CN-122023405-A - Fusion-assisted decision model construction method and system for pneumonia image interpretation

CN122023405ACN 122023405 ACN122023405 ACN 122023405ACN-122023405-A

Abstract

The application discloses a fusion auxiliary decision model construction method and system for pneumonia image interpretation, relates to the technical field of intelligent auxiliary diagnosis, and provides a scheme which comprises the steps of acquiring a pneumonia sample image and a healthy lung image, and constructing an interpreter and a positioner, extracting local textures of lung parenchyma from the healthy lung images, clustering and discretizing to obtain a discrete vector set representing the healthy textures as a healthy codebook, and identifying focus areas in the pneumonia sample images by using the positioner. According to the application, the health codebook is constructed, the anti-facts health image is generated, the current interpretation value is fed back to the locator parameter update through the inner layer iteration triggered by the preset health threshold, the defect that the interpretation conclusion in the existing scheme is loosely corresponding to the focus hot zone and is easily interfered by background clues irrelevant to the focus, so that the reliability is insufficient is overcome, and the interpretation conclusion can be checked by the anti-facts contrast and the interpretation output of consistent positioning evidence and interpretation basis is realized.

Inventors

LU YIJIA
RUI TAO
LI LEXI

Assignees

安徽大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The fusion-assisted decision model construction method for pneumonia image interpretation is characterized by comprising the following steps of: Acquiring a pneumonia sample image and a healthy lung image, and constructing an reader and a positioner; extracting local textures of lung parenchyma from the healthy lung images, and carrying out clustering discretization to obtain a discrete vector set representing the healthy textures as a healthy codebook; identifying a focus area in the pneumonia sample image by using a locator, extracting nearest discrete vectors of texture features projected to a health codebook, and reconstructing a health texture to cover the focus area to obtain a counterfactual health image; inputting the anti-fact health image into an interpreter to obtain a current interpretation value, and if the current interpretation value is lower than a preset health threshold, executing inner layer iteration, namely updating the positioner parameters according to the difference between the current interpretation value and the preset health threshold, adjusting the focus area and reconstructing the anti-fact health image to update the current interpretation value until the current interpretation value is not lower than the preset health threshold; Extracting focus texture features based on focus areas when inner layer iteration stops, and transferring the focus texture features to randomly selected healthy lung images for feature distribution alignment to generate a synthetic lesion image; Respectively inputting the anti-facts health image, the pneumonia sample image and the synthesized lesion image into a reader to obtain a first interpretation value, a second interpretation value and a third interpretation value, constructing a loss function by taking the minimized difference between the second interpretation value and the third interpretation value and the maximized difference between the first interpretation value and the third interpretation value as constraint, and updating the parameters of the reader based on the loss function; And alternately executing the updating of the parameters of the localizer and the updating of the parameters of the interpreter until the loss function value for updating the parameters of the interpreter meets the preset convergence condition or reaches the preset iteration upper limit, so that the interpreter and the localizer when the alternate execution is terminated construct a fusion auxiliary decision model.
2. The method of claim 1, wherein inputting the anti-factual health image into the interpreter to obtain a current interpretation value, and if the current interpretation value is lower than a preset health threshold, performing an inner layer iteration, specifically comprising: Setting the maximum circulation times and initializing a record snapshot, wherein the record snapshot is used for storing the current interpretation value with the smallest difference with a preset health threshold value in the inner layer iteration process, and a focus area and a counter fact health image corresponding to the current interpretation value; After each round of inner layer iteration updates the current interpretation value, if the difference between the current interpretation value and the preset health threshold value is smaller than the corresponding difference of the record snapshot, updating the record snapshot; Stopping inner layer iteration and outputting a current focus area and a negative fact health image when the current interpretation value is not lower than a preset health threshold and the area change rate of the focus area obtained by two adjacent iterations is smaller than the preset change threshold; if the iteration number reaches the maximum cycle number, the inner layer iteration is stopped, and the focus area corresponding to the record snapshot and the anti-reality health image are output.
3. The method according to claim 1, wherein updating the locator parameters based on the difference between the current interpretation value and the preset health threshold value comprises: calculating the difference between the current interpretation value and a preset health threshold value, and calculating the log likelihood loss according to the difference to serve as an countermeasure feedback item; in the focus area, calculating the distance of the texture feature projected to the nearest discrete vector in the healthy codebook as a manifold reconstruction deviation term; Extracting corresponding region characteristic representations of the pneumonia sample image and the anti-factual health image obtained by an interpreter in a lung parenchymal region outside the focus region, and calculating the characteristic distance between the pneumonia sample image and the anti-factual health image as a background keeping deviation term; Linearly combining the countermeasure feedback item, the manifold reconstruction deviation item and the background retention deviation item through preset weights to obtain a positioning guide target value, and constructing an area regular item based on sparsity measurement of a focus area; And updating the locator parameters with the locating guide target value and the area regular term and outputting the updated focus area.
4. A method according to claim 3, wherein the process of updating the locator parameters with the location guidance target value and the region regularization term further comprises the steps of constructing and introducing a spatial consistency constraint term, and specifically comprises: Dividing the pneumonia sample image into a plurality of local areas, and generating a corresponding disturbance image based on each local area, wherein the disturbance image is obtained by replacing the corresponding local area with a position area of the anti-reality health image; Inputting the pneumonia sample image into an interpreter to obtain a sample interpretation value, and inputting each disturbance image into the interpreter to obtain a corresponding disturbance interpretation value; Calculating the difference between each disturbance interpretation value and the sample interpretation value, and mapping each difference to a corresponding local area to form an interpretation response diagram; and calculating a spatial consistency metric based on the interpretation response graph and the focus region, obtaining a spatial consistency constraint term, and incorporating the spatial consistency constraint term into the region regularization term to update the locator parameters.
5. The method of claim 4, wherein the locator parameter update and the arbiter parameter update are performed alternately, comprising: setting a global training round including a locator optimization sub-period and an arbiter optimization sub-period, and configuring batch update steps of the locator optimization sub-period and the arbiter optimization sub-period in a single training round; Maintaining the parameters of the interpreter fixed in the optimization subcycle of the localizer, respectively converting each pneumonia sample image in the current input batch into a counter fact health image, and then executing inner layer iteration to obtain a focus area and a counter fact health image corresponding to the input batch; Aggregating error signals generated by each pneumonia sample image in the input batch, uniformly updating the parameters of the localizer, and repeatedly processing the input batch until the batch updating step number of the localizer is reached, wherein the error signals comprise a locating guide target value, a region regularization term and a space consistency constraint term; Maintaining the parameters of the localizer fixed in the optimizing sub-period of the interpreter, respectively calculating and summarizing the loss function values of each combined sample in the input batch, updating the parameters of the interpreter based on the summarized loss function values, and repeatedly processing the input batch until the batch updating step number of the interpreter is reached, wherein the combined samples comprise a counter-facts health image, a pneumonia sample image and a synthetic lesion image; and after finishing the locator optimization subcycle and the arbiter optimization subcycle in a single training round, entering the next training round until the loss function value for updating the arbiter parameters meets the preset convergence condition or reaches the preset iteration upper limit.
6. The method of claim 1, wherein constructing the loss function with the constraint of minimizing the difference between the second interpretation value and the third interpretation value and maximizing the difference between the first interpretation value and the third interpretation value comprises: Calculating the square difference measurement of the second interpretation value and the third interpretation value to be used as a consistency loss term for restricting the interpretation consistency of the synthesized lesion image and the pneumonia sample image; Calculating the opposite number of the absolute difference measurement of the first interpretation value and the third interpretation value, and taking the opposite number as a contrast loss term for restricting and increasing the interpretation difference between the anti-reality healthy image and the synthetic lesion image; introducing the classification cross entropy loss of the interpreter as a basic classification term, and carrying out linear weighted summation on the consistency loss term, the comparison loss term and the basic classification term to obtain the loss function value.
7. The method according to claim 1, wherein the constructing the fusion aid decision model with the interpreter and the localizer at the termination of the alternating execution, in particular comprises: Acquiring a lung image to be interpreted; Outputting a focus area to be interpreted as a lung image by using a positioner, and converting the focus area into a focus area heat map indicating the lesion position; outputting an original interpretation value of the lung image to be interpreted by using an interpreter, and outputting a counterfactual interpretation value of the counterfactual health image reconstructed by the focus area; Calculating the numerical deviation of the original interpretation value and the inverse fact interpretation value, and mapping the numerical deviation into a consistency confidence index; outputting an auxiliary decision result comprising the original interpretation value, the focus region heat map and the consistency confidence index.
8. The fusion-assisted decision model construction system for pneumonia image interpretation, which is characterized by being used for realizing the fusion-assisted decision model construction method for pneumonia image interpretation according to any one of claims 1-7, comprising: the data model construction module is used for acquiring a pneumonia sample image and a healthy lung image and constructing an reader and a positioner; The texture codebook generation module is used for extracting local textures of lung parenchyma from the healthy lung images and carrying out clustering discretization to obtain a discrete vector set representing the healthy textures as a healthy codebook; The anti-facts image generation module is used for identifying a focus area in the pneumonia sample image by using the localizer, extracting nearest discrete vectors of texture features projected to the health codebook, and reconstructing health textures to cover the focus area to obtain an anti-facts health image; The focus iteration optimization module is used for inputting the anti-fact health image into the interpreter to obtain a current interpretation value, and if the current interpretation value is lower than a preset health threshold, performing inner iteration, namely updating a positioner parameter according to the difference between the current interpretation value and the preset health threshold, adjusting a focus area and reconstructing the anti-fact health image to update the current interpretation value until the current interpretation value is not lower than the preset health threshold; The feature migration synthesis module is used for extracting focus texture features based on focus areas when inner layer iteration stops, and migrating the focus texture features to randomly selected healthy lung images for feature distribution alignment to generate a synthesized lesion image; The loss function optimizing module is used for respectively inputting the anti-facts health image, the pneumonia sample image and the synthetic lesion image into the interpreter to obtain a first interpretation value, a second interpretation value and a third interpretation value, constructing a loss function by taking the difference between the minimized second interpretation value and the maximized third interpretation value as constraint, and updating the parameters of the interpreter based on the loss function; And the model fusion construction module is used for alternately executing the updating of the parameters of the localizer and the updating of the parameters of the interpreter until the loss function value for updating the parameters of the interpreter meets the preset convergence condition or reaches the preset iteration upper limit so as to construct a fusion auxiliary decision-making model by the interpreter and the localizer when the alternate execution is terminated.
9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the method of any of claims 1-7 when executing the computer program.
10. A computer readable storage medium storing a computer program, which when executed by a processor performs the method according to any one of claims 1-7.

Description

Fusion-assisted decision model construction method and system for pneumonia image interpretation Technical Field The application relates to the technical field of intelligent auxiliary diagnosis, in particular to a fusion auxiliary decision model construction method and system for pneumonia image interpretation. Background Pulmonary infection diseases are characterized by high incidence and rapid progression in clinic, and chest imaging examinations are widely used for primary screening and typing because of convenient acquisition. With the development of artificial intelligence, an automatic image-based discrimination model gradually enters an auxiliary diagnosis process, and the core target of the automatic image-based discrimination model is not only to stably output the disease risk under complex individual differences, but also to provide clear and accurate pathological positioning to assist a doctor in decision making. The existing image interpretation scheme trains a classification network by a large number of marked images or superimposes a saliency visualization and region detection module outside the classification network to outline abnormal focus hot areas so as to enhance the interpretability. However, the real clinical image is affected by various variables such as scan parameters, combined underlying disease and noise artifacts, and the training data often only has coarse-grained image-level labels, which makes the model easily dependent on lesion-independent background cues or instrument traces to form shortcut features. The reliability of the auxiliary decision model is reduced due to the fact that high-risk decision output by the model is possibly not based on a real pathological basis, and marked hot areas are possibly deviated from a real pathogenic core, further, the existing training mode is usually only limited and correctly predicted, and the problem is solved due to the fact that constraint of 'when suspected pathological clues are weakened or removed, interpretation conclusion should correspondingly trend to health', and 'interpretation consistency of same pathological features should be kept in different image carriers', is lacked, and consistency and stability of interpretation results and positioning evidence are difficult to consider when imaging conditions of the model are changed and pathological forms are various. Disclosure of Invention In order to solve the technical problems, the technical scheme solves the problems in the background technology by providing a fusion-assisted decision model construction method and a fusion-assisted decision model construction system for pneumonia image interpretation. In order to achieve the above object, the technical scheme of the present invention is as follows: In a first aspect, the present application provides a fusion-assisted decision model construction method for pneumonia image interpretation, the method comprising: Acquiring a pneumonia sample image and a healthy lung image, and constructing an reader and a positioner; extracting local textures of lung parenchyma from the healthy lung images, and carrying out clustering discretization to obtain a discrete vector set representing the healthy textures as a healthy codebook; identifying a focus area in the pneumonia sample image by using a locator, extracting nearest discrete vectors of texture features projected to a health codebook, and reconstructing a health texture to cover the focus area to obtain a counterfactual health image; inputting the anti-fact health image into an interpreter to obtain a current interpretation value, and if the current interpretation value is lower than a preset health threshold, executing inner layer iteration, namely updating the positioner parameters according to the difference between the current interpretation value and the preset health threshold, adjusting the focus area and reconstructing the anti-fact health image to update the current interpretation value until the current interpretation value is not lower than the preset health threshold; Extracting focus texture features based on focus areas when inner layer iteration stops, and transferring the focus texture features to randomly selected healthy lung images for feature distribution alignment to generate a synthetic lesion image; Respectively inputting the anti-facts health image, the pneumonia sample image and the synthesized lesion image into a reader to obtain a first interpretation value, a second interpretation value and a third interpretation value, constructing a loss function by taking the minimized difference between the second interpretation value and the third interpretation value and the maximized difference between the first interpretation value and the third interpretation value as constraint, and updating the parameters of the reader based on the loss function; And alternately executing the updating of the parameters of the localizer and the updating of the parameters of the