CN-122023354-A - Idiopathic pulmonary fibrosis intelligent diagnosis method based on structure-texture decoupling and multi-scale state space modeling

Abstract

The invention discloses an intelligent diagnosis method for idiopathic pulmonary fibrosis based on structure-texture decoupling and multi-scale state space modeling. The method comprises the steps of respectively carrying out directional modeling on low-frequency anatomical structure information and high-frequency fibrosis texture information through structure-texture understanding coupling branches, realizing self-adaptive fusion by utilizing a dynamic gating mechanism so as to improve the distinguishing capability of different fibrosis phenotypes, constructing a multi-scale state space modeling module, efficiently describing long-range dependency relations of lung tissues on different spatial scales, realizing cooperative characterization of local details and global structures, setting a local context awareness attention mechanism in a feature fusion stage, generating spatial and channel joint modulation mapping based on local statistical characteristics, strengthening low-contrast focus area response and inhibiting background tissue interference. Experimental results show that the method has higher stability and robustness in intelligent diagnosis of idiopathic pulmonary fibrosis, and provides reliable technical support for auxiliary diagnosis.

Inventors

• SU SHUZHI
• ZHU KAI
• ZHU YANMIN
• YANG MENGYANG
• DAI YONG

Assignees

• 安徽理工大学

Dates

Publication Date

20260512

Application Date

20260202

Claims (4)

1. 1. An intelligent diagnosis method for idiopathic pulmonary fibrosis based on structure-texture decoupling and multi-scale state space modeling is characterized by comprising the following steps: (1) Acquiring and preprocessing idiopathic pulmonary fibrosis CT image data, acquiring chest CT image data of a patient suffering from pulmonary fibrosis and idiopathic pulmonary fibrosis from a medical institution, and performing spatial resampling, gray scale normalization and pulmonary parenchyma region extraction processing on an original image to eliminate scale difference and intensity offset caused by different scanning parameters and obtain a standardized input feature map; (2) Constructing a structure-texture decoupling feature module, respectively carrying out directional modeling on low-frequency anatomical structure information and high-frequency fibrosis texture information according to the characteristic of high coupling of anatomical structure information and pathological texture information in an input feature map, and carrying out self-adaptive fusion on the structure feature and the texture feature through a dynamic gating mechanism so as to reduce the interference of structure-texture aliasing on idiopathic pulmonary fibrosis parting discrimination; (3) Constructing a multi-scale state space feature module, modeling a long-range space dependence relation of lung tissue features by utilizing a plurality of state space modeling branches with different space scales on the basis of structure-texture understanding coupling features, and forming multi-scale context feature representation while keeping low computational complexity and considering the local focus details and global structure consistency of cross lung lobes; (4) The method comprises the steps of constructing a local context perception feature enhancement module, modeling local neighborhood statistical characteristics of multi-scale features, generating self-adaptive modulation weights in space dimension and channel dimension, carrying out key enhancement on a fibrosis focus area with low contrast and fuzzy boundary, simultaneously inhibiting invalid response of a non-focus background area, and fusing and mapping the features to obtain a corresponding idiopathic pulmonary fibrosis CT image typing result.
2. 2. The intelligent diagnosis method for idiopathic pulmonary fibrosis based on structure-texture decoupling and multi-scale state space modeling according to claim 1, wherein the structure-texture decoupling feature module in step (2) performs separation modeling on low-frequency anatomical structure information and high-frequency fibrotic texture information, and performs adaptive fusion of two types of features through a dynamic gating mechanism, so as to obtain a feature representation with stronger discriminant and purer semantics, and the steps are performed as follows: (2a) Low frequency anatomical feature modeling The input feature map is: , The height of the feature map is indicated, The width of the feature map is represented, For low-frequency structural information such as prominent lung lobe outline, bronchus running and lung parenchyma integral form, large receptive field space modeling is carried out on input characteristics under the condition that the channels are independent, and the structural characteristics are expressed as follows: Wherein, the A five-by-five spatial convolution kernel representing the channel independence, for modeling low frequency anatomy, Representing a channel-by-channel spatial convolution operation, Representing a point-by-point convolution, for channel linear mapping and information reorganization, Representing a batch normalization operation, for stabilizing the feature distribution, Representing a nonlinear activation function, for enhancing feature expression capabilities, Representing the extracted structural features; (2b) High frequency fibrotic texture feature modeling Aiming at the high-frequency edge and fine grain texture changes corresponding to fibrosis pathological manifestations such as honeycomb shadows, grid shadows, ground glass shadows and the like, a high-frequency texture modeling branch is constructed, texture information is emphasized through spatial convolution with edge sensitivity characteristics, and the texture characteristics are expressed as follows: Wherein, the Representing a three-by-three channel independent spatial convolution kernel, whose initial weights have edge enhancement properties for highlighting high frequency texture responses, Representing a point-wise convolution, for texture channel mapping, Representing extracted fibrotic texture features that are more sensitive to local contrast changes and microstructure anomalies; (2c) Structure-texture dynamic gating fusion mechanism In order to avoid feature redundancy and conflict caused by simple superposition, a dynamic gating mechanism is constructed, self-adaptive weight distribution is carried out on structural features and texture features, and channel-level splicing is carried out on two types of features: Wherein, the Representing a feature stitching operation in a channel dimension; then, generating fusion weights through a gating mapping function: Wherein, the A point-by-point convolution kernel representing the gating map, Representing a normalized activation function for constraining the gating weights to stable intervals, And on the basis, completing the weighted fusion of the structural features and the texture features to obtain the decoupled joint feature representation: Wherein, the Representing the output characteristics of the structure-texture decoupling, Representing an element-by-element weighting operation; (2d) Feature mapping and stabilization processing To further enhance consistency and discriminant of feature expression, nonlinear mapping and channel reforming are performed on the fusion features to obtain final output: Wherein, the Representing the output map convolution kernel, Representing the final output characteristics of the structure-texture decoupling module, which significantly improves the ability to distinguish between different pulmonary fibrosis phenotypes while reducing structure and texture aliasing interference.
3. 3. The intelligent diagnosis method for idiopathic pulmonary fibrosis based on structure-texture decoupling and multi-scale state space modeling according to claim 1, wherein the constructing a multi-scale state space feature module in step (3) aims at the problems that focus in pulmonary fibrosis CT images presents cross-scale distribution, long-range space dependence is obvious, and traditional convolution is difficult to consider the calculation complexity and modeling range, the constructing a multi-scale state space feature module in the invention uniformly models the context relation of pulmonary tissues under different spatial scales on the basis of structure-texture understanding coupling features so as to realize the cooperative expression of local detail perception and global structural consistency, and the steps are as follows: (3a) State space feature mapping and channel retention modeling The structure-texture decoupling characteristic of the output is provided as follows , And (3) with The spatial dimensions of the feature map are represented separately, To model the space dependence relationship on the premise of maintaining the channel independence, the state space mapping is carried out on the input features, and the single-scale state modeling process is expressed as follows: Wherein, the Represent the first A state space convolution kernel at a dimension that models channels independently in a spatial dimension for capturing local to mid-distance spatial dependencies, Representing a channel-by-channel spatial modeling operation, Representing a channel map convolution kernel for linearly recombining state features, Representing the non-linear mapping function, Represent the first State feature representation at the individual spatial scale; (3b) Multi-scale state space parallel modeling For simultaneously describing the fibrosis focus morphology and the space distribution characteristics under different scales, constructing a plurality of parallel state space branches, and modeling the same input characteristic under different space scales to obtain a multi-scale state characteristic set: Wherein, the Representing small-scale state characteristics, is used for capturing microscopic pathological characteristics such as tiny grid shadows, local texture anomalies and the like, Representing mesoscale state features for modeling changes in structural continuity within a lung segment or lobe, Representing large-scale state characteristics, and being used for representing the overall structure consistency and focus expansion trend of the cross-lung lobes, the parallel modeling mode realizes the unified depiction of multi-scale space dependence on the premise of not remarkably increasing the computational complexity; (3c) Multi-scale state feature fusion and compression In order to avoid multi-scale feature redundancy and enhance semantic consistency, channel-level fusion and compression processing are carried out on state features under different scales to obtain unified multi-scale context representation: Wherein, the Representing a feature stitching operation in the channel dimension, Representing a fusion map convolution kernel for compressing multi-scale features and accomplishing semantic alignment, Representing a non-linear activation function, Representing the characteristics of the fused multi-scale state space; (3d) Stabilization and residual modeling of multi-scale state features In order to ensure the numerical stability and deep feature transmission capability of the multi-scale modeling process, the fused state features and the input features are subjected to residual correlation to form final multi-scale state space output: Wherein, the The invention further constructs stable, continuous and globally consistent space characterization capability on the structure-texture decoupling characteristic by the multi-scale state space modeling mechanism, and effectively improves the characterization precision and parting robustness of the model on the complex idiopathic pulmonary fibrosis distribution mode.
4. 4. The intelligent diagnosis method for idiopathic pulmonary fibrosis based on structure-texture decoupling and multi-scale state space modeling according to claim 1, wherein the constructing a local context-aware feature enhancement module in step (4) is performed to solve the problem that a focus area in a pulmonary fibrosis CT image is generally low in contrast, blurred in boundary and easily confused with a normal blood vessel or pleura structure, by explicitly modeling local neighborhood statistical characteristics, the invention constructs a local context-aware feature enhancement module by modulating multi-scale state space features in spatial dimension and channel dimension in combination to enhance key focus area response and suppress non-focus background interference, and the steps are performed as follows: (4a) Local statistical property modeling Let the output multi-scale state space characteristics be expressed as , And (3) with The spatial dimensions of the feature map are represented separately, To describe contrast change characteristics of fibrosis focus in local neighborhood, local statistical modeling is carried out on the characteristic diagram in space dimension to obtain local mean characteristic and local variance characteristic, and the expression form is as follows: Wherein, the Representing the index of the spatial position in the feature map, The channel index is represented as a function of the channel index, Expressed in terms of Is a local neighborhood window that is centered, Representing the number of pixels within the window, Representing local mean features, for reflecting the overall response level of the region, The local variance information is represented and used for measuring the local structure and the texture change intensity; further, local standard deviation features are obtained by square root mapping: Wherein, the Is a numerical stability term and is used for avoiding the situation that the denominator is zero; (4b) Spatial dimension context enhancement To highlight focal areas with significant local variation, the local standard deviation features are normalized, and spatial attention weights are generated: Wherein, the Representing the normalized local contrast characteristics, The spatial projection mapping parameters are represented and, Representing the Sigmoid mapping function, Representing a spatial dimension attention weighting map; Based on the spatial attention weight, the original features are spatially enhanced: Wherein, the Representing an element-by-element multiplication operation, Representing the spatially context enhanced feature representation; (4c) Channel dimension context modulation To further model the contribution differences of different semantic channels to pulmonary fibrosis typing, global convergence is performed on local contrast features, and channel-level modulation weights are constructed: Wherein, the Representing the channel-level context description vector, And (3) with The channel map parameters are represented as such, Representing a non-linear activation function, Representing channel dimension modulation weights; And (3) completing feature recalibration based on channel weight: Wherein, the Representing enhanced features that fuse spatial and channel context information simultaneously; (4d) Residual fusion and discriminant feature output In order to keep original semantic information and stabilize feature distribution, residual fusion is carried out on the enhanced features and the input features, so that a final discrimination representation for pulmonary fibrosis typing is formed: Wherein, the The invention effectively strengthens the distinguishing response of the low-contrast fibrosis focus on the basis of the multi-scale state space characteristics through the local statistical modeling, the space-channel joint modulation and the residual error stabilization mechanism, inhibits the background structure interference and provides more stable, fine and clinically explanatory feature support for the intelligent diagnosis of the idiopathic pulmonary fibrosis.

Description

Idiopathic pulmonary fibrosis intelligent diagnosis method based on structure-texture decoupling and multi-scale state space modeling Technical Field The invention relates to an intelligent diagnosis method for idiopathic pulmonary fibrosis based on structure-texture decoupling and multi-scale state space modeling, and belongs to the field of intelligent analysis of medical images of pulmonary fibrosis. Background Pulmonary fibrosis (Pulmonary Fibrosis, PF) and idiopathic pulmonary fibrosis (Idiopathic Pulmonary Fibrosis, IPF) are chronic pulmonary diseases characterized by progressive pulmonary interstitial fibrosis and sustained decline of pulmonary function, and are characterized by hidden onset, rapid progression and very poor prognosis, which are one of the important causes of respiratory failure and death. High Resolution Computed Tomography (HRCT) is the main imaging means for current diagnosis and typing evaluation of pulmonary fibrosis, and clinically, imaging physicians are usually relied on to comprehensively judge characteristics such as glass grinding, grid imaging and honeycomb-like change in CT images. However, the pulmonary fibrosis image has complex phenotype, the focus presents diversified structural morphology and texture characteristics in different development stages, and early focus often has lower contrast and scattered distribution, so that the manual interpretation process is highly dependent on experience, and the problems of strong subjectivity, poor repeatability and the like exist. In recent years, the automatic lung fibrosis recognition and typing method based on deep learning improves the image analysis efficiency to a certain extent, but most of the existing methods are used for integrally modeling image features, and can not effectively distinguish lung anatomical structure information from fibrosis focus texture information, so that the structural features and the texture features are mutually interfered, and the problem of judging performance degradation easily occurs under complex background or weak focus scenes. Meanwhile, the pulmonary fibrosis focus has obvious multiscale spatial distribution characteristics, not only comprises local fine grain texture changes, but also relates to structural remodeling of lung levels and even whole lung ranges, and the traditional model has limited capacity in the aspects of cross-scale feature modeling and long-range spatial dependence depiction, and is difficult to simultaneously consider local detail and global information. Furthermore, the local context information of early stage pulmonary fibrosis lesions is insufficient, and the characterization capability of low contrast areas by existing methods is still limited. Aiming at the problems, it is necessary to provide an intelligent diagnosis method for idiopathic pulmonary fibrosis, which can effectively decouple structural features and fibrotic texture features of the lung and give consideration to multi-scale spatial characteristics and local context information, so as to improve the accuracy and stability of intelligent diagnosis for idiopathic pulmonary fibrosis. Disclosure of Invention Aiming at the problems of high coupling of focus structure and texture in idiopathic pulmonary fibrosis CT images, cross-scale spatial distribution, low local contrast of early focus and the like, the invention provides an idiopathic pulmonary fibrosis intelligent diagnosis method based on structure-texture decoupling and multi-scale state space modeling. The method realizes stable characterization and automatic typing of different pulmonary fibrosis phenotypes by decoupling the lung structure information and the fibrosis texture information and under the synergistic effect of multi-scale space state modeling and local context enhancement. The specific implementation steps of the invention are as follows: 1. The lung CT image acquisition and characteristic representation construction comprises the steps of acquiring a clinical lung CT image data set, covering lung fibrosis cases and control samples with different disease courses and different phenotypes, carrying out standardized pretreatment on original image data, firstly screening out effective sections with slice thickness less than or equal to 3mm and lung window parameters meeting clinical diagnosis standards, window level WL ranging from-800 to-200 HU and window width WW ranging from 500 to 2000HU, removing invalid data with parameters missing and serious artifacts, and then carrying out pixel interval standardization, unifying 1.0mm and size resampling to obtain an input image tensor with unified specification . Wherein, the AndThe spatial height and width of the image are respectively represented,Indicating the number of channels. Convolution mapping function through Stem moduleThe input image is subjected to feature extraction, and the module adopts 7×7 convolution to combine batch normalization and GELU activation functions, so that t