Search

CN-122023814-A - Pneumonia discernment detecting system based on image recognition

CN122023814ACN 122023814 ACN122023814 ACN 122023814ACN-122023814-A

Abstract

The invention relates to the technical field of medical image processing, in particular to a pneumonia recognition and detection system based on image recognition, which comprises an image preprocessing module, a feature fusion module, a focus segmentation module and a classification decision module which are connected in sequence. The image preprocessing module generates a dynamic pathological feature sequence containing time sequence dependency from the original image. And the feature fusion module verifies and fuses key features in the sequence. The focus segmentation module maps the features to a preset anatomical structure space to generate a pathological feature distribution map, outlines focus boundaries according to the pathological feature distribution map, and obtains a fine segmentation mask after optimization. The classification decision module extracts image histology characteristics based on the mask, and outputs a result after integrating classification and clinical knowledge calibration. According to the scheme, dynamic pathological features are constructed, segmentation is guided in the anatomical space, and the recognition accuracy and the clinical rationality of the result are improved.

Inventors

  • LIU ZHIWEI

Assignees

  • 首都医科大学附属北京积水潭医院

Dates

Publication Date
20260512
Application Date
20260413

Claims (10)

  1. 1. The pneumonia discernment detecting system based on image recognition, characterized by comprising: the image preprocessing module is used for acquiring original medical image data, performing image quality enhancement and standardization preprocessing, and generating a dynamic pathological feature sequence containing a time sequence dependency relationship; The feature fusion module is used for carrying out significance verification and fusion on key features in the dynamic pathological feature sequence based on the dynamic pathological feature sequence to generate verified fused pathological features; the focus segmentation module maps the verified fusion pathological features to a preset anatomical structure space, generates a pathological feature distribution diagram associated with a lung anatomical region, outlines boundary information of a suspected focus region according to the pathological feature distribution diagram, generates a preliminary focus segmentation result, executes morphological consistency verification, corrects boundary irregularity caused by image noise or artifact, and generates an optimized fine focus segmentation mask; And the classification decision module is used for extracting quantitative image histology characteristics of a focus area based on the optimized fine focus segmentation mask, forming an image histology characteristic vector for classification, inputting the image histology characteristic vector into the integrated classifier, calculating to obtain initial classification confidence coefficient about the type of pneumonia, combining a clinical priori knowledge base, calibrating the initial classification confidence coefficient, and finally outputting a detection result of pneumonia identification.
  2. 2. The image recognition-based pneumonia detection system according to claim 1, wherein said acquiring raw medical image data and performing image quality enhancement and normalization preprocessing generates a dynamic pathological feature sequence including a time-series dependency relationship, comprising Performing image quality enhancement and standardization preprocessing on the original medical image data to generate standardized lung image data; Performing multi-level image analysis on the standardized lung image data, extracting multi-scale imaging modes related to pathological features of pneumonia in the images, and constructing pathological feature maps; inputting the pathological feature map to a dynamic feature evolution model, simulating the evolution process of pathological features on a time sequence, and generating a dynamic pathological feature sequence containing a time sequence dependency relationship; The performing image quality enhancement and normalization preprocessing on the raw medical image data to generate normalized lung image data includes: processing the original medical image data by adopting a self-adaptive histogram equalization algorithm so as to enhance the local contrast of the image; carrying out noise reduction treatment on the enhanced image by applying a non-local mean denoising algorithm, and inhibiting noise while preserving tissue edges; The super-resolution reconstruction network based on deep learning is used for improving the resolution of the image after noise reduction, so that the definition of image details is increased; registering the image with the improved resolution to a standard anatomical template space to complete space standardization; and carrying out gray scale normalization on the images in the standardized space, eliminating gray scale differences brought by different scanning devices or protocols, and finally generating standardized lung image data.
  3. 3. The image recognition-based pneumonia recognition detection system according to claim 2, wherein said performing multi-level image analysis on said normalized lung image data, extracting multi-scale imaging patterns related to a lung pathology in an image, and constructing a pathology feature map, comprises: analyzing the gray distribution and texture characteristics of the standardized lung image data at a pixel level to identify an abnormal density region; Dividing different lung anatomical structures at the region level by applying a segmentation algorithm, and calculating statistics of imaging features in each region; Analyzing the overall pattern of the entire lung image at a global level, including symmetry analysis and volume fraction calculation; And fusing analysis results of the pixel level, the region level and the global level to construct a comprehensive pathological feature map integrated with multi-scale information.
  4. 4. The image recognition-based pneumonia recognition detection system according to claim 3, wherein said inputting the pathology feature map to a dynamic feature evolution model simulates a evolution process of a pathology feature over a time sequence, generating a dynamic pathology feature sequence including a time-series dependency relationship, comprises: inputting the pathological feature map as an initial state into a circulating neural network; the history characteristic state is saved by the aid of the circulating neural network memory unit; predicting the feature change of the next time step according to the currently input features and the history state stored in the memory unit in each time step; And performing iteration to generate a dynamic pathological feature sequence reflecting a potential evolution path of the pathological feature.
  5. 5. The image recognition-based pneumonia recognition detection system according to claim 4, wherein said performing saliency verification and fusion of key features in a dynamic pathology feature sequence based on said dynamic pathology feature sequence to generate a verified fused pathology feature comprises: calculating the significance score of each feature point in the dynamic pathological feature sequence; Screening out a key feature point set according to the significance score; Carrying out correlation analysis on the features in the key feature point set, and removing redundant features; Weighting and fusing the screened key features, wherein the weight is determined by the feature significance and clinical significance of the key features, and verified fused pathological features are generated; the calculating the significance score of each feature point in the dynamic pathological feature sequence specifically comprises the following steps: Extracting a change track of each feature point in the dynamic pathological feature sequence in the time dimension, and calculating Euclidean distance between the change track and a reference feature mode as a time significance component; the space neighborhood of the feature points is scanned through a convolution kernel, and the ratio of the local feature variance to the global feature variance is calculated to be used as a space significance component; Calculating cosine similarity between the feature points and a preset pneumonia typical feature template by using an attention mechanism, and generating a semantic significance component; weighting and fusing the time significance component, the space significance component and the semantic significance component, wherein the weight is dynamically adjusted according to the feature type; And carrying out normalization processing on the fused significance scores so as to enable the significance scores to be distributed in a numerical range from zero to one.
  6. 6. The image recognition-based pneumonia detection system of claim 5, wherein said mapping said validated fused pathology to a pre-set anatomical structure space generates a pathology profile associated with a lung anatomy region, comprising: Loading a predefined standard lung anatomy map defining a spatial positional relationship of a main lung region; Establishing a corresponding relation between the verified fusion pathological features and each region in the standard lung anatomy structure map; assigning the characteristic values to the corresponding anatomical regions according to the corresponding relations; Smoothing discrete feature assignment by adopting a spatial interpolation algorithm to generate a continuous pathological feature distribution map corresponding to an anatomical structure; the smoothing processing of discrete feature assignment by adopting a spatial interpolation algorithm comprises the following steps: Establishing a radial basis function interpolation model taking an anatomical structure key point as a center, wherein the radial basis function adopts a Gaussian kernel function form; Calculating an interpolation weight matrix according to the space coordinates of the discrete feature assignment points, wherein the weights and the point distances are in a negative correlation relationship; generating a characteristic value distribution function in a continuous space by solving a linear equation set of a radial basis function coefficient matrix; Extracting an isosurface from the eigenvalue distribution function to generate a smooth eigenvalue distribution curved surface which is matched with the boundary of the anatomical structure; And adopting an anisotropic diffusion algorithm to carry out edge-preserving smoothing treatment on the characteristic distribution curved surface, and preserving sharpness of an anatomical boundary.
  7. 7. The system of claim 6, wherein the delineating boundary information of a suspected lesion area according to the pathological feature distribution map, and generating a preliminary lesion segmentation result comprises: selecting points with characteristic values exceeding a preset threshold value from the pathological characteristic distribution diagram as seed points; taking the seed point as a center, and growing to surrounding areas according to a characteristic similarity criterion; Iteratively executing the growth process until a boundary point with a characteristic value lower than the growth threshold value is encountered; Marking all connected growing areas as suspected lesions, recording boundary coordinates of the suspected lesions, and generating a preliminary lesion segmentation result.
  8. 8. The image recognition-based pneumonia detection system according to claim 7, wherein said performing morphological consistency check to correct boundary irregularities due to image noise or artifacts, generating an optimized fine focus segmentation mask, comprises: Applying morphological closing operation to the preliminary lesion segmentation result to fill small holes inside the region; Applying morphological opening operations to eliminate small, isolated suspected lesion areas; smoothing and filtering the processed boundary to enable the processed boundary to be more in line with morphological characteristics of a real focus, and finally generating an optimized fine focus segmentation mask; The applying morphological closing operation to the preliminary lesion segmentation result to fill small holes inside the region includes: constructing a circular structural element, wherein the diameter of the circular structural element is adaptively determined according to the area of a focus area; expanding the circular structural element along the focus boundary to fill the boundary concave area and the internal hole; performing an etching operation on the expanded region to restore the approximate contour shape of the original lesion; calculating the difference value of the areas before and after the closing operation, and identifying the filled hole areas and the spatial distribution thereof; And carrying out connectivity analysis on the filled areas, and merging newly added focus areas connected by hole filling.
  9. 9. The image recognition-based pneumonia detection system according to claim 8, wherein said extracting quantitative image histology features of a lesion area based on said optimized fine lesion segmentation mask to form an image histology feature vector for classification comprises: Extracting first-order statistical features from a focus area defined by the optimized fine focus segmentation mask, and describing gray distribution characteristics of the first-order statistical features; extracting texture features and describing gray space relation inside the focus; Extracting morphological characteristics and describing geometric characteristics of a focus; And combining the extracted first-order statistical features, texture features and morphological features into a high-dimensional image histology feature vector.
  10. 10. The image recognition-based pneumonia recognition detection system according to claim 9, wherein said inputting the image histology feature vector into the integrated classifier, calculating an initial classification confidence about the pneumonia type, calibrating the initial classification confidence in combination with the clinical prior knowledge base, and finally outputting a detection result of the pneumonia recognition, includes: The image histology feature vector is simultaneously input into a support vector machine, a random forest and a plurality of basis classifiers of a neural network; aggregating the output results of the plurality of base classifiers, and obtaining initial classification confidence coefficient through voting or a weighted average mechanism; inquiring a clinical priori knowledge base to obtain epidemiological statistical data and typical image expression knowledge which are consistent with the information of the cases demographics; And carrying out Bayesian correction on the initial classification confidence according to the information in the clinical priori knowledge base to obtain a calibrated final classification result, and outputting the final classification result as a detection result of pneumonia recognition.

Description

Pneumonia discernment detecting system based on image recognition Technical Field The invention relates to the technical field of medical image processing, in particular to a pneumonia identification and detection system based on image identification. Background Current deep learning based pneumonia image recognition systems typically employ an end-to-end classification network or "split-classify" two-stage pipeline. According to the method, a single medical image is directly used as static input, spatial features are extracted through a convolutional neural network, and classification or segmentation is carried out. In the focus segmentation link, the mainstream method such as U-Net and its variant mainly carries out end-to-end learning in the image pixel space or the feature space, and outputs the segmentation mask in a pixel-by-pixel classification mode. These conventional solutions treat and parse an image as a static picture of one spatial dimension. The prior art solutions have drawbacks. The end-to-end processing mode ignores that the pathological change is a dynamic process containing a time sequence rule, and features extracted from single Zhang Jingtai images lack characterization of potential pathological evolution information, so that the recognition capability of atypical pneumonia or early-stage pneumonia is limited. Meanwhile, the segmentation is directly carried out in the pixel space, the decision of the segmentation is seriously dependent on the appearance characteristics of a local image, the explicit knowledge guidance of anatomical structures is lacking, and the segmentation is easily interfered by image noise, artifacts or similar textures of adjacent tissues, so that the segmentation boundary is inaccurate, does not accord with anatomical logic, or isolated false positive regions are generated. This reduces the reliability of subsequent feature extraction and the interpretability of the overall system. The invention aims to solve the problem of how to effectively capture and characterize implicit pathological dynamic characteristics from single Zhang Jingtai images so as to enrich identification characteristics. Meanwhile, the problem of how to deeply fuse the segmentation process with accurate priori knowledge of lung anatomy is solved, so that focus positioning and outlining are free from low-level image noise interference, and the real anatomical structure distribution rule is met. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a pneumonia identification and detection system based on image identification. In order to achieve the above purpose, the invention adopts the following technical scheme that the pneumonia identification and detection system based on image identification comprises: the image preprocessing module is used for acquiring original medical image data, performing image quality enhancement and standardization preprocessing, and generating a dynamic pathological feature sequence containing a time sequence dependency relationship; The feature fusion module is used for carrying out significance verification and fusion on key features in the dynamic pathological feature sequence based on the dynamic pathological feature sequence to generate verified fused pathological features; the focus segmentation module maps the verified fusion pathological features to a preset anatomical structure space, generates a pathological feature distribution diagram associated with a lung anatomical region, outlines boundary information of a suspected focus region according to the pathological feature distribution diagram, generates a preliminary focus segmentation result, executes morphological consistency verification, corrects boundary irregularity caused by image noise or artifact, and generates an optimized fine focus segmentation mask; And the classification decision module is used for extracting quantitative image histology characteristics of a focus area based on the optimized fine focus segmentation mask, forming an image histology characteristic vector for classification, inputting the image histology characteristic vector into the integrated classifier, calculating to obtain initial classification confidence coefficient about the type of pneumonia, combining a clinical priori knowledge base, calibrating the initial classification confidence coefficient, and finally outputting a detection result of pneumonia identification. As a further aspect of the present invention, the acquiring of the original medical image data and the performing of the image quality enhancement and the standardized preprocessing generate a dynamic pathological feature sequence including a time-series dependency relationship, including Performing image quality enhancement and standardization preprocessing on the original medical image data to generate standardized lung image data; Performing multi-level image analysis on the standardized lung image data, extrac