Search

CN-122000033-A - Machine learning-based coronary artery lipid spot identification method

CN122000033ACN 122000033 ACN122000033 ACN 122000033ACN-122000033-A

Abstract

The invention provides a machine learning-based coronary artery lipid spot identification method, which belongs to the field of image identification and aims at identifying a lipid spot structure region from a coronary artery image, the method fuses frequency domain and airspace characteristics through a spectrum space joint enhancement module to overcome illumination interference, the anisotropic structure characteristics are extracted through the asymmetric differential operator of the characteristic dimension analysis module, rotation invariant characteristic expression is established through the direction aggregation signature of the structure cooperative identification module, the problem that the identification precision of the non-uniform lipid spots is insufficient in a complex environment in the traditional method is effectively solved, and reliable technical support is provided for coronary heart disease screening and accurate auxiliary diagnosis.

Inventors

  • Sheng Deqiao
  • LIU YALI

Assignees

  • 三峡大学

Dates

Publication Date
20260508
Application Date
20260128

Claims (10)

  1. 1. A machine learning-based method for identifying coronary lipid plaques, comprising: Collecting an original image of a coronary artery blood vessel, preprocessing the original image through unified cutting, and constructing a coronary artery lipid spot image dataset; Providing high-precision color space transformation, constructing a spectrum space joint enhancement module, processing a coronary artery lipid spot image according to the spectrum space joint enhancement module, and outputting a multiphase characteristic fusion enhancement image; Constructing a feature dimension analysis module according to the asymmetric differential operator, and enabling the multi-phase feature fusion enhancement image to obtain a decoupling enhancement feature map through the feature dimension analysis module; Designing a structure cooperative operator and a direction aggregation signature strategy, and constructing a structure cooperative identification module, so that the coronary artery lipid spot image and the decoupling enhancement feature map obtain a lipid spot identification result through the structure cooperative identification module; Integrating the spectrum space joint enhancement module, the characteristic dimension analysis module and the structure cooperative identification module into a coronary artery lipid spot identification model; training the coronary artery lipid spot recognition model, inputting the coronary artery lipid spot image to be recognized into the recognition model after training to infer, and outputting a lipid spot recognition area of the coronary artery lipid spot image.
  2. 2. The machine learning based coronary artery lipid spot identification method of claim 1, wherein a color space transformation strategy is constructed based on a difference between an image RGB channel and a global color mean value, wherein different transformation rules are respectively adopted according to whether the difference exceeds a set threshold value, and the threshold value is determined by utilizing the image global color mean value and a proportional relation thereof, so as to generate a coronary artery lipid spot image after color space transformation.
  3. 3. The machine learning-based coronary artery lipid spot identification method according to claim 2, wherein the transverse component and the longitudinal component of the coronary artery lipid spot image are determined according to the coordinate value of each pixel position of the coronary artery lipid spot image after color space transformation, the adaptive standard deviation is built based on the components, the space sensitive filter kernel is generated, and the coupling filter characteristic is obtained by tensor product operation by combining a frequency domain set comprising low frequency, medium frequency and high frequency.
  4. 4. The machine learning-based coronary artery lipid spot identification method according to claim 3 is characterized in that phase feature bases corresponding to phases are constructed based on main diagonal length and phase offset of coronary artery lipid spot identification images after color space transformation, corresponding phase weights are introduced based on the phase feature bases, hadamard products are executed on coupling filter features, feature response fusion under each phase is completed sequentially, fusion results of all phases are integrated in a weighting mode, and a multi-phase feature fusion enhanced image is output.
  5. 5. The machine learning-based coronary artery lipid spot identification method according to claim 4, wherein L2 norm and infinity norm are calculated on a multi-phase feature fusion enhanced image respectively to generate adaptive weights corresponding to pixel positions, positive asymmetric differential calculated components of the enhanced image are extracted in a main diagonal direction based on the adaptive weights, and negative asymmetric differential calculated components of the enhanced image are extracted in a sub diagonal direction.
  6. 6. The machine learning-based coronary artery lipid spot recognition method according to claim 5 is characterized by calculating the Frobenius norms of the coronary artery lipid spot recognition method in the whole image range based on positive and negative asymmetric differential calculation components respectively, constructing a decomposition angle according to the ratio relation of the two norms, carrying out rotation combination on the positive and negative asymmetric differential calculation components based on the decomposition angle to construct a structure component and a texture component, carrying out nonlinear mapping on the texture component, introducing a structure-texture balance coefficient, completing joint enhancement of a structure and a texture in the image region, and outputting a decoupling enhancement feature map.
  7. 7. The machine learning-based coronary artery lipid spot identification method is characterized by obtaining image gradient direction information based on horizontal direction gradients and vertical direction gradients of a coronary artery lipid spot image, carrying out weighted combination on differences between the gradient direction information and a decoupling enhancement feature map to construct a structure synergistic operator, carrying out structure adjustment on residual information between original brightness of an image and the decoupling enhancement feature map based on the structure synergistic operator, and extracting a joint feature map.
  8. 8. The machine learning-based coronary artery lipid spot identification method of claim 7, wherein the joint feature map is converted into a polar coordinate representation form by using an inscribed circle radius of the joint feature map as an integral range, and the polar coordinate joint feature map is subjected to weighted integration based on a radial attenuation factor in the integral range to construct a global structural signature.
  9. 9. The machine learning-based coronary artery lipid spot identification method is characterized by comprising the steps of calculating a feature mean value and a feature variance of a global structural signature, carrying out standardization processing on the global structural signature based on the feature mean value and the feature variance, providing significant feature projection, projecting the standardized global structural signature to an image space to generate a significant heat map, constructing a self-adaptive identification strategy based on the significant heat map, and determining an identification mask of a target lipid spot area.
  10. 10. The machine learning-based coronary artery lipid spot identification method is characterized in that a multi-phase feature fusion enhancement image is obtained through a spectrum space joint enhancement module by a coronary artery lipid spot image, a decoupling enhancement feature map is obtained through a feature dimension analysis module by the multi-phase feature fusion enhancement image, and a final lipid spot identification result is obtained through a structural cooperative identification module by the coronary artery lipid spot image and the decoupling enhancement feature map.

Description

Machine learning-based coronary artery lipid spot identification method Technical Field The invention belongs to the field of image recognition, and particularly relates to a machine learning-based coronary artery lipid spot recognition method. Background The traditional coronary artery lipid spot identification mainly depends on manual film reading or basic image processing algorithm, has obvious defects that under complex imaging conditions, the traditional method is difficult to effectively extract the spectrum and texture characteristics of lipid spots in the blood vessel wall, in addition, the traditional algorithm has poor adaptability to non-ideal clinical scenes such as blood vessel bending angle change, local calcification shielding and the like, and the brightness and texture of the coronary artery wall and surrounding tissues are highly similar, so that the traditional method cannot accurately separate lipid spot areas from normal blood vessel wall structures, and the lesion screening and risk assessment efficiency is seriously affected. Although the current solution based on deep learning improves the recognition rate of part of clear images, the method still has obvious defects in actual clinical images, most algorithms directly process single-mode images, the single-mode images are not optimized for spectral attenuation characteristics in blood vessel cavities, signal differences caused by different illumination intensities and imaging depths cause the reduction of recognition stability, and secondly, the existing feature extraction method is used for mixing and processing blood vessel structural features and lipid speckle texture features, the false detection rate is higher in similar tissue type scenes, and most importantly, the existing method lacks uniform characterization capability for lipid spots under multi-angle imaging and is difficult to infer overall lesion distribution through local segment features. The invention provides a machine learning-based coronary artery lipid spot identification method, which realizes stable extraction of lipid spot characteristics under complex illumination and imaging conditions through a spectrum space joint enhancement module, solves the illumination sensitivity problem of the traditional method, adopts a characteristic dimension analysis module to separate a blood vessel structure and lipid textures and comprehensively enhance, obviously improves the geometric expression of a key lesion region, and finally integrates multi-angle imaging characteristics through a structure cooperative identification module to realize robust identification of multi-angle scenes. Disclosure of Invention The invention provides a machine learning-based coronary artery lipid spot identification method, which aims to realize the identification of coronary artery lipid spots through the synergistic effect of a spectrum space joint enhancement module, a feature dimension analysis module and a structure synergistic identification module. The invention aims at providing a coronary artery lipid spot identification model, and provides a machine learning-based coronary artery lipid spot identification method, which comprises the following steps: s1, acquiring an original medical image of a coronary artery lipid spot, and preprocessing the original image through conversion format and unified cutting to generate a coronary artery lipid spot image data set; S2, providing high-precision color space transformation, constructing a spectrum space joint enhancement module, and obtaining a multiphase characteristic fusion enhancement image by the spectrum space joint enhancement module from the coronary artery lipid spot image; s3, designing an asymmetric differential operator, constructing a feature multidimensional analysis module, and obtaining a decoupling enhancement feature map by the multi-phase feature fusion enhancement image through the feature multidimensional analysis module; s4, a structure cooperative operator and a direction aggregation signature strategy are proposed, a structure cooperative identification module is constructed, and a coronary artery lipid spot image and a decoupling enhancement feature map obtain a lipid spot identification result through the structure cooperative identification module; s5, integrating a spectrum space joint enhancement module, a characteristic dimension analysis module and a structure cooperative identification module to construct a coronary artery lipid spot identification model; S6, training a coronary artery lipid spot recognition model, inputting a coronary artery lipid spot image to be recognized into the trained coronary artery lipid spot recognition model for recognition, and outputting a lipid spot recognition area in the coronary artery lipid spot image. Preferably, in S1, constructing the coronary artery lipid spot image dataset specifically includes the steps of using a clinical imaging device to perform multi-section and mult