition based on dynamic rhythmic feature allocation and output the image recognition results. The impedance particle flow analysis module is used to deconstruct and extract impedance identification based on fluid physics mechanisms and output the impedance identification results. The Riemannian manifold fusion arbitration module is used to fuse image recognition results and impedance recognition results based on manifold space mapping and curvature guidance mechanism, and output the final tumor recognition result.

Images