Search

CN-122023988-A - Tumor identification method based on multi-mode information coding

CN122023988ACN 122023988 ACN122023988 ACN 122023988ACN-122023988-A

Abstract

The invention provides a tumor identification method based on multi-modal information coding, which aims to solve the problems of Ji Queshi, insufficient mining of time sequence features and weak semantic relevance in the prior art, and comprises the steps of firstly synchronously collecting structural image data and electrical impedance time sequence data of a tissue to be detected, realizing time-space synchronization of the dual-modal data through time stamp alignment and anatomical marker mapping, reconstructing a dual-path self-adaptive enhancement path, respectively carrying out targeted preprocessing on the two types of data, distributing fusion weights based on data quality scores, then extracting multi-scale space features and multi-directional time sequence cyclic coding features through a dual-path modal adaptive coding architecture, finally integrating the dual-modal features based on a cross-modal attention fusion mechanism, sequentially completing tumor existence judgment, benign and malignant nature classification and malignant degree classification by means of a multi-stage progressive classifier, and deeply fusing structural morphology and electrophysiological metabolic information, improving the accuracy and robustness of tumor identification, and meeting the requirements of clinical rapid screening and accurate diagnosis.

Inventors

  • YAN ZIBAO
  • HOU JINXUAN
  • XIAO JIABO

Assignees

  • 武汉中针智诊科技有限公司

Dates

Publication Date
20260512
Application Date
20260130

Claims (10)

  1. 1. A tumor identification method based on multi-mode information coding is characterized by comprising the following steps: s1, synchronously acquiring structural image data and electrical impedance time sequence data of tissue to be detected, and realizing space-time synchronization of bimodal data through time stamp alignment and space marker mapping; S2, constructing a dual-path self-adaptive enhancement path, respectively performing blocking dynamic enhancement processing on the structural image data, performing filtering and conductivity conversion on the electrical impedance time sequence data, and distributing fusion weights based on the data quality scores; S3, designing a dual-path mode adaptive coding framework, extracting multi-scale space features of the enhanced image data, performing multi-directional time sequence cyclic coding on the preprocessed electrical impedance time sequence data, and outputting mode adaptive features; s4, integrating the dual-path coding features based on a cross-modal attention fusion mechanism, and sequentially executing tumor existence judgment, benign and malignant nature assessment and malignancy degree assessment through a multi-stage progressive classifier.
  2. 2. The tumor recognition method based on multi-modal information encoding according to claim 1, wherein step S1 specifically comprises: recording reference time stamps by means of a hardware synchronization signal And combining the time offset of each frame sampling point Calculating absolute time Realizing time alignment; by labeling anatomical markers on the tissue of the subject, spatial synchronization is achieved by locking the correspondence of the coordinates of the markers in the image to physical space.
  3. 3. The method for identifying tumor based on multi-modal information encoding according to claim 1, wherein the processing of the structural image data in step S2 includes: Dividing the image into non-overlapping blocks, calculating contrast measure Noise level Edge strength ; Based on characteristic index Learnable weight parameters Bias term Respectively calculating the weight of the contrast self-adaptive histogram equalization algorithm Weighting of bilateral filtering denoising algorithm Steering filter sharpening algorithm weights And performs weighted fusion.
  4. 4. The tumor recognition method based on multi-modal information encoding according to claim 1, wherein the processing of the electrical impedance time series data in step S2 includes: by setting window step length Filtering the original signal to obtain an output signal; by constructing a virtual grid, based on boundary voltage measurements And input current Iterative solution of conductivity distribution 。
  5. 5. The method for identifying a tumor based on multi-modal information encoding according to claim 1, wherein step S2 further comprises: Calculating structural image quality scores And electrical impedance data quality scoring ; Using quality scoring And (3) with And normalizing to determine the fusion weight of each mode characteristic.
  6. 6. The tumor recognition method based on multi-modal information encoding according to claim 1, wherein the multi-scale spatial feature extraction in step S3 includes: parallel extraction of fine granularity, medium and coarse granularity feature maps using convolution kernels with different receptive fields ; Adaptive scaling weights ; Dimension reduction of spatial dimensions by global averaging pooling layer, followed by execution of Normalizing to obtain structural image coding characteristics 。
  7. 7. The tumor recognition method based on multi-modal information encoding according to claim 1, wherein the multi-directional time-sequence cyclic encoding in step S3 includes: Constructing a forward coding unit to record a polarization dynamic balance process; Constructing a reverse coding unit capture end relaxation property; constructing an interval coding unit to filter physiological pulsation; the construction segment coding unit explicitly extracts phase characteristics of the initial response segment, the platform stabilizing segment and the attenuation release segment.
  8. 8. The method for identifying a tumor based on multi-modal information encoding according to claim 7, wherein the multi-directional time-series cyclic encoding further comprises: Computing hidden state vectors for multi-directional coding features Importance scoring of (2) ; Calculating the weight coefficient of each step And generates a comprehensive feature vector ; Transforming through a fully connected neural network, mapping the transformed fully connected neural network into a feature vector with fixed dimension, and screening feature information with more discrimination to obtain a final feature vector after electrical impedance coding 。
  9. 9. The tumor recognition method based on multi-modal information encoding according to claim 1, wherein step S4 specifically comprises: by calculating the overall distribution statistical information pairs of each characteristic vector And (3) with Performing a distribution calibration; metabolic abnormalities by imaging features Index of impedance feature metabolism Impedance characteristic value Respectively carrying out importance evaluation on the two standardized modal feature vectors, and calculating the cross-modal attention weight And (3) with ; Linear weighted summation is carried out on the standardized feature vectors of the two modes to obtain initial fusion features And performing two-stage feature optimization to obtain final fusion features 。
  10. 10. The method for identifying a tumor based on multi-modal information encoding according to claim 1, wherein step S4 further comprises: Judging the existence probability of tumor by adopting linear full-connection layer to cooperate with normalization operator ; For lesions determined to be present, the same fusion features Mapping to a higher dimensional property discrimination space to compute benign Malignant and malignant Junction property Probability; deep characterization in the fusion characteristics is extracted, and a four-classification probability model is input to calculate the malignancy degree classification probability 。

Description

Tumor identification method based on multi-mode information coding Technical Field The invention relates to the technical field of tumor recognition, in particular to a tumor recognition method based on multi-mode information coding. Background The difficulty in tumor recognition is to accurately capture the morphology and physiological characteristics of the focus and effectively distinguish normal tissues from lesion areas. However, the existing identification system is difficult to realize high-precision identification and fine classification in clinical diagnosis due to limited information dimension and lack of fusion logic. At present, the tumor recognition technology mainly depends on structural imaging examination or electrophysiological index monitoring, and the two single-mode means have obvious information deficiency problems that the structural imaging mode captures visual information by extracting spatial distribution, boundary characteristics and the like of a focus, but cannot be related to biochemical metabolic characteristics of tissues, and the method of leading functional modes such as electrical impedance and the like can reflect abnormality of tissue metabolic activity, but has lower spatial resolution, is sensitive to physiological noise and lacks spatial structural support and positioning accuracy of the focus. In the existing multi-mode fusion technology, patent CN101984462a discloses a method for synthesizing an electrical impedance tomography functional image and an anatomical image, and the method realizes the visual fusion display of the electrical impedance tomography functional image and the anatomical image by acquiring electrical impedance imaging data and the anatomical image and utilizing coordinate mapping. However, the method still has the defects of space-time alignment calibration deficiency, insufficient time sequence characteristic mining, weak semantic relevance and the like in practical application, the special metabolic evolution and biomembrane capacitance characteristics of tumors cannot be represented by fully utilizing multidirectional coding logics such as forward, reverse, interval, segmentation and the like, attention mechanisms and self-adaptive enhancement weights are not introduced, key characteristics are easily submerged by redundant noise, and the accurate tracing requirement of clinical diagnosis is difficult to meet. Therefore, there is an urgent need to design a tumor recognition method based on multi-mode information coding, which is based on realizing high-precision time-space synchronization, extracts multi-dimensional time sequence coding features in depth, and combines progressive classification logic to solve the problems in the prior art. Disclosure of Invention The invention mainly aims to provide a tumor recognition method based on multi-mode information coding, which solves the problems of lack of time-space alignment calibration, insufficient mining of time sequence characteristics and weak semantic relevance in a tumor recognition technology. In order to solve the technical problems, the technical scheme adopted by the invention is that the tumor identification method based on multi-mode information coding comprises the following steps: s1, acquiring structural image data by using standard medical image equipment, synchronously acquiring electrical impedance time sequence data by using a multi-electrode array system, and realizing time-space synchronization of two modal data by time stamp alignment; S2, constructing a dual-path self-adaptive feature enhancement and filtering module, and respectively carrying out targeted pretreatment on structural image data and electrical impedance data; S3, designing a dual-path mode adaptive coding framework, and respectively extracting depth features of the preprocessed dual-mode data; s4, integrating double-path coding features based on a cross-modal attention fusion mechanism, and realizing accurate identification and classification of tumors through a multistage progressive classifier. In a preferred embodiment, the step S1 specifically includes the following steps: The TTL level signal is sent to the impedance acquisition upper computer at the image scanning starting moment through the hardware synchronous trigger and recorded as a reference time stamp Each frame of sampling point of impedance data is accompanied by time offsetCorresponding to absolute timeRealizing time alignment; spatial synchronization is achieved by labeling anatomical markers visualized under a structural image on the subject tissue. In a preferred embodiment, the step S2 specifically includes the following steps: s21, constructing a self-adaptive feature enhancement path of the structural image, and dynamically adjusting an enhancement strategy based on local block features; S22, constructing an electrical impedance data filtering and feature purifying path, eliminating noise interference and extracting space electromagnetic