Pneumoconiosis grading recognition method based on lesion distribution and density point cloud learning

Abstract

This invention pertains to intelligent medical image processing technology, specifically a method for grading and identifying pneumoconiosis based on lesion distribution and density point cloud learning. The method includes the following steps: acquiring chest CT data and performing lung field segmentation and connectivity-aware vascular suppression; converting voxels within the suppressed lung fields into candidate points; obtaining an evidence-enriched point cloud through joint sampling based on the spatial structural features and HU attenuation attributes of the candidate points; jointly encoding the three-dimensional coordinates and HU attenuation attributes of the evidence-enriched point cloud to obtain regional lesion features and a whole-lung descriptor; obtaining regional lesion features and whole-lung distribution features through distribution-guided correction; constructing a density-aware attention model based on self-attention to obtain density-enhanced regional lesion features; and fusing and classifying the density-enhanced regional lesion features and whole-lung distribution features for graded identification of pneumoconiosis. This invention can reduce vascular interference and synergistically express the whole-lung lesion distribution and lesion density.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image processing technology, specifically involving a method for graded identification of pneumoconiosis based on learning of lesion distribution and density point cloud. Background Technology

[0002] With the continuous development of occupational health monitoring and medical imaging computer processing technologies, the timely staging and classification of pneumoconiosis has received much attention due to its importance for disease monitoring and clinical management. On the other hand, chest computed tomography (CT) scan, as one of the important technologies in the imaging assessment of pneumoconiosis, provides three-dimensional anatomical structures of the lungs and attenuation information based on Hounsfield Units (HUs) through continuous tomographic imaging, providing a technical basis for the spatial observation and staging analysis of small opacities related to pneumoconiosis.

[0003] However, in chest computed tomography (CT) data, pneumoconiosis-related small opacities are usually sparse and may resemble pulmonary vascular cross-sections in morphology and attenuation range, making their identification and staging still insufficient and easily affected by vascular structures and numerous normal lung parenchyma voxels. Furthermore, the location and extent of lesions vary significantly across different cases, and the density of small opacities changes continuously between adjacent stages. The identification and staging of pneumoconiosis based on chest CT still faces the following challenges:

[0004] 1) Vascular interference problem. In the lung field, pulmonary vessels and small shadows associated with pneumoconiosis may have similar local morphology and HU attenuation properties. Fixed threshold extraction may easily select vascular structures into the candidate region as well, while directly suppressing all tubular structures may delete small shadows related to the stage, thus interfering with the lesion evidence in the point cloud.

[0005] 2) Problems in modeling the distribution of pneumoconiosis lesions. Small opacities associated with pneumoconiosis may be distributed in a scattered, widespread or diffuse manner. Early lesions are particularly sparse and their locations are not fixed. Processing by slice or only aggregating local neighborhoods will weaken the continuity between slices and the connection between different lung regions, making it difficult to form a case-level representation of the distribution of lesions in the whole lung.

[0006] 3) Complexity of density discrimination. Since the number of small shadows, the degree of aggregation, and the HU attenuation may only change slightly between adjacent pneumoconiosis stages, and this change is a continuous transition, the extraction of density-related differences becomes particularly complex when aggregation is based solely on the learned similarity between regional features.

[0007] Existing technologies are unable to effectively address these challenges, resulting in an impact on the density of stage-related evidence, the distribution of lung lesions, and the ability to distinguish adjacent stages in chest CT data.

[0008] Although some studies have attempted to address these issues, most have employed two-dimensional slice analysis, dense three-dimensional volume data processing, or local aggregation of point clouds for general shapes. These approaches are insufficient to simultaneously meet the requirements for pneumoconiosis classification and identification, which include suppressing continuous vascular interference, preserving evidence of sparse lesions, characterizing lesion distribution patterns, and identifying subtle changes in the density of small shadows. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention proposes a method for classifying and identifying pneumoconiosis based on lesion distribution and density point cloud learning. It utilizes the three-dimensional coordinates and HU attenuation attributes of chest CT data to construct an evidence-enriched point cloud, builds a density-aware attention model, and enhances staging-related features through collaborative learning of whole-lung lesion distribution and regional small shadow density. This improves the utilization of sparse lesion evidence and the ability to distinguish adjacent stages, effectively solving the problems of vascular structure interference, difficulty in expressing whole-lung distribution, and subtle density differences. It achieves graded identification of normal, stage I, stage II, and stage III pneumoconiosis.

[0010] On one hand, the pneumoconiosis grading and identification method based on lesion distribution and density point cloud learning in this embodiment of the invention includes the following steps:

[0011] S1. Acquire chest CT volume data and perform lung field segmentation, calculate multi-scale vascular response, and perform connectivity-sensing vascular inhibition based on multi-scale vascular response; convert voxels in the inhibited lung field into candidate points, and perform joint sampling based on the spatial structural features and HU attenuation properties of the candidate points to obtain evidence-enriched point cloud.

[0012] S2. Jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-rich point cloud to obtain regional lesion features and whole-lung descriptors; then, correct the regional lesion features and whole-lung descriptors through a two-level distribution to obtain the two-level distribution-guided correction of regional lesion features and whole-lung distribution features.

[0013] S3. Align the regional lesion features after two-level distribution guided correction with the corresponding HU attenuation attributes in intensity, and construct a density-aware attention model based on the self-attention mechanism. The constructed density-aware attention model includes:

[0014] The HU attribute encoding and modulation module is used to map the intensity-aligned HU attenuation attribute to the latent space and generate intensity gating with restricted values ​​and intensity bias with positive and negative adjustment directions.

[0015] The perceptual attention aggregation module is used to project the intensity-aligned regional lesion features into query features. Key features Sum value characteristics Simultaneously, based on intensity gating and intensity bias modulation attention calculations, density-aware attention weights are obtained; these density-aware attention weights are then used to log the value features. Weighted aggregation is performed to obtain attention-aggregated features;

[0016] The differential feature enhancement module is used to subtract the intensity-aligned regional lesion features from the attention aggregation features to obtain differential features; the differential features are then processed through convolution, batch normalization, and activation functions to enhance the differential information, resulting in enhanced differential features.

[0017] The residual back-injection module is used to add the enhanced differential feature residuals to the intensity-aligned regional lesion features to obtain the density-enhanced regional lesion features;

[0018] S4. The lesion features of the density-enhanced region and the whole lung distribution features are fused through the classification module to obtain the staged aggregate features. The staged aggregate features are then classified to obtain the category probability of pneumoconiosis identification and staging, as well as the corresponding graded identification results.

[0019] On the other hand, the pneumoconiosis grading and identification system based on lesion distribution and density point cloud learning provided in this embodiment of the invention includes:

[0020] The evidence-enriched point cloud generation unit is used to acquire chest CT volume data and perform lung field segmentation, calculate multi-scale vascular response, and perform connectivity-sensing vascular inhibition based on the multi-scale vascular response; voxels in the inhibited lung field are converted into candidate points, and joint sampling is performed based on the spatial structural features and HU attenuation properties of the candidate points to obtain the evidence-enriched point cloud.

[0021] A two-level distribution-guided correction unit is used to jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-rich point cloud to obtain regional lesion features and whole-lung descriptors; the regional lesion features and whole-lung descriptors are then corrected by a two-level distribution to obtain the two-level distribution-guided correction regional lesion features and whole-lung distribution features.

[0022] The density-aware attention model aligns the intensity of the two-level distribution-guided correction regional lesion features with the corresponding HU attenuation attributes of the region, and performs density perception based on the self-attention mechanism to obtain density-enhanced regional lesion features.

[0023] The classification module fuses the lesion features of the densely concentrated area and the whole lung distribution features to obtain the staging aggregate features, and classifies the staging aggregate features to obtain the category probability of pneumoconiosis identification and staging, as well as the corresponding grade identification results;

[0024] Among them, density-aware attention models include:

[0025] The HU attribute encoding and modulation module is used to map the intensity-aligned HU attenuation attribute to the latent space and generate intensity gating with restricted values ​​and intensity bias with positive and negative adjustment directions.

[0026] The perceptual attention aggregation module is used to project the intensity-aligned regional lesion features into query features. Key features Sum value characteristics Simultaneously, based on intensity gating and intensity bias modulation attention calculations, density-aware attention weights are obtained; these density-aware attention weights are then used to log the value features. Weighted aggregation is performed to obtain attention-aggregated features;

[0027] The differential feature enhancement module is used to subtract the intensity-aligned regional lesion features from the attention aggregation features to obtain differential features; the differential features are then processed through convolution, batch normalization, and activation functions to enhance the differential information, resulting in enhanced differential features.

[0028] The residual back-injection module is used to add the enhanced differential feature residuals to the intensity-aligned regional lesion features to obtain densely enhanced regional lesion features.

[0029] The technical solution of this invention employs a point cloud construction for enriched evidence, bidirectional correction of lesion distribution, and density-aware attention modeling. The constructed density-aware attention model effectively reduces voxel and vascular interference unrelated to staging, enhancing the synergistic expression of whole-lung distribution evidence and regional small opacity density evidence. Compared to existing technologies, the specific technical effects achieved by this invention include:

[0030] 1. Constructing an evidence-enriched point cloud: Chest CT volume data contains a large number of normal lung parenchyma and continuous vascular voxels. The purpose of point cloud construction is to preserve the three-dimensional structure and attenuated evidence related to the identification of pneumoconiosis in sparse point clouds.

[0031] Existing point cloud construction methods mainly include random sampling and fixed HU threshold sampling. Random sampling may miss sparse small shadows with variable locations, while fixed HU threshold sampling may select vascular structures that overlap with the attenuation range of small shadows. It cannot simultaneously achieve lesion preservation and vascular interference suppression.

[0032] The evidence-enrichment point sampling strategy employed in this invention can adapt to vascular-like structures of different diameters through multi-scale Hessian-Frangi responses, and form two types of connected components using candidate thresholds and more stringent anchoring thresholds. When a candidate connected component intersects with a size-filtered vascular anchoring set, it is suppressed. Then, the lung point cloud is formed by combining structural saliency and HU attenuation attribute scores with farthest point sampling, thereby reducing continuous vascular interference and increasing the proportion of graded identification-related evidence in the fixed point set.

[0033] 2. Two-way correction for lesion distribution: Small opacities associated with pneumoconiosis can be distributed in different lung regions. Regional lesion characteristics need to be linked with the case-level whole-lung context in order to describe the scattered, widespread or diffuse spatial distribution.

[0034] Common point cloud feature aggregation methods are mainly aimed at local neighborhoods or object-level shapes. These methods can gather information from neighboring points, but it is difficult to continuously update the whole lung distribution representation with sparse regional lesion evidence. In addition, it is difficult to establish mutual constraints between regional evidence and the whole lung context by unidirectionally adding global vectors to regional features.

[0035] This invention proposes a two-level distribution-guided correction technique. Distribution tokens are used to query regional lesion tokens to aggregate lesion evidence from different lung regions. Simultaneously, regional lesion tokens are used to query distribution tokens to correct regional responses using case-level context. Furthermore, local abstraction is performed between the two levels of correction, and bidirectional mutual correction is repeated within each level. Through iterative interaction between regional evidence and the whole-lung context, whole-lung lesion distribution characteristics are formed, improving the expressive power of sparse lesion spatial distribution.

[0036] 3. Dense-perceived attention modeling: The HU attenuation attribute in chest CT body data reflects the attenuation characteristics of tissue to X-rays. The density of small shadows in adjacent pneumoconiosis stages changes continuously. If regional self-attention is established based solely on the similarity of learned features, subtle differences in intensity and aggregation can easily be smoothed out.

[0037] To address the difficulty in extracting density differences, this invention aligns the lesion tokens of each region with the HU attenuation attribute of the corresponding region center. It generates a value-restricted intensity gating and an intensity bias with positive and negative adjustment directions from the same HU input, enabling them to participate in regional self-attention. Then, it learns and enhances the difference between the original regional response and the attention aggregation context. After reinjecting the enhanced result residual, it fuses it with the whole lung distribution characteristics, thereby highlighting the regional response related to the density of small shadows. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the pneumoconiosis grading and identification method based on lesion distribution and density point cloud learning in an embodiment of the present invention. Detailed Implementation

[0039] The pneumoconiosis grading identification method of this invention first addresses the problem of easy confusion between vascular structures and small pneumoconiosis-related shadows in chest CT data in terms of morphology and HU attenuation attributes. It constructs an evidence-enriched point cloud through lung field segmentation, multi-scale vascular estimation, dual-threshold connectivity analysis, and joint sampling of structural saliency and HU attenuation attributes. Then, it performs bidirectional mutual correction between regional lesion features and whole-lung context through two-level distribution-guided correction to characterize the spatial distribution of pneumoconiosis-related lesions. Finally, it enhances subtle density differences between adjacent stages through intensity gating, intensity bias, and density-aware attention. After fusing whole-lung distribution features, it outputs the category probability of normal, pneumoconiosis stage I, pneumoconiosis stage II, or pneumoconiosis stage III and the corresponding grading identification result.

[0040] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0041] Example

[0042] This embodiment provides a method for classifying and identifying pneumoconiosis based on lesion distribution and density point cloud learning, applicable to computer-aided identification and staging of chest CT cases, such as... Figure 1 As shown, the specific implementation steps include:

[0043] S1. Acquire chest CT volume data and segment the lung field, calculate multi-scale vascular response, and perform connectivity-sensing vascular inhibition based on the multi-scale vascular response; convert voxels in the inhibited lung field into candidate points, and perform joint sampling based on the spatial structural features and HU attenuation properties of the candidate points to obtain evidence-enriched point clouds.

[0044] Overall, this step reduces continuous vascular interference through multi-scale vascular computation and connectivity-sensing vascular inhibition, and then constructs an evidence-rich point cloud for the lungs based on the structural salience of candidate points and HU attenuation properties.

[0045] Specifically, this step first calculates the vascular response of each voxel in the lung field at multiple spatial scales to obtain multi-scale vascular responses; then, candidate masks and anchoring masks are formed based on candidate thresholds and anchoring thresholds, respectively, and three-dimensional connected component analysis is performed on the candidate masks and anchoring masks to obtain candidate connected components and anchored connected components. A vascular anchoring set is formed based on the size of the anchored connected components, and candidate connected components that intersect with the vascular anchoring set are suppressed to inhibit vascular interference; then, based on the structural salience of the candidate points and the HU attenuation properties, the candidate points are sampled, scored, pre-selected, and farthest point sampled to obtain the evidence-enriched point cloud of the lungs.

[0046] This step enriches the grading-related evidence in chest CT data. Small opacities associated with pneumoconiosis and pulmonary vessels may share similar local morphology and HU attenuation ranges. Directly retaining all high-response voxels would introduce continuous vascular structures, while directly deleting all high-response voxels might inadvertently delete sparse small opacities. This step employs multi-scale Hessian-Frangi vascular estimation, dual thresholding for candidate and anchor points, anchor size filtering, and connected component intersection determination, suppressing only candidate vascular-like components connected to high-confidence vascular anchor sets. Subsequently, candidate points are jointly evaluated based on structural significance and HU attenuation properties, and farthest-point sampling is performed within the pre-selected point set.

[0047] In this embodiment, the specific process of constructing evidence enrichment points is as follows:

[0048] S11. For the segmented chest CT volume data of the lung field, calculate the scale-normalized Hessian matrix at multiple spatial scales, perform eigenvalue decomposition on the scale-normalized Hessian matrix, and take the maximum value among the Frangi vascular responses at each scale to obtain the multi-scale vascular response, as follows:

[0049] ;

[0050] in, This represents chest CT scan data after lung field segmentation. Represents the three-dimensional coordinates of a voxel. Indicates spatial scale Indicates spatial scale Gaussian kernel, convolution symbol " " represents spatial convolution, This represents the Hessian operator for calculating the second-order spatial derivative. Represents the scale-normalized Hessian matrix; Represents a scale set, , and This represents the Hessian eigenvalues ​​arranged in ascending order of absolute value; This represents the Frangi vascular response function, which generates multi-scale vascular responses based on local curvature relationships. , These represent the adjustment parameters in the Frangi vascular response function used to control the suppression of plate-like structures, mass-like structures, and background noise, respectively.

[0051] S12. Based on the distribution of multi-scale vascular responses within the lung fields, determine candidate thresholds and anchoring thresholds, where the percentile level corresponding to the anchoring threshold is higher than the percentile level corresponding to the candidate threshold; and use the candidate threshold and anchoring threshold to form candidate masks and anchoring masks respectively, specifically:

[0052] ;

[0053] in, Indicates the candidate threshold. This indicates the anchoring threshold, which is more stringent than the candidate threshold. , Both are percentile functions, representing the th percentile, Indicates the percentile level of the candidate threshold. This represents the percentile level of the anchoring threshold and satisfies... ; This indicates a candidate mask used to preserve a wide range of vascular-like structures; This indicates an anchoring mask used to provide a high-confidence continuous vascular segment.

[0054] S13. Perform three-dimensional connected component analysis on the candidate mask and the anchoring mask respectively. Define spatially connected voxels in the candidate mask as candidate connected components, and define spatially connected voxels in the anchoring mask as anchored connected components. Remove isolated components with sizes smaller than the minimum anchoring size from the anchored connected components. Merge the anchored connected components whose sizes meet the anchoring conditions into a vessel anchoring set, as shown in the following equation:

[0055] ;

[0056] Next, determine whether each candidate connected component intersects with the vessel anchoring set:

[0057] ;

[0058] in, Indicates anchored connected components, Indicates candidate connected components, Indicates the minimum anchoring size. Indicates vascular anchoring set, A connectivity index representing candidate connected components. Indicates intersection, Indicates the empty set; This indicates that the union operation is performed on all anchored connected components whose size is not less than the minimum anchor size; This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise.

[0059] For each candidate connected component, suppression is only applied when it intersects with the set of vascular anchors.

[0060] S14. When a candidate connected component intersects with the vessel anchoring set, the corresponding voxel is assigned a background sentinel value to form a background voxel. The remaining non-background voxels are converted into candidate points containing three-dimensional coordinates and HU attenuation attributes. Eigenvalue analysis of the covariance matrix is ​​performed on the nearest neighbor coordinates of each candidate point to determine structural saliency. The structural saliency is weighted with the HU attenuation attribute to obtain a sampling score. Candidate points are pre-selected based on the sampling score to obtain a pre-selected point set. The farthest point sampling is performed on the pre-selected point set to obtain an evidence-rich point cloud, specifically:

[0061] ;

[0062] in, Represents the candidate point set, From three-dimensional coordinates and HU attenuation properties constitute; Indicates the first The eigenvalues ​​of the nearest neighbor coordinate covariance matrix of each candidate point, and having ; Indicates structural saliency. Indicates the first Sampling scores of candidate points Indicates the structural significance contribution coefficient. Evidence-enriched point clouds representing a fixed number of points; This represents the number of candidate points retained after sampling from the farthest point, i.e., the number of fixed points in the evidence-enriched point cloud.

[0063] In this embodiment, the scale set Pick Individuals, parameters , and The values ​​were taken as 0.5, 0.8, and 30.0 respectively; the structural significance contribution coefficient. A value of 0.6 is used to retain the top 50,000 candidate points based on the sampling score, and 32,768 points are sampled from the farthest point. The candidate threshold percentile, anchoring threshold percentile, minimum anchoring size, and number of nearest neighbors can be set according to the response distribution and point density of the training data.

[0064] This step reduces continuous vascular interference by using multi-scale vascularity and vascular anchoring connectivity, and increases the proportion of staging-related evidence in the fixed-point set by using joint sampling of structure-HU attenuation properties.

[0065] In practical applications, chest CT volume data can come from a single received case. After verifying the slice position, voxel spacing, and rescaling parameters of the chest CT volume data, three-dimensional volume data is obtained. The HU attenuation properties of each voxel are obtained based on the rescaling slope and rescaling intercept in the DICOM (Digital Imaging and Communications in Medicine) file.

[0066] S2. Jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-rich point cloud to obtain regional lesion features and whole-lung descriptors; then, correct the regional lesion features and whole-lung descriptors through a two-level distribution to obtain the two-level distribution-guided correction of regional lesion features and whole-lung distribution features.

[0067] In the two-level distribution-guided correction of this embodiment, the regional lesion features and the whole lung descriptor are sequentially subjected to the first-level distribution-guided correction, local abstraction and the second-level distribution-guided correction. The whole lung descriptor is mapped into multiple distribution tokens, and the regional lesion features are converted into regional lesion tokens. The distribution tokens gather lesion evidence from different lung regions, and at the same time, the regional lesion tokens are corrected using the case-level whole lung context (i.e., the whole lung descriptor), so as to realize the bidirectional mutual correction between regional lesion tokens and distribution tokens.

[0068] Specifically, the backbone network is used to jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-enriched lung point cloud to obtain the lesion features of the first scale region and the first whole lung descriptor; the lesion features of the first scale region and the first whole lung descriptor are subjected to first-level distribution-guided correction to obtain the lesion features of the first corrected region and the distribution features of the first whole lung; the lesion features of the first corrected region are locally abstracted to obtain the lesion features of the second scale region; the lesion features of the second scale region and the distribution features of the first whole lung are subjected to second-level distribution-guided correction to obtain the lesion features of the second corrected region and the distribution features of the second whole lung.

[0069] Each level of distributed guided correction is achieved through mutual correction of the core and output mapping, specifically including:

[0070] S21. Let the target token sequence be... The source token sequence is At each attention head, a query matrix is ​​formed from the target token, and a key matrix and a value matrix are formed from the source token. Multi-head cross-attention correction features from the source token to the target token are calculated, and residual connections, layer normalization, and feedforward network processing are performed on the correction features to obtain the updated target token sequence for the current direction, specifically represented as follows:

[0071] ;

[0072] in, Note the header index. This indicates the total number of heads, and , Indicates the first The attention head includes the query, key, and value vector dimensions. Indicates the target token sequence In the The query matrix obtained by projecting from the attention head. This represents the corresponding target query projection matrix; and These represent the source token sequence. In the The key matrix and value matrix obtained by projecting from the attention head. and These represent the corresponding source bond projection matrix and source value projection matrix, respectively. This represents the transpose of the key matrix. This represents the relevance score between the target token and the source token. This represents the factor used to scale the relevance score. This indicates that normalization is performed along the source token dimension to obtain attention weights; Indicates the first The source-to-target attention output of each attention head, i.e., the attention weights paired with the value matrix. The corrected features obtained by weighting. This indicates that the outputs of each attention head are concatenated along the feature dimension. This represents the output projection matrix that maps the splicing result back to the target feature space. This indicates a multi-head cross-attention correction feature where the source token is injected into the target token. The symbol "+" indicates the addition of residuals. This represents the intermediate target token sequence after attention sublayer residual connections and normalization. This indicates the layer normalization of the attention sublayer. This indicates a feedforward network where each target token is executed independently. This represents the updated target token sequence after residual connection and normalization in the feedforward sublayer. This indicates the layer normalization of the feedforward sublayer.

[0073] When the distribution token is used as the target, it gathers evidence of regional lesions; when the regional lesion token is used as the target, it receives the whole-lung distribution context. The two directions are computed in parallel on the same round of input. That is, in one correction direction, the distribution token is used as the target token and the regional lesion token is used as the source token; in the other correction direction, the distribution token is used as the source token and the regional lesion token is used as the target token.

[0074] When the distribution token is the target token and the regional lesion token is the source token, it is used to gather lesion evidence from the regional lesion token to the distribution token; when the regional lesion token is the target token and the distribution token is the source token, it is used to inject whole-lung distribution context from the distribution token to the regional lesion token.

[0075] S22. Repeat the preset number of correction iterations for the bidirectional mutual correction of the regional lesion token and the distribution token; perform mean pooling and linear mapping on the corrected distribution token, and perform residual fusion with the whole lung descriptor; at the same time, transpose the regional lesion token and the channel dimension after the iteration is completed, and restore it to the corrected regional lesion feature tensor.

[0076] The iterative process of mutual correction and the output mapping of whole-lung distribution characteristics and lesion characteristics in the correction area are specifically represented as follows:

[0077] ;

[0078] in, Indicates the current mutual correction round. Indicates the first The distribution tokens of the round input, Indicates the first The regional lesion token input in the round, Indicates the first The distributed tokens of the round output, Indicates the first The regional lesion tokens output in the round; This represents the mutual correction core that performs bidirectional cross-attention, residual connection, layer normalization, and feedforward network processing on distributed tokens and regional lesion tokens. This indicates the preset number of correction iteration rounds. Indicates the number of distributed tokens. Indicates the distributed token index; Indicates completion After the first round of correction Features of a distributed token across all batch samples Indicates to Sum of distributed tokens, Indicates to All distributed tokens are pooled. This indicates the current distribution characteristics of the entire lungs. This indicates the dimension of the distribution characteristics of the whole lung. This represents the channel or feature dimension of the regional lesion token. Indicates will dimensional pooled distributed token features linearly mapped to Dimensional operations, This represents the layer normalization of the output distribution throughout the lungs. This represents the updated whole-lung distribution feature obtained after fusing it with the current whole-lung distribution feature residual. Indicates completion Regional lesion tokens after round correction This indicates the transpose operation between the token dimension and the channel dimension. This indicates the lesion features in the corrected area after restoring the feature map arrangement; express The real tensor space to which it belongs, where Indicates batch size. Indicates the number of tokens for regional lesions.

[0079] This embodiment takes For 2, 4. The value is 4, and the number of tokens for the first and second level regional lesions are 512 and 128, respectively, with the configured shared correction weights used in both directions.

[0080] Through two-level distribution-guided correction, the lesion features of the first-scale region and the first whole-lung descriptor first form the lesion features of the first-corrected region and the distribution features of the first whole lung. Local abstraction then forms the lesion features of the second-scale region. The second-level correction thus obtains higher-level lesion features of the corrected region and the distribution features of the whole lung, so that the evidence of sparse lesions and the context of the whole lung distribution are linked step by step.

[0081] S3. Align the regional lesion features after two-level distribution guided correction with the corresponding HU attenuation attributes in intensity, and construct a density-aware attention model based on the self-attention mechanism. The constructed density-aware attention model includes:

[0082] The HU attribute encoding and modulation module is used to map the intensity-aligned HU attenuation attribute to the latent space and generate intensity gating with restricted values ​​and intensity bias with positive and negative adjustment directions.

[0083] The perceptual attention aggregation module is used to project the intensity-aligned regional lesion features into query features. Key features Sum value characteristics Simultaneously, based on intensity gating and intensity bias modulation attention calculations, density-aware attention weights are obtained; these density-aware attention weights are then used to log the value features. Weighted aggregation is performed to obtain attention-aggregated features;

[0084] The differential feature enhancement module is used to subtract the intensity-aligned regional lesion features from the attention aggregation features to obtain differential features; the differential features are then processed through convolution, batch normalization, and activation functions to enhance the differential information, resulting in enhanced differential features.

[0085] The residual back-injection module is used to add the enhanced differential feature residuals to the intensity-aligned regional lesion features to obtain densely enhanced regional lesion features.

[0086] Since the density of small shadows changes continuously between adjacent pneumoconiosis stages (i.e., pneumoconiosis stage I, pneumoconiosis stage II, and pneumoconiosis stage III), conventional regional self-attention mainly aggregates information based on the similarity of learned features, which easily smooths out subtle differences in HU attenuation attributes and aggregation between local regions, making it difficult to extract density difference information. Therefore, this embodiment innovatively introduces the physical attenuation cue (i.e., HU attenuation attribute) corresponding to the region directly into the attention relationship modeling.

[0087] In summary, to address the challenge of extracting density differences, this step constructs a density-aware attention model. This model aligns regional lesion features with the HU attenuation attribute of the corresponding regional center. The regional lesion features, after two-level distribution-guided correction, serve as the model input. Furthermore, the HU attenuation attribute of the corresponding region is introduced to generate intensity bias and intensity gating. Through intensity gating, intensity bias, difference enhancement, and residual backinjection, density discriminative features are learned, thereby modulating the correlation between regional lesion features and enhancing the model's ability to represent different small shadow densities. Therefore, the constructed model can extract small shadow density differences using density-aware attention.

[0088] The density-aware attention model constructed in this embodiment is described in detail below:

[0089] (1) HU attribute encoding and modulation module

[0090] In the HU attribute encoding and modulation module, intensity gating parameters with limited values ​​are generated based on the intensity-aligned HU attenuation attributes. and retain the intensity bias in the positive and negative adjustment directions Specifically, it is expressed as:

[0091] ;

[0092] in, Indicates spatial scale This indicates the HU attenuation attribute aligned with the token intensity of the regional lesion. and This represents the convolution mapping in the intensity-gated branch. This represents the convolution mapping in the intensity bias branch. Represents a non-linear activation function. This represents the hyperbolic tangent function.

[0093] Through nonlinear activation functions (such as...) The function will gate the intensity. Limiting the range to 0 to 1, the intensity is biased using the hyperbolic tangent function. Both positive and negative regulation are retained.

[0094] (2) Perception and Attention Aggregation Module

[0095] The perceptual attention aggregation module projects the intensity-aligned regional lesion features into query features, key features, and value features, respectively. It then modulates the key features element by element with intensity gating to obtain the gated key features. The region correlation score between the transpose of the query feature and the gated key features, and the correlation score between the transpose of the query feature and the intensity bias are added together and then normalized to obtain the density perceptual attention weight matrix.

[0096] Gated modulation key features and density-aware attention weight matrix Specifically, it is expressed as follows:

[0097] ;

[0098] Among them, the symbol " " indicates element-wise multiplication, which enables intensity gating to modulate bond features in a multiplicative manner; This indicates the transpose of the query feature. This represents the region relevance score between the transpose of the query feature and the gated key feature. The relevance score represents the relationship between the transpose of the query feature and the strength bias. This represents the normalized exponential function; normalization is performed on the sum of the two types of related terms to obtain the density-aware attention weight matrix between regions.

[0099] The density-aware attention weight matrix is ​​used to perform weighted aggregation of value features, specifically as follows:

[0100] ;

[0101] In the formula For density-aware attention weight matrix transpose; for value features Denseness-aware attention weight matrix Weighted aggregation based on regional relationships yields attention aggregation features. .

[0102] Attention-aggregated features reflect the regional context modulated according to the HU attenuation attribute, and are used to compare with the original regional lesion response (i.e., the intensity-aligned regional lesion features) to further extract differential features related to small shadow density.

[0103] (3) Differential Feature Enhancement Module

[0104] The differential feature enhancement module subtracts the intensity-aligned regional lesion features from the attention aggregation features, and then processes them through differential enhancement convolution, batch normalization, and ReLU activation function to enhance the differential information.

[0105] The differential features and their enhancement process are specifically represented as follows:

[0106] ;

[0107] in, This indicates the regional lesion characteristics after intensity alignment; This represents the difference between the original region response (i.e., the region lesion features after intensity alignment) and the attention convergence features; This indicates differential augmentation convolution. Indicates batch normalization, This indicates the enhanced differential features.

[0108] This embodiment uses a difference feature enhancement module to nonlinearly enhance the difference between the intensity-aligned regional lesion features and the attention aggregation features, which can suppress irrelevant residuals and highlight the density discrimination response.

[0109] (4) Residual back-injection module

[0110] The residual back-injection module injects the enhancement result residuals back into the intensity-aligned regional lesion features to obtain the lesion features of the densely enhanced region, as shown below:

[0111] ;

[0112] in, This indicates the characteristics of lesions in areas with increased density.

[0113] S4. The lesion features of the density-enhanced region and the whole lung distribution features are fused through the classification module to obtain the staged aggregate features. The staged aggregate features are then classified to obtain the category probability of pneumoconiosis identification and staging, as well as the corresponding graded identification results.

[0114] Specifically, step S4 includes the following sub-steps:

[0115] S41. Replicate the whole-lung distribution features after two-level distribution-guided correction along the regional token dimension. Next, the replicated whole-lung distribution features and the lesion features in areas of increased density are made to have the same number of regional tokens; the two are then spliced ​​along the channel dimension to obtain the staged aggregation features.

[0116] ;

[0117] in, This indicates the distribution characteristics of the whole lungs after two-level distribution guidance correction. The number of tokens for regional lesions after hierarchical distribution-guided correction. Indicates replication along the region token dimension Subsequent lung distribution characteristics This indicates a splicing operation along the channel dimension. This indicates a staging aggregation feature that combines evidence of regional density and evidence of whole-lung distribution.

[0118] S42. Perform convolution projection, global pooling, fully connected mapping and normalized exponent processing on the staged aggregated features in sequence to obtain the category probability of pneumoconiosis identification and stage, and output the corresponding hierarchical identification results.

[0119] ;

[0120] in, Indicates the category header, This represents a category probability vector; the categories include normal, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis.

[0121] S5. Train the model, calculate the cross-entropy loss based on the true class labels of the training samples and the class probabilities output in step S4, and update the model parameters based on the cross-entropy loss.

[0122] During the model training phase, the cross-entropy loss is calculated based on the true stage labels of the samples and the classification probabilities output in step S4. The parameters of the joint encoding of the evidence-enriched point cloud, the two-level distribution guided correction, the density-aware attention model, and the classification module are updated through backpropagation.

[0123] In this embodiment, the function of cross-entropy loss is expressed as:

[0124] ;

[0125] in, Indicates the number of categories. Indicates the first The true label of the class, Indicates the first Predicted probability of class This represents the cross-entropy loss.

[0126] By iteratively minimizing the cross-entropy loss, the model learns the relationship between the distribution and density of lesions in the evidence-rich point cloud. The trained model is used to sequentially execute steps S1 to S4 on the input chest CT body data, and outputs the category probability of normal, stage I pneumoconiosis, stage II pneumoconiosis, or stage III pneumoconiosis and the corresponding classification identification results.

[0127] Based on the same inventive concept, this embodiment also provides a pneumoconiosis grading and identification system based on lesion distribution and density point cloud learning, specifically including:

[0128] The evidence-enriched point cloud generation unit is used to acquire chest CT volume data and perform lung field segmentation, calculate multi-scale vascular response, and perform connectivity-sensing vascular inhibition based on the multi-scale vascular response; voxels in the inhibited lung field are converted into candidate points, and joint sampling is performed based on the spatial structural features and HU attenuation properties of the candidate points to obtain the evidence-enriched point cloud.

[0129] A two-level distribution-guided correction unit is used to jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-rich point cloud to obtain regional lesion features and whole-lung descriptors; the regional lesion features and whole-lung descriptors are then corrected by a two-level distribution to obtain the two-level distribution-guided correction regional lesion features and whole-lung distribution features.

[0130] The density-aware attention model aligns the intensity of the two-level distribution-guided correction regional lesion features with the corresponding HU attenuation attributes of the region, and performs density perception based on the self-attention mechanism to obtain density-enhanced regional lesion features.

[0131] The classification module fuses the lesion features of the densely concentrated area and the whole lung distribution features to obtain the staging aggregate features, and classifies the staging aggregate features to obtain the category probability of pneumoconiosis identification and staging, as well as the corresponding grade identification results;

[0132] Among them, density-aware attention models include:

[0133] The HU attribute encoding and modulation module is used to map the intensity-aligned HU attenuation attribute to the latent space and generate intensity gating with restricted values ​​and intensity bias with positive and negative adjustment directions.

[0134] The perceptual attention aggregation module is used to project the intensity-aligned regional lesion features into query features. Key features Sum value characteristics Simultaneously, based on intensity gating and intensity bias modulation attention calculations, density-aware attention weights are obtained; these density-aware attention weights are then used to log the value features. Weighted aggregation is performed to obtain attention-aggregated features;

[0135] The differential feature enhancement module is used to subtract the intensity-aligned regional lesion features from the attention aggregation features to obtain differential features; the differential features are then processed through convolution, batch normalization, and activation functions to enhance the differential information, resulting in enhanced differential features.

[0136] The residual back-injection module is used to add the enhanced differential feature residuals to the intensity-aligned regional lesion features to obtain densely enhanced regional lesion features.

[0137] The specific implementation process of each unit and module mentioned above is described in the preceding steps and will not be repeated here.

[0138] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for classifying and identifying pneumoconiosis based on lesion distribution and density point cloud learning, characterized in that, Includes the following steps: S1. Acquire chest CT volume data and perform lung field segmentation, calculate multi-scale vascular response, and perform connectivity-sensing vascular inhibition based on multi-scale vascular response; convert voxels in the inhibited lung field into candidate points, and perform joint sampling based on the spatial structural features and HU attenuation properties of the candidate points to obtain evidence-enriched point cloud. S2. Jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-rich point cloud to obtain regional lesion features and whole-lung descriptors; then, correct the regional lesion features and whole-lung descriptors through a two-level distribution to obtain the two-level distribution-guided correction of regional lesion features and whole-lung distribution features. S3. Align the regional lesion features after two-level distribution guided correction with the corresponding HU attenuation attributes in intensity, and construct a density-aware attention model based on the self-attention mechanism. The constructed density-aware attention model includes: The HU attribute encoding and modulation module is used to map the intensity-aligned HU attenuation attribute to the latent space and generate intensity gating with restricted values ​​and intensity bias with positive and negative adjustment directions. The perceptual attention aggregation module is used to project the intensity-aligned regional lesion features into query features. Key features Sum value characteristics Simultaneously, based on intensity gating and intensity bias modulation attention calculations, density-aware attention weights are obtained; these density-aware attention weights are then used to log the value features. Weighted aggregation is performed to obtain attention-aggregated features; The differential feature enhancement module is used to subtract the intensity-aligned regional lesion features from the attention aggregation features to obtain differential features; the differential features are then processed through convolution, batch normalization, and activation functions to enhance the differential information, resulting in enhanced differential features. The residual back-injection module is used to add the enhanced differential feature residuals to the intensity-aligned regional lesion features to obtain the density-enhanced regional lesion features; S4. The lesion features of the density-enhanced region and the whole lung distribution features are fused through the classification module to obtain the staged aggregate features. The staged aggregate features are then classified to obtain the category probability of pneumoconiosis identification and staging, as well as the corresponding graded identification results.

2. The method for classifying and identifying pneumoconiosis according to claim 1, characterized in that, In step S1: First, the vascular response of each voxel in the lung field is calculated at multiple spatial scales to obtain the multi-scale vascular response. Then, candidate masks and anchoring masks are formed based on candidate thresholds and anchoring thresholds, respectively. Three-dimensional connected component analysis is performed on the candidate masks and anchoring masks to obtain candidate connected components and anchored connected components. A vascular anchoring set is formed based on the size of the anchored connected components, and candidate connected components that intersect with the vascular anchoring set are suppressed. Then, based on the structural significance and HU attenuation properties of the candidate points, the candidate points are sampled, scored, pre-selected, and sampled at the farthest point to obtain the evidence-enriched point cloud of the lungs.

3. The method for classifying and identifying pneumoconiosis according to claim 2, characterized in that, Step S1 includes: S11. For the chest CT volume data after lung field segmentation, calculate the scale normalization matrix at multiple spatial scales, perform eigenvalue decomposition on the scale normalization matrix, and take the maximum value in the vascular response at each scale to obtain the multi-scale vascular response. S12. Based on the distribution of multi-scale vascular responses in the lung fields, determine the candidate threshold and the anchoring threshold, where the percentile level corresponding to the anchoring threshold is higher than the percentile level corresponding to the candidate threshold; and use the candidate threshold and the anchoring threshold to form the candidate mask and the anchoring mask, respectively. S13. Perform three-dimensional connected component analysis on the candidate mask and the anchoring mask respectively. Divide the spatially interconnected voxels in the candidate mask into candidate connected components and divide the spatially interconnected voxels in the anchoring mask into anchoring connected components. Filter out isolated components with sizes smaller than the minimum anchoring size from the anchoring connected components and merge the anchoring connected components with sizes that meet the anchoring conditions into a vascular anchoring set. S14. When a candidate connected component intersects with the vessel anchoring set, the corresponding voxel is assigned a background sentinel value to form a background voxel. The remaining non-background voxels are converted into candidate points containing three-dimensional coordinates and HU attenuation attributes. Eigenvalue analysis of the covariance matrix is ​​performed on the nearest neighbor coordinates of each candidate point to determine the structural saliency. The structural saliency is weighted with the HU attenuation attribute to obtain a sampling score. Candidate points are pre-selected based on the sampling score to obtain a pre-selected point set. The farthest point is sampled in the pre-selected point set to obtain an evidence-enriched point cloud.

4. The method for classifying and identifying pneumoconiosis according to claim 1, characterized in that, In the two-level distribution correction in step S2, the regional lesion features and the whole lung descriptor are sequentially subjected to first-level distribution-guided correction, local abstraction and second-level distribution-guided correction, the whole lung descriptor is mapped to multiple distribution tokens, and the regional lesion features are converted into regional lesion tokens; The backbone network is used to jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-enriched lung point cloud to obtain the lesion features of the first scale region and the first whole lung descriptor. The lesion features of the first scale region and the first whole lung descriptor are subjected to first-level distribution-guided correction to obtain the lesion features of the first corrected region and the distribution features of the first whole lung. The lesion features of the first corrected region are locally abstracted to obtain the lesion features of the second scale region. The lesion features of the second scale region and the distribution features of the first whole lung are subjected to second-level distribution-guided correction to obtain the lesion features of the second corrected region and the distribution features of the second whole lung.

5. The method for classifying and identifying pneumoconiosis according to claim 4, characterized in that, In step S2, each level of distribution-guided correction specifically includes: S21. Let the target token sequence be... The source token sequence is At each attention head, a query matrix is ​​formed from the target token, and a key matrix and a value matrix are formed from the source token. The multi-head cross-attention correction features from the source token to the target token are calculated, and residual connection, layer normalization and feedforward network processing are performed on the correction features to obtain the updated target token sequence in the current direction. When the distribution token is the target token and the regional lesion token is the source token, it is used to gather lesion evidence from the regional lesion token to the distribution token; when the regional lesion token is the target token and the distribution token is the source token, it is used to inject whole-lung distribution context from the distribution token to the regional lesion token. S22. Repeat the preset number of correction iterations for the bidirectional mutual correction of the regional lesion token and the distribution token; perform mean pooling and linear mapping on the corrected distribution token, and perform residual fusion with the whole lung descriptor; at the same time, transpose the regional lesion token and the channel dimension after the iteration is completed, and restore it to the corrected regional lesion feature tensor.

6. The method for classifying and identifying pneumoconiosis according to claim 1, characterized in that, In the perception and attention aggregation module of step S3, the intensity-aligned regional lesion features are projected into query features, key features and value features respectively, and the key features are modulated element by element with intensity gating to obtain the gated modulated key features. The density-aware attention weight matrix is ​​obtained by adding the region correlation score between the transpose of the query feature and the gated key feature, and the correlation score between the transpose of the query feature and the intensity bias, and then normalizing the result. We use a density-aware attention weight matrix to weight and aggregate the value features to obtain attention-aggregated features.

7. The method for classifying and identifying pneumoconiosis according to claim 1, characterized in that, Step S4 includes: S41. Replicate the whole-lung distribution features after two-level distribution-guided correction along the regional token dimension. Next, the replicated whole-lung distribution features and the lesion features in the densely enhanced areas have the same number of regional tokens; the replicated whole-lung distribution features and the lesion features in the densely enhanced areas are spliced ​​along the channel dimension to obtain the staged aggregation features; S42. Perform convolution projection, global pooling, fully connected mapping and normalized exponent processing on the staged aggregate features in sequence to obtain the category probability of pneumoconiosis identification and stage.

8. The method for classifying and identifying pneumoconiosis according to claim 1, characterized in that, The method further includes the following steps: S5. Calculate the cross-entropy loss based on the true class labels of the training samples and the class probabilities output in step S4. Update the parameters of the joint encoding of the evidence-enriched point cloud, the two-level distribution guided correction, the density-aware attention model, and the classification module based on the cross-entropy loss.

9. The method for classifying and identifying pneumoconiosis according to claim 8, characterized in that, In step S5, the function of cross-entropy loss is expressed as: ； in, Indicates the number of categories. Indicates the first The true label of the class, Indicates the first Predicted probability of class This represents the cross-entropy loss.

10. A pneumoconiosis grading and identification system based on lesion distribution and density point cloud learning, characterized in that, include: The evidence-enriched point cloud generation unit is used to acquire chest CT volume data and perform lung field segmentation, calculate multi-scale vascular response, and perform connectivity-sensing vascular inhibition based on the multi-scale vascular response; voxels in the inhibited lung field are converted into candidate points, and joint sampling is performed based on the spatial structural features and HU attenuation properties of the candidate points to obtain the evidence-enriched point cloud. A two-level distribution-guided correction unit is used to jointly encode the three-dimensional coordinates and HU attenuation attributes of the evidence-rich point cloud to obtain regional lesion features and whole-lung descriptors; the regional lesion features and whole-lung descriptors are then corrected by a two-level distribution to obtain the two-level distribution-guided correction regional lesion features and whole-lung distribution features. The density-aware attention model aligns the intensity of the two-level distribution-guided correction regional lesion features with the corresponding HU attenuation attributes of the region, and performs density perception based on the self-attention mechanism to obtain density-enhanced regional lesion features. The classification module fuses the lesion features of the densely concentrated area and the whole lung distribution features to obtain the staging aggregate features, and classifies the staging aggregate features to obtain the category probability of pneumoconiosis identification and staging, as well as the corresponding grade identification results; Among them, density-aware attention models include: The HU attribute encoding and modulation module is used to map the intensity-aligned HU attenuation attribute to the latent space and generate intensity gating with restricted values ​​and intensity bias with positive and negative adjustment directions. The perceptual attention aggregation module is used to project the intensity-aligned regional lesion features into query features. Key features Sum value characteristics Simultaneously, based on intensity gating and intensity bias modulation attention calculations, density-aware attention weights are obtained; these density-aware attention weights are then used to log the value features. Weighted aggregation is performed to obtain attention-aggregated features; The differential feature enhancement module is used to subtract the intensity-aligned regional lesion features from the attention aggregation features to obtain differential features; the differential features are then processed through convolution, batch normalization, and activation functions to enhance the differential information, resulting in enhanced differential features. The residual back-injection module is used to add the enhanced differential feature residuals to the intensity-aligned regional lesion features to obtain densely enhanced regional lesion features.

Images