A CT image intelligent analysis system for pneumonia auxiliary screening

Abstract

The present application relates to the technical field of medical image processing, in particular to a CT image intelligent analysis system for pneumonia auxiliary screening. The present application firstly pre-processes chest CT images and detects candidate lesion regions; then extracts topological features, deep convolution features and texture statistical features based on persistent homology theory for each candidate lesion, performs feature fusion through a multi-head self-attention mechanism to generate a unified lesion representation vector; then maps the lesion representation vector to a pre-constructed radiology knowledge graph, performs reasoning using a graph neural network, and outputs the pneumonia suspected probability and lesion classification for each lesion; finally, the evidence theory is used to fuse and quantify the uncertainty of the analysis results of multiple lesions to generate a comprehensive screening report. The present application effectively identifies complex morphological lesions through topological features, accurately identifies lesion types through knowledge graph reasoning, and provides quantitative diagnosis uncertainty through the evidence theory.

Description

Technical Field

[0001] This invention relates to the field of medical image artificial intelligence analysis technology, specifically to an intelligent CT image analysis system for pneumonia auxiliary screening. Background Technology

[0002] Pneumonia is a common disease that seriously threatens human health. Accurate and rapid imaging screening is of great significance for early diagnosis and treatment planning. Chest CT, as an important means of diagnosing pneumonia, can clearly show the morphology, distribution, and density characteristics of lesions. However, it relies on the doctor's experience and judgment, which has the problems of high subjectivity and low efficiency.

[0003] Existing intelligent analysis methods for pneumonia CT scans mainly suffer from the following technical deficiencies: First, traditional feature extraction methods, such as texture analysis, struggle to accurately describe the spatial topology of complex lesions, resulting in low accuracy in identifying irregular lesions. Second, deep learning-based methods lack prior medical knowledge, making it difficult to accurately differentiate between lesions with similar imaging manifestations, such as bacterial and viral pneumonia, and resulting in insufficient medical interpretability. Third, existing methods generally lack uncertainty quantification mechanisms, failing to provide diagnostic reliability assessments for multi-lesion cases and posing a risk of misdiagnosis. Summary of the Invention

[0004] This invention provides an intelligent CT image analysis system for auxiliary screening of pneumonia, which improves the accuracy and reliability of pneumonia lesion identification by integrating topological feature extraction, knowledge graph reasoning and evidence theory.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention discloses an intelligent CT image analysis system for auxiliary screening of pneumonia, comprising:

[0007] The lesion identification and localization module is used to preprocess chest CT images and detect candidate lesion areas;

[0008] The feature extraction and fusion module is used to extract topological features, deep convolutional features, and texture statistical features based on persistent homology theory for each candidate lesion region, and generate a unified lesion representation vector through the feature fusion network.

[0009] The knowledge graph reasoning module maps lesion representation vectors to a pre-constructed radiology knowledge graph, uses a graph attention network for reasoning, and outputs the pneumonia-suspected probability and lesion classification for each lesion.

[0010] The evidence fusion and quantification module is used to fuse and quantify the uncertainty of the analysis results of multiple lesions using evidence theory, and generate a comprehensive screening report.

[0011] As a preferred embodiment of the present invention, the preprocessing of chest CT images specifically includes:

[0012] CT images are analyzed, and HU values ​​are standardized through window width and window level adjustment and linear mapping.

[0013] A threshold-based segmentation and morphological manipulation method was used to separate the lung field regions from the chest wall and mediastinal structures.

[0014] Image noise and motion artifacts are removed using median filtering and Gaussian filtering.

[0015] As a preferred embodiment of the present invention, the topological feature extraction step based on persistent homology theory includes:

[0016] A multi-layer simple complex structure is constructed based on the gray-level distribution of the lesion region, and the topological connection relationship between pixels is established.

[0017] Calculate the generation and disappearance process of homology groups in each dimension to construct a persistent barcode that reflects morphological changes;

[0018] Persistent barcodes are converted into low-dimensional topologically invariant features using vectorization methods to capture the geometric morphological characteristics of lesions.

[0019] As a preferred embodiment of the present invention, the deep convolutional feature extraction step includes: using a pre-trained ResNet-50 network to perform deep convolutional feature extraction on the lesion region, and compressing the feature map into a fixed-dimensional feature vector through a global average pooling layer.

[0020] As a preferred embodiment of the present invention, the step of generating a unified lesion representation vector through a feature fusion network specifically includes:

[0021] Topological features, depthwise convolutional features, and texture statistical features are respectively processed by dimensionality transformation and standardization through fully connected layers;

[0022] A multi-head self-attention mechanism is used to dynamically calculate the importance weights of each modality feature;

[0023] Multimodal features are fused into a fixed-dimensional lesion representation vector through weighted join operations.

[0024] As a preferred embodiment of the present invention, the pre-constructed radiological knowledge graph specifically includes:

[0025] Construct a heterogeneous graph structure that includes pneumonia lesion nodes, imaging feature nodes, and anatomical location nodes;

[0026] Establish semantic association edges between nodes to represent the causal relationship between lesions and features;

[0027] An initial feature vector is assigned to each node, and pre-training is performed using medical literature and expert-annotated data.

[0028] As a preferred embodiment of the present invention, the output of the pneumonia-like probability and lesion classification for each lesion specifically includes:

[0029] A graph attention network is used to perform multi-layer message passing and feature aggregation on nodes in the knowledge graph.

[0030] The features of lesion nodes are mapped to a probability distribution of suspected pneumonia by using a probabilistic activation function;

[0031] A multi-classifier structure is used to output lesion classification and corresponding confidence scores.

[0032] As a preferred technical solution of the present invention, the step of fusing and quantifying the uncertainty of the analysis results of multiple lesions using evidence theory specifically includes:

[0033] Construct a basic probability assignment function based on the suspected pneumonia probability and confidence level of each lesion;

[0034] Evidence from each lesion was combined and fused using evidence synthesis rules, and the overall confidence and likelihood were calculated.

[0035] The consistency among diagnoses of multiple lesions is quantified by measuring the conflict of evidence, thereby assessing the level of uncertainty in the overall analysis.

[0036] As a preferred embodiment of the present invention, the generation of the comprehensive screening report specifically includes:

[0037] Statistical analysis and visualization of spatial location information, morphological features, and imaging characteristics of each detected lesion;

[0038] The overall pneumonia imaging score is calculated based on the imaging severity and spatial distribution pattern of the lesions;

[0039] A standardized template is used to generate a structured report that includes lesion imaging descriptions, quantitative analysis results, uncertainty assessments, and auxiliary screening recommendations.

[0040] The beneficial effects of this invention are:

[0041] 1. This invention introduces the persistent homology theory of mathematical topology into the field of medical image analysis, and for the first time realizes the accurate mathematical description of the geometric topology of pneumonia lesions. Through the synergistic fusion of multimodal features, a significant nonlinear synergistic effect is generated, which effectively solves the technical problem that traditional methods cannot accurately identify lesions with complex shapes, and provides a brand-new mathematical tool for the intelligent identification of irregular lesions.

[0042] 2. This invention constructs a medical semantic-driven knowledge graph reasoning framework, which deeply integrates prior knowledge of imaging with deep learning technology, and realizes causal reasoning diagnosis based on pathological mechanisms. It effectively solves the problem of differentiating similar lesions such as bacterial and viral pneumonia, and has stronger medical interpretability compared with traditional classification methods.

[0043] 3. By employing the DS evidence theory to integrate the results of multifocal analysis, and by integrating information through the basic probability allocation function and evidence synthesis rules, an uncertainty quantification mechanism is introduced into pneumonia screening. This provides clinical diagnostic credibility assessment, effectively reduces the risk of misdiagnosis, and improves the safety and reliability of clinical decision-making. Attached Figure Description

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0045] Figure 1 This is a schematic diagram of the results of a CT image intelligent analysis system for pneumonia auxiliary screening according to the present invention;

[0046] Figure 2 This is a schematic diagram of the BPA construction and DS evidence fusion process of the present invention;

[0047] Figure 3 This is a flowchart illustrating a CT image intelligent analysis system for auxiliary screening of pneumonia, as described in this invention. Detailed Implementation

[0048] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0049] Example 1: As Figure 1 As shown, the present invention provides an intelligent CT image analysis system for auxiliary screening of pneumonia, comprising:

[0050] The lesion identification and localization module is used to preprocess chest CT images and detect candidate lesion areas;

[0051] Furthermore, the preprocessing of the chest CT images specifically includes:

[0052] CT images are analyzed, and HU values ​​are standardized through window width and window level adjustment and linear mapping.

[0053] A threshold-based segmentation and morphological manipulation method was used to separate the lung field regions from the chest wall and mediastinal structures.

[0054] Image noise and motion artifacts are removed using median filtering and Gaussian filtering.

[0055] Specifically, the input raw chest CT image data is read using a DICOM parser to extract pixel data and metadata. The CT images are then processed with window width and level adjustments. Lung window parameters are set (window width 1500 HU, window level -600 HU), where a window width of 1500 HU covers the complete lesion density range from emphysema to consolidation (-900 HU to -100 HU), and a window level of -600 HU corresponds to the average CT value of normal lung tissue. This combination of parameters optimally balances image contrast and detail display. Mediastinal window parameters are set (window width 400 HU, window level 40 HU). This parameter is suitable for displaying mediastinal soft tissue structures, a window level of 40 HU is close to the average density of soft tissue, and a window width of 400 HU can simultaneously display mediastinal structures such as blood vessels and lymph nodes. HU values ​​are standardized to the [0, 255] grayscale range using linear mapping to eliminate imaging differences between different devices.

[0056] A threshold-based segmentation method was used to separate lung field regions. The threshold range was set from -1024 HU to -200 HU for lung parenchyma extraction. The lower limit of -1024 HU was the standard air density value of the CT system, and the upper limit of -200 HU was the empirical boundary between aerated lung tissue and soft tissue. This range effectively excluded non-lung field tissues such as the chest wall and mediastinum. Morphological opening and closing operations were performed on the segmented lung field regions to remove boundary spurs and fill internal cavities. A 5×5 median filter and a Gaussian filter with a standard deviation of 1.0 were used for image denoising to eliminate noise and motion artifacts during the scanning process.

[0057] It should be noted that a 5×5 median filter can effectively remove salt-and-pepper noise while preserving lesion edge information. If the size is too small, the noise reduction effect will be insufficient, and if the size is too large, the lesion boundary will be blurred. A Gaussian filter with a standard deviation of 1.0 can achieve a balance between removing high-frequency noise and maintaining image clarity. If the standard deviation is too small, the filtering effect will be weak, and if it is too large, the lesion details will be lost.

[0058] The preprocessed data is used to obtain candidate lesion regions through a target detection model. The target detection model is preferably a mature three-dimensional target detection algorithm such as 3D RCNN, nnU-Net, or 3D RetinaNet.

[0059] The preprocessing steps described above effectively eliminate imaging differences between different CT devices, accurately separate lung field regions, and remove noise interference. The target detection model enables automatic localization of candidate lesions, avoiding computational redundancy in whole-lung scanning and providing accurate lesion regions for subsequent feature extraction.

[0060] The feature extraction and fusion module is used to extract topological features, deep convolutional features, and texture statistical features based on persistent homology theory for each candidate lesion region, and generate a unified lesion representation vector through the feature fusion network.

[0061] Furthermore, the topological feature extraction step based on persistent homology theory includes:

[0062] A multi-layer simple complex structure is constructed based on the gray-level distribution of the lesion region, and the topological connection relationship between pixels is established.

[0063] Calculate the generation and disappearance process of homology groups in each dimension to construct a persistent barcode that reflects morphological changes;

[0064] Persistent barcodes are converted into low-dimensional topologically invariant features using vectorization methods to capture the geometric morphological characteristics of lesions.

[0065] Specifically, for each candidate lesion region, a multi-layered simple complex structure is constructed based on its gray-level distribution. The three-dimensional voxels of the lesion region are hierarchically decomposed according to gray-level values, and a gray-level threshold sequence is set. ,in , , Let n be the set of gray values ​​for the lesion region, and n be the threshold number. and These represent the minimum and maximum grayscale values ​​of the lesion area, respectively, determining the dynamic range of the threshold, with the threshold interval set to [value missing]. It should be noted that 20 threshold levels are sufficient to capture the evolution of lesion morphology at different density levels. Fewer than 15 levels will miss key topological changes, while more than 25 levels significantly increase computational complexity but limit information gain. Validation experiments have determined that 20 levels are the optimal balance point.

[0066] For each threshold Construct the corresponding simple complex The connectivity between voxels is determined by the 6-adjacency criterion, i.e., the voxel value is greater than 6. Furthermore, spatially adjacent voxels form 0-simplexes, and adjacent 0-simplexes constitute 1-simplexes.

[0067] Calculate the homology groups in each dimension The generation and disappearance processes of topological features are tracked using an incremental algorithm to monitor the evolution of connected components, loops, and cavity structures at different thresholds. The generation threshold for each topological feature is recorded. and disappearance threshold Construct persistent intervals Durability is defined as All persistent regions were sorted in descending order of persistence, and the top 10 most significant persistent regions were selected to construct a persistent barcode map. It should be noted that the top 10 most significant persistent regions contain the main topological features of the lesions and can distinguish more than 90% of typical lesion morphological patterns. More persistent regions would introduce noise features and reduce classification accuracy.

[0068] The persistent barcode is converted into a 128-dimensional topologically invariant feature vector using a vectorization method. Specifically, each persistent interval... Mapped to two-dimensional points Persistent landscape function Functional representation of topological features:

[0069] ;

[0070] Among them, for k=1 represents the most significant topological feature landscape, and k=2 represents the second most significant feature landscape; In the interval The continuous variable represents the grayscale threshold level currently being examined; To calculate the threshold The minimum distance to the point of origin and the point of disappearance.

[0071] In the interval Calculate using 64 uniformly sampled points. and The function values ​​are used to concatenate the sampled values ​​of the two landscape functions to form a 128-dimensional topologically invariant feature vector. 64 sampling points can maintain the continuity of the landscape function, while fewer than 50 points will result in undersampling distortion.

[0072] The introduction of persistent cohomology theory enables a mathematically quantitative description of the geometric topology of lesions, effectively capturing complex and irregular morphological information that is difficult to describe using traditional methods. This method has significant advantages in identifying lesions with complex morphologies, such as ground-glass opacities in viral pneumonia.

[0073] Furthermore, the deep convolutional feature extraction step includes: using a pre-trained ResNet-50 network to perform deep convolutional feature extraction on the lesion region, and compressing the feature map into a fixed-dimensional feature vector through a global average pooling layer.

[0074] Specifically, a pre-trained ResNet-50 network is used to extract features from the lesion region. The ResNet-50 network is a deep convolutional neural network model pre-trained on a large-scale natural image dataset. It contains 50 convolutional layers and employs a residual connection structure. The network consists of one 7×7 convolutional layer, one max-pooling layer, four residual block groups, and one global average pooling layer. The pre-training process enables the network to learn hierarchical representations from low-level edge and texture features to high-level shape and structural features, providing a solid feature foundation for medical image feature extraction.

[0075] The three-dimensional lesion area was sliced ​​along the axis, and five slices were selected, including the central layer of the lesion and two layers before and after it. Each slice was adjusted to 224×224 pixels (the standard input size of ResNet-50) by bilinear interpolation and copied into a three-channel format to adapt to the input requirements of the pre-trained model. The five consecutive slices can capture the three-dimensional information of the lesion, with a coverage area of ​​about 6.25mm (1.25mm×5), which meets the requirement of complete display of small lesions.

[0076] The five slices are input into the ResNet network to extract the fourth residual block. The output feature map has a size of 14×14×1024. The fourth residual block can identify complex lesion patterns while retaining sufficient location information. Global average pooling is applied to each feature map to compress the spatial dimension to 1×1, resulting in a 1024-dimensional feature vector. The feature vectors corresponding to the five slices are then fused using average pooling to obtain a 1024-dimensional depthwise convolutional feature vector. .

[0077] The pre-trained ResNet-50 network fully leverages the feature representation capabilities of large-scale pre-training, achieving good feature representations even with limited medical imaging data. Choosing five consecutive slices instead of a single slice preserves the three-dimensional information of the lesion, improving the ability to identify three-dimensional lesion structures.

[0078] Furthermore, gray-level co-occurrence matrix texture features are calculated for the lesion region. Pixel spacing is set within the lesion region. ,angle Calculate the corresponding gray-level co-occurrence matrix in four directions. Four texture parameters—contrast, correlation, energy, and homogeneity—are extracted from each co-occurrence matrix, resulting in a total of 16-dimensional texture feature vectors. .

[0079] Furthermore, the generation of a unified lesion representation vector through a feature fusion network specifically includes:

[0080] Topological features, depthwise convolutional features, and texture statistical features are respectively processed by dimensionality transformation and standardization through fully connected layers;

[0081] A multi-head self-attention mechanism is used to dynamically calculate the importance weights of each modality feature;

[0082] Multimodal features are fused into a fixed-dimensional lesion representation vector through weighted join operations.

[0083] Specifically, the topologically invariant eigenvectors Depth-convolutional feature vectors and texture feature vector Dimensional transformations are performed using fully connected layers. Topological features are transformed from 128 to 256 through fully connected layers, depthwise convolutional features from 1024 to 256 through fully connected layers, and texture features from 16 to 256 through fully connected layers. Each layer employs ReLU activation and batch normalization.

[0084] A multi-head self-attention mechanism is used to calculate the importance weights of each modality feature. The number of attention heads is set to H=8, and the dimension of each head is... It should be noted that 256 dimensions represent a trade-off between feature representation power and computational efficiency, effectively preserving modal feature information while avoiding the curse of dimensionality. Grid search experiments showed that 128 dimensions were insufficient in expressive power, while 512 dimensions were computationally too expensive and prone to overfitting. Eight attention heads can learn diverse correlation patterns between features from different subspaces; fewer than four heads limit expressive power, and more than 12 heads significantly increase computational complexity with limited performance improvement.

[0085] The three 256-dimensional feature vectors are concatenated into a 768-dimensional input vector, and then a linear transformation matrix is ​​applied. (All dimensions are 768×256) Calculate the query Q, key K, and value V matrices respectively. The attention weight calculation formula is:

[0086] ;

[0087] in, The query matrix represents the query representation of three feature modalities (topology, convolution, and texture). The key representation represents the three feature modalities and is used to calculate the similarity between features; This is a matrix transpose operation; The scaling factor is used to prevent the gradient from vanishing due to an excessively large dot product.

[0088] Multimodal features are fused using a weighted join operation, and a weighted feature vector is calculated:

[0089] ;

[0090] in The attention weights are normalized and satisfy the following conditions: . These are topologically invariant eigenvectors; These are depthwise convolutional feature vectors; This is the texture feature vector.

[0091] Finally, a fully connected layer of 256 to 512 is used to map the weighted features into a 512-dimensional lesion representation vector. .

[0092] The application of a multi-head self-attention mechanism enables intelligent weight allocation of topological, depthwise convolutional, and texture features, adaptively adjusting the importance of each modality feature according to different lesion types. This dynamic fusion strategy effectively avoids information redundancy and feature conflicts that may result from simple feature concatenation, thus improving classification accuracy.

[0093] The knowledge graph reasoning module is used to map lesion representation vectors onto a pre-constructed radiology knowledge graph, and uses a graph attention network for reasoning to output the pneumonia suspicion probability and lesion classification for each lesion.

[0094] Furthermore, the pre-constructed radiological knowledge graph specifically includes:

[0095] Construct a heterogeneous graph structure that includes pneumonia lesion nodes, imaging feature nodes, and anatomical location nodes;

[0096] Establish semantic association edges between nodes to represent the causal relationship between lesions and features;

[0097] An initial feature vector is assigned to each node, and pre-training is performed using medical literature and expert-annotated data.

[0098] Specifically, a radiological knowledge graph was constructed comprising three types of heterogeneous nodes: pneumonia lesion nodes, imaging feature nodes, and anatomical location nodes. The pneumonia lesion nodes include 12 common pneumonia types, such as bacterial pneumonia, viral pneumonia, mycoplasma pneumonia, and fungal pneumonia. The imaging feature nodes include 32 CT imaging features, such as ground-glass opacity, consolidation, cavitation, and pleural thickening. The anatomical location nodes include 18 pulmonary anatomical regions, such as the upper lobe, middle lobe, lower lobe, and left and right lungs, for a total of 62 nodes.

[0099] Semantic association edges between nodes are established to represent the causal relationship between lesions and features. Based on relevant medical literature, associations between lesion types and imaging features are constructed. For example, a strong association edge (weight 0.8) is established between bacterial pneumonia nodes and consolidation nodes, and a strong association edge (weight 0.9) is established between viral pneumonia nodes and ground-glass opacity nodes. Imaging feature nodes and anatomical location nodes are associated based on the predilection sites of the lesions, resulting in a total of 156 directed edges.

[0100] Each node is assigned an initial 512-dimensional feature vector. Semantic vectors are obtained by embedding disease description texts from medical literature using the efficient Word2Vec word vector training model. The word vector model is trained using a corpus containing 500,000 radiology articles, with a window size of 5, a negative sampling count of 10, and 100 iterations to obtain the initial feature representations for each node. A window size of 5 captures the local semantic relationships of medical terms; a window that is too small lacks contextual information, while a window that is too large introduces irrelevant noise. A negative sampling count of 10 strikes a balance between training efficiency and model quality. 100 iterations ensure sufficient model convergence; the validation set loss continues to decrease after 50 iterations, and the performance improvement becomes negligible after 150 iterations.

[0101] Furthermore, the output of the pneumonia-like probability and lesion classification for each lesion specifically includes:

[0102] A graph attention network is used to perform multi-layer message passing and feature aggregation on nodes in the knowledge graph.

[0103] The features of lesion nodes are mapped to a probability distribution of suspected pneumonia by using a probabilistic activation function;

[0104] A multi-classifier structure is used to output lesion classification and corresponding confidence scores.

[0105] Specifically, the 512-dimensional lesion representation vector As a feature of the query node, it is matched with 62 nodes in the knowledge graph. A graph attention network (GAT) is used for multi-layer message passing and feature aggregation. The GAT network consists of 3 layers, each with 8 attention heads, and the hidden layer dimension is 256.

[0106] In the l-th layer GAT, nodes The feature update formula is:

[0107] ;

[0108] in, Represents a node The set of neighboring nodes, For attention weights, Let be the weight matrix of the l-th layer. For the first Features of layer node j This is the LeakyReLU activation function. Attention weights. Calculated using the attention mechanism:

[0109] ;

[0110] in, For attention parameter vectors, This indicates a feature concatenation operation. Represents a node Through weight matrix The transformed features; Representing neighboring nodes Through weight matrix The transformed features; This indicates that the transformation features of two nodes are concatenated along the dimension.

[0111] After three layers of GAT propagation, the lesion type nodes obtain updated feature representations that integrate imaging features and anatomical location information. The cosine similarity between the lesion representation vector and the final feature representations of the 12 lesion type nodes is calculated, and the probability distribution of each lesion type is obtained by normalization using the softmax function. .

[0112] in, Each element represents the probability of the corresponding lesion type; This is the lesion representation vector; The node characteristics represent the lesion type after propagation through 3 layers of GAT, where j=1, 2, ..., 12 correspond to 12 lesion types.

[0113] A binary classifier structure is used to output the probability of suspected pneumonia, and the lesion representation vector is used. The input is fed into a fully connected network with two hidden layers (512→256→128), and finally the probability of suspected pneumonia is output through a sigmoid activation function.

[0114] ;

[0115] in, Indicates the probability of suspected pneumonia. It consists of three fully connected layers: 512→256→128→1, with each layer followed by a ReLU activation function (except for the last layer).

[0116] Simultaneously, the lesion classification results are output, selecting the lesion type with the highest probability as the predicted label, and the corresponding confidence score is output. .

[0117] The construction of a radiology knowledge graph deeply integrates medical prior knowledge with deep learning technology. A graph attention network, through multi-layer message passing, enables semantic association reasoning between lesion types, imaging features, and anatomical locations. This method effectively solves the challenge of differentiating lesions with similar imaging manifestations, such as bacterial and viral pneumonia. The reasoning process exhibits good medical interpretability, and the output probability distribution and confidence score provide clinicians with reliable decision-making support.

[0118] The evidence fusion and quantification module is used to fuse and quantify the uncertainty of the diagnostic results of multiple lesions using evidence theory, and generate a comprehensive screening report.

[0119] Furthermore, the fusion of analysis results from multiple lesions and the quantification of uncertainty using evidence theory specifically include:

[0120] Construct a basic probability assignment function based on the suspected pneumonia probability and confidence level of each lesion;

[0121] Evidence from each lesion was combined and fused using evidence synthesis rules, and the overall confidence and likelihood were calculated.

[0122] The consistency among diagnoses of multiple lesions is quantified by measuring the conflict of evidence, thereby assessing the level of uncertainty in the overall analysis.

[0123] Specifically, such as Figure 2 As shown, a basic probability assignment function (BPA) is first constructed based on the classification probability and confidence level of each lesion. For the th... Given a lesion, let the probability of it being suspected to be pneumonia be [value missing]. The confidence level for lesion classification is Construct a three-element recognition framework The basic probability assignment function is defined as follows:

[0124]

[0125]

[0126] ;

[0127] in, This indicates uncertainty due to insufficient evidence.

[0128] Dempster's evidence fusion rules were used to combine and fuse evidence from each lesion. For evidence fusion from n lesions, the fusion rule is as follows:

[0129] ;

[0130] in, Elements belonging to the identification frame Θ={pneumonia, non-pneumonia, uncertain}; The i-th lesion provides evidence supporting the proposition. Normalization constant:

[0131] ;

[0132] Among them, when This indicates that the evidence for each lesion is completely consistent and there is no conflict. A value close to 0 indicates highly conflicting evidence from different lesions, making the evidence synthesis unreliable.

[0133] By synthesizing evidence step by step, the evidence for the first two lesions is first synthesized, then combined with the evidence for the third lesion, and so on up to all n lesions.

[0134] Calculate the overall trust level:

[0135] ;

[0136] ;

[0137] in, Trust level indicates the strength of evidence supporting a diagnosis of pneumonia. This indicates the strength of evidence that does not oppose a diagnosis of pneumonia. The degree of conflict of evidence is measured by the conflict coefficient. When the degree of conflict is greater than 0.3, the diagnostic results for multiple lesions are considered to have significant inconsistencies. The uncertainty of the overall analysis is defined as follows:

[0138] ;

[0139] When the degree of conflict of evidence or overall uncertainty When this occurs, the system automatically triggers a secondary review mechanism: it re-analyzes suspected lesions using feature extraction methods with different parameters and marks the case as requiring manual expert review. Conflict handling strategies include: prioritizing the single source of evidence with the highest confidence level in cases of high conflict, and clearly indicating the level of uncertainty in the report to provide risk assessment reference for clinical decision-making. It should be noted that the uncertainty threshold is determined based on statistical analysis of the training data; this embodiment only provides a preferred example, and equivalent substitutions or adjustments can be made according to the specific scenario.

[0140] Furthermore, the generation of the comprehensive screening report specifically includes:

[0141] Statistical analysis and visualization of spatial location information, morphological features, and imaging characteristics of each detected lesion;

[0142] Based on the fusion of evidence theories and the level of uncertainty, a tiered screening recommendation is generated.

[0143] A standardized template is used to generate a structured report that includes lesion imaging descriptions, quantitative analysis results, uncertainty assessments, and auxiliary screening recommendations.

[0144] Specifically, the spatial localization information of each detected lesion is statistically analyzed and visualized. The three-dimensional bounding box coordinates of each lesion are... Converted to anatomical description, the anatomical region to which the lesion belongs (right upper lobe, right middle lobe, right lower lobe, left upper lobe, left lower lobe) is determined using a lung lobe segmentation template. The coordinates of the starting column of the lesion boundary within the scanning plane (cross section); The coordinates of the starting row of the lesion boundary within the scanning plane (cross section); This refers to the starting slice number of the lesion boundary in the cephalic (axial) direction, corresponding to the CT scan slice number. Layer slice.

[0145] Calculate the morphological characteristics of the lesion: volume , These represent the maximum span of the lesion along the X, Y, and Z axes, respectively. Sphericity This value measures the regularity of the lesion's shape; the closer the value is to 1, the closer it is to a sphere.

[0146] ;

[0147] in, The volume of the lesion; The normalization coefficient is... This is a normalization transformation for volume, used to convert three-dimensional volume quantities into dimensions comparable to two-dimensional surface areas. This represents the surface area of ​​the lesion. This formula quantifies the regularity of the shape by comparing the surface area ratio of an actual lesion to that of a sphere of equal volume. When the lesion is a perfect sphere... The more irregular the shape of the lesion The smaller the value, the more effectively this parameter can distinguish between regular bacterial consolidation and irregular viral ground-glass opacities.

[0148] Aspect Ratio Assess the degree of stretching in the lesion's morphology:

[0149] ;

[0150] in, This represents the maximum span of the lesion in three spatial directions. The minimum span of the lesion in three spatial directions.

[0151] The description of imaging characteristics includes quantitative parameters such as mean CT value, standard deviation of CT value, and edge sharpness.

[0152] A tiered screening recommendation is generated based on the fusion of evidence theories and the level of uncertainty. This recommendation is then adjusted according to the overall trust level. and level of uncertainty Based on the combination of factors, a three-tiered screening recommendation is formulated:

[0153] when and At that time, the system determined it to be a highly credible suspected case of pneumonia and recommended that you seek medical attention promptly for further diagnosis.

[0154] when or At that time, the system judged it as a medium confidence anomaly and suggested further examinations such as sputum testing or serological testing;

[0155] when and If the system determines the information to be of low credibility or high uncertainty, regular follow-up observation is recommended.

[0156] It should be noted that the trust threshold is determined based on statistical analysis of the training data. This embodiment only provides a preferred example, and equivalent substitutions or adjustments can be made according to the specific scenario.

[0157] A structured report is generated using a standardized XML template. The report comprises four core parts: a lesion imaging description section detailing the anatomical location, three-dimensional dimensions, morphological features, and CT density distribution of each lesion; a quantitative analysis results section including the comprehensive pneumonia suspicion probability (Bel) calculated based on evidence theory, the specific classification results for each lesion, and the corresponding confidence level; an uncertainty assessment section reporting the degree of evidence conflict and the overall uncertainty level, providing clinicians with a reference for decision-making credibility; and an auxiliary screening recommendation section providing personalized follow-up treatment suggestions based on the evidence fusion results. The report also records the analysis timestamp, algorithm version, and the total number of lesions involved in the analysis to ensure the traceability of the results and the safety of clinical applications.

[0158] Example 2: A hospital's radiology department faced the following problems in pneumonia CT screening: complex and irregular lesions were difficult to identify accurately; the differentiation between bacterial and viral pneumonia was challenging; and the reliability of comprehensive diagnosis of multiple lesions lacked assessment. To address these problems, the hospital introduced a CT image intelligent analysis system for pneumonia auxiliary screening, as described in this invention, for verification and application. The system implementation process is as follows: Figure 3 As shown.

[0159] We selected chest CT data from 120 cases collected by the hospital in 2024, including 78 patients diagnosed with pneumonia (34 cases of bacterial pneumonia, 26 cases of viral pneumonia, and 18 cases of mycoplasma pneumonia) and 42 normal controls. The CT scan slice thickness was 1.25 mm, and the reconstruction matrix was 512×512. All cases were confirmed by etiological examination.

[0160] Specifically, the following three typical cases are presented:

[0161] Case 1: An irregular ground-glass opacity in the right lower lobe of a patient. This invention's persistent homology analysis extracted three connected components with a 0-dimensional Betty number and detected two ring structures with a 1-dimensional Betty number. The persistent landscape function exhibited a typical distribution pattern of viral pneumonia within a specific interval. Knowledge graph inference output a probability of 0.82 for viral pneumonia, which was subsequently confirmed by PCR testing. Traditional texture feature methods output a probability of 0.71 for bacterial pneumonia, resulting in a false positive.

[0162] Case 2: A patient had three lesions in the upper lobe of the left lung. The evidence allocation for each lesion was as follows: Lesion 1 m (pneumonia) = 0.43, Lesion 2 m (pneumonia) = 0.38, Lesion 3 m (pneumonia) = 0.61. After fusion using Dempster's synthesis rules, the overall confidence level Bel (pneumonia) was 0.56, and the uncertainty was 0.31. The system recommended further investigation, and subsequent serological testing confirmed mycoplasma pneumonia.

[0163] Case 3: A patient presented with multiple nodules and cavitation in both lungs. During the graph neural network inference process, the association weight (0.85) between the cavity feature node and the bacterial pneumonia node was activated, with particular attention paid to the lower lobe distribution pattern. The final output showed a bacterial pneumonia probability of 0.89, and sputum culture confirmed Streptococcus pneumoniae infection.

[0164] The comparison results on a dataset of 120 cases are shown in Table 1:

[0165] Table 1. Experimental Comparison Table

[0166]

[0167] Of the 34 cases of bacterial pneumonia, this invention correctly identified 29 cases (accuracy rate 85.3%), while the traditional method correctly identified 23 cases (accuracy rate 67.6%). Of the 26 cases of viral pneumonia, this invention correctly identified 22 cases (accuracy rate 84.6%), while the traditional method correctly identified 18 cases (accuracy rate 69.2%).

[0168] In 120 tests, the system output a high uncertainty label (uncertainty > 0.4) for 23 cases. Among them, 21 cases were later confirmed to require additional etiological examination or were cases with atypical imaging manifestations. The system's warning prompts provided a reference for clinical decision-making.

[0169] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1.A CT image intelligent analysis system for pneumonia auxiliary screening, characterized in that, The method comprises the following steps: a lesion identification and positioning module for preprocessing chest CT images and detecting candidate lesion regions; a feature extraction and fusion module for extracting topological features based on persistent homology theory, deep convolutional features, and texture statistical features for each candidate lesion region, and generating a unified lesion representation vector through a feature fusion network; a knowledge graph reasoning module for mapping the lesion representation vector to a pre-constructed radiology knowledge graph, using a graph attention network for reasoning, and outputting the pneumonia suspected probability and lesion classification for each lesion; an evidence fusion and quantification module for using evidence theory to fuse and quantify the uncertainty of the analysis results of multiple lesions, and generating a comprehensive screening report. The topological feature extraction step based on persistent homology theory comprises: constructing a multi-layer simplicial complex structure based on the gray scale distribution of the lesion region, and establishing the topological connection relationship between pixels; calculating the generation and disappearance process of each dimension homology group, and constructing persistent barcodes reflecting morphological changes; converting the persistent barcodes into low-dimensional topological invariant features through vectorization method to capture the geometric morphological characteristics of the lesion; The generation of a unified lesion representation vector through a feature fusion network specifically comprises: dimensional transformation and standardization processing of topological features, deep convolutional features, and texture statistical features through a fully connected layer respectively; using a multi-head self-attention mechanism to dynamically calculate the importance weight of each modal feature; fusing multi-modal features into a fixed-dimensional lesion representation vector through weighted connection operation. 2.The CT image intelligent analysis system for pneumonia auxiliary screening of claim 1, wherein, The preprocessing of chest CT images specifically comprises: parsing the CT image, and realizing HU value standardization processing through window width and window level adjustment and linear mapping; using a threshold segmentation and morphological operation based method to separate the lung field region from the chest wall, mediastinal structure; removing image noise and motion artifacts through median filtering and Gaussian filtering. 3.The CT image intelligent analysis system for pneumonia auxiliary screening of claim 1, wherein, The deep convolutional feature extraction step comprises: using a pre-trained ResNet-50 network to extract deep convolutional features of the lesion region, and compressing the feature map into a fixed-dimensional feature vector through a global average pooling layer. 4.The CT image intelligent analysis system for pneumonia auxiliary screening of claim 1, wherein, The pre-constructed radiology knowledge graph specifically comprises: constructing a heterogeneous graph structure containing pneumonia lesion nodes, imaging feature nodes, and anatomical location nodes; establishing semantic association edges between nodes to represent the causal relationship between lesions and features; assigning an initial feature vector to each node, and pre-training through medical literature and expert annotation data. 5.The CT image intelligent analysis system for pneumonia auxiliary screening of claim 1, wherein, The output of the pneumonia suspected probability and lesion classification for each lesion specifically comprises: using a graph attention network to perform multi-layer message passing and feature aggregation on the nodes in the knowledge graph; mapping the features of the lesion nodes to the pneumonia suspected probability distribution through a probability activation function; using a multi-classifier structure to output the lesion classification and corresponding confidence. 6.The CT image intelligent analysis system for pneumonia auxiliary screening of claim 1, wherein, The fusion and uncertainty quantification of the analysis results of multiple lesions using evidence theory specifically comprises: constructing a corresponding basic probability assignment function based on the pneumonia suspected probability and confidence of each lesion; combining and fusing the evidence of each lesion using evidence synthesis rules, calculating the comprehensive trustworthiness and likelihood; quantifying the consistency between the diagnosis results of multiple lesions through evidence conflict, and evaluating the uncertainty level of the overall analysis. 7.The CT image intelligent analysis system for pneumonia auxiliary screening of claim 1, wherein, The generating a comprehensive screening report specifically includes: statistically and visually describing spatial localization information, morphological features and imaging characteristics of each detected lesion; calculating an overall pneumonia imaging score based on the imaging severity and spatial distribution pattern of the lesions; generating a structured report containing imaging description of the lesions, quantitative analysis results, uncertainty assessment and auxiliary screening recommendations using standardized templates.

Images