Lung injury area automatic segmentation system and method based on deep learning

Abstract

The invention provides a lung injury area automatic segmentation system and method based on deep learning, and aims to improve the accuracy of automatic segmentation of a lung injury area. Firstly, denoising, edge detection and region enhancement are carried out on an original lung image, and the processed original lung image is output; thirdly, performing feature extraction and image segmentation on the processed original lung image by adopting a deep convolutional neural network (CNN), optimizing a segmentation result in combination with a medical map, and outputting an optimized lung injury region segmentation image; and finally, through an augmented reality technology and an interactive correction function, a doctor is helped to adjust a segmentation result in real time, and through morphological operation and smooth processing, a final lung injury area segmentation image is output. According to the method, through a multi-level technical means, the accuracy of automatic segmentation of the lung injury area is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of automatic segmentation of lung injury regions, and specifically to an automatic segmentation system and method for lung injury regions based on deep learning. Background Technique

[0002] The automatic segmentation technology of lung injury regions is of great significance in medical imaging, especially in the early diagnosis and treatment of lung diseases. Currently, traditional lung injury segmentation methods usually rely on manual annotation and manual judgment, which are not only time-consuming and laborious, but also easily affected by doctors' experience and subjective factors, resulting in inaccurate or inconsistent segmentation results. In addition, existing automatic segmentation methods have problems such as insufficient accuracy and poor adaptability. Especially when dealing with complex lung images, it is often unable to effectively segment fine lung injury regions.

[0003] In recent years, deep learning technology, especially convolutional neural networks (CNNs), has made remarkable progress in the field of medical image analysis. Deep learning methods can automatically learn features from a large number of medical images and perform effective segmentation, but still face problems such as noise influence, image quality differences, and blurred boundaries. Although some methods attempt to improve these problems through image preprocessing and postprocessing techniques, how to accurately identify and optimize lung injury regions remains a challenge.

[0004] In the prior art, deep learning and convolutional neural networks (CNNs) are used for image segmentation and optimization. By training a deep network model, features are extracted from the original image and segmented. However, in terms of image noise processing and boundary optimization, it fails to be fine enough, resulting in a low segmentation accuracy of the injury region in a complex image background. In addition, these methods often lack effective postprocessing steps, and the segmentation boundary may be serrated, affecting the accuracy of the diagnostic result. Summary of the Invention

[0005] In order to solve the technical problems mentioned in the current background technique, the present invention proposes an automatic segmentation system and method for lung injury regions based on deep learning.

[0006] For this reason, the technical solution adopted by the present invention is as follows:

[0007] An automatic segmentation system for lung injury regions based on deep learning, characterized in that the system includes:

[0008] M1. An image processing module, which receives the original lung image from a medical imaging device, performs denoising processing, edge detection, and region enhancement processing on the original lung image, and outputs the processed original lung image;

[0009] M2, a segmentation and optimization module, performs multi-layer convolution operations on the processed original lung image through a deep convolutional neural network, extracts low-level and high-level features, segments the processed original lung image based on the low-level and high-level features to generate a segmented image of the lung injury area, and optimizes the segmented image of the lung injury area using medical atlas knowledge to output an optimized segmented image of the lung injury area;

[0010] M3, a visualization and correction module, overlays and displays the optimized segmented image of the lung injury area with the original lung image for comparison, presents the comparison result to the doctor for viewing through augmented reality technology, and the doctor interactively corrects the optimized segmented image of the lung injury area through a virtual interface to output a corrected segmented image of the lung injury area;

[0011] M4, a post-processing module, performs morphological operations on the corrected segmented image of the lung injury area and smooths the boundaries of the segmented image through an interpolation algorithm to output a final segmented image of the lung injury area.

[0012] Further, the denoising process is performed through bilateral filtering; the edge detection is implemented using the Canny edge detection algorithm, including gradient calculation, gradient magnitude, and edge direction.

[0013] The formula for the gradient calculation is:

[0014]

[0015] where G x is the gradient of the original lung image I in the horizontal direction; G y is the gradient of the original lung image I in the vertical direction; I(x, y) is the pixel value of the original lung image I at the position (x, y).

[0016] The formula for the gradient magnitude is:

[0017]

[0018] where G is the gradient magnitude, that is, the clarity of the edge.

[0019] The formula for the gradient direction is:

[0020]

[0021] where θ is the gradient direction, indicating the direction of the edge of the original lung image.

[0022] Further, the region enhancement includes the calculation of local mean and standard deviation, and the formula is:

[0023]

[0024] Among them, N(x, y) is the neighborhood area centered on the pixel (x, y); μ local (x, y) is the mean value of the area N(x, y), that is, the average value of all pixels in this area; σ local (x, y) is the standard deviation of the area N(x, y); I(x′, y′) is the pixel value of the pixel (x′, y′) in the neighborhood;

[0025] According to the local mean and standard deviation, calculate the enhancement value C(x, y) of the pixels of the original lung image, and the formula is:

[0026]

[0027] Among them, α is the enhancement coefficient, which controls the intensity of enhancement;

[0028] According to the denoising process, edge detection and region enhancement, output the processed original lung image J.

[0029] Furthermore, the operation steps of the segmentation and optimization module include feature extraction, image segmentation, and medical knowledge graph optimization.

[0030] The feature extraction uses a deep convolutional neural network to extract low-level and high-level features in the processed original lung image through multi-layer convolution operations. The low-level features include edges, textures, and corners, and the high-level features include shapes, structures, and regions. The low-level feature y low is extracted through multi-layer convolution of the processed original lung image by a convolution kernel, and the formula is:

[0031]

[0032] Among them, J(m, n) is the pixel value of the processed original lung image J at the position (m, n); K low (x - m, y - n) is the convolution kernel;

[0033] According to the low-level features, extract high-level features y high , and the formula is:

[0034]

[0035] Among them, y high (x, y) is the output of the high-level feature;

[0036] The low-level features and high-level features constitute the output feature f.

[0037] Furthermore, the image segmentation judges whether it belongs to the lung injury area by classifying the output features, and the formula is:

[0038]

[0039] Among them, W and b are the weights and biases of the deep convolutional network respectively; σ is the Sigmoid activation function, which is used to compress the output of the deep convolutional network to between [0,1], indicating that the feature belongs to the lung injury area. is the segmented image of the lung injury area;

[0040] The medical knowledge graph optimizes and utilizes the medical graph to optimize the segmented image of the lung injury area, corrects unreasonable segmented areas, identifies incorrect areas in the segmented image of the lung injury area according to the medical graph, and optimizes by adjusting the segmentation boundary, and outputs the optimized segmented image Y of the lung injury area.

[0041] Furthermore, the overlapping display and comparison are realized by augmented reality technology, and the optimized segmented image Y of the lung injury area is overlapped and displayed with the original lung image I. The formula is:

[0042] I overlay (x,y) = α·I(x,y) + (1 - α)·Y(x,y)

[0043] where I overlay (x,y) is the overlapping image; α is the weighting coefficient, which controls the contribution ratio of the original lung image and the optimized segmented image of the lung injury area. α ∈ [0,1], and the display effect can be adjusted according to needs.

[0044] Furthermore, the interactive correction is specifically that the doctor selects the area to be corrected through the interactive interface, adjusts the segmented image of the lung injury area according to the doctor's operation, and the corrected segmented image Y of the lung injury area corrected The calculation formula is:

[0045]

[0046] where R corrected is the coordinate set of the area manually corrected by the doctor, representing the updated boundary of the segmented image of the lung injury area.

[0047] Furthermore, the morphological operations include dilation and erosion operations. The dilation operation expands the boundary of the corrected segmented image of the lung injury area, enlarges the foreground area, and fills small holes.

[0048] The erosion operation contracts the boundary of the foreground area of the corrected segmented image of the lung injury area and removes small noise areas.

[0049] The smoothing process is realized by the interpolation algorithm. The interpolation algorithm calculates a smooth curve by weighted averaging of boundary pixels to adjust the boundary. The formula is:

[0050] Y smooth (x, y) = f(Y corrected (x, y))

[0051] where Y smooth (x, y) is the segmented image after boundary smoothing; f is a smoothing algorithm that is adjusted according to the boundary of the segmented image;

[0052] Through the morphological operation and smoothing process, the final segmented image Y of the lung injury area is output smooth .

[0053] An automatic segmentation method for lung injury area based on deep learning, characterized in that the method comprises the following steps:

[0054] S1. Receive the original lung image from a medical imaging device and perform denoising, edge detection, and region enhancement processing, and output the processed original lung image;

[0055] S2. Perform multi-layer convolution operations on the processed original lung image through a deep convolutional neural network, extract low-level and high-level features, based on the low-level and high-level features, segment the processed original lung image, and optimize the segmentation result using a medical atlas to generate an optimized segmented image of the lung injury area;

[0056] S3. Overlay and display the optimized segmented image of the lung injury area with the original lung image, and the doctor performs interactive correction on the optimized segmented image of the lung injury area through a virtual interface, and outputs the corrected segmented image of the lung injury area

[0057] S4. Perform morphological repair and boundary smoothing on the corrected segmented image of the lung injury area to generate the final segmented image of the lung injury area.

[0058] Compared with the prior art, the advantages of the present invention are:

[0059] 1. High precision in image processing and optimization: The present invention adopts a bilateral filtering denoising technique, which can effectively retain the detailed edges of the lungs and the injury area. Through the Canny edge detection algorithm and region enhancement, the contrast and clarity of the segmented area are further improved, ensuring that the injury area can be presented more prominently.

[0060] 2. Efficient feature extraction and accurate segmentation: The present invention uses a deep convolutional neural network (CNN) and combines a medical atlas to optimize the segmentation result. Through the effective abstraction and fusion of low-level and high-level features by the deep network, it not only has strong image understanding ability, but also can identify error areas in the segmentation according to the medical atlas and perform boundary adjustment, thereby realizing accurate segmentation of the lung injury area.

[0061] 3. Interactive correction function: The present invention introduces augmented reality technology. By overlapping and displaying the original image and the optimized segmented image, doctors can interactively correct the segmentation result through a virtual interface. This function ensures that the segmentation result is further optimized with the participation of doctors, improving the reliability of the system.

[0062] 4. Improvement of post-processing technology: Through morphological operations and boundary smoothing, small errors in the segmentation result are effectively repaired to ensure that the final image meets the standards of medical images and is convenient for clinical applications. Brief Description of the Drawings

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0064] Figure 1 It is the flowchart of the automatic segmentation system for lung injury area of the present invention;

[0065] Figure 2 It is the flowchart of the image processing module of the present invention;

[0066] Figure 3 It is the flowchart of the segmentation and optimization module of the present invention. Detailed Embodiment

[0067] To achieve the above objectives, the present invention is realized through the following technical solutions. The present invention provides an automatic segmentation system for lung injury area based on deep learning. The system includes:

[0068] M1. Image processing module, which receives the original lung image from a medical imaging device, performs denoising, edge detection, and region enhancement processing on the original lung image, and outputs the processed original lung image;

[0069] The original lung image is denoised through bilateral filtering to remove high-frequency noise in the original lung image, improve the image quality, and ensure subsequent segmentation and enhancement effects. Bilateral filtering can not only remove noise but also retain the edge details of the image.

[0070] Bilateral filtering can effectively retain the edge information of the lung image while removing background noise. For particularly complex noise, median filtering can be combined for additional denoising to further enhance the noise suppression effect. The denoising process removes the noise in the original lung image through bilateral filtering, especially retaining the detailed edges of the lung and injury areas, providing a clean basic image for edge detection and region enhancement.

[0071] The purpose of edge detection is to extract the boundary information of the original lung image and provide structural information for the segmentation task. The Canny edge detection algorithm is used to identify regions with large changes in the image by calculating the gradient magnitude of the original lung image. These regions correspond to the boundaries in the image. The calculation steps include gradient calculation, gradient magnitude, and edge direction.

[0072] Gradient calculation formula:

[0073]

[0074] where G x is the gradient of image I in the horizontal direction, reflecting the change of the image in the horizontal direction; G y is the gradient of image I in the vertical direction, reflecting the change of the image in the vertical direction; I(x,y) is the pixel value of image I at position (x,y);

[0075] Gradient magnitude calculation formula:

[0076]

[0077] where G is the gradient magnitude, representing the change intensity of the image at a certain position, that is, the obviousness of the edge;

[0078] Gradient direction formula:

[0079]

[0080] where θ is the gradient direction, representing the direction of the image edge;

[0081] Through edge detection, non-maximum suppression and double-threshold processing are completed. Local maximum suppression is performed on the gradient magnitude to remove redundant edge responses and only retain the most important edge information. Edge connection is performed through thresholds. Strong edges are marked as boundaries, and weak edges will also be retained as boundaries when meeting certain conditions; The Canny algorithm can effectively extract the detailed edges in the image. Especially in images with more noise, it can identify the main structural boundaries in the image, such as the contour of the lung and the boundary of the damaged area;

[0082] After edge detection, the lung edges and damaged areas identified in the image can be used as target areas for further processing. Region enhancement is mainly to improve the visibility of low-contrast regions and make the damaged areas more prominent. In this process, edge information can be used as a reference for enhancing the target area. The enhancement algorithm preferentially processes regions near the edge, which usually contain the most important structural information. By enhancing the contrast, region enhancement provides a clearer visual effect for the segmentation of the damaged area;

[0083] Regional enhancement is performed after edge detection. The lung edges and damaged areas determined using the edge information are used to enhance the contrast, including the calculation of local mean and standard deviation. The formula is as follows:

[0084]

[0085]

[0086] where N(x, y) is the neighborhood area centered on the pixel (x, y), including pixels within a certain range around it; μ local (x, y) is the mean of the area N(x, y), that is, the average value of all pixels in this area; σ local (x, y) is the standard deviation of the area N(x, y), representing the degree of variation of pixel values in this area; I(x′, y′) is the pixel value of the pixel (x′, y′) in the neighborhood;

[0087] According to the local mean and standard deviation, the enhancement value C(x, y) of the image pixels is calculated. The formula is as follows:

[0088]

[0089] where α is the enhancement coefficient, controlling the intensity of enhancement;

[0090] The original lung image is preprocessed through denoising, edge detection, and regional enhancement technical steps, so as to provide a high-quality input image for the subsequent segmentation and optimization modules.

[0091] M2, Segmentation and Optimization Module: Through a deep convolutional neural network, perform multi-layer convolutional operations on the processed original lung image, extract low-level and high-level features. Based on the low-level and high-level features, segment the processed original lung image to generate a segmented image of the lung damage area, and use medical atlas knowledge to optimize the segmented image of the lung damage area, and output the optimized segmented image of the lung damage area;

[0092] The main task of the segmentation and optimization module is to use a deep convolutional neural network (CNN) to extract features and segment the preprocessed original lung image, and optimize the preliminary segmentation results based on medical atlas knowledge, and finally generate a high-precision segmented image of the lung damage area. The technical steps include feature extraction, image segmentation, and medical knowledge atlas optimization.

[0093] In the feature extraction stage, the deep convolutional neural network (CNN) extracts low-level and high-level features in the image through multi-layer convolutional operations. The low-level features include edges, textures, and corners in the image, and the high-level features include shapes, structures, and regions in the image. The low-level features y low are extracted through multi-layer convolution of the image by the convolution kernel. The formula is as follows:

[0094]

[0095] Among them, J(m,n) is the pixel value of the processed original lung image J at the position (m,n); K low (x - m, y - n) is the convolution kernel;

[0096] According to the low - level features, high - level features y are extracted through a deep convolutional network high , and the formula is:

[0097]

[0098] Among them, y high (x,y) is the output of the high - level features;

[0099] The final output feature f includes low - level features and high - level features;

[0100] Image segmentation classifies the features to determine whether they belong to the lung injury area, and the formula is:

[0101]

[0102] Among them, W and b are the weights and biases of the deep convolutional network respectively, and σ is the Sigmoid activation function, which is used to compress the network output to between [0,1], indicating that the feature belongs to the lung injury area, is the image segmentation result of the lung injury area;

[0103] The medical knowledge graph optimizes and utilizes the knowledge of the medical graph to optimize the segmentation result of the lung injury area, correct the unreasonable segmentation areas. The medical graph provides the anatomical structure, pathological information and common locations of the injury areas of the lungs. According to the medical graph, the wrong areas in the segmentation result are identified and optimized by adjusting the segmentation boundary, and the final optimized segmentation image Y is output.

[0104] M3, the visualization and correction module, overlays and displays the optimized segmentation image of the lung injury area and the original lung image for comparison, and presents the comparison result to the doctor for viewing through augmented reality technology. The doctor makes interactive corrections to the optimized segmentation image of the lung injury area through a virtual interface, and outputs the corrected segmentation image of the lung injury area;

[0105] The main task of the visualization and correction module is to overlay and display the segmentation result and the original image through augmented reality technology, and provide an interactive interface that allows the doctor to make manual corrections. This module ensures that after the segmentation result is corrected by the doctor, a more accurate annotation of the lung injury area is obtained, thereby improving the reliability of automatic segmentation. The technical steps include overlay display and interactive correction,

[0106] The optimized segmented image Y is overlapped with the original lung image I to help doctors visually see the differences between the segmentation result and the original image. The image content is superimposed through augmented reality technology to obtain a new image for display, so that doctors can evaluate and adjust more clearly. The formula is:

[0107] I overlay (x,y) = α·I(x,y) + (1 - α)·Y(x,y)

[0108] Among them, I overlay (x,y) is the overlapped image shown to the doctor, which combines the original image and the segmented image; α is the weighting coefficient that controls the contribution ratio of the original image and the segmented image, and α ∈ [0,1] so as to adjust the display effect as needed;

[0109] Through the overlapping display, doctors can visually see the differences between the segmentation result and the original image, especially in the distribution and morphology of the lung injury area, thus helping doctors make more accurate judgments;

[0110] Based on the overlapping display, an interactive interface is provided to allow doctors to correct the segmented image through a virtual interface. The specific operation is as follows:

[0111] Doctors select the area to be corrected through the interactive interface, and the system will automatically update the segmentation result of this area, adjust the image according to the doctor's operation. The correction process involves merging the boundary of the area selected by the doctor with the segmented image. The area manually adjusted by the doctor is marked as R corrected , and the corrected segmented image Y corrected The calculation formula is:

[0112]

[0113] Among them, R corrected is the coordinate set of the area manually corrected by the doctor, representing the updated boundary of the lung injury area;

[0114] Through the interactive correction, doctors can adjust the segmentation result in real time to ensure that the segmented area better conforms to the actual injury area. This process further improves the accuracy of the automatic segmentation result, especially in the case of complex or blurred boundaries.

[0115] M4, the post-processing module, performs morphological operations on the corrected segmented image of the lung injury area and smooths the boundary of the segmented image through an interpolation algorithm, and outputs the final segmented image of the lung injury area.

[0116] The task of the post - processing module is to perform final optimization on the corrected segmented image, ensuring that the final output segmentation result has higher precision and meets the standards of medical images. The key objectives of post - processing are to repair any small segmentation errors, smooth the segmentation boundaries, and ultimately provide an image that meets clinical requirements. The technical steps include morphological operations and boundary smoothing;

[0117] The purpose of morphological operations is to eliminate small holes, noise, and missing regions in the segmented image, ensuring the integrity of the segmented region. Morphological operations include dilation and erosion operations, which are used to repair small - area errors.

[0118] The dilation operation expands the boundaries of the segmented image, making the foreground region in the image larger and filling small gaps. By selecting a structuring element to process the image, the center of the structuring element is aligned with each pixel in the image, and the pixel value is updated to the maximum value in the neighborhood.

[0119] The erosion operation shrinks the boundaries of the foreground region, removing small noise regions in the image. By selecting a structuring element to process the image, the pixel value is updated to the minimum value in the neighborhood.

[0120] Dilation and erosion are used alternately to repair small - area errors in the image. By first dilating and then eroding, it can effectively fill the holes in the region and remove small noise, ensuring the coherence and integrity of the segmented region;

[0121] The purpose of boundary smoothing is to make the boundaries of the segmented image smoother, avoiding overly sharp and irregular edges. Through boundary smoothing, the contour of the segmented image will be more natural and conform to the actual morphology of medical images. An interpolation algorithm is used for boundary smoothing. The specific operation is as follows:

[0122] The interpolation algorithm calculates a smooth curve to adjust the boundary by performing weighted averaging on the boundary pixels. The formula is:

[0123] Y smooth (x,y)=f(Y corrected (x,y))

[0124] where, Y smooth [x,y) is the segmented image after boundary smoothing; f is the smoothing algorithm, which is adjusted according to the boundary of the segmented image;

[0125] Boundary smoothing ensures that the boundaries in the segmented image are more natural and smooth, removing overly sharp edges, making the contour of the damaged area more in line with the actual morphology of medical images, and enhancing the clinical applicability of the segmented image;

[0126] The final output of the post - processing module is the segmented image of the lung injury area after morphological repair and boundary smoothing. This image will be saved as the final result and provided to clinicians for diagnosis and subsequent treatment.

[0127] An automatic segmentation method for lung injury area based on deep learning, characterized in that the method comprises the following steps:

[0128] S1. Receive the original lung image from a medical imaging device and perform denoising, edge detection, and region enhancement processing, and output the processed original lung image;

[0129] S2. Perform multi - layer convolution operations on the processed original lung image through a deep convolutional neural network, extract low - level and high - level features, segment the processed original lung image based on the low - level and high - level features, and optimize the segmentation result using a medical atlas to generate an optimized segmented image of the lung injury area;

[0130] S3. Overlay and display the optimized segmented image of the lung injury area with the original lung image, and the doctor makes interactive corrections to the optimized segmented image of the lung injury area through a virtual interface, and outputs the corrected segmented image of the lung injury area

[0131] S4. Perform morphological repair and boundary smoothing on the corrected segmented image of the lung injury area to generate the final segmented image of the lung injury area.

[0132] An automatic segmentation system and method for lung injury area based on deep learning proposed by the present invention combines an image processing module, a segmentation and optimization module, a visualization and correction module, and a post - processing module, aiming to improve the accuracy and reliability of lung injury area segmentation. First, the original lung image is processed by techniques such as bilateral filtering denoising, Canny edge detection, and region enhancement to improve the image quality. Then, a deep convolutional neural network (CNN) is used to segment the image and combined with a medical atlas for optimization to improve the accuracy of the segmentation result. Finally, doctor interactive correction is realized through augmented reality technology, and morphological operations and smoothing processing are used to optimize the segmentation result to ensure that the final output segmented image of the lung injury area meets medical standards.

[0133] In summary, the advantages of the present invention are as follows: by adopting advanced image processing technology, the accuracy and clarity of the segmentation results are improved, especially when dealing with complex images; secondly, the segmentation results are optimized by combining with medical atlases, making the segmentation more in line with clinical needs and enhancing the practical application value; thirdly, by introducing augmented reality technology and an interactive correction function, it is ensured that the segmentation results can be further optimized by doctors, increasing the reliability of the system; finally, the morphological operations and boundary smoothing processing of the post-processing module ensure the smoothness and integrity of the segmented images, ultimately achieving automatic segmentation of the lung injury area with high precision and having important clinical application prospects.

[0134] As mentioned above, the above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described.

Claims

1. An automatic segmentation system for lung injury regions based on deep learning, characterized in that, The system includes: M1, an image processing module, which receives the original lung images from a medical imaging device, performs denoising, edge detection, and region enhancement on the original lung images, and outputs the processed original lung images; M2, a segmentation and optimization module, which performs multi-layer convolution operations on the processed original lung images through a deep convolutional neural network, extracts low-level and high-level features, segments the processed original lung images based on the low-level and high-level features to generate a segmented image of the lung injury region, and optimizes the segmented image of the lung injury region using medical atlas knowledge, and outputs the optimized segmented image of the lung injury region; M3, a visualization and correction module, which overlays and displays the optimized segmented image of the lung injury region and the original lung images for comparison, presents the comparison result to a doctor for viewing through augmented reality technology, and the doctor performs interactive correction on the optimized segmented image of the lung injury region through a virtual interface, and outputs the corrected segmented image of the lung injury region; M4, a post-processing module, which performs morphological operations on the corrected segmented image of the lung injury region and smooths the boundaries of the segmented image through an interpolation algorithm, and outputs the final segmented image of the lung injury region.

2. The automatic segmentation system for lung injury regions based on deep learning according to claim 1, wherein The denoising is performed through bilateral filtering; the edge detection is implemented using the Canny edge detection algorithm, including gradient calculation, gradient magnitude, and edge direction, The formula for the gradient calculation is: Among them, G x is the gradient of the original lung image I in the horizontal direction; G y is the gradient of the original lung image I in the vertical direction; I(x, y) is the pixel value of the original lung image I at the position (x, y); The formula for the gradient magnitude is: where G is the gradient magnitude, that is, the obviousness of the edge; The formula for the gradient direction is: where θ is the gradient direction, indicating the direction of the edge of the original lung image.

3. The automatic segmentation system for lung injury regions based on deep learning according to claim 2, characterized in that, The region enhancement includes the calculation of local mean and standard deviation, and the formula is: where N(x, y) is the neighborhood region centered on the pixel (x, y); μ local (x, y) is the mean value of the region N(x, y), that is, the average value of all pixels in this region; σ local (x, y) is the standard deviation of the region N(x, y); I(x′, y′) is the pixel value of the pixel (x′, y′) in the neighborhood; According to the local mean and standard deviation, calculate the enhancement value C(x, y) of the pixels of the original lung image, and the formula is: where α is an enhancement coefficient, which controls the intensity of enhancement; According to the denoising, edge detection, and region enhancement, output the processed original lung image J.

4. The automatic segmentation system for lung injury regions based on deep learning according to claim 3, wherein, The operation steps of the segmentation and optimization module include feature extraction, image segmentation, and medical knowledge atlas optimization, The feature extraction uses a deep convolutional neural network to extract low-level and high-level features from the processed original lung image through multiple convolutional operations. The low-level features include edges, textures, and corners, and the high-level features include shapes, structures, and regions. The low-level feature y low is extracted by performing multiple convolutions on the processed original lung image using a convolutional kernel. The formula is: Among them, J(m,n) is the pixel value of the processed original lung image J at the position (m,n); K low (x - m, y - n) is the convolution kernel; Extract the high-level feature y through a deep convolutional network according to the low-level feature high , the formula is: Among them, y high (x, y) is the output of the high-level feature; The low-level features and high-level features form the output feature f.

5. The automatic segmentation system for lung injury regions based on deep learning according to claim 4, wherein The image segmentation is performed by classifying the output feature to determine whether it belongs to the lung injury region, and the formula is: Among them, W and b are the weights and biases of the deep convolutional network respectively; σ is the Sigmoid activation function, which is used to compress the output of the deep convolutional network to between [0,1], indicating that the feature belongs to the lung injury area, is the segmented image of the lung injury area; The medical knowledge atlas optimization uses the medical atlas to optimize the segmented image of the lung injury region, corrects unreasonable segmented regions, identifies incorrect regions in the segmented image of the lung injury region according to the medical atlas, and optimizes it by adjusting the segmentation boundary, and outputs the optimized segmented image Y of the lung injury region.

6. The automatic segmentation system for lung injury regions based on deep learning according to claim 5, characterized in that, The overlay display and comparison are realized through augmented reality technology, and the optimized segmented image Y of the lung injury region and the original lung image I are overlaid and displayed, and the formula is: I overlay (x,y) = α·I(x,y) + (1 - α)·Y(x,y) Among them, I overlay (x, y) is the overlapping image; α is the weighting coefficient that controls the contribution ratio of the original lung image and the optimized segmented image of the lung injury area. α ∈ [0, 1], and the display effect can be adjusted as needed.

7. An automatic segmentation system for lung injury regions based on deep learning according to claim 6, characterized in that, The interactive correction specifically means that the doctor selects the area to be corrected through the interactive interface, adjusts the segmented image of the lung injury area according to the doctor's operation, and the segmented image Y of the corrected lung injury area corrected The calculation formula of which is: Among them, R corrected is the coordinate set of the region manually corrected by the doctor, representing the updated boundary of the segmented image of the lung injury region.

8. The automatic segmentation system for lung injury regions based on deep learning according to claim 7, characterized in that, The morphological operations include dilation and erosion operations. The dilation operation expands the boundary of the corrected segmented image of the lung injury region, expands the foreground region, and fills small holes, The erosion operation shrinks the boundary of the foreground region of the corrected segmented image of the lung injury region and removes small noise regions. The smoothing process is achieved through a difference algorithm. The interpolation algorithm calculates a smooth curve by weighted averaging of boundary pixels to adjust the boundary. The formula is: Y smooth (x, y) = f(Y corrected (x, y)) Among them, Y smooth (x, y) is the segmented image after boundary smoothing; f is the smoothing algorithm, which is adjusted according to the boundary of the segmented image; Through the morphological operation and smoothing process, the final segmented image Y of the lung injury area is output smooth .

9. An automatic segmentation method for lung injury regions based on deep learning, characterized in that, This method includes the following steps: S1. Receive the original lung image from a medical imaging device, perform denoising, edge detection, and regional enhancement processing, and output the processed original lung image; S2. Perform multi-layer convolution operations on the processed original lung image through a deep convolutional neural network, extract low-level and high-level features, segment the processed original lung image based on the low-level and high-level features, and optimize the segmentation result using a medical atlas to generate an optimized segmented image of the lung injury area; S3. Overlay and display the optimized segmented image of the lung injury area with the original lung image. The doctor makes interactive corrections to the optimized segmented image of the lung injury area through a virtual interface and outputs the corrected segmented image of the lung injury area S4. Perform morphological repair and boundary smoothing on the corrected segmented image of the lung injury area to generate the final segmented image of the lung injury area.

Images