Lung image recognition method and system for clinical diagnosis of respiratory medicine department

Abstract

The invention provides a lung image recognition method and system for clinical diagnosis of the respiratory medicine department. The method comprises the steps that a binary mask containing complete lung parenchyma is obtained from a lung image; segmenting a target lung field image from the lung image based on the binary mask, and extracting a texture saliency map from the target lung field image; performing pixel-level fusion on the texture saliency map and the target lung field image to obtain a lung feature image, and performing multi-resolution pyramid decomposition on the lung feature image to obtain a Gaussian pyramid layer and a Laplacian pyramid layer; direction gradient histogram features and local binary pattern features are extracted from the Gaussian pyramid layer and the Laplacian pyramid layer respectively, and then a multi-resolution joint feature vector is constructed; and inputting the multi-resolution joint feature vector into a pre-trained image analysis network, and positioning a focus area in the lung image. According to the technical scheme provided by the invention, the lesion area in the lung image can be identified under the coupling interference of the anatomical structure and the pathological features.

Description

Technical Field

[0001] This application relates to the field of lung image recognition technology, and more specifically, to a lung image recognition method and system for clinical diagnosis in respiratory medicine. Background Technology

[0002] Lung image recognition is one of the core applications in the field of medical artificial intelligence. Lung image recognition is mainly based on deep learning, especially the development of convolutional neural networks (CNN). Traditional diagnosis relies on doctors manually interpreting images, which is time-consuming and easily affected by subjective factors. This technology trains on large-scale, high-quality lung image datasets (such as X-rays and CT scans), enabling the model to automatically learn and extract deep features from subtle textures to complex structures, thereby achieving accurate localization, segmentation, and classification of lung nodules, pneumonia, tuberculosis, and tumors. Therefore, lung image recognition is gradually becoming an important tool for modern medical image diagnosis.

[0003] In existing lung image recognition, which is based on deep learning, particularly convolutional neural networks (CNNs), lung X-ray image data is first input into the network. Then, the network automatically and hierarchically extracts image features through multiple convolutional and pooling layers, from low-level edges and textures to mid-level local structures and high-level semantically meaningful lesion areas. Finally, these abstract features are passed to fully connected layers, and the classifier outputs the corresponding recognition results. However, in lung image recognition, the pulmonary vascular tree, as a high-brightness anatomical structure, is distributed throughout the entire lung field. The complex texture features it generates can create significant background noise. Traditional methods directly extract features from the entire lung field image, making it difficult to fundamentally separate the coupling interference between anatomical structures and pathological features. This leads to inaccurate identification of lesion areas in lung images. Therefore, how to identify lesion areas in lung images under the coupling interference of anatomical structures and pathological features has become a challenge for the industry. Summary of the Invention

[0004] This application provides a lung image recognition method and system for clinical diagnosis in respiratory medicine, which can identify lesion areas in lung images under the coupling interference of anatomical structure and pathological features.

[0005] In a first aspect, this application provides a method for lung image recognition in respiratory medicine clinical diagnosis, comprising the following steps: The lung image to be identified is acquired, and the lung image is segmented into lung field regions to obtain a binary mask containing complete lung parenchyma. Based on the binary mask, the target lung field image is segmented from the lung image, and texture saliency detection based on local entropy is performed on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma. The texture saliency map is fused with the target lung field image at the pixel level to obtain a pixel-enhanced lung feature image. The lung feature image is then decomposed into a multi-resolution pyramid to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. Oriented gradient histogram features and local binary pattern features are extracted from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively. Then, a multi-resolution joint feature vector is constructed based on the oriented gradient histogram features and the local binary pattern features. The multi-resolution joint feature vector is input into a pre-trained image analysis network to locate the lesion region in the lung image.

[0006] In some embodiments, segmenting the lung image to obtain a binary mask body containing intact lung parenchyma includes: The lung images are subjected to grayscale normalization processing to obtain grayscale normalized lung images; The gray-level normalized lung image was segmented using adaptive threshold segmentation to obtain a preliminary lung field mask; The preliminary lung field mask is morphologically optimized to obtain a morphologically optimized lung field mask; The morphologically optimized lung field mask is binarized to generate a binary mask containing the complete lung parenchyma.

[0007] In some embodiments, segmenting a target lung field image from the lung image based on the binary mask specifically includes: The binary mask is subjected to morphological dilation to obtain the dilated binary mask; The expanded binary mask is multiplied pixel by pixel with the lung image to segment the target lung field image from the lung image.

[0008] In some embodiments, performing texture saliency detection based on local entropy values ​​on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma specifically includes: Preprocessing of the target lung field image yields a preprocessed target lung field image. The local neighborhood block corresponding to each pixel is obtained from the preprocessed target lung field image by using a preset sliding window; Calculate the gray-level histogram for each local neighborhood block and solve for the probability distribution to obtain the local entropy value of each local neighborhood block; All local entropy values ​​are integrated to form a local entropy value map, and the local entropy value map is normalized to obtain a normalized local entropy value map. Texture saliency enhancement is performed on the normalized local entropy map to generate a texture saliency map with complex texture patterns of lung parenchyma.

[0009] In some embodiments, performing pixel-level fusion of the texture saliency map with the target lung field image to obtain a pixel-enhanced lung feature image specifically includes: Spatial rigid registration and grayscale normalization are performed on the texture saliency map and the target lung field image to obtain the aligned normalized texture saliency map and the normalized target lung field image. Pixel-level fusion weights are calculated based on the entropy distribution of the aligned normalized texture saliency map to obtain a texture-guided fusion weight map. The aligned, normalized target lung field image and the normalized texture saliency map are then subjected to pixel-wise weighted summation according to the texture-guided fusion weight map to obtain a pixel-enhanced lung feature image.

[0010] In some embodiments, performing multi-resolution pyramid decomposition on the lung feature image to obtain a Gaussian pyramid layer containing global contextual information and a Laplacian pyramid layer containing local detailed features specifically includes: Gaussian filtering is applied to the lung feature images to obtain filtered lung images; The filtered lung images are iteratively downsampled and Gaussian smoothed to generate multiple Gaussian pyramid layers; Upsampling and Gaussian smoothing are performed on each Gaussian pyramid layer to obtain the reconstructed Gaussian image at the corresponding scale; Each Gaussian pyramid layer is subjected to pixel-by-pixel difference operation with the corresponding reconstructed Gaussian image to generate multiple Laplacian pyramid layers. By integrating all Gaussian pyramid layers and Laplacian pyramid layers, we obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detailed features.

[0011] In some embodiments, images of the lungs to be identified are acquired using a CT scanner.

[0012] Secondly, this application provides a lung image recognition system for clinical diagnosis in respiratory medicine, comprising: The acquisition module is used to acquire the lung image to be identified and to segment the lung image into lung field regions to obtain a binary mask containing complete lung parenchyma. The processing module is used to segment the target lung field image from the lung image based on the binary mask, and perform texture saliency detection based on local entropy value on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma. The processing module is further configured to perform pixel-level fusion of the texture saliency map and the target lung field image to obtain a pixel-enhanced lung feature image, and to perform multi-resolution pyramid decomposition on the lung feature image to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. The processing module is further configured to extract directional gradient histogram features and local binary pattern features from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively, and then construct a multi-resolution joint feature vector based on the directional gradient histogram features and the local binary pattern features. The execution module is used to input the multi-resolution joint feature vector into a pre-trained image analysis network to locate the lesion region in the lung image.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described method for lung image recognition in respiratory medicine clinical diagnosis.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for lung image recognition in respiratory medicine clinical diagnosis.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The lung image recognition method and system for clinical diagnosis in respiratory medicine provided in this application first acquires a lung image to be identified and segments the lung field region of the lung image to obtain a binary mask containing the complete lung parenchyma. Second, based on the binary mask, a target lung field image is segmented from the lung image, and texture saliency detection based on local entropy is performed on the target lung field image to generate a texture saliency map with complex texture patterns of the lung parenchyma. Further, the texture saliency map and the target lung field image are fused at the pixel level to obtain a pixel-enhanced lung feature image, and the lung feature image is decomposed into a multi-resolution pyramid to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. Then, directional gradient histogram features and local binary pattern features are extracted from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively, and a multi-resolution joint feature vector is constructed based on the directional gradient histogram features and the local binary pattern features. Finally, the multi-resolution joint feature vector is input into a pre-trained image analysis network to locate the lesion region in the lung image.

[0016] Therefore, this application can identify lesion regions in lung images under the coupling interference of anatomical structures and pathological features. First, by segmenting the lung field region to obtain a binary mask containing the complete lung parenchyma, non-lung parenchyma background interference in the lung image can be effectively eliminated, accurately focusing on the core analysis object, providing an effective data foundation for subsequent feature extraction and lesion identification. Second, based on the binary mask segmentation of the target lung field image and texture saliency detection using local entropy values, complex texture patterns related to lesions in the lung parenchyma can be accurately captured. The generated texture saliency map can intuitively distinguish between pathological textures and normal textures, providing a clear target for subsequent feature enhancement and avoiding lesion localization deviation caused by the coupling interference of anatomical structures and pathological features due to the lung anatomical structures being distributed throughout the entire lung field region. Furthermore, pixel-level fusion enhances the grayscale of pathological areas and normal tissues. By combining multi-resolution pyramid decomposition to obtain Gaussian and Laplacian pyramid layers, the global structural information of lung parenchyma at different scales is preserved while accurately capturing local detail features, achieving comprehensive coverage of lesions of different sizes. Then, by extracting directional gradient histogram features (representing global structure and edge contours) and local binary pattern features (depicting local texture details) and constructing a multi-resolution joint feature vector, the information bias caused by a single feature or single resolution is effectively avoided, significantly improving the recognition accuracy of lesion features. Finally, the joint feature vector is input into a pre-trained image analysis network, leveraging the network's deep association learning ability for high-dimensional features to achieve accurate localization of lesion regions. In summary, the technical solution provided in this application can identify lesion regions in lung images under the coupling interference of anatomical structure and pathological features. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an application scenario architecture for a lung image recognition method for clinical diagnosis in respiratory medicine, as shown in some embodiments of this application. Figure 2 This is an exemplary flowchart of a lung image recognition method for clinical diagnosis in respiratory medicine, according to some embodiments of this application; Figure 3 This is an exemplary flowchart illustrating the determination of a texture saliency map according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a lung image recognition system for clinical diagnosis in respiratory medicine, according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device for implementing a lung image recognition method for clinical diagnosis in respiratory medicine, according to some embodiments of this application. Detailed Implementation

[0018] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] refer to Figure 1 This figure is a schematic diagram of an application scenario architecture for a lung image recognition method for clinical diagnosis in respiratory medicine, according to some embodiments of this application. The application scenario architecture includes a data acquisition terminal, a communication network, and a server. The data acquisition terminal and the server are directly or indirectly connected through the communication network. The data acquisition terminal acquires the lung image to be identified and uploads it to the server. The server segments the lung image to obtain a binary mask containing the complete lung parenchyma. Based on the binary mask, a target lung field image is segmented from the lung image, and texture saliency detection based on local entropy is performed on the target lung field image to generate a texture display with complex texture patterns of the lung parenchyma. The texture saliency map is then fused with the target lung field image at the pixel level to obtain a pixel-enhanced lung feature image. The lung feature image is then decomposed into a multi-resolution pyramid to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. Oriented gradient histogram features and local binary pattern features are extracted from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively. A multi-resolution joint feature vector is then constructed based on the oriented gradient histogram features and the local binary pattern features. This multi-resolution joint feature vector is then input into a pre-trained image analysis network to locate the lesion region in the lung image.

[0020] refer to Figure 2 The figure is an exemplary flowchart of a lung image recognition method for clinical diagnosis in respiratory medicine, according to some embodiments of this application. This lung image recognition method for clinical diagnosis in respiratory medicine mainly includes the following steps: In step 101, the lung image to be identified is acquired, and the lung image is segmented into lung field regions to obtain a binary mask containing complete lung parenchyma.

[0021] In practice, lung images to be identified are obtained through a CT scanner. These lung images refer to chest tomographic images of the human body in the form of a pixel matrix, which are obtained through digital medical imaging equipment. By acquiring lung images, structured and standardized data input can be provided for subsequent identification of lesion areas.

[0022] In some embodiments, segmenting the lung field region of the lung image to obtain a binary mask containing the complete lung parenchyma is achieved by the following steps: The lung images are subjected to grayscale normalization processing to obtain grayscale normalized lung images; The gray-level normalized lung image was segmented using adaptive threshold segmentation to obtain a preliminary lung field mask; The preliminary lung field mask is morphologically optimized to obtain a morphologically optimized lung field mask; The morphologically optimized lung field mask is binarized to generate a binary mask containing the complete lung parenchyma.

[0023] In specific implementation, firstly, existing linear gray-level normalization methods are used to map pixel values ​​of different gray-level ranges in lung images to a standard gray-level range of 0-255, resulting in a gray-level normalized lung image. This gray-level normalized lung image refers to an image where the gray-level values ​​of the original lung image are unified to a fixed range. Secondly, the Otsu adaptive thresholding method in image processing is used to automatically calculate the optimal segmentation threshold in the gray-level normalized lung image that maximizes the inter-class variance between foreground and background. Regions in the gray-level normalized lung image with gray-level values ​​higher than the optimal segmentation threshold are marked as candidate foreground for the lung field, and regions with gray-level values ​​lower than the optimal segmentation threshold are marked as non-lung field background, resulting in a preliminary lung field mask. This preliminary lung field mask refers to a labeled image that marks the lung field area. Then, a morphological closing operation (i.e., first performing a 3×3 structuring element dilation operation on the initial lung field mask to fill the internal holes, and then performing an erosion operation on the same structuring element to restore the lung field contour) combined with connected component analysis is used to traverse all connected regions in the initial lung field mask, and remove non-lung field regions with an area smaller than the preset minimum connected region area (i.e., the specific area can be set according to actual needs, and is not limited here) to obtain a morphologically optimized lung field mask. The morphologically optimized lung field mask refers to the mask image after morphological artifact removal. Finally, a fixed threshold binarization method is used to set the lung field foreground pixel value to 1 and the non-lung field background pixel value to 0 in the morphologically optimized lung field mask to generate a binary mask containing complete lung parenchyma.

[0024] It should be noted that in this application, the binary mask refers to a binary image containing only two gray values, 0 and 1. Gray value 1 strictly corresponds to the complete lung parenchyma region, while gray value 0 corresponds to non-lung parenchyma tissue and background. By determining the binary mask, the spatial range and boundary of the complete lung parenchyma in the lung image can be accurately defined, providing a clear spatial mask basis for subsequent segmentation of the target lung field image and effectively eliminating interference from non-lung parenchyma tissue and imaging background.

[0025] In step 102, a target lung field image is segmented from the lung image based on the binary mask, and texture saliency detection based on local entropy is performed on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma.

[0026] In some embodiments, segmenting the target lung field image from the lung image based on the binary mask is achieved by the following steps: The binary mask is subjected to morphological dilation to obtain the dilated binary mask; The expanded binary mask is multiplied pixel by pixel with the lung image to segment the target lung field image from the lung image.

[0027] In specific implementation, firstly, a morphological dilation algorithm is used, selecting circular structuring elements to dilate the binary mask to compensate for any minor missing edges of the lung parenchyma that may exist during the binary mask segmentation process, resulting in an expanded binary mask. The expanded binary mask refers to a mask image that covers the edges of the lung parenchyma and adjacent minor lung tissue. Secondly, the expanded binary mask is multiplied by the corresponding pixel values ​​of the lung image, that is, the original pixel values ​​of the lung image are retained at the positions with a gray value of 1 in the binary mask, and the pixel values ​​of the lung image are set to 0 at the positions with a gray value of 0, thus obtaining the target lung field image.

[0028] It should be noted that the target lung field image in this application refers to a pure grayscale image containing a complete lung parenchyma region. Determining the target lung field image can eliminate non-lung tissue interference and residual background artifacts in the lung image, retaining a complete and pure lung parenchyma region. This provides a clutter-free analysis carrier for subsequent extraction of the lung vascular network through vascular morphology features, ensuring that vascular extraction focuses only on the vascular structure within the lung parenchyma, avoiding misidentification of non-lung tissue artifacts as blood vessels, providing a standardized grayscale image of the lung parenchyma, reducing the interference of redundant information on feature representation, and improving the accuracy of subsequent image analysis network in locating lesion regions.

[0029] In some embodiments, reference Figure 3 As shown, this figure is an exemplary flowchart of determining a texture saliency map according to some embodiments of this application. In this embodiment, performing texture saliency detection based on local entropy values ​​on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma can be achieved by the following steps: First, in step 1021, the target lung field image is preprocessed to obtain a preprocessed target lung field image; Secondly, in step 1022, a local neighborhood block corresponding to each pixel is obtained from the preprocessed target lung field image through a preset sliding window; Furthermore, in step 1023, a gray-level histogram is calculated for each local neighborhood block and the probability distribution is solved to obtain the local entropy value of each local neighborhood block; Then, in step 1024, all local entropy values ​​are integrated to form a local entropy value map, and the local entropy value map is normalized to obtain a normalized local entropy value map. Finally, in step 1025, the normalized local entropy map is subjected to texture saliency enhancement to generate a texture saliency map with complex texture patterns of lung parenchyma.

[0030] In specific implementation, firstly, a Gaussian filtering algorithm is used to preprocess the target lung field image to obtain a preprocessed target lung field image. Alternatively, other filtering algorithms can be used for preprocessing; this is not limited here. The preprocessed target lung field image refers to a clean lung field image after Gaussian smoothing and denoising. Secondly, a set sliding window (specifically a 3×3 sliding window, this is not limited here) is obtained, and the preprocessed target lung field image is scanned pixel-by-pixel, making each pixel the center pixel of the corresponding sliding window. This results in the division of the local neighborhood block corresponding to each pixel. The local neighborhood block refers to the set of local pixels centered on a single pixel and selected by the sliding window, used to characterize the local texture environment of the center pixel. Further, a gray-level histogram statistical method is used for each local neighborhood block to count the number of pixels at different gray levels within that local neighborhood. Then, by dividing the number of pixels at each gray level by the total number of pixels in the local neighborhood, the probability distribution of each gray level is obtained. This is based on the entropy calculation formula in information theory (entropy = -Σ( The entropy value of each local neighborhood block is calculated using the probability multiplied by log2. This yields the local entropy value of each local neighborhood block. The magnitude of the local entropy value directly reflects the complexity of the local texture; a higher local entropy value indicates a more complex texture. The local entropy value refers to the value that characterizes the local texture complexity of the corresponding central pixel. Then, the local entropy values ​​corresponding to each pixel are arranged one-to-one according to their spatial positions in the preprocessed target lung field image to form a local entropy map. Linear normalization (e.g., min-max normalization) is then used to normalize the local entropy values ​​in the local entropy map. The local entropy value is mapped to the standard interval of 0-1 to obtain a normalized local entropy map, which refers to the lung field image reflecting the complexity of local texture. Finally, a nonlinear gray-scale stretching algorithm (specifically, logarithmic transformation can be used, i.e., the formula s=log(1+k*r) / log(1+k), where k is an adjustment coefficient, usually taken as 255 to adapt to subsequent gray-scale mapping, and r is the input local entropy value) is used to enhance the texture saliency of the normalized local entropy map, generating a texture saliency map with complex texture patterns of lung parenchyma.

[0031] It should be noted that, in this application, texture saliency map refers to a visual image that can intuitively present the distribution of lung parenchyma texture. By determining the texture saliency map, the texture complexity of lung parenchyma in the target lung field image can be quantitatively marked and visually distinguished, providing precise guidance for subsequent pixel-level fusion to prioritize the enhancement of pathological texture. This allows the fused lung feature image to directly amplify the difference between pathological texture and surrounding normal tissue, thereby improving the accuracy of identifying suspicious lesion areas.

[0032] In step 103, the texture saliency map is fused with the target lung field image at the pixel level to obtain a pixel-enhanced lung feature image, and the lung feature image is decomposed into a multi-resolution pyramid to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features.

[0033] In some embodiments, the pixel-level fusion of the texture saliency map with the target lung field image to obtain a pixel-enhanced lung feature image is achieved through the following steps: Spatial rigid registration and grayscale normalization are performed on the texture saliency map and the target lung field image to obtain the aligned normalized texture saliency map and the normalized target lung field image. Pixel-level fusion weights are calculated based on the entropy distribution of the aligned normalized texture saliency map to obtain a texture-guided fusion weight map. The aligned, normalized target lung field image and the normalized texture saliency map are then subjected to pixel-wise weighted summation according to the texture-guided fusion weight map to obtain a pixel-enhanced lung feature image.

[0034] In specific implementation, firstly, a well-known rigid spatial registration algorithm based on mutual information is adopted. Using the spatial coordinate system of the target lung field image as a reference, the mutual information value between the texture saliency map and the target lung field image is calculated, and the translation and rotation parameters are iteratively optimized to ensure a one-to-one correspondence between the pixel positions of the texture saliency map and the target lung field image. Then, linear gray-level normalization is used to uniformly map the registered texture saliency map and the target lung field image to the 0-1 interval, resulting in an aligned normalized texture saliency map and a normalized target lung field image. The aligned normalized texture saliency map refers to a texture saliency representation image whose spatial position completely matches the target lung field image after spatial registration and gray-level normalization. The aligned normalized target lung field image refers to a target lung field image with a gray-level range normalized to 0-1. Then, the normalized local entropy value of each pixel in the aligned normalized texture saliency map is used as the fusion weight for the corresponding position, thereby obtaining a texture-guided fusion weight map. Texture-guided fusion weight map refers to the grayscale image used to guide the priority enhancement of pathological texture regions. Finally, a pixel-by-pixel weighted summation formula is used to multiply each pixel value of the aligned normalized target lung field image by the corresponding fusion weight, and simultaneously multiply each pixel value of the aligned normalized texture saliency map by (1-fusion weight). The two results are then superimposed pixel by pixel to obtain the fused pixel value, thereby obtaining a pixel-enhanced lung feature image. In this way, the high-weight pathological texture regions retain the original details of the target lung field image while the saliency features of the texture saliency map are superimposed to enhance the contrast, while the low-weight normal regions mainly retain the original grayscale information of the target lung field image. Among them, (1-fusion weight) is a complementary coefficient used to balance the degree of preservation of the original information of the target lung field image during the pixel-level fusion process. Its sum with the fusion weight is always 1, and its function is to dynamically adjust the retention ratio of the original grayscale information of the target lung field image according to the lung tissue attributes corresponding to the pixels.

[0035] It should be noted that the pixel-enhanced lung feature images in this application refer to lung parenchymal feature enhancement images that enhance the pathological texture region pixel by pixel. The determination of pixel-enhanced lung feature images can be based on preserving the complete structure of normal lung tissue in the target lung field image. The pathological texture marked by the texture saliency map is accurately enhanced through pixel-level weighted fusion, which significantly improves the gray-level contrast and texture detail recognition between the pathological area and normal tissue. This provides a high signal-to-noise ratio feature source for subsequent multi-resolution pyramid decomposition, allowing the decomposed Gaussian pyramid layer to more clearly capture the global shape of the pathological area and the Laplacian pyramid layer to accurately preserve the local details of the pathological texture. This ensures that the subsequent extraction of oriented gradient histogram and local binary pattern features focuses only on key information directly related to the lesion.

[0036] In some embodiments, performing multi-resolution pyramid decomposition on the lung feature image to obtain a Gaussian pyramid layer containing global contextual information and a Laplacian pyramid layer containing local detailed features is achieved through the following steps: Gaussian filtering is applied to the lung feature images to obtain filtered lung images; The filtered lung images are iteratively downsampled and Gaussian smoothed to generate multiple Gaussian pyramid layers; Upsampling and Gaussian smoothing are performed on each Gaussian pyramid layer to obtain the reconstructed Gaussian image at the corresponding scale; Each Gaussian pyramid layer is subjected to pixel-by-pixel difference operation with the corresponding reconstructed Gaussian image to generate multiple Laplacian pyramid layers. By integrating all Gaussian pyramid layers and Laplacian pyramid layers, we obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detailed features.

[0037] In specific implementation, firstly, a Gaussian filtering algorithm is used to perform Gaussian filtering on the lung feature image to obtain a filtered lung image. The filtered lung image refers to the lung feature image after Gaussian filtering. Secondly, using the filtered lung image as the initial bottom layer image of the Gaussian pyramid, a row-and-column downsampling method is used to select pixels according to a preset scaling rule (i.e., retaining pixels in odd-numbered rows and columns of the filtered lung image and discarding the rest) to achieve the first reduction in the size of the filtered lung image. Gaussian smoothing is then performed on the downsampled image to obtain the first Gaussian pyramid layer. This first Gaussian pyramid layer is used as the next iteration. The input is used to repeat the standardized process of "alternating row and column downsampling - Gaussian smoothing". Each iteration performs size reduction and smoothing based on the previous layer image, and the parameters of the Gaussian kernel used for Gaussian smoothing remain consistent in each iteration. A series of scale images with progressively smaller scales and progressively lower resolutions are generated through a preset number of iterations (the specific number of iterations can be set according to actual needs, and is not limited here). Finally, the initial filtered lung image (as the bottom layer of the pyramid) and all the iteratively generated images of each scale are arranged in descending order of scale to form a multi-layer Gaussian pyramid. The multi-layer Gaussian pyramid refers to the structure from the bottom to the top. The image sequence consists of progressively decreasing layer scales and progressively decreasing resolutions. Further, for each Gaussian pyramid layer, a bilinear interpolation algorithm is first used for upsampling. For each pixel in each Gaussian pyramid layer, the positions of its four nearest neighbors are determined in the corresponding output image (i.e., the target scale of the output image is the scale of the next Gaussian pyramid layer). By calculating the weighted sum of the grayscale values ​​of these four neighboring pixels (the weights are inversely proportional to the spatial distance between pixels, with closer pixels having larger weights), all blank pixels in the output image are filled, thus accurately restoring the scale of the original Gaussian pyramid layer to that of the next Gaussian pyramid layer. The size of the pyramid layer is used to obtain an upsampled image. After obtaining the upsampled image, Gaussian smoothing with the same parameters as the downsampling stage is performed to eliminate jagged artifacts introduced by interpolation, resulting in a reconstructed Gaussian image at the corresponding scale. The reconstructed Gaussian image refers to the reconstructed image whose scale matches the original next-level Gaussian image after upsampling and Gaussian smoothing. Then, pixel-by-pixel difference operation is used to subtract the pixel value of each Gaussian pyramid layer from the pixel value of the reconstructed Gaussian image at the corresponding scale to obtain the grayscale difference between the two images, thereby generating a multi-layer Laplacian pyramid layer. The multi-layer Laplacian pyramid layer refers to an image sequence composed of grayscale differences at various scales.Finally, taking the multi-layered Gaussian pyramid (labeled as layer 0, layer 1... layer n according to scale from largest to smallest, with the bottom layer having the largest scale and the top layer the smallest) and the corresponding multi-layered Laplacian pyramid (each layer corresponds one-to-one with a Gaussian pyramid layer, labeled as layer L0, layer L1... layer Ln) as the processing objects, first, layer matching is performed according to the principle of scale identification consistency, that is, ensuring that the k-th Laplacian pyramid layer (k is an integer from 0 to n) is always consistent with the k-th Gaussian pyramid layer. Then, all Gaussian pyramid layers are structurally organized and arranged in descending order of scale (layer 0 to layer n). Next, all Laplacian pyramid layers are synchronously organized and arranged in the scale order of the corresponding Gaussian pyramid layers (layer L0 to layer Ln), resulting in Gaussian pyramid layers containing global context information and Laplacian pyramid layers containing local detailed features.

[0038] It should be noted that the Gaussian pyramid layer containing global contextual information in this application refers to an image set that can cover the overall morphology and spatial distribution of lung parenchyma at different scales; the Laplacian pyramid layer containing local detail features in this application refers to an image set that can fully present the edges, texture details, and minute features of lung parenchyma at various scales. The Gaussian pyramid layer can provide global contextual information of lung parenchyma at different scales. By progressively reducing the size of the images at multiple scales, the overall morphology, spatial distribution, and large-scale structural features of lung parenchyma at each level are preserved, providing global perspective support for subsequent feature extraction. The Laplacian pyramid layer can capture local detail features at each corresponding scale, focusing on the edges of lung parenchyma, subtle differences in texture, and minute structures of lesions, making up for the high-frequency detail information lost by the Gaussian pyramid layer. The two form a complementary multi-resolution feature system, ensuring that subsequent directional gradient histogram and local binary pattern feature extraction can both cover global structural correlations and accurately capture local lesion details, providing comprehensive and three-dimensional feature input for image analysis networks and improving the accuracy and robustness of lung lesion region identification.

[0039] In step 104, directional gradient histogram features and local binary pattern features are extracted from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively, and then a multi-resolution joint feature vector is constructed based on the directional gradient histogram features and the local binary pattern features.

[0040] In some embodiments, the extraction of directional gradient histogram features and local binary pattern features from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively, is achieved by the following steps: The images at each scale in the Gaussian pyramid layer are subjected to grayscale normalization to obtain a normalized Gaussian pyramid layer. The gradient magnitude and direction of each pixel are calculated based on the normalized Gaussian pyramid layer to obtain the Gaussian layer gradient feature map. Histogram statistics are performed on the gradient feature map of the Gaussian layer according to the preset cell and block division to obtain the directional gradient histogram features; Local neighborhood normalization is performed on the images at each scale in the Laplacian pyramid layer to obtain the normalized Laplacian pyramid layer. The local binary pattern value of each pixel's neighborhood window is calculated based on the normalized Laplacian pyramid layer to obtain the local binary pattern feature.

[0041] In specific implementation, firstly, a linear normalization algorithm (i.e., min-max normalization) is applied to the images at each scale in the Gaussian pyramid layer to map the gray values ​​of each scale image to a unified standard range (i.e., 0-255), eliminating the gray scale differences between images of different scales, thus obtaining a normalized Gaussian pyramid layer. The normalized Gaussian pyramid layer refers to a set of Gaussian pyramid layers with standardized grayscale ranges for each scale image. Secondly, the Sobel operator in image processing is used to calculate the gradient response values ​​in the x and y directions for each scale image in the normalized Gaussian pyramid layer, and the gradient magnitude is then used to determine the gradient response values. The formula calculates the gradient magnitude of each pixel, and the gradient direction of each pixel is calculated using the gradient direction formula. The gradient magnitudes and directions of each pixel are then arranged in their original positions to obtain a Gaussian gradient feature map. This Gaussian gradient feature map refers to a feature map reflecting changes in image edges and structures. Further, the Gaussian gradient feature map is divided into non-overlapping cells according to a preset fixed size (e.g., 8×8 pixels), with each cell serving as the basic unit for gradient statistics. Then, the gradient direction of each pixel is divided into a preset number of direction intervals (e.g., 9 direction intervals, each corresponding to 20°) according to a uniform discretization rule. The angle range is defined as 0-20°, 20-40°...160-180°. Simultaneously, the gradient magnitude of each pixel is retained. For each cell, all pixels within it are traversed, and the direction interval to which it belongs is determined based on the gradient direction of each pixel. The gradient magnitude of that pixel is accumulated into the statistical value of the corresponding direction interval. The statistical values ​​of all direction intervals within each cell together constitute the gradient histogram of that cell (i.e., the horizontal axis of the gradient histogram represents the discretized direction intervals, and the vertical axis represents the sum of the gradient magnitudes of the corresponding direction intervals). Then, according to preset rules (e.g., 2×2 cells form a block), multiple adjacent cells are... The blocks are integrated into a single block, and the blocks are divided by sliding according to a preset overlap ratio (e.g., 50%) (i.e., the next block slides half a block size relative to the previous block in both the horizontal and vertical directions). The gradient histograms of all cells in each block are concatenated to form a block-level joint histogram. Then, L2 normalization is used to numerically calibrate the block-level joint histogram. The entire Gaussian layer gradient feature map is traversed, and the normalized block-level joint histograms of all blocks are concatenated in the spatial order of the blocks to obtain the directional gradient histogram feature. The directional gradient histogram feature refers to a high-dimensional feature vector that represents the global structure and edge contour of the image.Then, a local neighborhood normalization method is applied to the images at each scale in the Laplacian pyramid layer. A local neighborhood of a fixed size (e.g., 3×3) is defined centered on each pixel. The mean and variance of the grayscale values ​​of the pixels within the local neighborhood are calculated. Each pixel is then normalized using the formula (normalized grayscale value = (original grayscale value - local neighborhood mean) / local neighborhood standard deviation) to obtain the normalized Laplacian pyramid layer. The normalized Laplacian pyramid layer refers to the set of Laplacian pyramid layers after local neighborhood grayscale calibration. Finally, based on the normalized Laplacian pyramid layer... A local binary pattern (LDB) calculation method is employed. For each pixel in the image at each scale, a circular neighborhood window with a fixed radius (specifically, one pixel) is defined, and a preset number of neighborhood sampling points are selected. The gray value of each sampling point is compared with the gray value of the center pixel. If the gray value of the sampling point is greater than or equal to the gray value of the center pixel, it is marked as 1; otherwise, it is marked as 0. These binary marks are concatenated clockwise to form a binary number. This binary number is then converted to decimal as the LDB value for the corresponding center pixel. After traversing all pixels, the LDB feature is obtained.

[0042] It should be noted that, in this application, the local binary pattern feature refers to the feature vector characterizing the local texture details and gray-level distribution pattern of the image. In this embodiment, the determination of the directional gradient histogram feature is based on the gradient magnitude and direction information of the Gaussian pyramid layer. Through cell and block-level histogram statistics, the global structural contour, edge direction, and morphological distribution features of lesions (such as nodules and inflammatory areas) of the lung parenchyma at various scales of the lung image are accurately captured. It has strong robustness to geometric deformation and illumination changes of the image and can effectively characterize the macroscopic morphological information of the lesion. The determination of the local binary pattern feature is based on the local neighborhood gray-level contrast relationship of the Laplacian pyramid layer. Through binary encoding, the local texture details, gray-level distribution pattern of the lung parenchyma and the micro-texture differences of the lesion (such as the texture roughness of the fibrotic area and the texture abrupt change features of the nodule edge) are accurately depicted. It can efficiently extract local high-frequency detail information not covered by the Gaussian pyramid layer. The complementary feature system formed by the two integrates the global structure and local details under multiple resolutions, providing highly recognizable core feature inputs for subsequent lesion localization and significantly improving the accuracy of lung lesion identification.

[0043] In some embodiments, constructing a multi-resolution joint feature vector based on the directional gradient histogram features and the local binary pattern features is achieved through the following steps: The directional gradient histogram features and the local binary pattern features are standardized to obtain standardized directional gradient histogram features and standardized local binary pattern features, respectively. Based on the pyramid hierarchical index, the standardized directional gradient histogram features and the standardized local binary pattern features are dimensionally aligned to obtain hierarchically aligned feature pairs. The alignment feature pairs at each level are concatenated along the channel dimension to generate a single-resolution joint feature set. The single-resolution joint feature sets are combined in ascending order of pyramid resolution to form a multi-resolution feature tensor. The multi-resolution feature tensor is then normalized to obtain the multi-resolution joint feature vector.

[0044] In specific implementation, firstly, the Z-score normalization algorithm is used to normalize the histogram of oriented gradients (HOR) features and the local binary pattern features, respectively, to obtain normalized HOR features and normalized local binary pattern features. Secondly, based on the pyramid hierarchy index (i.e., the scale identifiers corresponding one-to-one between Gaussian pyramid layers and Laplacian pyramid layers), the normalized HOR features and normalized local binary pattern features at the same scale are paired. If the dimensions of the two types of features differ, zero-padding (i.e., zero-padding the tail of the feature vector with the shorter dimension) is used to make the dimensions of the paired feature vectors completely consistent, obtaining hierarchically aligned feature pairs. The hierarchically aligned feature pairs refer to the combination of normalized HOR features and normalized local binary pattern features representing global structure and local details at the same scale. Then, a general approach is adopted... The method of concatenating dimensions involves sequentially concatenating the standardized directional gradient histogram feature vectors and standardized local binary pattern feature vectors from each level-aligned feature pair in element-wise order to form a high-dimensional feature vector (i.e., single-resolution joint feature) containing global-local complementary information at the corresponding scale. The high-dimensional feature vectors from all scales together constitute a single-resolution joint feature set, which refers to the set of high-dimensional feature vectors at each scale level. Finally, in order of pyramid resolution from low to high (i.e., from the single-resolution joint features corresponding to the top-level scale image to the single-resolution joint features corresponding to the bottom-level scale image), all single-resolution joint features in the single-resolution joint feature set are sequentially combined to form a multi-resolution feature tensor. The multi-resolution feature tensor is then normalized using the L2 normalization algorithm to obtain the multi-resolution joint feature vector.

[0045] It should be noted that the multi-resolution joint feature vector in this application refers to a high-dimensional feature that integrates global structural and local detail information across all scales. The reason for determining the multi-resolution joint feature vector is that a single-resolution feature cannot capture the features of lesions at different scales (such as large-scale inflammatory areas and small nodules), which can easily lead to one-sided feature information or omission of key details. By integrating global structural features (orientation gradient histogram features) and local detail features (local binary pattern features) at each scale of the multi-resolution pyramid, the low-resolution features capture global contextual information such as the overall morphology and spatial distribution of the lesion, while the high-resolution features accurately capture local details such as the edge texture and microstructure of the lesion. Ultimately, this provides core feature inputs that are rich in dimensions, complete in information, and robust for subsequent lung lesion localization, effectively improving the accuracy of identifying lung lesion regions of different sizes, shapes, and textures.

[0046] In step 105, the multi-resolution joint feature vector is input into a pre-trained image analysis network to locate the lesion region in the lung image.

[0047] In some embodiments, the multi-resolution joint feature vector is input into a pre-trained image analysis network to locate the lesion region in the lung image using the following steps: The multi-resolution joint feature vector is input into a pre-trained image analysis network to generate a probability map of the lesion region. An adaptive threshold segmentation algorithm is used to process the probability map of the lesion region to obtain a binary mask of the lesion candidate region; Based on the pixel coordinate system of the lung image, the binary mask of the candidate lesion region is analyzed for connected components, and the lesion region in the lung image is output.

[0048] It should be noted that the pre-trained image analysis network in this application adopts a convolutional neural network based on an encoder-decoder architecture. The encoder part includes four downsampling stages, each consisting of two 3×3 convolutional layers (each followed by batch normalization and activation functions) and a 2×2 max-pooling layer, progressively halving the input feature map size while doubling the number of channels (from the initial 64 channels to 512 channels) to extract deep semantic features layer by layer. The decoder part restores the feature map size through four upsampling stages. Each stage includes a 2×2 transposed convolutional layer for upsampling, skip connections to the corresponding stage feature map in the encoder, two 3×3 convolutional layers (containing batch normalization and activation functions), and finally a 1×1 convolutional layer and a sigmoid function. The activation function outputs a probability map of lesion regions with the same size as the input. During the training phase, this image analysis network uses a dataset containing 2000 labeled lung images (1500 for training and 500 for validation). It employs a stochastic gradient descent optimizer (initial learning rate 0.01, momentum 0.9), a composite loss function combining the Dice loss function and binary cross-entropy, and iteratively adjusts the network's convolutional kernel weights and bias parameters until the composite loss function converges (i.e., the Dice coefficient reaches 0.82). The Dice loss function is a metric used to measure the similarity between two sample sets; in segmentation, it measures the overlap between the predicted region and the true labeled region. This completes the pre-training of the image analysis network, which will not be elaborated further.

[0049] In specific implementation, firstly, the joint feature vectors of multiple resolutions are input into a pre-trained image analysis network, which then generates a lesion region probability map that is completely consistent with the spatial dimensions of the original lung image. This lesion region probability map refers to a grayscale image whose spatial coordinates correspond one-to-one with the original lung image, and where each pixel value reflects the probability that the corresponding location is a lesion. Secondly, the Otsu adaptive threshold segmentation algorithm from image processing is used to process the lesion region probability map. Specifically, the Otsu adaptive threshold segmentation algorithm automatically calculates the segmentation threshold that maximizes the inter-class variance between lesion-class pixels and background-class pixels by statistically analyzing the grayscale distribution histogram of the lesion region probability map. Positions with pixel values ​​greater than the segmentation threshold are marked as 1 (identified as suspected lesion areas), and positions with pixel values ​​less than or equal to the segmentation threshold are marked as 0 (identified as background areas), thus obtaining a binary mask for candidate lesion areas. The binary mask for candidate lesion areas refers to a binary image that distinguishes between background and suspected lesions using only 0 and 1 binary values. Finally, based on the pixel coordinate system of the lung image (i.e., the coordinate system in which the binary mask is completely aligned with the pixel positions of the original lung image), the eight-neighbor connected component analysis algorithm is used to traverse the binary mask for candidate lesion areas, grouping adjacent pixels (up, down, left, right, and four diagonal directions) marked as 1 into the same connected region (i.e., candidate lesion block), thereby outputting the lesion area in the lung image.

[0050] It should be noted that in this application, the lesion area refers to the lesion image area with its spatial location marked.

[0051] In another aspect, in some embodiments, this application provides a lung image recognition system for clinical diagnosis in respiratory medicine, referring to... Figure 4 The figure is a schematic diagram of the structure of a lung image recognition system for clinical diagnosis in respiratory medicine, according to some embodiments of this application. The lung image recognition system for clinical diagnosis in respiratory medicine includes: an acquisition module 201, a processing module 202, and an execution module 203, which are described below: The acquisition module 201 in this application is mainly used to acquire the lung image to be identified and to segment the lung image into lung field regions to obtain a binary mask containing complete lung parenchyma. Processing module 202, in this application, is mainly used to segment the target lung field image from the lung image based on the binary mask, and to perform texture saliency detection based on local entropy value on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma. The processing module 202 is further configured to perform pixel-level fusion of the texture saliency map and the target lung field image to obtain a pixel-enhanced lung feature image, and to perform multi-resolution pyramid decomposition on the lung feature image to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. In addition, the processing module 202 is also used to extract directional gradient histogram features and local binary pattern features from the Gaussian pyramid layer and the Laplacian pyramid layer respectively, and then construct a multi-resolution joint feature vector based on the directional gradient histogram features and the local binary pattern features; The execution module 203 in this application is mainly used to input the multi-resolution joint feature vector into a pre-trained image analysis network to locate the lesion area in the lung image.

[0052] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described lung image recognition method for clinical diagnosis in respiratory medicine.

[0053] In some embodiments, reference Figure 5 This figure is a schematic diagram of the structure of a computer device implementing a lung image recognition method for clinical diagnosis in respiratory medicine, according to some embodiments of this application. The lung image recognition method for clinical diagnosis in respiratory medicine described in the above embodiments can be... Figure 5The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0054] The processor 301 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the lung image recognition method for clinical diagnosis in respiratory medicine as described in this application.

[0055] The communication bus 302 can be used to transmit information between the aforementioned components.

[0056] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0057] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the lung image recognition method for clinical diagnosis in respiratory medicine can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0058] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0059] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0060] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0061] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described lung image recognition method for clinical diagnosis in respiratory medicine.

[0062] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0063] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for lung image recognition in clinical diagnosis of respiratory medicine, characterized in that, Includes the following steps: The lung image to be identified is acquired, and the lung image is segmented into lung field regions to obtain a binary mask containing complete lung parenchyma. Based on the binary mask, the target lung field image is segmented from the lung image, and texture saliency detection based on local entropy is performed on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma. The texture saliency map is fused with the target lung field image at the pixel level to obtain a pixel-enhanced lung feature image. The lung feature image is then decomposed into a multi-resolution pyramid to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. Oriented gradient histogram features and local binary pattern features are extracted from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively. Then, a multi-resolution joint feature vector is constructed based on the oriented gradient histogram features and the local binary pattern features. The multi-resolution joint feature vector is input into a pre-trained image analysis network to locate the lesion region in the lung image.

2. The method as described in claim 1, characterized in that, The lung images are segmented to obtain a binary mask body containing complete lung parenchyma, comprising: The lung images are subjected to grayscale normalization processing to obtain grayscale normalized lung images; The gray-level normalized lung image was segmented using adaptive threshold segmentation to obtain a preliminary lung field mask; The preliminary lung field mask is morphologically optimized to obtain a morphologically optimized lung field mask; The morphologically optimized lung field mask is binarized to generate a binary mask containing the complete lung parenchyma.

3. The method as described in claim 1, characterized in that, Segmenting the target lung field image from the lung image based on the binary mask specifically includes: The binary mask is subjected to morphological dilation to obtain the dilated binary mask; The expanded binary mask is multiplied pixel by pixel with the lung image to segment the target lung field image from the lung image.

4. The method as described in claim 1, characterized in that, Performing texture saliency detection based on local entropy on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma specifically includes: Preprocessing of the target lung field image yields a preprocessed target lung field image. The local neighborhood block corresponding to each pixel is obtained from the preprocessed target lung field image by using a preset sliding window; Calculate the gray-level histogram for each local neighborhood block and solve for the probability distribution to obtain the local entropy value of each local neighborhood block; All local entropy values ​​are integrated to form a local entropy value map, and the local entropy value map is normalized to obtain a normalized local entropy value map. Texture saliency enhancement is performed on the normalized local entropy map to generate a texture saliency map with complex texture patterns of lung parenchyma.

5. The method as described in claim 1, characterized in that, The pixel-level fusion of the texture saliency map with the target lung field image to obtain a pixel-enhanced lung feature image specifically includes: Spatial rigid registration and grayscale normalization are performed on the texture saliency map and the target lung field image to obtain the aligned normalized texture saliency map and the normalized target lung field image. Pixel-level fusion weights are calculated based on the entropy distribution of the aligned normalized texture saliency map to obtain a texture-guided fusion weight map. The aligned, normalized target lung field image and the normalized texture saliency map are then subjected to pixel-wise weighted summation according to the texture-guided fusion weight map to obtain a pixel-enhanced lung feature image.

6. The method as described in claim 1, characterized in that, Performing multi-resolution pyramid decomposition on the lung feature images to obtain a Gaussian pyramid layer containing global contextual information and a Laplacian pyramid layer containing local detailed features specifically includes: Gaussian filtering is applied to the lung feature images to obtain filtered lung images; The filtered lung images are iteratively downsampled and Gaussian smoothed to generate multiple Gaussian pyramid layers; Upsampling and Gaussian smoothing are performed on each Gaussian pyramid layer to obtain the reconstructed Gaussian image at the corresponding scale; Each Gaussian pyramid layer is subjected to pixel-by-pixel difference operation with the corresponding reconstructed Gaussian image to generate multiple Laplacian pyramid layers. By integrating all Gaussian pyramid layers and Laplacian pyramid layers, we obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detailed features.

7. The method as described in claim 1, characterized in that, Images of the lungs to be identified are obtained using a CT scanner.

8. A lung image recognition system for clinical diagnosis in respiratory medicine, characterized in that, The system includes: The acquisition module is used to acquire the lung image to be identified and to segment the lung image into lung field regions to obtain a binary mask containing complete lung parenchyma. The processing module is used to segment the target lung field image from the lung image based on the binary mask, and perform texture saliency detection based on local entropy value on the target lung field image to generate a texture saliency map with complex texture patterns of lung parenchyma. The processing module is further configured to perform pixel-level fusion of the texture saliency map and the target lung field image to obtain a pixel-enhanced lung feature image, and to perform multi-resolution pyramid decomposition on the lung feature image to obtain a Gaussian pyramid layer containing global context information and a Laplacian pyramid layer containing local detail features. The processing module is further configured to extract directional gradient histogram features and local binary pattern features from the Gaussian pyramid layer and the Laplacian pyramid layer, respectively, and then construct a multi-resolution joint feature vector based on the directional gradient histogram features and the local binary pattern features. The execution module is used to input the multi-resolution joint feature vector into a pre-trained image analysis network to locate the lesion region in the lung image.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the lung image recognition method for clinical diagnosis in respiratory medicine as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the lung image recognition method for clinical diagnosis in respiratory medicine as described in any one of claims 1 to 7.

Images