CN-116580242-B - Method for classifying CT images of nontuberculosis mycobacteria lung diseases and phthisis

Abstract

The invention discloses a method for classifying CT images of nontuberculous mycobacteria lung diseases and phthisis, which mainly comprises the following steps of extracting an image feature map through an encoder and dividing the image feature map into a plurality of examples, wherein each example represents feature information in a certain area of an image, calculating the correlation weight of the examples and classification tasks, aggregating the examples into package feature representations according to the weight to express the features of CT images, predicting the package feature representations into CT image categories through a classifier, selecting important examples according to the correlation weight, evaluating through the classifier to represent the correlation degree of learning content and expected content of a network, and constructing a loss function which comprises package classification loss and correlation degree loss and respectively represents the difference between a CT predicted value and a true value and the difference between model learning content and expected content. The method can realize the two classification of the non-tuberculosis mycobacterium lung disease and the pulmonary tuberculosis CT image, and is superior to the classification accuracy of the existing method.

Inventors

• WAN LIANG
• MA HAODONG
• Xing Zhihang
• Tian yuntong

Assignees

• 天津大学

Dates

Publication Date

20260512

Application Date

20230522

Claims (4)

1. 1. A method for classifying CT images of nontuberculosis mycobacterial lung diseases and phthisis is characterized by comprising the following steps: Step S1, extracting features of a non-tuberculosis mycobacterium lung disease and a tuberculosis CT image through a neural network, and taking each feature pixel of the extracted feature image divided along the length, width and height as an example, wherein each example represents feature information in a specific area of the image; step S2, calculating the correlation weights of the examples obtained in the step S1) and the classification tasks, and weighting and aggregating the correlation weights into a package feature representation for representing the features of the CT image; step S3, predicting the packet characteristic representation obtained in the step S2) as the category of the CT image through a classifier; Step S4), an example with high weight value is selected according to the example correlation weight obtained in the step S2), and is evaluated through a classifier, so that the association degree of the learning content and the expected content of the network is represented; Step S5, constructing a loss function, wherein the loss function comprises packet classification loss and association loss, the packet classification loss represents the difference between a CT predicted value and a true value, the association loss represents the difference between model learning content and expected content, and an association loss formula is as follows: 。
2. 2. the method for classifying non-tubercular mycobacterial lung diseases and tubercular CT images according to claim 1, wherein the specific steps of step S1) are as follows: Step S11, data in the training set sequentially flows through each convolution module in the network, an input image firstly enters an improved ResNet coder to extract characteristics, the fourth stage is removed by the ResNet coder ResNet, and the stacking layers of basic blocks in the first three stages are adjusted to be 3, 3 and 3 layers; step S12, the features of step S11) are sent to a local-global feature extraction module LGE to extract a feature map, the LGE module comprising 1 and 1 1 1. Packet convolution, consisting of two 3 3 3. A convolution branch consisting of a group convolution module and consisting of 3 3 3. Roll up volume and two 1 1 1. Self-attention branch and 1 formed by grouping convolution 1 1. Convolving the branches; step S13, dividing the feature map obtained in step S12) into examples along each feature pixel of length, width and height.
3. 3. The method of claim 1, wherein step S2) uses a feature aggregation module FA based on an attention mechanism.
4. 4. The method according to claim 1, wherein step S4) the classifier for evaluating the association degree of learning content and expected content of the network is pre-trained by lesion labeling and has fixed parameters.

Description

Method for classifying CT images of nontuberculosis mycobacteria lung diseases and phthisis Technical Field The invention relates to the technical fields of computer vision, medical image analysis and the like, and particularly relates to a CT image classification method for classifying non-tuberculosis mycobacterial lung diseases and tuberculosis. Background With the rapid improvement of computer computing power and the extensive research of deep learning, image classification technology is continuously developed, and the speed and accuracy are continuously leaved. However, these image classification methods are mainly applied to natural image classification tasks where the background is distinct from the foreground. The non-tuberculosis mycobacterial lung disease is highly similar to a tuberculosis CT image, and various imaging symptoms such as cavities, bronchiectasis, solid changes, nodules and the like are contained in the images of the non-tuberculosis mycobacterial lung disease and the tuberculosis CT image, part of the symptoms are highly similar to normal tissues of a human body (such as pleural thickening), and part of the symptoms are undersized and difficult to identify (such as single-shot nodules and calcification foci). The complex signs of non-tubercular mycobacterial lung disease and tubercular CT images make the current mainstream image classification algorithm difficult to exert its effective performance. The accurate classification of non-tubercular mycobacterial lung disease and tubercular CT images is of great value to many tasks. In view of the development of a multi-example learning framework, the concept of considering the category of an image as being commonly determined by a plurality of examples in the image is suitable for processing the image classification problem that a plurality of target categories in the image and each target commonly determine the category of the image. The multi-example learning architecture breaks through the one-to-one correspondence between the traditional supervised learning image categories and the targets in the images, can effectively establish the one-to-many correspondence between the image categories and the targets in the images, and is suitable for the condition that the non-tuberculosis mycobacterium lung diseases and the pulmonary tuberculosis CT images are multiple and complex. However, the current multi-example learning method is mainly applied to the classification task of the 2D pathological image, and no algorithm is used for solving the task of the 3DCT image and the complex sign. Aiming at the CT image classification problem of patients suffering from non-tuberculosis mycobacterial pulmonary diseases and tuberculosis patients, the invention provides a CT image classification method for the non-tuberculosis mycobacterial pulmonary diseases and tuberculosis patients, which realizes the accurate classification of the CT image classification of the non-tuberculosis mycobacterial pulmonary diseases and tuberculosis. Reference to the literature [1]Ratnatunga C N,Lutzky V P,Kupz A,et al.The rise of non-tuberculosis mycobacterial lung disease[J].Frontiers in immunology,2020,11:303. [2]Nasiri M J,Dabiri H,Darban-Sarokhalil D,et al.Prevalence of non-tuberculosis mycobacterial infections among tuberculosis suspects in Iran:systematic review and meta-analysis[J].PloS one,2015,10(6):e0129073. [3]Chu H Q,Li B,Zhao L,et al.Chest imaging comparison between non-tuberculous and tuberculosis mycobacteria in sputum acid fast bacilli smear-positive patients[J].Eur Rev Med Pharmacol Sci,2015,19(13):2429-2439.. [4] Wu Xiaoguang, gaomanqiu Ma Liping. 50 cases of non-tubercular mycobacteriosis clinical analysis [ J ]. J. Of Chinese anti-tuberculosis, 2009,31 (8): 481. [5] Lai Yanfen, wu Dongling, yanglin. Non-tuberculosis Mycobacterium pulmonary disease 50 cases misdiagnosis analysis [ J ]. Proc. National medical college, 2014,36 (3): 374-375. Disclosure of Invention In order to solve the problem of classification of CT images of nontuberculous mycobacteria lung diseases and phthisis in the prior art, the invention provides a classification method of CT images of nontuberculous mycobacteria lung diseases and phthisis, which is used for realizing classification of nontuberculous mycobacteria lung diseases and phthisis. The technical scheme of the invention is as follows: a method for classifying CT images of nontuberculosis mycobacterial lung diseases and phthisis comprises the following steps: Step S1, extracting features of a non-tuberculosis mycobacterium lung disease and a tuberculosis CT image through a neural network, and taking each feature pixel of the extracted feature image divided along the length, the width and the height as an example, wherein each example represents feature information in a certain area of the image. Step S2), calculating the correlation weights of the examples obtained in the step S1) and the classification tasks, and performin