A lung abnormal nodule recognition processing method and system

Abstract

This invention discloses a method and system for identifying and processing abnormal lung nodules. The method includes the following steps: acquiring lung CT image data; preprocessing the lung CT image data to obtain a lung image to be detected; performing pixel analysis on the lung image to be detected using a discrete fractional differential operator to obtain a fractional-order enhanced image; acquiring pixels of the fractional-order enhanced image at multiple scales and in multiple directions to calculate local energy and local entropy, and constructing a high-dimensional fusion feature matrix; reducing the dimensionality of the high-dimensional fusion feature matrix to extract the lung nodule region; effectively extracting nodule regions from lung CT images; and ultimately identifying abnormal lung nodule regions through pixel analysis, calculation of local energy and entropy, and construction and dimensionality reduction of the high-dimensional feature matrix, thereby improving the detection accuracy and efficiency of lung nodules.

Description

Technical Field

[0001] This invention relates to the field of lung detection, and more particularly to a method and system for identifying and processing abnormal lung nodules. Background Technology

[0002] Lung cancer is one of the leading causes of death from malignant tumors worldwide. Pulmonary nodules, as an early manifestation of lung cancer, are crucial for detection and diagnosis. Early detection and accurate assessment of the benign or malignant nature of pulmonary nodules can effectively improve patient survival rates. With the development of medical imaging technology, CT (computed tomography) has become an important tool for detecting pulmonary nodules. However, traditional manual diagnostic methods not only rely on the experience and skill level of physicians but are also susceptible to interference from subjective human factors when processing complex imaging data, resulting in low accuracy and efficiency in nodule detection.

[0003] Currently, various methods have been proposed for the detection and identification of nodules in lung CT images. For example, WO2022063199A1 discloses an automatic lung nodule detection method. This method first performs filtering and enhancement processing on the CT image to be detected to obtain an enhanced CT image sequence of the lung. Then, it uses a thresholding method for segmentation to obtain an image containing only the lung parenchyma. A U-Net network model with multi-scale feature fusion is used to obtain the region of interest. Finally, a 3D CNN model is used to automatically detect and identify the region of interest. CN113643279A proposes a lung nodule detection device based on CT images. This device smoothly segments the original CT image to be detected to obtain lung images. Then, based on the Laplacian Gaussian operator, it performs multi-scale spatial lung nodule spot identification on the lung images to obtain candidate lung nodule regions. A lung nodule identification model is then used to perform feature recognition on the candidate lung nodule regions.

[0004] CN107301640A discloses a method for detecting small lung nodules based on convolutional neural networks. This method uses an edge detection algorithm and two-dimensional Gaussian random sampling to obtain suspected regions to be detected in lung CT images, and employs a hybrid convolutional neural network based on unsupervised training of a single layer and supervised global fine-tuning to extract features representing small lung nodules. CN113129314B provides an intelligent image processing method for lung shadows, including extracting the original lung image, preprocessing it to segment it into lung nodules, and performing feature extraction and feature selection on the lung nodules.

[0005] However, the aforementioned existing technologies still have some shortcomings in the identification and processing of abnormal lung nodules: First, existing image enhancement methods mostly employ traditional filtering or convolution operations, which have limited effectiveness in enhancing the edge and texture details of small nodules in lung CT images (currently, common image processing methods include edge detection, image enhancement, feature extraction, and dimensionality reduction techniques. Edge detection methods often identify nodule edges by detecting grayscale changes in the image; image enhancement techniques aim to improve the visibility of nodule regions and reduce the impact of noise on nodule identification. However, these traditional methods are often limited by factors such as image quality, noise, nodule morphology, and size when processing small details in images, resulting in unsatisfactory detection results); Second, existing feature extraction methods often only focus on features at a single scale or in a single direction, failing to comprehensively capture the texture features of lung nodules at different scales and directions, resulting in insufficient feature representation capabilities.

[0006] Therefore, there is an urgent need for a method that can effectively enhance the detailed features of nodules in lung CT images and comprehensively analyze the texture features of lung nodules from multiple scales and directions, so as to improve the accuracy and reliability of identifying abnormal lung nodules. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for identifying and processing abnormal lung nodules, which solves the above-mentioned technical problems pointed out in the prior art.

[0008] This invention provides a method for identifying and processing abnormal lung nodules, comprising the following steps:

[0009] Acquire lung CT image data, preprocess the lung CT image data to obtain lung images to be detected;

[0010] The discrete fractional differential operator is used to perform pixel analysis on the lung image to be detected to obtain a fractional enhanced image; the pixels of the fractional enhanced image are collected at multiple scales and in multiple directions to calculate local energy and local entropy, and a high-dimensional fusion feature matrix is ​​constructed; the high-dimensional fusion feature matrix is ​​then reduced in dimensionality to extract the lung nodule region.

[0011] Accordingly, the present invention also proposes a lung abnormal nodule identification and processing system, comprising: a data acquisition module; and an identification module;

[0012] The acquisition module is used to acquire lung CT image data and preprocess the lung CT image data to obtain lung images to be detected.

[0013] The recognition module is used to perform pixel analysis on the lung image to be detected using a discrete fractional differential operator to obtain a fractional enhanced image; to collect pixels of the fractional enhanced image at multiple scales and in multiple directions to calculate local energy and local entropy, and to construct a high-dimensional fusion feature matrix; and to reduce the dimensionality of the high-dimensional fusion feature matrix to extract the lung nodule region.

[0014] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages:

[0015] Analysis of the lung abnormality nodule identification and processing method and system provided by this invention reveals that, in specific applications, convolution operations using discrete fractional differential operators are employed. The fractional-order parameter α and the spacing between adjacent pixels are used to construct the operator, thereby enhancing the local texture details of the image and obtaining a fractional-order enhanced image. Then, based on Gabor wavelet transform, a custom wavelength is set, and multi-scale and multi-directional transformations are combined to extract the local texture of the image. Lung nodules typically possess certain scale and directional characteristics, and Gabor wavelets can effectively capture these details. Through multi-scale and multi-directional transformations, texture information in the image is revealed, effectively improving the accuracy of nodule identification. By analyzing each scale and direction in the fractional-order enhanced image and calculating local energy and entropy, lung nodule regions can be better identified, helping to reveal the intensity and complexity of the texture in the fractional-order enhanced image.

[0016] The aforementioned local energy reflects the overall intensity of image texture, while entropy measures the complexity of texture. Higher local energy and entropy usually indicate that a certain region in the image has a more complex texture structure, which helps to improve the accurate localization of nodule regions. At the same time, the technical solution of this application combines multi-scale and multi-directional local energy and entropy to fuse into local statistical features, constructing a high-dimensional fusion feature matrix. This matrix can comprehensively and meticulously describe the texture features of the image. By assigning different feature weights, the expressive power of nodule region features can be improved, thereby improving the accuracy of feature selection and effectively enhancing the detection and differentiation capabilities of lung nodule regions, especially showing good advantages when processing complex texture regions.

[0017] Furthermore, L1 sparse coding is used to extract key features from the image. Through sparse representation, the feature vector of each pixel is concatenated with the sparse coefficients of each pixel into a low-dimensional matrix A. This compression method not only reduces the data dimensionality but also makes it easier to process. By mapping the sparse coefficients to the activation map, potential abnormal regions can be clearly revealed. The activation map highlights regions with high local sparse coefficients (usually associated with abnormal features such as lung nodules), further enhancing the ability to identify abnormal regions. By dividing the image into grids based on the average distance between adjacent pixels (i.e., the average distance between adjacent pixels), accurate analysis of each pixel within its local region is ensured, adapting to different features in the image. The method identifies nodules or abnormal regions (i.e., second candidate abnormal regions) based on their shape and size. Simultaneously, it employs a global Otsu method combined with sparse coding to generate an activation map, automatically selecting a segmentation threshold to separate the background from the abnormal region (i.e., the first candidate abnormal region). This is particularly effective when dealing with irregular lung nodules of varying sizes. Through repeated segmentation of the activation map and threshold adjustment, the entire method continuously optimizes the segmentation effect during the iteration process. The ratio difference between the first and second candidate abnormal regions effectively determines whether to fuse global and local segmentation results, thereby obtaining more accurate detection results. In this way, not only is the detection accuracy of nodules improved, but the risk of false positives and false negatives is also reduced. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for identifying and processing abnormal lung nodules according to Embodiment 1;

[0019] Figure 2 This is a flowchart illustrating the lung nodule region for enhanced detection in an image of the lung to be detected, as described in Embodiment 1 of a lung abnormal nodule identification and processing method.

[0020] Figure 3 A schematic diagram of the local energy distribution in a lung abnormal nodule identification and processing method according to Example 1;

[0021] Figure 4 A case sample image of ground-glass nodules in the middle part of a method for identifying and processing abnormal lung nodules in Example 1;

[0022] Figure 5 A case sample image of a solid nodule in an embodiment of a method for identifying and processing abnormal pulmonary nodules;

[0023] Figure 6 Flowchart of dimensionality reduction and segmentation of a high-dimensional fusion feature matrix in an embodiment of a lung abnormal nodule identification and processing method;

[0024] Figure 7 The overall flow of the main scheme of a method for identifying and processing abnormal lung nodules in Example 2;

[0025] Figure 8 A flowchart of a lung abnormal nodule identification and processing system according to Example 3;

[0026] Labels: Acquisition module 10; Recognition module 20. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0028] Example 1

[0029] like Figure 1 As shown, this application provides a method for identifying and processing abnormal lung nodules, including the following steps:

[0030] S10: Acquire lung CT image data, preprocess the lung CT image data to obtain the lung image to be detected;

[0031] S20: Pixel analysis is performed on the lung image to be detected using the discrete fractional differential operator to obtain a fractional-order enhanced image; local energy and local entropy are calculated for the pixels of the fractional-order enhanced image at multiple scales and in multiple directions to construct a high-dimensional fusion feature matrix; the high-dimensional fusion feature matrix is ​​reduced in dimensionality to extract the lung nodule region;

[0032] It should be noted that the discrete fractional differential operator is an image processing method based on fractional calculus. Unlike traditional integer-order differential operators, fractional-order differential operators can enhance high-frequency information in images, better revealing details such as edges and textures, especially performing better in noisy images. Therefore, in step S20 (and in the execution of specific steps S21-S25), by using the discrete fractional differential operator to perform pixel-by-pixel analysis on the lung image to be detected, a fractional-order enhanced image can be obtained. This makes subtle changes and details in the image clearer, especially the boundaries of nodules, which are more prominent (the regularity and clarity of nodule boundaries directly affect the recognition accuracy and quality of the image).

[0033] Furthermore, by processing fractional-order enhanced images at different scales and in different directions, local energy and local entropy in fractional-order enhanced images are captured. Local energy usually represents the brightness intensity information of a certain region in a fractional-order enhanced image, while local entropy reflects the complexity or uncertainty of a region in the image. In lung nodule images, the local energy and local entropy of the nodule region are often different from those of the background region. These features help to distinguish between nodule and non-nodule regions.

[0034] Subsequently, by constructing a high-dimensional fusion feature matrix, features extracted from different scales, orientations, local energies, and local entropies are integrated to form a high-dimensional feature matrix. These high-dimensional features can describe the nodule region and other structural features in the image in detail, which helps to improve the accuracy of classification and recognition. Finally, by processing these dimensionality-reduced feature matrices, the lung nodule region is extracted, providing accurate regional information for nodule detection and segmentation. The above dimensionality reduction operation reduces redundant information by reducing the dimension of the feature matrix, and retains the most valuable features for nodule identification.

[0035] Specifically, such as Figure 2 As shown, in step S20, the discrete fractional differential operator is used to perform pixel analysis on the lung image to be detected to obtain a fractional-order enhanced image; the pixels of the fractional-order enhanced image are collected at multiple scales and in multiple directions (i.e., in a specific multi-directional extraction scenario) to calculate local energy and local entropy, and a high-dimensional fusion feature matrix is ​​constructed; the high-dimensional fusion feature matrix is ​​then dimensionality reduced to extract the lung nodule region. The specific operation steps are as follows:

[0036] S21: Using a discrete fractional differential operator, perform multi-level fractional differential processing on the lung image to be detected: calculate the adjacent distance of the adjacent domain for each pixel in the lung image to be detected, and calculate the average distance between adjacent pixels using the adjacent distances of the adjacent domains of all pixels.

[0037] A fractional-order parameter α is preset, and a discrete fractional differential operator is constructed using the fractional-order parameter α and the average distance between adjacent pixels. (i.e., convolution kernel); perform discrete fractional differential operator on each pixel in the image to be detected. The convolution operation extracts the local texture of pixel grayscale values ​​and generates a fractional-order enhanced image.

[0038] ;

[0039] in, Represented as a fractional-order enhanced image, used to highlight local texture edge information;

[0040] Represented as the image to be detected;

[0041] It is represented as a discrete fractional differential operator (convolution kernel), where α represents the fractional-order parameter;

[0042] It should be noted that the fractional differential operator can control the intensity of image detail enhancement by adjusting the value of the fractional parameter α; by using the spacing between adjacent pixels, it can adaptively perform enhancement processing on different image regions, especially for detailed regions such as edges and textures; enhancement can highlight subtle texture information in lung tissue. Since lung nodules are usually small and irregular in shape, local texture enhancement of the image helps to improve the contrast between the nodules and the surrounding healthy tissue, and increase the identifiability of the nodules.

[0043] By enhancing image details through discrete fractional differential operators, nodule edges and texture information become more prominent, which helps to improve the detection accuracy of nodules. Fractional enhancement can improve the contrast of fractional enhanced images through local adjustment and multi-scale characteristics, making the details in the fractional enhanced images more clearly visible. The fractional method can enhance the edge features, texture and details of the image, especially in areas with low contrast or blurred details. Different enhancement effects can be controlled by fine-tuning the α value.

[0044] S22: Use Gabor wavelets to enhance the fractional-order image. Set a custom wavelength for local texture; perform multi-scale and multi-directional acquisition of pixels in the fractional-order enhanced image according to the set custom wavelength;

[0045] Wavelet transform is performed on the pixels in the fractional-order enhanced image at multiple scales and in multiple directions to obtain multiple scales of the pixels. With multiple directions wavelet coefficients (That is, the coefficients reflect the local texture information of the image at different resolutions and directions);

[0046] It should be noted that multi-scale, multi-directional wavelet transform is used to obtain local texture information of images at different resolutions and directions. Gabor wavelets can extract texture information from images at different scales and directions, handling texture changes at different scales and directions. For lung nodule identification, nodules usually have certain directional and scale characteristics, and Gabor wavelet transform can effectively capture these characteristics. Through multi-scale and multi-directional transform, the texture features of each pixel at different scales and directions can be obtained, generating wavelet coefficients. Subsequently, local energy is calculated based on these coefficients, which helps to reveal the texture distribution and characteristics of the nodule region. Gabor wavelet transform, by extracting local texture information at different scales and directions, helps to reveal the detailed features of lung nodules, improving the robustness and accuracy of identification. The wavelet coefficient vector of any pixel o in the fractional-order enhanced image is... ;

[0047] S23: Calculate the local energy based on the wavelet coefficients of each scale and each direction of the pixels in the fractional-order enhanced image;

[0048] ;

[0049] in, Represented as the wavelet coefficient of the 0th pixel in the fractional-order enhanced image;

[0050] It is represented as the multi-directional aggregated energy (i.e., energy superposition) of the upper wavelet coefficients at each scale and in each direction at the o-th pixel.

[0051] It is represented as the wavelet coefficient vector of the o-th pixel in the fractional-order enhanced image (that is, the wavelet coefficients of all scales and directions at pixel o are combined into a vector, which reflects the local information of the o-th pixel in the fractional-order enhanced image).

[0052] It is represented as the vector of the i-th pixel in the fractional-order enhanced image (that is, the wavelet coefficients of all scales and directions at pixel i are combined into a vector, which reflects the local information of the i-th pixel in the fractional-order enhanced image).

[0053] Represented as in the adaptive neighborhood The sum of distances between the vectors of the o-th pixel and the i-th neighboring pixel (or the pixels in the remaining neighboring pixels) is used to reflect the discreteness (or difference) of the energy distribution of the wavelet coefficients between the pixels; i represents the i-th pixel; o represents the o-th pixel.

[0054] The above calculation process involves normalizing the local energy of each pixel at each scale and in each direction, calculating the sum of the local energies at all scales and in all directions, and then calculating the probability distribution of each pixel at each scale and in each direction using the sum of the local energies and the local energy (i.e., the proportion of energy at each scale and in each direction; this is obtained by dividing the local energy at one scale and in one direction by the sum of the local energies at all scales and in each direction). Finally, the local entropy is calculated using the probability distribution of the local energy.

[0055] It should be noted that in step S23 above, the local energy and entropy of the pixels in the fractional-order enhanced image are calculated at various scales and directions to describe the distribution characteristics of the image texture. The energy distribution of this local energy reflects the texture intensity of local regions in the fractional-order enhanced image. Changes in local energy help identify prominent areas in the image, such as the area where lung nodules are located. Entropy values ​​can measure the complexity and uniformity of texture in the fractional-order enhanced image. Higher entropy values ​​usually mean a more uniform distribution of texture information, reflecting a more complex structure in the image region. By aggregating the energy of wavelet coefficients of each pixel in different directions, this energy superposition reflects the texture intensity and local features of the pixels. It can extract texture information of the image in various directions, thus more accurately reflecting the local structure of the image, especially the intensity and differences of texture details, such as... Figure 3 As shown;

[0056] The above process normalizes the local energy at all scales and directions, calculating the ratio of the local energy at each scale and direction to the sum of the local energies at all scales and directions. This allows the texture features of the image to be standardized at different resolutions and directions. The normalization of local energy helps to reveal the relative importance of different regions in the image, highlighting regions with more complex textures and higher energy, thereby helping to identify abnormal regions (such as nodules) in the image. This calculation method helps to distinguish between complex textures and relatively simple regions.

[0057] For nodular regions, changes in local entropy help distinguish nodules from normal tissue; the uniformity of energy distribution at different scales and directions—the more uniform the energy distribution, the higher the entropy value, and vice versa—provides a quantitative description of the texture features of lung nodules, which helps improve segmentation and recognition accuracy.

[0058] From the detailed features of lung nodules on CT images;

[0059] In CT images, pulmonary nodules typically appear as dense, well-defined masses with irregular borders. After fractional-order differential enhancement, the edges and internal texture of the nodule are enhanced. The texture response of the nodule region produces strong coefficient values ​​in multiple directions and scales. Therefore, after energy superposition, the local energy value of the nodule region will be significantly higher than that of the surrounding homogeneous tissue. That is, the lung parenchyma surrounding the pulmonary nodule may only have a strong response in a few directions (such as blood vessel orientation) or at a specific scale, while the response of the nodule region is more extensive (multi-directional) and more intense (especially at the enhanced edges and texture).

[0060] In nodular regions, due to their complex internal texture (possibly including ground-glass opacities, calcifications, and other mixed structures), energy is relatively uniformly distributed across multiple scales and directions. In contrast, homogeneous tissues such as fat or fluid may concentrate energy in a few directions. Therefore, the local entropy value of nodular regions tends to be higher. Specifically, healthy lung parenchyma (alveolar regions, i.e., homogeneous lung parenchyma) typically has a relatively simple and uniform texture. Its energy is likely mainly distributed at specific scales and directions characterizing the background texture, resulting in a lower entropy value. The energy of linear structures (blood vessels) is highly concentrated in directions perpendicular to their orientation (strongest response at the edges), while the response is weak in parallel directions. This high directional concentration of energy leads to lower local entropy. However, the energy distribution in nodular regions is relatively more dispersed and uniform across directions and scales, therefore the calculated entropy value Ho is higher than that of homogeneous lung parenchyma and linear structures (blood vessels).

[0061] Based on the density analysis of pulmonary nodules, pulmonary nodular lesions can be classified into three categories: pure ground-glass nodules; partially ground-glass nodules; and solid nodules. These three types of nodules have different characteristics in terms of nature and doubling time.

[0062] See Figure 4 See the imaging in Case 1 for details; some ground-glass nodules (arrows) sometimes show small vacuolar signs. These partially ground-glass nodules typically progress very slowly, or remain unchanged for years, or only gradually become denser. Pathologically, this imaging feature often corresponds to adenocarcinoma in situ or atypical adenoid hyperplasia. Some ground-glass nodules may also be accompanied by vacuolar signs and high internal density, with the solid component often being invasive adenocarcinoma.

[0063] See Figure 5 See the imaging of Case 2 for details; a nodular shadow in the lower lobe of the right lung (arrow), about 10 mm in diameter, with clear borders and close to the pleura, and uniform internal density (with a relatively high internal density), is a solid nodule.

[0064] S24: Combine the local energy and entropy of the fractional-order enhanced image to obtain local statistical features; assign weights to the local statistical features to construct a high-dimensional fusion feature matrix. ;

[0065] The high-dimensional fusion feature matrix The middle row represents m pixels, and the column represents n feature vectors (i.e., local statistical features).

[0066] The explanation explains that multi-directional and multi-scale local texture features of pixels are fused to construct a high-dimensional feature matrix to describe the texture features of pixels. By combining texture features from different directions and scales, a high-dimensional fused feature matrix is ​​obtained. Each column represents the features of a local region of a pixel. Local statistical features describe the entropy (i.e., reflecting the complexity of gray-level distribution within the region) and energy (i.e., the overall intensity of the local texture) of the fractional-order enhanced image. It also includes different aspects of nodules in the fractional-order enhanced image, such as edges, structure, and texture (studies have found that the texture and edges of lung nodule areas are more complex and have typical characteristics compared to normal areas).

[0067] The above technical solution can further improve the expressive power of nodule region features, reduce redundant information, and improve the accuracy of feature selection by assigning different weights to each local statistical feature. The high-dimensional fusion feature matrix can comprehensively and meticulously describe the texture features of the image, and improve the detection and discrimination capabilities of nodule regions. The high-dimensional fusion feature matrix includes texture features of multiple directions and multiple scales in the fractional-order enhanced image. Therefore, the high-dimensional fusion feature matrix has n feature vectors (i.e., local statistical features) and m sampling points (i.e., the local regions of each pixel at each scale and in the corresponding direction, and the surrounding regions of the pixel). Each column represents the descriptive vector of the texture features of a local region.

[0068] S25: The high-dimensional fusion feature matrix is ​​reduced in dimensionality using the L1 sparse coding algorithm to generate an activation map; the activation map is segmented to obtain candidate abnormal regions; the candidate abnormal regions are thresholded to obtain abnormal lung nodule regions.

[0069] It should be noted that, through L1 sparse coding, we can transform the high-dimensional feature matrix in an image into a low-dimensional feature matrix; sparse coding can extract local features in an image, rather than relying solely on the representation of global grayscale values; this allows the details of candidate abnormal regions (such as lung nodules) in an image to be better preserved, because these candidate abnormal regions are usually accompanied by changes in local texture and morphology.

[0070] An activation map is generated by mapping the sparse coefficients in the low-dimensional feature matrix to the corresponding positions (i.e., coordinates) in the lung image to be detected. The activation map can reflect the activation intensity of the local sparse coefficients in the image (i.e., the sparse coefficient values ​​represent different activation intensities). Higher activation intensities correspond to abnormal regions of texture or morphology in the image, which are usually potential lung nodules. The activation map helps to highlight candidate abnormal regions in the image. Then, an accurate segmentation threshold is obtained by combining a global threshold and an adaptive threshold, thereby obtaining the nodule region.

[0071] In medical image analysis, a large number of feature vectors are typically extracted, potentially involving tens of thousands of pixels and multiple features for each pixel. This data exhibits very high dimensionality, leading to enormous computational resource consumption and low analysis efficiency. Dimensionality reduction aims to compress the feature dimension while preserving as much key information as possible, thus reducing the computational burden. L1 sparse coding is particularly suitable for extracting local features. By representing the feature vector of each pixel in the image with a small number of dictionary atoms, sparse coding effectively highlights local structural features of the image (such as texture and shape) while ignoring unimportant parts. This representation helps to better identify abnormal regions like lung nodules in the dimensionality-reduced feature space.

[0072] Meanwhile, the sparse coefficients after dimensionality reduction can better reflect the features of local regions, rather than simply global grayscale information; abnormal regions such as lung nodules usually exhibit local texture changes or morphological abnormalities, and sparse coding can capture these features more accurately, thereby improving the detection accuracy; therefore, it is necessary to perform dimensionality reduction analysis on the high-dimensional fused feature matrix using the L1 sparse coding algorithm to obtain the final lung nodule region.

[0073] Specifically, such as Figure 4 As shown, in step S25, the high-dimensional fusion feature matrix is ​​reduced in dimensionality using the L1 sparse coding algorithm to generate an activation map; the activation map is segmented to obtain candidate abnormal regions; and the candidate abnormal regions are thresholded to obtain abnormal lung nodule regions. The specific operation steps are as follows:

[0074] S251: Use the L1 sparse coding algorithm to select n feature vectors from m sampled pixels of the high-dimensional fusion feature matrix;

[0075] K-Means clustering is used to cluster the n feature vectors of m sampled pixels into k clusters (i.e., pixel clustering is performed based on the similarity between features, and Euclidean distance can be used for clustering), which serve as k initial dictionary atoms;

[0076] The k preliminary dictionary atoms are arranged into a dictionary matrix D by columns (that is, each column of the matrix is ​​a dictionary atom, and then sparse encoding is performed on them).

[0077] The sparse coefficient of each sampled pixel is calculated using the dictionary matrix D and the n feature vectors in the high-dimensional fusion feature matrix.

[0078] The sparse coefficients of each sampled pixel are concatenated to form a low-dimensional feature matrix A;

[0079] It should be noted that L1 sparse coding can extract key features from images, especially local features. Through sparse representation, the feature vector of each pixel can be effectively represented by a small number of dictionary atoms (basis). This preserves the key information of the image while ignoring unimportant parts, thereby improving computational efficiency. Concatenating the sparse coefficients of each pixel into a low-dimensional matrix A, this compression method not only reduces the data dimensionality but also makes it easier to process. Sparse coefficients reflect the features of local regions rather than global gray values, thus accurately describing the structural features in the image, such as texture and shape, making it suitable for subsequent image analysis (such as anomaly detection).

[0080] Sparse coefficients represent the features of a pixel (i.e., the region surrounding the pixel, i.e., the local region), rather than directly encoding the pixel's grayscale value. By calculating the L1 norm and the energy distribution of the sparse coefficients, the algorithm quantifies the distribution of the sparse coefficients, further optimizing the objective function. This allows the algorithm to accurately extract key features and avoid redundant information interference. As iterations proceed, the algorithm adaptively adjusts the regularization parameter λ according to changes in the objective function. This dynamic adjustment helps to further balance reconstruction error and sparsity, thereby improving the accuracy and stability of feature extraction. By setting a convergence threshold and a maximum number of iterations, the algorithm ensures that the final sparse coefficients are stable and effectively optimized. Each iteration avoids excessive computation while ensuring accuracy, ultimately obtaining sparse coefficients that meet the requirements and effectively support the task of lung nodule identification.

[0081] S252: Map the sparse coefficients of each sampled pixel in the low-dimensional feature matrix A according to the position coordinates in the lung image to be detected to generate an activation map;

[0082] It should be noted that by mapping the sparse coefficients in the low-dimensional feature matrix A to the lung image to be detected, each pixel represents the sparse coefficient activation intensity of that region. In the activation map, the regions with higher local sparse coefficients usually correspond to parts of the image with abnormal texture or morphological features. In this way, potential abnormal regions in the image (such as lung nodules, tumors, etc.) can be displayed on the activation map, thereby helping to detect these regions.

[0083] Since lung nodules often manifest as localized texture changes or morphological abnormalities, activation maps can effectively highlight these areas of change, providing very useful visual cues. Compared with analysis methods based directly on grayscale, activation maps based on sparse coding perform better in terms of resolution against noise interference and background areas, resulting in more accurate detection results.

[0084] In the activation map, the regions with significantly prominent local sparse coefficient values ​​usually correspond to areas with abnormal texture and morphological features in the image; each pixel represents the sparse coefficient activation intensity of that region, and abnormal regions are usually characterized by regions with high local values.

[0085] S253: Divide the pixels into grid image blocks according to the average distance between adjacent pixels. Each grid image block contains one pixel, and the grid image block corresponding to the pixel is a local region.

[0086] It should be noted that the expansion in step S2 is for illustrative purposes. The surrounding area of ​​each pixel at each scale and direction is considered as a local region. That is, regardless of the pixel's scale and direction, or the wavelet coefficients, they all represent the local area surrounding a pixel. The average distance between adjacent pixels describes the average distance between adjacent pixels. Using this average distance to divide the neighborhood of a pixel into half positions, a local region with a length equal to the average distance between adjacent pixels is formed (i.e., half of the average distance between adjacent pixels is used as the dividing line, one scale and one direction of the pixel are half, and adding half of the average distance between adjacent pixels at the corresponding scale and direction results in a length or diameter equal to the complete average distance between adjacent pixels. Due to the irregularity of pixels, the average distance between adjacent pixels may form a circle or a rectangle).

[0087] S254: Calculate a global histogram for the activation map using the global Otsu method, and statistically analyze the distribution of all activation values ​​(i.e., sparse coefficients) of the pixels in the global histogram; and preset a global segmentation threshold;

[0088] Determine whether the activation value of the pixel is greater than or equal to the global segmentation threshold, and generate a preliminary binary image;

[0089] If not, the local region corresponding to the activation value of the pixel is determined as the background image;

[0090] If so, the local region corresponding to the activation value of the pixel is determined as the first candidate abnormal region;

[0091] It should be noted that by calculating the global histogram using the Otsu method and setting a segmentation threshold, abnormal regions in the activation map can be separated from the background. The Otsu method can automatically select the optimal threshold to maximize the difference between the foreground and background of the image, which helps to accurately segment possible lung nodule regions.

[0092] If the activation value of a pixel is greater than the set threshold, the region is considered to be an abnormal region (i.e., the first candidate abnormal region, lung nodule). In this way, potential nodule regions can be preliminarily screened out, providing a basis for subsequent fine segmentation.

[0093] S255: Set a fixed window for the activation image, and re-divide the activation image through the fixed window to obtain the segmented region;

[0094] For each segmented region, calculate the mean and standard deviation of the corresponding activation values; then, calculate the adaptive segmentation threshold for each segmented region using the mean and standard deviation of the activation values.

[0095] Determine whether the activation value of the pixel is greater than or equal to the adaptive segmentation threshold;

[0096] If not, the local region corresponding to the activation value of the pixel is determined as the background image;

[0097] If so, the segmentation region corresponding to the activation value of that pixel is determined as the second candidate abnormal region;

[0098] It should be noted that by setting a fixed window to re-segment the activation map and calculating the adaptive segmentation threshold by combining the mean and standard deviation of the activation values ​​of each segmented region, the segmentation accuracy is improved, especially for nodules of varying sizes and irregular shapes. In this embodiment, an initial screening is first performed, and then possible abnormal regions (i.e., second candidate abnormal regions) are reconfirmed. The threshold is automatically adjusted adaptively based on the feature changes of the local region, which can more finely segment nodules from the background and avoid errors that may be caused by the global threshold.

[0099] Because the characteristics of lung nodules can vary significantly across different patients' CT images (e.g., shape, texture, and size), traditional segmentation methods may struggle to adapt uniformly to all situations. Adaptive segmentation methods, however, dynamically adjust thresholds to automatically adjust segmentation criteria based on specific features of different regions (e.g., local texture changes or morphological abnormalities), ensuring accurate segmentation for various nodule features. Furthermore, by introducing local adjustments on top of global methods, the segmentation process not only relies on global information but also allows for precise local optimization at the detail level. This combination of global background information and local detailed features improves the final segmentation performance.

[0100] S256: Calculate the proportional difference between all first-candidate anomaly regions and all second-candidate anomaly regions;

[0101] A preset difference threshold is set; it is then determined whether the proportional difference is greater than or equal to the difference threshold.

[0102] If so, the global segmentation threshold and the adaptive segmentation threshold are fused to calculate a new global threshold;

[0103] If not, the second candidate abnormal region obtained by the adaptive segmentation threshold is determined as the final abnormal lung nodule region.

[0104] It should be noted that the difference in the proportion between the first and second candidate abnormal regions is calculated. If the difference is large, the global and adaptive thresholds are fused to obtain a more robust and accurate threshold for abnormal detection. This strategy can effectively reduce false positives and false negatives, and ultimately obtain a more accurate lung nodule region.

[0105] When the difference between the first and second candidate anomaly regions is significant, it means that the anomaly regions identified by the global threshold and the adaptive threshold are quite different. This may be because the local features of the image (such as the shape and size of lung nodules) differ greatly from the global background, causing the global segmentation method to fail to accurately handle these local differences. To improve segmentation accuracy, it is necessary to fuse the global threshold and the adaptive segmentation threshold to obtain a more robust segmentation standard. This fusion can take into account the advantages of both, combining global information and local features, avoiding the coarsening of the global method or the limitations of the adaptive method, resulting in a more accurate final segmentation.

[0106] When the difference between the first and second candidate anomaly regions is small, it indicates that the results of global segmentation and adaptive segmentation are relatively consistent. This usually means that the anomaly region features in the image are relatively obvious and are not affected by noise or local changes. Both thresholding methods can capture these anomaly regions well. In this case, the second candidate anomaly region (based on the result of adaptive segmentation) can be directly used as the final lung nodule region. The adaptive segmentation method is more refined in handling local details and can accurately capture the shape and texture of the nodule without further adjusting the global threshold.

[0107] S257: The activation map is re-segmented by setting a fixed window to obtain a new segmentation region. Steps S255-S257 are repeated iteratively until the ratio difference between the first candidate anomaly region and the second candidate anomaly region of the new global threshold does not exceed the limit.

[0108] It should be noted that the above steps continuously optimize the identification of abnormal regions through multiple iterations (updating the global threshold and adjusting the local segmentation threshold); each iteration makes the segmentation results more accurate and reduces the problem of missegmentation caused by inaccurate initial thresholds.

[0109] Typically, sparse coding aims to represent raw data (such as images) with fewer dictionary atoms (i.e., feature vectors). In medical images, especially for the identification of lung nodules, images often contain a lot of background information and noise, while the nodules themselves may be sparse, that is, they only exist in certain areas of the image. By calculating sparsity coefficients, the algorithm can selectively activate a few dictionary atoms, thereby focusing on important features in the image and ignoring irrelevant parts.

[0110] Furthermore, in each iteration, the sparse coefficients determine the contribution of dictionary atoms to the reconstructed data. In lung nodule identification, by calculating and updating the sparse coefficients, the algorithm can gradually reduce the reconstruction error of the image, making the reconstructed image as close as possible to the original image. This process helps improve the recognition accuracy of nodule regions in the image because sparse representation can better capture the key features of the nodules and avoid interference from background noise. Therefore, Embodiment 2 of this invention provides the following special explanation regarding the encoding of sparse coefficients:

[0111] Example 2

[0112] The main technical solution implemented in Embodiment 2 of the present invention is the same as that implemented in Embodiment 1 above;

[0113] Embodiment 2 of the present invention provides a method for identifying and processing abnormal lung nodules, comprising the following steps:

[0114] S10: Acquire lung CT image data, preprocess the lung CT image data to obtain the lung image to be detected;

[0115] S20: Pixel analysis is performed on the lung image to be detected using the discrete fractional differential operator to obtain a fractional-order enhanced image; local energy and local entropy are calculated for the pixels of the fractional-order enhanced image at multiple scales and in multiple directions to construct a high-dimensional fusion feature matrix; the high-dimensional fusion feature matrix is ​​reduced in dimensionality to extract the lung nodule region;

[0116] The above embodiment two also involves steps S21-S25 and S251-S257 which are the same as those in the above embodiment one. The difference is that the second embodiment of the present invention provides a further extended explanation for "calculating the sparsity coefficient of each sampled pixel" in step S251.

[0117] Specifically, in step S251, the sparse coefficients of each sampled pixel are calculated using the dictionary matrix D and the n feature vectors in the high-dimensional fusion feature matrix, including:

[0118] The initial iteration parameters (i.e., initial sparse coefficients and step size, initial regularization parameter, and maximum iteration count b) are set using an iterative threshold algorithm. The initial objective function is calculated using the dictionary matrix D and the n eigenvectors in the high-dimensional fusion feature matrix, combined with the initial iteration parameters. The initial objective function is then iteratively updated, and the sparse coefficient vector for the current iteration count is calculated. The proportion of non-zero elements in the sparse coefficient vector for the current iteration count is calculated. The energy distribution of the sparse coefficient vector for the current iteration count is calculated based on this proportion. The initial objective function is updated based on the energy distribution of the sparse coefficient vector for the current iteration count to obtain the objective function for the current iteration count. The objective function for the current iteration count is then converged to obtain the final sparse coefficients. The specific steps are as follows:

[0119] S2511: Use the iterative threshold algorithm to set the initial sparse coefficients, step size, and maximum number of iterations b; use the initial sparse coefficients as the zero vector;

[0120] An initial regularization parameter is introduced to constrain the initial sparse coefficients;

[0121] It should be noted that step S2511 is mainly responsible for setting the initial sparse coefficients, step size, and maximum number of iterations b. The sparse coefficients are usually initialized to zero vectors. The zero vectors, as the starting point, do not favor any dictionary atoms. This means that at the beginning of the iterative thresholding algorithm, all feature vectors are equal for reconstruction. The zero vectors at the time of initialization represent an "unbiased" starting point, which helps to avoid the influence of initial bias on subsequent iterations and ensures the stability of the algorithm.

[0122] The initial regularization parameter λ can balance data reconstruction error and sparsity in sparse coding. By controlling the size of the sparsity coefficient, λ prevents overfitting and ensures that important features are extracted while irrelevant features are ignored. This regularization term is particularly important for processing noisy data, as it can suppress the influence of noise and reduce interference with the final result.

[0123] In the task of lung nodule recognition, images often contain noise. This initialization step can effectively reduce the interference of noise on lung nodule feature extraction by setting reasonable regularization parameters λ and sparsity coefficients. Furthermore, by balancing reconstruction error and sparsity through initialization of regularization parameters, the algorithm can ensure that it finds the most representative lung nodule features in noisy image data.

[0124] S2512: Using the dictionary matrix D and the initial sparse coefficients, perform linear calculations to obtain the reconstructed vector. ;

[0125] The initial objective function is calculated by combining the n eigenvectors in the high-dimensional fusion feature matrix with the initial sparse coefficients, step size, and reconstructed vector.

[0126] t;

[0127] in, This is represented as reconstruction error;

[0128] Represented as a reconstructed vector;

[0129] This is represented as a sparsity penalty (which is the adjustment parameter of the initial regularization parameter to the initial sparsity coefficient).

[0130] λ represents the initial regularization parameter;

[0131] Represented as the initial sparsity coefficients;

[0132] t represents the step size;

[0133] It should be noted that at the beginning of an iterative algorithm (such as ISTA), an initial sparsity coefficient needs to be set for the iterative process. (That is, usually set as a vector of all zeros or an initial value set according to prior information); then, by using this initial sparsity coefficient Substitute the values ​​into the objective function and calculate the initial objective function value;

[0134] The reconstructed vector is obtained by linear computation using the dictionary matrix D and the initial sparse coefficients. In sparse coding, the dictionary matrix D contains possible feature atoms, while the sparse coefficients determine the contribution of these atoms in the reconstruction process. The goal of linear computation is to reconstruct the input data (e.g., image feature vector) into a vector that closely approximates the original data. To evaluate the reconstruction effect, an objective function is usually computed, which includes reconstruction error and sparsity penalty terms.

[0135] The objective function is generally a weighted sum of the squares of the reconstruction error and the sparsity regularization term. The reconstruction error measures the difference between the original and reconstructed data, while the sparsity penalty term penalizes overly complex solutions and encourages sparse coefficients; the regularization parameter λ controls the balance between the two; by calculating the objective function, the current sparse coefficients can be evaluated and feedback can be provided for subsequent optimization steps.

[0136] The initialized objective function value can be used to evaluate the performance at the starting point (i.e., the initial sparsity coefficients). This provides a reference standard for subsequent iterations; a new objective function value is calculated in each iteration, and when F(α) no longer decreases significantly (that is, the final sparse coefficients are obtained through the objective function), the algorithm can be considered to have converged; the initial objective function value can sometimes also help to determine whether the parameters (such as λ) or step size t are reasonable, so as to make adaptive adjustments when needed;

[0137] The objective function is a scalar, representing the current sparse coding (i.e., the current initial sparse coefficients). The sum of data reconstruction error and sparsity regularization term; the objective function is initialized using initial sparsity coefficients (e.g., ...). The calculated F( The sparse coefficients are not equal to the sparse coefficients themselves; the sparse coefficients are the parameter vector that we hope to obtain throughout the iteration process to minimize the objective function.

[0138] The application of sparse coding technology in medical images can help extract key features related to nodules, and by adjusting the sparse coefficients to optimize the reconstruction error, the accuracy of nodule region identification can be improved.

[0139] S2513: Iteratively update the initial objective function described above;

[0140] The reconstruction residual for the current iteration is calculated by comparing the reconstruction vector of the current iteration with the n feature vectors in the high-dimensional fusion feature matrix using the gradient descent method.

[0141] ;

[0142] in, Let p represent the reconstructed vector for the current iteration number;

[0143] S2514: Calculate the gradient of the reconstruction residual for the current iteration number to obtain the reconstruction error for the current iteration number;

[0144] The reconstruction error at the current iteration number and the sparsity coefficient of the step size with respect to the current iteration number are used. Perform gradient descent to update the intermediate variables;

[0145] S2515: Apply the soft threshold operator to adjust the regularization parameter for the current iteration number. The effective threshold factor is calculated by incorporating it into the step size;

[0146] The sparse coefficient vector for the current iteration number is obtained by calculating using each of the intermediate variables and the effective threshold factor.

[0147] It should be noted that in steps S2513 to S2515, the algorithm optimizes the objective function using gradient descent. During gradient descent, the sparse coefficients are adjusted based on the current reconstruction error and gradient information, gradually optimizing the objective function. The core idea of ​​gradient descent is to update the sparse coefficients based on the gradient of the objective function with respect to the sparse coefficients, thereby reducing the value of the objective function and improving the accuracy of the result.

[0148] Next, the soft thresholding operator is applied to the current sparse coefficients. The soft thresholding operator is a commonly used tool for handling sparsity; its function is to set some sparse coefficients to zero, thereby achieving a sparse representation. The soft thresholding operation determines a threshold based on the product of the regularization parameter λ and the step size t (i.e., the soft threshold of the soft thresholding operator, usually obtained by multiplying the step size and the regularization parameter λ, which is the effective threshold factor). The soft threshold operator effectively controls the sparsity of the model and reduces redundant information by zeroing out coefficients smaller than the threshold.

[0149] The sparse coefficient vector contains multiple sparse coefficients, and each sparse coefficient corresponds to a dictionary atom in the dictionary matrix D.

[0150] In lung nodule identification, gradient descent reduces image reconstruction error by progressively updating sparse coefficients, ensuring more accurate extraction of nodule region features; soft thresholding operator improves sparsity and removes irrelevant features, helping to remove noise and redundant information, thus focusing more on the key features of lung nodules.

[0151] S2516: Determine whether the sparse coefficients in the sparse coefficient vector of the current iteration number are less than the effective threshold factor;

[0152] If so, then the sparse coefficients in the sparse coefficient vector of the current iteration number that are less than the effective threshold factor are taken as non-zero elements;

[0153] Count the number of non-zero elements; calculate the ratio of the number of non-zero elements to the sparse coefficient vector of the current iteration number to obtain the proportion of non-zero elements;

[0154] It should be noted that in sparse representation or sparse coding, each value in a vector is usually called an "element". After the soft thresholding operator, most coefficients may become zero due to the thresholding operation and are considered to be inactive or non-contributing parts. Non-zero elements can improve sparsity. Statistical analysis of these non-zero elements can help determine the quality of the final sparse representation and whether the regularization parameter λ needs to be adjusted to achieve a better balance between reconstruction and sparsity (i.e., when the objective function does not meet the requirements and iteration is performed again, it serves as the basis for subsequent adjustment of the regularization parameter λ).

[0155] The source of the non-zero number is mainly obtained after the soft thresholding operation by counting each non-zero sparse coefficient in the sparse coefficient vector of the current iteration. It can not only reflect the sparsity of the sparse coefficients obtained at the current iteration (i.e., convergence), but also serve as the basis for adjusting the regularization parameter λ when the final calculated objective function does not meet the requirements and iterates again.

[0156] In the image processing of lung nodules, by statistically analyzing the proportion of non-zero elements, it is possible to effectively determine whether the feature extraction is sufficiently concise, avoid overfitting, and identify features that make significant contributions, thereby improving the efficiency and accuracy of image analysis.

[0157] S2517: Calculate the L1 norm of the sparse coefficient vector of the current iteration number to obtain the energy distribution (i.e., the distribution of the sparse coefficient magnitudes) of the sparse coefficient vector of the current iteration number.

[0158] S2518: Update the initial objective function using the reconstruction error of the current iteration number, the regularization parameter λ of the current iteration number, and the energy distribution of the sparse coefficient vector of the current iteration number to obtain the objective function of the current iteration number;

[0159] It should be noted that steps S2517 to S2518 mainly involve calculating the L1 norm of the sparse coefficients and updating the objective function based on the distribution of the sparse coefficients. The L1 norm is an important indicator of sparsity; by calculating the L1 norm, we can understand the energy distribution of the sparse coefficients and thus determine the contribution of the features. The objective function is updated based on the energy distribution of the sparse coefficients to further optimize sparsity. The calculation of the L1 norm helps to evaluate the current sparsity of the sparse coefficients. By optimizing the objective function, the above steps ensure that the reconstruction error is minimized and the accuracy of key features is maintained in the lung nodule image. After updating the objective function, the identification of lung nodules can be more accurate, avoiding excessive redundant information from affecting the identification results.

[0160] S2519: Preset convergence threshold v; Determine whether the objective function of the current iteration number is less than the preset convergence threshold v, and simultaneously less than or equal to the maximum iteration number b;

[0161] If so, then determine the sparse coefficients in the sparse coefficient vector of the current iteration number as the final sparse coefficients;

[0162] If not, the sparse coefficient vector of the current iteration number and the objective function of the current iteration number are retained, and the process returns to step S2513 to continue iterating until the judgment result is met or the maximum iteration number b is greater than the maximum iteration number.

[0163] When re-iteratio calculating the objective function using the sparse coefficient vector of the current iteration number and the objective function of the current iteration number, the regularization parameter λ is adaptively adjusted using the sparse coefficient vector of the current iteration number and the objective function of the current iteration number.

[0164] It should be noted that the convergence threshold and the maximum number of iterations are set to determine whether the algorithm has converged. If the objective function has met the convergence condition or the maximum number of iterations has been reached, the algorithm is considered to be complete and the final sparse coefficients are returned. By terminating the iteration in a timely manner, unnecessary calculations and overfitting are avoided, and the processing efficiency of lung nodule identification is improved.

[0165] Simultaneously, the data from the current iteration (i.e., the sparse coefficient vector of the current iteration number and the objective function of the current iteration number) is retained, and the regularization parameter λ is adjusted based on the data from the current iteration to prepare for the next iteration, ensuring that the final objective function converges and improving the accuracy and efficiency of convergence; according to... Figure 5 The diagram illustrates the overall process for identifying lung nodules.

[0166] Example 3

[0167] like Figure 6 As shown, this application also provides a lung abnormal nodule identification and processing system, including: an acquisition module 10; and an identification module 20. The acquisition module 10 is used to acquire lung CT image data and preprocess the lung CT image data to obtain a lung image to be detected. The identification module 20 is used to perform pixel analysis on the lung image to be detected using a discrete fractional differential operator to obtain a fractional-order enhanced image. The pixels of the fractional-order enhanced image are acquired at multiple scales and in multiple directions to calculate local energy and local entropy, and a high-dimensional fusion feature matrix is ​​constructed. The high-dimensional fusion feature matrix is ​​dimensionality reduced to extract the lung nodule region.

[0168] Analysis shows that feature extraction is a crucial step in lung nodule detection. Traditional feature extraction methods often rely on local image features, such as texture, grayscale, and shape. However, for complex structures and subtle differences in images, traditional methods often struggle to obtain effective high-dimensional feature representations. This invention proposes a lung abnormal nodule identification and processing method based on fractional calculus image processing, which can better capture image details and perform more refined image analysis.

[0169] Processing lung CT images using discrete fractional differential operators enhances high-frequency information, better revealing edges, textures, and other details, making subtle changes and details clearer, and highlighting nodule boundaries. Compared to traditional integer-order differentials, fractional-order differentials perform better in noisy images, effectively suppressing noise interference while preserving image details. Multi-scale and multi-directional feature extraction from the enhanced fractional-order images improves recognition accuracy. Dimensionality reduction of the high-dimensional fusion feature matrix reduces redundant information, retaining the most valuable features for nodule identification while reducing computational complexity. Furthermore, this invention's method enhances image details using discrete fractional differential operators, making nodule and texture information more prominent, improving nodule identifiability, and contributing to higher detection accuracy.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and processing abnormal lung nodules, characterized in that, The following steps are included: Acquire lung CT image data, preprocess the lung CT image data to obtain lung images to be detected; The discrete fractional differential operator is used to perform pixel analysis on the lung image to be detected to obtain a fractional enhanced image; multi-scale and multi-directional pixels are collected from the fractional enhanced image, and local energy and local entropy are further calculated to construct a high-dimensional fusion feature matrix; The high-dimensional fusion feature matrix is ​​reduced in dimensionality to extract the lung nodule region; The fractional-order enhanced image is sampled at multiple scales and in multiple directions, and local energy and local entropy are further calculated to construct a high-dimensional fusion feature matrix. The specific operation steps are as follows: The fractional-order enhanced image was processed using Gabor wavelets. A custom wavelength is set for local texture; pixels in the fractional-order enhanced image are acquired at multiple scales and in multiple directions according to the set custom wavelength; wherein... This is represented as a fractional-order enhanced image, where α represents a preset fractional-order parameter; Wavelet transform is performed on the pixels in the fractional-order enhanced image at multiple scales and in multiple directions to obtain multiple scales of the pixels. Multiple directions wavelet coefficients ; Local energy is calculated based on wavelet coefficients of pixels at each scale and in each direction in the fractional-order enhanced image. The local energy of each pixel at each scale and in each direction is normalized, and the sum of the local energy at all scales and in all directions is calculated. The probability distribution of each pixel at each scale and in each direction is obtained by calculating the sum of the local energy and the local energy. The local entropy is obtained by calculating the probability distribution of the local energy. The local energy and entropy of the fractional-order enhanced image are combined to obtain local statistical features; weights are assigned to these local statistical features to construct a high-dimensional fusion feature matrix. ; The high-dimensional fusion feature matrix The middle row is represented by m pixels, and the column is represented by n feature vectors; The high-dimensional fusion feature matrix is ​​reduced in dimensionality to extract the lung nodule region. The specific steps are as follows: The high-dimensional fusion feature matrix is ​​reduced in dimensionality using the L1 sparse coding algorithm to generate an activation map; the activation map is segmented to obtain candidate abnormal regions; and the candidate abnormal regions are thresholded to obtain abnormal lung nodule regions.

2. The method for identifying and processing abnormal lung nodules according to claim 1, characterized in that, The discrete fractional differential operator is used to perform pixel-level analysis on the lung image to be detected, resulting in a fractional-order enhanced image. The specific operation steps are as follows: The discrete fractional differential operator is used to perform multi-level fractional differential processing on the lung image to be detected: the neighboring distance of the neighboring domain is calculated for each pixel in the lung image to be detected, and the average distance between neighboring pixels is calculated by the neighboring distance of the neighboring domains of all pixels. A fractional-order parameter α is preset, and a discrete fractional differential operator is constructed using the fractional-order parameter α and the average distance between adjacent pixels. ; Perform discrete fractional differential operator on each pixel in the image to be detected The convolution operation extracts the local texture of pixel grayscale values ​​and generates a fractional-order enhanced image.

3. The method for identifying and processing abnormal lung nodules according to claim 2, characterized in that, The high-dimensional fused feature matrix is ​​reduced in dimensionality using the L1 sparse coding algorithm to generate an activation map. The specific steps are as follows: The L1 sparse coding algorithm is used to select n feature vectors from m sampled pixels in the high-dimensional fusion feature matrix; K-Means clustering is used to cluster the n feature vectors of m sampled pixels into k clusters, which serve as k initial dictionary atoms; Arrange the k initial dictionary atoms into a dictionary matrix D by columns; The sparse coefficient of each sampled pixel is calculated using the dictionary matrix D and the n feature vectors in the high-dimensional fusion feature matrix. The sparse coefficients of each sampled pixel are concatenated to form a low-dimensional feature matrix A; The sparse coefficients of each sampled pixel in the low-dimensional feature matrix A are mapped according to the position coordinates in the lung image to be detected to generate an activation map.

4. The method for identifying and processing abnormal lung nodules according to claim 3, characterized in that, The activation map is segmented to obtain candidate abnormal regions; a threshold judgment is performed on the candidate abnormal regions to obtain abnormal lung nodule regions. The specific operation steps are as follows: The pixels are divided into grid image blocks based on the average distance between adjacent pixels. Each grid image block contains one pixel, and the grid image block corresponding to that pixel is considered a local region. The global Otsu method is used to calculate a global histogram of the activation map, and the distribution of pixels with all activation values ​​in the global histogram is statistically analyzed; and a global segmentation threshold is preset. Determine whether the activation value of the pixel is greater than or equal to the global segmentation threshold, and generate a preliminary binary image; If so, the local region corresponding to the activation value of the pixel is determined as the first candidate abnormal region; The activation image is set to a fixed window, and the activation image is re-divided through the fixed window to obtain the segmented region; For each segmented region, calculate the mean and standard deviation of the corresponding activation values; then, calculate the adaptive segmentation threshold for each segmented region using the mean and standard deviation of the activation values. Determine whether the activation value of the pixel is greater than or equal to the adaptive segmentation threshold; If so, the segmentation region corresponding to the activation value of that pixel is determined as the second candidate abnormal region.

5. The method for identifying and processing abnormal lung nodules according to claim 4, characterized in that, The activation map is segmented to obtain candidate abnormal regions; a threshold judgment is performed on the candidate abnormal regions to obtain abnormal lung nodule regions. The specific operation steps are as follows: Calculate the proportional difference between all first-candidate anomaly regions and all second-candidate anomaly regions; A preset difference threshold is set; it is then determined whether the proportional difference is greater than or equal to the difference threshold. If so, the global segmentation threshold and the adaptive segmentation threshold are fused to calculate a new global threshold; If not, the second candidate abnormal region obtained by the adaptive segmentation threshold is determined as the final abnormal lung nodule region. The activation map is re-segmented by setting a fixed window to obtain a new segmentation region. The above steps are repeated iteratively until the ratio difference between the first candidate anomaly region and the second candidate anomaly region of the new global threshold is less than the difference threshold.

6. The method for identifying and processing abnormal lung nodules according to claim 5, characterized in that, The sparse coefficients of each sampled pixel are calculated using the dictionary matrix D and the n feature vectors in the high-dimensional fusion feature matrix. The specific steps are as follows: The initial iteration parameters are set using an iterative threshold algorithm. The initial objective function is calculated by combining the dictionary matrix D with the n feature vectors in the high-dimensional fusion feature matrix and the initial iteration parameters. The initial objective function is iteratively updated, and the sparse coefficient vector of the current iteration number is calculated. Calculate the proportion of non-zero elements in the sparse coefficient vector of the current iteration number; calculate the energy distribution of the sparse coefficient vector of the current iteration number based on the proportion of non-zero elements. The initial objective function is updated based on the energy distribution of the sparse coefficient vector at the current iteration number to obtain the objective function at the current iteration number. The objective function for the current iteration number is converged to obtain the final sparse coefficients.

7. A lung abnormal nodule identification and processing system, characterized in that, include: Data acquisition module; Recognition module; The acquisition module is used to acquire lung CT image data and preprocess the lung CT image data to obtain lung images to be detected. The recognition module is used to perform pixel analysis on the lung image to be detected using the discrete fractional differential operator to obtain a fractional enhanced image; to collect pixels of the fractional enhanced image at multiple scales and in multiple directions, and to further calculate local energy and local entropy to construct a high-dimensional fusion feature matrix. The high-dimensional fusion feature matrix is ​​reduced in dimensionality to extract the lung nodule region; The fractional-order enhanced image is sampled at multiple scales and in multiple directions, and local energy and local entropy are further calculated to construct a high-dimensional fusion feature matrix. The specific operation steps are as follows: The fractional-order enhanced image was processed using Gabor wavelets. A custom wavelength is set for local texture; pixels in the fractional-order enhanced image are acquired at multiple scales and in multiple directions according to the set custom wavelength; wherein... This is represented as a fractional-order enhanced image, where α represents a preset fractional-order parameter; Wavelet transform is performed on the pixels in the fractional-order enhanced image at multiple scales and in multiple directions to obtain multiple scales of the pixels. Multiple directions wavelet coefficients ; Local energy is calculated based on wavelet coefficients of pixels at each scale and in each direction in the fractional-order enhanced image. The local energy of each pixel at each scale and in each direction is normalized, and the sum of the local energy at all scales and in all directions is calculated. The probability distribution of each pixel at each scale and in each direction is obtained by calculating the sum of the local energy and the local energy. The local entropy is obtained by calculating the probability distribution of the local energy. The local energy and entropy of the fractional-order enhanced image are combined to obtain local statistical features; weights are assigned to these local statistical features to construct a high-dimensional fusion feature matrix. ; The high-dimensional fusion feature matrix The middle row is represented by m pixels, and the column is represented by n feature vectors; The high-dimensional fusion feature matrix is ​​reduced in dimensionality to extract the lung nodule region. The specific steps are as follows: The high-dimensional fusion feature matrix is ​​reduced in dimensionality using the L1 sparse coding algorithm to generate an activation map; the activation map is segmented to obtain candidate abnormal regions; and the candidate abnormal regions are thresholded to obtain abnormal lung nodule regions.

Images