Lung ventilation-perfusion development area image segmentation and quantitative analysis system

Abstract

The invention discloses a pulmonary ventilation-perfusion development area image segmentation and quantitative analysis system, and relates to the technical field of medical image analysis, and the system comprises a data collection module which obtains SPECT V / Q image data and medical record information corresponding to the SPECT V / Q image data, and generates a structured data set; the feature extraction module is used for performing format conversion and multi-scale feature extraction on the SPECT V / Q image data in the structured data set to generate a compressed feature map data set; the lung lobe segmentation module is used for inputting the compressed feature map data set into a U-Net model to segment a lung lobe region and carrying out post-processing to generate a lung lobe region mask image; and the image fusion module is used for performing alpha blending fusion on the lung lobe region mask image and SPECT V / Q image data to generate a color fusion image. According to the method, the alveolar-to-lung-lobe-level structure is captured by fusing the features of the four branches through the cavity pyramid network, and the GPU video memory consumption is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a system for segmenting and quantifying lung ventilation-perfusion imaging regions. Background Technology

[0002] Significant progress has been made in the field of medical image analysis in recent years, driven by deep learning technology, especially in the diagnosis of lung diseases, where segmentation models such as U-Net have achieved automatic segmentation of lung lobe regions. Existing technologies mainly extract features from single image modalities and combine them with traditional clustering algorithms to quantify the ventilation-perfusion ratio (V / Q).

[0003] However, conventional methods rely on isolated image data (such as SPECT V / Q imaging) and do not integrate clinical medical record information, resulting in a disconnect between functional analysis and pathological background. Feature extraction networks lack multi-scale fusion mechanisms, have insufficient interaction between shallow edge features and deep semantic features, and ignore feature distillation and compression, resulting in wasted computational resources and missed detection of small lesions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a lung ventilation-perfusion imaging region image segmentation and quantitative analysis system to solve the problem of limited accuracy in quantitative analysis of lung function regions caused by insufficient collaboration between multimodal medical imaging and clinical information.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a system for segmenting and quantifying lung ventilation-perfusion imaging regions, comprising, The data acquisition module acquires SPECT V / Q image data and the corresponding medical record information, and generates a structured dataset. The feature extraction module performs format conversion and multi-scale feature extraction on the SPECT V / Q image data in the structured dataset to generate a compressed feature map dataset. The lung lobe segmentation module inputs the compressed feature map dataset into the U-Net model to segment the lung lobe region and performs morphological post-processing to generate a lung lobe region mask image. The image fusion module performs alpha blending on the masked image of the lung lobe region and the SPECT V / Q image data to generate a color fused image; The feature enhancement module performs Min-Max normalization processing and brightness feature value calculation on a channel-by-channel basis based on the color fusion image. By standardizing the pixel values ​​of each channel and performing a nonlinear transformation based on brightness, it highlights useful information, suppresses noise interference, and performs feature enhancement to generate an enhanced feature tensor. The clustering analysis module uses the K-means clustering algorithm to divide the enhanced feature tensor into sub-regions, calculates the V / Q ratio of each sub-region and maps it to the CT image, and generates a quantitative analysis report.

[0007] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the step of performing format conversion and multi-scale feature extraction on the SPECT V / Q image data to generate a compressed feature map dataset includes the following steps: Convert the SPECT V / Q image data to generate a PNG format image; Input a PNG format image into a hollow pyramid network to capture multi-scale receptive fields and output multi-scale feature maps. The shallow edge feature map and the deep semantic feature map are extracted from the multi-scale feature map and then concatenated through the feature fusion channel to obtain the fused feature map; The fused feature map is compressed by feature distillation through depthwise separable convolutional layers to generate a compressed feature map dataset.

[0008] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the step of inputting the compressed feature map dataset into the U-Net model to segment the lung lobe region and performing morphological post-processing to generate a lung lobe region mask image includes the following steps: The encoder path processes the compressed feature map dataset to generate a dimensionality-reduced feature map; The decoder path upsamples the dimensionality-reduced feature map and fuses it with the encoder feature skip connections to output a probability distribution map. Morphological post-processing is performed on the probability distribution map to output a masked image of the lung lobe region.

[0009] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the morphological post-processing includes: An erosion operation is performed on the masked image of the lung lobe region to remove isolated noise points, resulting in a denoised mask image; Perform a dilation operation on the denoised mask image to fill the internal holes and output a continuous region mask image; Edge patching is used to repair boundary breaks in continuous region mask images.

[0010] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the generation of the color fusion image includes the following steps. Convert the masked image of the lung lobe region into a multi-channel grayscale image to generate a grayscale mask image; The grayscale mask image is spatially registered with the SPECT V / Q image by feature matching, and the registered image pair is output. Based on the anatomical structure of the lung lobes, the registered image pairs are dynamically fused in different regions. Adaptive fusion weights are calculated according to the spatial positions of the left and right lung lobes. The left lung region is weighted using the first weight coefficient, and the right lung region is weighted using the second weight coefficient to generate a color fused image.

[0011] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the step of performing Min-Max normalization processing and brightness feature value calculation includes the following steps. Calculate the normalization parameters independently for each channel of the color fusion image and output the normalized channel data; Luminance feature values ​​are calculated based on normalized channel data to generate a luminance feature map.

[0012] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the feature enhancement, generating enhanced feature tensors, includes the following steps: Invalid pixel data in the brightness feature map are excluded, and channel-wise Min-Max normalization is performed on the valid pixel data to obtain a normalized pixel matrix. Multi-scale convolution operations are applied to the standardized pixel matrix to extract local texture features at different scales and integrate them to generate a texture feature vector; Multi-scale global context features are captured from the normalized pixel matrix using spatial pyramid pooling operations, and a context feature vector is output. The texture feature vector and the context feature vector are fused, and an enhanced feature tensor is generated by highlighting useful information and suppressing noise interference through nonlinear transformation.

[0013] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the step of using the K-means clustering algorithm to divide the enhanced feature tensor into sub-regions includes the following steps. Extract the spatial coordinates, brightness values, and texture feature vectors from the enhanced feature tensor to construct a multidimensional feature vector set; The number of clusters is set according to the symmetry of the left and right lung anatomical structures, and the cluster centers are initialized to be evenly distributed in space. Calculate the Euclidean distance from each feature vector in the multidimensional feature vector set to each cluster center, assign cluster labels based on the nearest neighbor principle, and generate an initial sub-region partitioning map; Iteratively update the cluster center position based on the mean of the feature vectors within the cluster according to the initial sub-region partitioning map, and optimize the cluster boundary distribution; The calculation terminates when the cluster boundary distribution of consecutive iterations falls below the boundary threshold, and the cluster label map of the left and right lung function sub-regions is output.

[0014] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantification analysis system of the present invention, the calculation of the V / Q ratio of each region and mapping it to the CT image includes the following steps: The average ventilation and irrigation intensity of each sub-region is calculated based on the cluster label map, and the V / Q ratio is output. The V / Q ratio is mapped to the CT image to generate a discrete color map. The discrete color map is converted into a continuous heatmap using an interpolation algorithm.

[0015] As a preferred embodiment of the lung ventilation-perfusion imaging region image segmentation and quantitative analysis system of the present invention, the CT image refers to a computed tomography image that provides a baseline of lung anatomical structure.

[0016] The beneficial effects of this invention are as follows: SPECT V / Q imaging is dynamically linked with electronic medical records to construct a structured dataset. A gated attention mechanism is used to screen pathology-related features, significantly improving the accuracy of identifying pulmonary embolism and COPD. A hollow pyramid network integrates four-branch features to capture alveolar to lobar structures, and depthwise separable convolution compresses feature dimensions, preserving microlesions such as ground-glass nodules while reducing GPU memory consumption. Based on enhanced feature tensors, functional subregions are automatically divided, and RANSAC-optimized registration and cubic spline interpolation generate anatomically calibrated V / Q heatmaps, eliminating manual setting errors. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a lung ventilation-perfusion imaging region image segmentation and quantitative analysis system.

[0019] Figure 2 A schematic diagram for generating structured datasets.

[0020] Figure 3 This is a schematic diagram of feature extraction.

[0021] Figure 4 Example 1 of a binary graph after morphological processing.

[0022] Figure 5 Example 2 of a binary graph after morphological processing.

[0023] Figure 6 Example 3 is a binary graph after morphological processing.

[0024] Figure 7 This is a schematic diagram of color fusion image generation.

[0025] Figure 8 This is a schematic diagram of enhanced feature tensor generation.

[0026] Figure 9 This is a schematic diagram of K-means cluster analysis of the three-dimensional lung V / Q ratio. Detailed Implementation

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0029] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0030] Reference Figures 1-9 This is one embodiment of the present invention, which provides a lung ventilation-perfusion imaging region image segmentation and quantification analysis system, including the following steps: The data acquisition module obtains SPECT V / Q image data and corresponding medical record information, and generates a structured dataset.

[0031] Specifically, see Figure 2 It receives raw image data transmitted by SPECT device in real time through the DICOM protocol interface, including DICOM sequence files of lung ventilation phase and lung perfusion phase, and outputs uncompressed raw DICOM dataset; Based on the patient ID, the system automatically retrieves relevant fields from the electronic medical record, including basic information, medical history characteristics, and laboratory indicators, and outputs a structured case JSON file. When a field is missing, it is automatically filled with the default value NULL.

[0032] The basic information includes age, gender, height / weight (e.g., triggering the marker when BMI ≥ 18.5); medical history characteristics include chronic obstructive pulmonary disease (COPD) stage and pulmonary embolism (PE) history markers; laboratory indicators include arterial partial pressure of oxygen (PaO2) and D-dimer concentration (unit: μg / L).

[0033] All of the above information has been agreed to by the user and is used for legitimate purposes.

[0034] Perform data consistency checks on the original DICOM dataset and structured medical record JSON files, including checks on image-medical record correlation, completeness, and timeliness; Among them, the image-medical record correlation verification refers to the complete match between the patient ID in the DICOM file header and the electronic medical record; the integrity verification refers to the consistency of the number of images in the ventilation phase / perfusion phase; and the timeliness verification refers to the difference between the image acquisition time and the medical record recording time being ≤24 hours. If the verification passes, the status code 200_OK will be output. If it fails, an error type code (such as E101: ID mismatch) will be returned.

[0035] The original DICOM dataset that has passed verification is linked to the structured medical record JSON file with the check number as the primary key to create an index, and the output is a structured dataset containing image data and medical record information.

[0036] The feature extraction module performs format conversion and multi-scale feature extraction on the SPECT V / Q image data in the structured dataset to generate a compressed feature map dataset.

[0037] Specifically, see Figure 3 The system reads the original DICOM file from the structured dataset, converts each DICOM slice into PNG format, preserves the original grayscale resolution (e.g., 16-bit to 8-bit linear mapping) during the conversion, and writes the key metadata (e.g., window width / window level) from the DICOM file header into the Exif information area of ​​the PNG file to generate a standardized PNG image sequence. The standardized PNG image sequence is input into the Hollow Pyramid (ASPP) feature extraction network, which processes the standardized PNG image sequence through four parallel branches. Branch 1 uses a 3×3 convolutional layer with a hole ratio of 6 to capture large-scale contextual features. Branch 2 uses a 3×3 convolutional layer with a hole ratio of 12 to extract ultra-large-scale semantic features; Branch 3 uses a 3×3 convolutional layer with a porosity of 18 to perceive the global anatomical structure; Branch 4 generates image-level features by compressing spatial dimensions through a global average pooling layer; The outputs of the four parallel branches are channel aligned by a 1×1 convolutional layer and then spliced ​​and fused to generate a multi-scale fused feature map. The shallow edge feature map (128×128 resolution) of the third residual block is extracted from the multi-scale fused feature map, and the deep semantic feature map (32×32 resolution) at the end of the network is extracted simultaneously. The shallow edge feature map is upsampled to 32×32 resolution through bilinear interpolation and concatenated with the deep semantic feature map along the channel dimension. The concatenated feature map is then fused through a 3×3 convolutional layer to output an enhanced feature map. The enhanced feature map is input into a depthwise separable convolutional layer, which includes a depthwise convolution stage and a pointwise convolution stage. The depthwise convolution stage uses a 3×3 convolutional kernel for spatial filtering, while the pointwise convolution stage uses a 1×1 convolutional kernel to perform channel transformation. By separating spatial and channel operations, efficient computation is achieved, the dimension of the feature map is compressed, and the final output is a compressed feature map dataset.

[0038] The lung segmentation module inputs the compressed feature map dataset into the U-Net model to segment the lung lobe region and performs morphological post-processing to generate a lung lobe region mask image.

[0039] Specifically, see Figures 4-6 The masked image of the lung lobe region is input into the encoder path of the U-Net model. The encoder contains a five-level downsampling structure. Each level passes through two consecutive 3×3 convolutional layers, extracts local features through the ReLU activation function, and then passes through a 2×2 max pooling layer to halve the spatial resolution while doubling the number of channels, gradually capturing the global semantic features of the lung lobe anatomical structure. After five levels of downsampling, the output is a low-resolution feature map with a spatial size reduced to 1 / 32 of the original image. The low-resolution feature map contains a deep semantic representation of the lung lobe region. In the decoder path of the U-Net model, a five-level upsampling operation is performed on the low-resolution feature map. At each level, the size of the low-resolution feature map is first doubled by transposed convolution (2×2 kernel, stride 2), and then concatenated with the feature map of the corresponding layer of the encoder through skip connections to achieve channel-dimensional concatenation. The concatenated feature map is then processed by two 3×3 convolutional layers, and multi-scale information is fused through the ReLU activation function to gradually restore the spatial resolution to the original size. Finally, a probability distribution map of the same size as the input image is output, where each pixel value in the probability distribution map represents the probability of belonging to the lung lobe region (0-1 interval). The output probability distribution map is binarized using a fixed threshold: pixels with a probability ≥ the fixed threshold are marked as 1 (lung lobe region), and the rest are marked as 0 (background), generating an initial binary mask image; morphological opening operations, including erosion and dilation operations, are performed on the initial binary mask image; It should be noted that the fixed threshold setting is based on the bimodal characteristics of the output probability distribution of the U-Net model. The segmentation critical value is dynamically determined by maximizing the inter-class variance between the lung lobe region and the background, ensuring robust separation of anatomical structures from noise.

[0040] The erosion operation uses a 3×3 circular structuring element to traverse the mask image and eliminate isolated noise points (such as tracheal artifacts); the dilation operation uses the same structuring element to fill the tiny gaps caused by erosion (such as blood vessel interruption areas) and outputs a lung lobe mask image with optimized connectivity. The boundary break regions of the optimized mask image are detected and repaired using a region growing algorithm: with the break endpoint as the center, the nearest neighbor boundary point is searched along the 8-neighbor direction, and a linear connection path is generated within a maximum distance of 5 pixels. The closed region formed by the linear connection path is filled and smoothly transitioned, and a lung lobe region mask image with anatomical structural integrity is output to ensure the continuity of the boundaries of the lung apex, lung base and interlobar fissure.

[0041] The image fusion module performs alpha blending on the masked image of the lung lobe region and the SPECT V / Q image data to generate a color fused image.

[0042] Specifically, see Figure 7 The lung lobe region mask image is converted into a three-channel grayscale image using a channel expansion algorithm: lung lobe region pixels with a mask value of 1 are mapped to pure white, and background region pixels are mapped to pure black, generating a multi-channel grayscale mask image of the same size as the SPECT V / Q image. This preserves the geometric accuracy and spatial information of the original mask, ensuring the consistency of anatomical structures in subsequent registration. A feature-point-based registration algorithm is used to align the multi-channel grayscale mask image with the original SPECT V / Q image. Lung lobe contour corner points are extracted as feature points in the grayscale mask image, and texture feature points are extracted at the corresponding anatomical locations in the SPECT V / Q image. The cosine similarity between the lung lobe contour corner points and texture feature points is calculated using a FLANN (Fast Approximate Nearest Neighbor) matcher to establish the correspondence between feature points. Simultaneously, the RANSAC algorithm is used to estimate the optimal affine transformation parameters, forming an affine transformation matrix to eliminate interference from outlier matching points. Furthermore, estimating the optimal affine transformation parameters using the RANSAC algorithm involves, after establishing the correspondence between feature points using the FLANN matcher, randomly selecting the smallest sample subset (e.g., 3 pairs of points) from the feature point pairs using the RANSAC algorithm to calculate the initial affine transformation parameters and generate candidate transformation matrices. Subsequently, the spatial position deviation of all feature points under the candidate transformation matrices is calculated, and points with deviations less than the deviation threshold are identified as inliers and their numbers are counted. The sampling and inlier counting process is iteratively executed, and the calculation is terminated when the proportion of inliers no longer increases in N consecutive iterations. The candidate transformation matrix with the largest number of inliers is selected as the optimal solution. Finally, based on the inlier set of the optimal solution, the accurate affine transformation parameters are fitted using the least squares method, outputting a six-degree-of-freedom affine transformation matrix that eliminates outlier interference, including rotation, scaling, and translation parameters.

[0043] It should be noted that the deviation threshold is determined statistically based on the average edge alignment error of the lung parenchyma anatomical structure.

[0044] An affine transformation matrix is ​​applied to resample the grayscale mask image to achieve complete spatial alignment with the SPECT V / Q image, outputting a precisely registered image pair. For this precisely registered image pair, a partitioned dynamic alpha transparency fusion process is performed: based on the anatomical structure of the lung lobes, the spatial positions of the left and right lung lobes are identified. The left lung region is weighted using a first weight coefficient, and the right lung region is weighted using a second weight coefficient. The SPECT V / Q image is used as the base layer, and the registered mask image is used as the overlay layer. An offset of 0 is set to ensure no brightness shift. The initial fused pixel values ​​are calculated according to a pixel-by-pixel weighting formula. Channel enhancement processing is performed on the initial fused pixel values: the display of the lung lobe region is enhanced by the red channel gain, so that the ventilation / perfusion status is superimposed on the anatomical structure in a semi-transparent red, while the original grayscale information is maintained in the non-lung lobe region, thus generating a channel-enhanced image. Post-processing optimization of the enhanced images was performed: the Laplacian operator was used to enhance the contours of the lung lobes to improve the recognition of anatomical structures, and the contrast was optimized by adaptive histogram equalization to highlight areas of abnormal perfusion. Scale bars and direction markers were added to finally generate a 24-bit true-color fused image, in which low-perfusion areas are shown as dark red patches and high-ventilation areas are shown as bright red uniformly distributed.

[0045] The feature enhancement module performs Min-Max normalization processing and brightness feature value calculation on a channel-by-channel basis based on the color fusion image. By standardizing the pixel values ​​of each channel and performing a brightness-based nonlinear transformation, it highlights useful information, suppresses noise interference, and enhances features to generate an enhanced feature tensor.

[0046] Specifically, see Figure 8The color fusion image is decomposed into three independent channels: red (R), green (G), and blue (B). An independent normalization operation is performed on each channel: the effective pixel value range of each channel is calculated (excluding the background area), and the minimum and maximum reference points of the channel are determined. The pixel values ​​are mapped to the [0,1] interval through linear transformation, the relative density gradient characteristics of the lung region are preserved, and the normalized three-channel floating-point data matrix is ​​output. For the normalized three-channel floating-point data matrix, brightness feature transformation is performed: based on the distribution characteristics of photoreceptor cells in the human retina, different weight coefficients are assigned to the R, G, and B channels; the weighted calculation result is used as the pixel brightness value to generate a single-channel brightness feature map, and a background threshold judgment mechanism is established at the same time: the average pixel brightness of the image edge region is calculated, and regions with an average brightness value lower than the average threshold are marked as invalid regions (including the background outside the chest cavity and device artifacts), and the brightness feature map with invalid region markings is output; Based on the invalid region markers in the brightness feature map, a binary mask is created to identify the valid lung regions: all pixels are traversed, and regions with brightness values ​​greater than or equal to the background threshold are identified as valid regions; morphological closing operations are applied to fill the tiny gaps within the valid regions to ensure the continuity of lung tissue; the original data is filtered based on the binary mask to remove invalid region pixels, and a normalized pixel matrix containing only lung parenchyma is output. Multi-scale convolution operations are performed on the normalized pixel matrix to capture texture features at different anatomical levels, as follows: Phase 1: Multi-scale convolution captures anatomical hierarchical features; Using lung CT images in a standardized pixel matrix as input, different sized "lenses" (convolution kernels) are used to observe different levels of anatomical structures. Subtle grayscale changes are captured by small-scale scanning with 3×3 small-scale convolution kernels, and microscopic texture features at the alveolar level are extracted, including ground-glass nodule texture. Ground-glass nodule texture appears as a blurry shadow like frosted glass on CT, which is an early sign of many lung diseases. Small convolution kernels are best suited to capture such subtle texture changes. A 5×5 mesoscale convolution kernel was used for medium-range scanning to capture the mesoscopic structural features at the pulmonary lobule level, with a focus on extracting vascular branching patterns and accompanying structures. The pulmonary lobule is the basic structural and functional unit of the lung, containing bronchioles and their accompanying blood vessels. The 5x5 convolution kernel can clearly show whether the morphology, course, and distribution of these small blood vessels are normal. Using 7×7 large-scale convolutional kernels for large-area scanning, we can perceive the macroscopic distribution characteristics at the lung lobe level, paying particular attention to the gradient change pattern between ventilation and perfusion. Pulmonary embolism can cause obstruction of pulmonary blood flow (perfusion), while airflow (ventilation) may be normal. This "mismatch" will form a specific density distribution gradient on CT. Large convolutional kernels can perceive macroscopic distribution changes. The three sets of convolutional outputs are processed by nonlinear activation functions to enhance feature representation capabilities and are combined into a multi-scale texture feature vector set to comprehensively characterize the structural heterogeneity of lung tissue at multiple scales.

[0047] Phase Two: Spatial Pyramid Pooling Integration of Global Context; Multi-scale texture feature vector sets are input into the spatial pyramid pooling layer to achieve multi-level global context information integration: The first-level max pooling divides the feature map in the multi-scale texture feature vector set into a 4×4 grid, extracts the maximum response value of each local region, captures the context features of the medium range and highlights the salient regions (such as the region most suspected of being a lesion). The second-level average pooling uses a 2×2 grid to divide the feature map in the multi-scale texture feature vector set into four large blocks, and takes the average value of each block to reflect the overall situation of the corresponding region and capture the contextual distribution characteristics of a wider range. The third-level global average pooling performs a global averaging operation, averaging the entire feature map in the multi-scale texture feature vector set to generate an image-level summary vector that represents the overall attributes of the entire image. The pooling outputs of the three levels are concatenated to form a high-dimensional global context feature vector that contains the full lung function distribution characteristics.

[0048] Phase 3: Cross-modal fusion and gated attention mechanism; The high-dimensional global context feature vector (knowing the "overall situation") interacts with the multi-scale texture feature vector set (containing "local details"), and automatically determines "which local texture features are more worthy of attention in a certain global context" through the feature cross-attention mechanism. For example, in the vascular distribution area, the texture of the alveoli does not need much attention, but the morphological features of the blood vessels themselves are very important, thus strengthening the response to pathological areas. A gated attention network is established to dynamically adjust the contribution ratio of global and local features in the final result through learnable parameters, and to assign higher weights to high-risk areas of pulmonary embolism (such as the distribution area of ​​pulmonary segmental arteries). Residual connections are introduced during the fusion process to preserve the original feature information and ensure effective gradient propagation, and to output an enhanced feature tensor, in which pathological areas (such as perfusion defects caused by pulmonary embolism) show high-intensity feature activation, which significantly improves the segmentation and recognition performance of downstream tasks.

[0049] It should be noted that the enhanced feature tensor contains multi-scale information of micro, meso and macro features, achieving a unified global and local context that knows both details and the overall picture. Due to the attention mechanism and gating weighting, pathological areas such as perfusion defects caused by pulmonary embolism will show high-intensity feature activation in the enhanced feature tensor, as if they are highlighted, making the pathological areas significantly enhanced and more suitable for downstream tasks. This is because the features that the network needs to judge have been made very obvious and prominent. As input to subsequent classifiers or segmentation networks, it can greatly improve the accuracy and reliability of pulmonary embolism detection.

[0050] The clustering analysis module uses the K-means clustering algorithm to divide the enhanced feature tensor into sub-regions, calculates the V / Q ratio of each sub-region and maps it to the CT image, and generates a quantitative analysis report.

[0051] Specifically, see Figure 9 Based on enhanced feature tensors, spatial coordinates, brightness values, and texture feature vectors are extracted to construct a multidimensional feature vector set. Eighteen cluster centers (nine zones in the left lung / nine zones in the right lung) are pre-defined according to the anatomical symmetry of the left and right lungs, and the center point positions are initialized using a spatially uniform distribution strategy. The Euclidean distance from the feature vectors to each center is calculated, and cluster labels are assigned using the nearest neighbor principle to generate an initial sub-region partitioning map. The cluster center positions to the mean feature values ​​within the clusters are iteratively updated to dynamically optimize the cluster boundary distribution. The calculation terminates when the mean squared error of the cluster boundary displacement is ≤ the variance threshold pixels for multiple consecutive iterations, and the final cluster label map conforming to the anatomical functional zoning is output. Pixel sets of each sub-region are extracted based on the final cluster label map, and the mean ventilation intensity (V) and mean perfusion intensity (Q) are calculated respectively. The V / Q ratio matrix of each region is output; Jet chromatographic mapping rules are constructed. Low V / Q areas (e.g., <0.8) are mapped to deep red, indicating perfusion defects; normal areas (e.g., 0.8-1.2) are mapped to yellow; high V / Q areas (e.g., >1.2) are mapped to deep blue, indicating ventilation impairment. Discrete ratios are converted into continuous color gradients through cubic spline interpolation, which are then superimposed on CT images to generate a continuous heatmap that integrates anatomical localization and functional evaluation. Here, CT images refer to computed tomography images that provide a baseline of lung anatomy. A visual diagnostic framework is constructed by integrating continuous heatmaps and V / Q ratio matrices: heatmap profiles are rendered in the coronal, sagittal, and axial planes, and dynamic tables are embedded to display the V / Q statistics (mean / standard deviation / extreme values) of each lung lobe. Interactive sliders are added to adjust the window width and window level to assist in the precise localization of lesions, and an HTML5 diagnostic analysis framework that supports multi-dimensional browsing is output. The HTML5 diagnostic framework is input into the Playwright headless browser engine and automatically converted into a PDF document conforming to the DICOM SR standard. Key quantitative indicators in the PDF are extracted through OCR and dynamically linked with the medical record database: abnormal D-dimer values ​​in pulmonary embolism patients are associated with low V / Q regions; the correlation between emphysema regions and high V / Q regions in COPD patients is marked, and risk grading prompts (high / intermediate / low risk) are inserted. A structured initial diagnosis report with clinical decision support annotations is output. The system integrates structured initial diagnosis reports, patient basic information, and laboratory indicators to construct a final diagnostic report. A patient basic information matrix (age / gender / BMI) is embedded in the header of the final diagnostic report. The image analysis area links heatmaps with abnormal laboratory indicators (e.g., PaO2 < 80 mmHg is highlighted in red for warning). The diagnostic conclusion area generates a natural language description, such as "A 2.1cm × 1.8cm perfusion defect area (V / Q = 0.63) is seen in the lower lobe of the left lung, combined with elevated D-dimer (624 μg / L), consistent with pulmonary embolism imaging findings," and outputs a quantitative analysis report.

[0052] In summary, this invention constructs a structured dataset by dynamically linking SPECT V / Q imaging with electronic medical records, and filters pathology-related features through a gated attention mechanism, significantly improving the accuracy of identifying pulmonary embolism and COPD. A hollow pyramid network integrates four-branch features to capture alveolar to lobar structures, and compresses feature dimensions through depthwise separable convolution, preserving microlesions such as ground-glass nodules while reducing GPU memory consumption. Based on enhanced feature tensors, functional subregions are automatically divided, and RANSAC-optimized registration and cubic spline interpolation are used to generate anatomically calibrated V / Q heatmaps, eliminating manual setting errors.

[0053] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A system for segmenting and quantifying lung ventilation-perfusion imaging regions, characterized in that: include, The data acquisition module acquires SPECT V / Q image data and the corresponding medical record information, and generates a structured dataset. The feature extraction module performs format conversion and multi-scale feature extraction on the SPECT V / Q image data to generate a compressed feature map dataset; The lung lobe segmentation module inputs the compressed feature map dataset into the U-Net model to segment the lung lobe region and performs morphological post-processing to generate a lung lobe region mask image. The image fusion module performs alpha blending on the masked image of the lung region and the SPECT V / Q image data to generate a color fused image; The feature enhancement module performs Min-Max normalization processing and brightness feature value calculation on a channel-by-channel basis based on the color fusion image. By standardizing the pixel values ​​of each channel and performing a nonlinear transformation based on brightness, it highlights useful information, suppresses noise interference, and performs feature enhancement to generate an enhanced feature tensor. The clustering analysis module uses the K-means clustering algorithm to divide the enhanced feature tensor into sub-regions, calculates the V / Q ratio of each sub-region and maps it to the CT image, and generates a quantitative analysis report.

2. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The process of converting the format of the SPECT V / Q image data and extracting multi-scale features to generate a compressed feature map dataset includes the following steps: Convert the SPECT V / Q image data to generate a PNG format image; Input a PNG format image into a hollow pyramid network to capture multi-scale receptive fields and output multi-scale feature maps. The shallow edge feature map and the deep semantic feature map are extracted from the multi-scale feature map and then concatenated through the feature fusion channel to obtain the fused feature map; The fused feature map is compressed by feature distillation through depthwise separable convolutional layers to generate a compressed feature map dataset.

3. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The process of inputting the compressed feature map dataset into the U-Net model to segment the lung lobe region and performing morphological post-processing to generate a lung lobe region mask image includes the following steps: The encoder path processes the compressed feature map dataset to generate a dimensionality-reduced feature map; The decoder path upsamples the dimensionality-reduced feature map and fuses it with the encoder feature skip connections to output a probability distribution map. Morphological post-processing is performed on the probability distribution map to output a masked image of the lung lobe region.

4. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 3, characterized in that: The morphological post-processing includes, An erosion operation is performed on the masked image of the lung lobe region to remove isolated noise points, resulting in a denoised mask image; Perform a dilation operation on the denoised mask image to fill the internal holes and output a continuous region mask image; Edge patching is used to repair boundary breaks in continuous region mask images.

5. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The generation of the color fused image includes the following steps: Convert the masked image of the lung lobe region into a multi-channel grayscale image to generate a grayscale mask image; The grayscale mask image is spatially registered with the SPECT V / Q image by feature matching, and the registered image pair is output. Based on the anatomical structure of the lung lobes, the registered image pairs are dynamically fused in different regions. Adaptive fusion weights are calculated according to the spatial positions of the left and right lung lobes. The left lung region is weighted using the first weight coefficient, and the right lung region is weighted using the second weight coefficient to generate a color fused image.

6. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The process of performing Min-Max normalization and calculating luminance feature values ​​includes the following steps: Calculate the normalization parameters independently for each channel of the color fusion image and output the normalized channel data; Luminance feature values ​​are calculated based on normalized channel data to generate a luminance feature map.

7. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The process of performing feature enhancement and generating enhanced feature tensors includes the following steps: Invalid pixel data in the brightness feature map are excluded, and channel-wise Min-Max normalization is performed on the valid pixel data to obtain a normalized pixel matrix. Multi-scale convolution operations are applied to the standardized pixel matrix to extract local texture features at different scales and integrate them to generate a texture feature vector; Multi-scale global context features are captured from the normalized pixel matrix using spatial pyramid pooling operations, and a context feature vector is output. The texture feature vector and the context feature vector are fused, and an enhanced feature tensor is generated by highlighting useful information and suppressing noise interference through nonlinear transformation.

8. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The process of using the K-means clustering algorithm to partition the enhanced feature tensor into sub-regions includes the following steps. Extract the spatial coordinates, brightness values, and texture feature vectors from the enhanced feature tensor to construct a multidimensional feature vector set; The number of clusters is set according to the symmetry of the left and right lung anatomical structures, and the cluster center points are initialized to be evenly distributed in space. Calculate the Euclidean distance from each feature vector in the multidimensional feature vector set to each cluster center, assign cluster labels based on the nearest neighbor principle, and generate an initial sub-region partitioning map; Iteratively update the cluster center position based on the mean of the feature vectors within the cluster according to the initial sub-region partitioning map, and optimize the cluster boundary distribution; The calculation terminates when the cluster boundary distribution of consecutive iterations falls below the boundary threshold, and the cluster label map of the left and right lung function sub-regions is output.

9. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The calculation of the V / Q ratio for each region and its mapping to the CT image includes the following steps. The average ventilation and irrigation intensity of each sub-region is calculated based on the cluster label map, and the V / Q ratio is output. The V / Q ratio is mapped to the CT image to generate a discrete color map. The discrete color map is converted into a continuous heatmap using an interpolation algorithm.

10. The lung ventilation-perfusion imaging region image segmentation and quantification analysis system as described in claim 1, characterized in that: The CT images refer to computed tomography images that provide a baseline of lung anatomy.

Images