Chest CT image segmentation method and system based on deep learning

Abstract

The invention relates to a medical image processing technology, and discloses a chest CT image segmentation method and system based on deep learning. The method comprises the steps of obtaining a first image marked with organ contour information, dividing the first image into a plurality of first areas, setting first labels, and defining an image set composed of the first areas with the same first labels as a target data set; performing principal component analysis on each target data set to generate a feature space corresponding to each target data set; obtaining a second image to be segmented, extracting a second region from the second image, obtaining pixel information of the second region, selecting an optimal feature space from all the feature spaces, and obtaining a target contour and an organ type of a target organ in the second image; and establishing a verification model based on deep learning, determining whether the organ type of the target organ is correct based on the verification model, and if yes, segmenting the target organ based on the target contour. According to the invention, the segmentation efficiency of multiple organs in the chest CT image is improved.

Description

Technical Field

[0001] This application relates to the technical field of medical image processing, and particularly to a method and system for segmenting chest CT images based on deep learning. Background Art

[0002] In the field of medical image processing, especially in the analysis of chest CT (Computed Tomography) images, the accurate segmentation of target organs and lesions is of great significance for disease diagnosis and treatment. Traditional image segmentation methods mainly rely on manually designed features and threshold segmentation techniques. These methods have many limitations when dealing with complex medical images, such as being sensitive to noise, difficult to handle changes in the shape and size of target organs, and having insufficient generalization ability for images of different patients. With the development of deep learning technology, methods based on convolutional neural networks (CNNs) have gradually become the mainstream technology for medical image segmentation.

[0003] Existing methods for segmenting target organs using deep learning models, such as the Chinese patent application with publication number CN117808820A, propose a method for segmenting pulmonary embolism based on LSCU-net, which solves the problem that using only the U-net model to segment the lesion area cannot meet the requirements for segmentation accuracy, segmentation speed, segmentation accuracy, etc. due to the diversity of the shape of the lesion area and the differences caused by different target organ structures, and belongs to the field of deep learning semantic segmentation. The invention includes: proposing a method for segmenting pulmonary embolism based on LSCU-net. The invention improves the U-net by reasonably designing and adding a CBAM attention module and a Bi-LSTM bidirectional long short-term memory module, which can realize the fusion of different shape feature information and the extraction of sequence information between CTPA slices of pulmonary embolism.

[0004] However, this invention mainly focuses on the segmentation of a single target organ. For the simultaneous segmentation of multiple target organs, the complexity and computational cost of the model will increase significantly. Therefore, a method for segmenting chest CT images based on deep learning is needed, which can efficiently and accurately segment the contours of multiple target organs, while reducing the data annotation cost and improving the generalization ability of the model. Summary of the Invention

[0005] To solve the above technical problems, this application provides a method and system for segmenting chest CT images based on deep learning, which is used to improve the efficiency of chest CT image segmentation.

[0006] In the first aspect, this application provides a method for segmenting chest CT images based on deep learning, and the method includes:

[0007] Step S1: Obtain a chest CT image marked with contour information of multiple target organs, defined as the first image. Divide the first image into multiple first regions based on the contour information, and set a first label for each first region to label the organ type corresponding to the first region. Define the image set composed of the first regions with the same first label as the target data set;

[0008] Step S2: Obtain the first pixel information and contour information of the target organs in each target data set. Perform principal component analysis on each target data set based on the first pixel information and the contour information to generate a feature space corresponding to each target data set;

[0009] Step S3: Obtain the chest CT image to be segmented, defined as the second image. Extract the rectangular region of the target organ from the second image, defined as the second region. Obtain the second pixel information of the second region. Select the optimal feature space from all the feature spaces based on the second pixel information to obtain the target contour and organ type of the target organ in the second image;

[0010] Step S4: Establish a verification model based on deep learning. Judge whether the organ type of the target organ in the second image is correct based on the verification model. If so, output the organ type of the target organ, and segment the target organ based on the target contour.

[0011] Combined with the first aspect, in the first implementation manner of the first aspect of the present application, performing principal component analysis on each target data set based on the first pixel information and contour information includes:

[0012] Arrange the first pixel information in the first region in a preset order to generate a first vector. Take multiple feature points on the contour information, and combine the position coordinates of all the feature points to generate a second vector. Fuse the first vector and the second vector to generate a third vector. Perform principal component analysis on the third vector to obtain the feature vector and the corresponding eigenvalue corresponding to each target data set. Calculate the cumulative explained variance of each feature vector, and define the first M feature vectors when the cumulative explained variance is greater than the first threshold as the main feature vectors. Construct and generate the corresponding feature space based on the main feature vectors and the mean vectors of each target organ. Each feature space matches the first label of the corresponding target organ.

[0013] Combined with the first aspect, in the second implementation manner of the first aspect of the present application, extracting the rectangular region of the target organ from the second image includes:

[0014] Generate a pixel value histogram of the second image based on the second pixel information of the second image, identify multiple peak points in the pixel value histogram, obtain the position coordinates of the corresponding pixels at each peak point in the second image, set a rectangular range to expand outward with the position coordinates as the central nodes, and simultaneously identify the edge of the target organ based on an edge detection algorithm until all pixels of the target organ are within the rectangular range. The formed rectangular range is defined as the second region of the target organ.

[0015] Combined with the first aspect, in the third implementation manner of the first aspect of the present application, selecting the optimal feature space from all the feature spaces based on the second pixel information includes:

[0016] Convert the second pixel information of each second region in the second image into a fourth vector, and project it into all the feature spaces respectively to obtain the projected first data. The first data is a combination of multiple feature vectors of the target organ in the second region, and back-project the first data into the original image space to obtain the second data. Calculate the mean square error between the fourth vector and each second data, and define the feature space with the mean square error less than the second threshold as the optimal feature space of the second region.

[0017] Combined with the first aspect, in the fourth implementation manner of the first aspect of the present application, obtaining the target contour and organ type of the target organ in the second image includes:

[0018] Take the first label corresponding to the optimal feature space as the organ type of the target organ in the second image. Define the vector composed of multiple feature points in the target contour of the target organ in the second image as the fifth vector. Perform matrix calculation using the optimal feature space based on the first formula to obtain the original feature vector b of the target organ in the second image. The first formula is: b = R(R T BR) -1 R T b', where b includes a combination of the fourth vector representing the second pixel information and the fifth vector representing the target contour. R is the vector matrix generated by all the main feature vectors in the optimal feature space. B is the diagonal matrix formed when the fourth vector and the fifth vector are 0. b' is the input vector composed of the fourth vector and the fifth vector when they are 0. Extract the fifth vector from the original feature vector, and obtain the position information of the feature points that make up the target contour of the target organ based on the fifth vector.

[0019] Combined with the first aspect, in the fifth implementation manner of the first aspect of the present application, determining whether the organ type of the target organ in the second image is correct based on the verification model includes:

[0020] The verification model is a deep convolutional neural network. The third image containing multiple target organs is input into the verification model for training to obtain the organ type and position information of each target organ. The second image is input into the trained verification model to obtain the organ type of each target organ in the second image, and it is judged whether it is the same as the organ type obtained from the optimal feature space. If so, it is determined that the organ type identified by the optimal feature space is correct, and the verification model outputs the organ type of the target organ. Otherwise, repeat step S3 to continue searching for the optimal feature space until the generated organ type is the same as the organ type output by the verification model.

[0021] In a second aspect, the present application provides a chest CT image segmentation system based on deep learning. The system includes:

[0022] An acquisition module, configured to acquire a chest CT image marked with contour information of multiple target organs, defined as a first image. Based on the contour information, the first image is divided into multiple first regions, and a first label is set for each first region to label the organ type corresponding to the first region. An image set composed of first regions with the same first label is defined as a target data set;

[0023] An analysis module, configured to acquire the first pixel information and contour information of the target organs in each target data set, and perform principal component analysis on each target data set based on the first pixel information and contour information to generate a feature space corresponding to each target data set;

[0024] A contour generation module, configured to acquire a chest CT image to be segmented, defined as a second image, extract a rectangular region of the target organ from the second image, defined as a second region, acquire the second pixel information of the second region, and select an optimal feature space from all the feature spaces based on the second pixel information to obtain the target contour and organ type of the target organ in the second image;

[0025] A verification module, configured to establish a verification model based on deep learning, and judge whether the organ type of the target organ in the second image is correct based on the verification model. If so, output the organ type of the target organ, and segment the target organ based on the target contour.

[0026] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0027] In the technical solution provided by this application, by annotating the contour information of multiple target organs, the image is divided into multiple first regions and tags are set, and the regions with the same tag are combined into a target data set, which can generate a dedicated feature space for each organ to better capture the unique features of each organ; generating an independent feature space for each target data set can provide a unique feature representation for each organ, and this feature space can be used for subsequent image segmentation and organ recognition tasks. In multi-organ segmentation tasks, the generation of the feature space can help the model better understand the unique features of each organ, thereby achieving more accurate segmentation.

[0028] By extracting the rectangular region of the target organ from the second image, the position of the target organ can be quickly located; by matching the pixel information of the second region with all feature spaces, the most suitable feature space can be selected, which can ensure the highest accuracy of subsequent organ contour extraction and type recognition. Through feature matching, the contour and type of the target organ can be initially identified. This initial identification method can provide important reference information for subsequent deep learning verification. By establishing a verification model through deep learning, the initially identified organ type can be further verified; the deep learning model has strong feature learning ability and generalization ability, which can significantly improve the accuracy of organ type recognition. After confirming the organ type, organ segmentation based on the target contour can achieve high-precision organ segmentation. This segmentation method can provide clear organ images for subsequent medical analysis and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0030] Figure 1 It is a schematic diagram of an embodiment of the method for segmenting chest CT images based on deep learning in the embodiments of this application;

[0031] Figure 2 It is a schematic diagram of feature points in the embodiments of this application;

[0032] Figure 3 It is a flow chart for selecting the optimal feature space for each second region in the second image in the embodiments of this application;

[0033] Figure 4 It is a schematic diagram of an embodiment of the system for segmenting chest CT images based on deep learning in the embodiments of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The embodiments of the present application provide a method and system for chest CT image segmentation based on deep learning. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and the above-mentioned drawings of the present application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order different from that illustrated or described here. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily limit to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0035] For ease of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1 , an embodiment of the method for chest CT image segmentation based on deep learning in the embodiments of the present application includes:

[0036] Step S1: Obtain a chest CT image labeled with contour information of multiple target organs, defined as the first image. Divide the first image into multiple first regions based on the contour information, and set a first label for each first region to label the organ type corresponding to the first region. Define the image set composed of the first regions with the same first label as the target data set.

[0037] Specifically, obtain a chest CT image from an existing medical image database. The chest CT image is a 2D image or 3D image including one or more target organs. The target organs are, for example, the heart, left lung, right lung, trachea, etc. Use a professional medical image annotation tool to draw the contour of each target organ on each slice image using a polygon tool.

[0038] Assign a unique first label to each first region to identify the organ type. To ensure the integrity of each region's pixels, verify the divided first regions to ensure that there are no overlapping pixels between different regions and that all target organs are correctly divided and labeled. Combine the first regions with the same first label to generate a target data set. For example, the first data set: the heart region with the first label of 1.

[0039] Step S2: Obtain the first pixel information and contour information of the target organs in each target data set. Perform principal component analysis on each target data set based on the first pixel information and contour information to generate a feature space corresponding to each target data set.

[0040] Specifically, the original pixel values within the first region are extracted. The contour information is usually represented as a series of continuous pixel point position coordinates. The first pixel information and the contour information of each target organ are linearly combined to generate a data vector, and principal component analysis is performed on the data vector, which will be elaborated later.

[0041] The original features of the target organ in each target dataset, such as the centroid coordinates of the target organ, the average distance from the centroid to the contour, the distance variance, geometric features (the area, perimeter, circularity, eccentricity, and major axis direction of the target organ), and the shape details of the target organ contour (Fourier coefficients), etc. The main features are extracted from the above-mentioned extracted features through PCA analysis. The main features are linear combinations of the original features and can retain the variance in the data to the greatest extent, corresponding to the important change directions in the data. A new multi-dimensional feature space is constructed by selecting the first k main features, and the distribution and clustering of the data are visually observed in the new feature space.

[0042] Step S3: Obtain the chest CT image to be segmented, defined as the second image. Extract the rectangular region of the target organ from the second image, defined as the second region. Obtain the second pixel information of the second region. Based on the second pixel information, select the optimal feature space from all the feature spaces to obtain the target contour and the organ type of the target organ in the second image.

[0043] Specifically, the target organ in the second image is initially located to obtain the minimum bounding rectangle containing all the pixel information of the target organ, defined as the second region. The gray values of each pixel in the second region are extracted to form a pixel value matrix and input into all the feature spaces. The second pixel information in each second region is matched with the constructed feature space library containing the features of different organ types and different morphologies. The feature space with the highest similarity to the second pixel information is selected as the optimal feature space. The contour information in the optimal feature space is used, combined with the pixel information of the second region, to obtain the accurate contour information, and the specific method will be elaborated later.

[0044] Step S4: Establish a verification model based on deep learning. Based on the verification model, determine whether the organ type of the target organ in the second image is correct. If so, output the organ type of the target organ and segment the target organ based on the target contour.

[0045] Specifically, first, use deep learning techniques (such as deep convolutional neural network, CNN) to establish a verification model. This verification model can predict the position information of different organs. The second image to be processed is input into the trained CNN verification model, and the predicted position information of each target organ is output. The organ type corresponding to the predicted position information is queried in the database.

[0046] Compare the organ type obtained from the database with the organ type obtained by the optimal feature space method. If the two are the same, it is determined that the organ type identified by the optimal feature space is correct, and the CNN verification model outputs the organ type of the target organ. Through the deep learning model, complex features and patterns can be learned from a large amount of data, thereby improving the accuracy of organ type recognition. However, if the deep learning model is used for segmentation, usually a large amount of training data is required and it is easily affected by noise. The present invention uses the method of selecting the optimal feature space to simultaneously identify the contours of multiple target organs, relies on less training data and has high stability, is not easily affected by data noise and outliers, uses the characteristics of the feature space to make up for the data requirements of deep learning, and at the same time combines the deep model for further verification to obtain a more convincing result.

[0047] In a specific embodiment, the principal component analysis of each target data set based on the first pixel information and the contour information specifically includes the following steps:

[0048] Arrange the first pixel information in the first region in a preset order to generate a first vector. Take multiple feature points on the contour information, and combine the position coordinates of all feature points to generate a second vector. Fuse the first vector and the second vector to generate a third vector. Perform principal component analysis on the third vector to obtain the feature vector and the corresponding eigenvalue corresponding to each target data set. Calculate the cumulative explained variance of each feature vector, and define the first M feature vectors when the cumulative explained variance is greater than the first threshold as the main feature vectors. Construct and generate the corresponding feature space based on the main feature vectors and the mean vectors of each target organ, and each feature space matches the first label of the corresponding target organ.

[0049] Specifically, for example, the first region is the left atrium region of the heart. Extract the gray values of all pixels from the left atrium region, and arrange the pixel information in a preset order, such as the raster scan order, to form a first vector, denoted as the first vector v1 = [p 1,1 , p 1,2 , p 1,3 , p 1,4 ,..., p m,n , where p m,n is the pixel value of the i-th row and j-th column. Use the equidistant sampling method to select multiple feature points on the contour line, extract the position coordinates (x, y) of each feature point, and combine the position coordinates of all feature points into a vector, denoted as the second vector v2. As Figure 2 shown, it is a schematic diagram of feature points. Assume that 16 feature points are selected, that is, p1 - p16, then v2 = [x 1 , y 1 , x 2 , y 2 ,..., xk , y k , concatenate the first vector v1 and the second vector v2 to generate a longer vector, denoted as the third vector v3 = [v 1 , v 2 . Combine the third vectors of all target data sets into a matrix, where each row corresponds to the third vector of a sample, and centralize the data matrix, that is, subtract the mean of each feature.

[0050] According to the formula: Calculate the covariance matrix of the centralized data matrix, perform eigenvalue decomposition on the covariance matrix C to obtain eigenvectors and corresponding eigenvalues. The eigenvalues represent the variance contribution of each principal component, and the eigenvectors represent the change directions of the principal components. Arrange the eigenvalues in descending order, and the corresponding eigenvectors are also sorted accordingly. According to the formula: Calculate the cumulative explained variance q of the first M eigenvalues M , and select the first M eigenvectors when the cumulative explained variance is greater than the first threshold as the main features.

[0051] The mean vector represents the center point of all sample data in the original feature space. It is the average of all sample third vectors and reflects the overall trend of the data. The main eigenvectors are obtained by performing eigenvalue decomposition on the covariance matrix and represent the directions with the largest variance in the data, that is, the main directions of data change. By projecting the original data onto the main eigenvectors, a low-dimensional representation can be obtained to achieve dimensionality reduction.

[0052] In a specific embodiment, extracting the rectangular region of the target organ from the second image specifically includes the following steps:

[0053] Generate a pixel value histogram of the second image based on the second pixel information of the second image, identify multiple peak points in the pixel value histogram, obtain the position coordinates of the pixels corresponding to each peak point in the second image, set a rectangular range to expand outward with the position coordinates as the central nodes, and at the same time identify the edge of the target organ based on the edge detection algorithm until all pixels of the target organ are within the rectangular range. The formed rectangular range is defined as the second region of the target organ.

[0054] Specifically, a peak detection algorithm (such as finding local maxima) is used to identify multiple peak points in the pixel value histogram. The peak points are the pixel value positions of local maxima in the histogram, which usually correspond to the points in the image where the pixel values are significantly different from the surrounding areas. This is often a typical feature of organs or lesions in medical images. By identifying these peak points, the area that may contain the target organ can be quickly located. Each peak point corresponds to a pixel position coordinate. Taking the pixel position coordinate where each peak point is located as the central node, a rectangular range is set, and a certain pixel distance (such as 5 pixels) is extended outward from the central node. An edge detection algorithm (such as Canny edge detection, Sobel operator) is used to identify the edge of the target organ. According to the edge detection result, the rectangular range is adjusted to include all the pixels of the target organ, and the formed rectangular range is defined as the second area of the target organ.

[0055] In a specific embodiment, the selection of the optimal feature space from all feature spaces based on the second pixel information specifically includes the following steps:

[0056] The second pixel information of each second area in the second image is converted into a fourth vector and projected into all feature spaces respectively to obtain the projected first data. The first data is a combination of multiple feature vectors of the target organ in the second area, and the first data is back-projected into the original image space to obtain the second data. The mean square error between the fourth vector and each second data is calculated, and the feature space with the mean square error less than the second threshold is defined as the optimal feature space of the second area.

[0057] Specifically, the second image to be segmented is obtained from a medical image database. The second image is an image with unknown contour and unknown type of the target organ, such as Figure 3As shown in the figure, it is a flowchart for selecting the optimal feature space for each second region in the second image. The pixel values of all pixels are extracted from each second region, and the extracted pixel information is organized into a fourth vector. Assuming that each second region has N pixels, the second pixel information can be represented as an N-dimensional fourth vector. All pre-constructed feature spaces are obtained. These feature spaces are obtained by performing principal component analysis (PCA) on different datasets. Each feature space corresponds to a dataset. The fourth vector of each second region is projected into all feature spaces respectively. The projection vectors of each second region in all feature spaces are combined to form first data. The first data is back-projected into the original image space to obtain second data. The second data is the data information reconstructed using each feature space. Calculate the mean square error between the fourth vector and each second data. This mean square error refers to the reconstruction error between the data reconstructed based on the feature space and the original data. The smaller the reconstruction error, the better the feature space can represent the pixel value information of the second image, that is, the higher the similarity between the feature space and the second image. Select the feature space with a reconstruction error less than the second threshold as the optimal feature space for the second region. Each second region in the second image corresponds to an optimal feature space, but the second image may correspond to multiple optimal feature spaces because an image may contain multiple target organs.

[0058] In a specific embodiment, obtaining the target contour and organ type of the target organ in the second image specifically includes the following steps:

[0059] Taking the first label corresponding to the optimal feature space as the organ type of the target organ in the second image, defining the vector composed of multiple feature points in the target contour of the target organ in the second image as the fifth vector, and performing matrix calculation using the optimal feature space based on the first formula to obtain the original feature vector b of the target organ in the second image. The first formula is: b = R(R T BR) -1 R T b', where b includes the combination of the fourth vector representing the second pixel information and the fifth vector representing the target contour. R is the vector matrix generated by all the main feature vectors in the optimal feature space. B is the diagonal matrix formed when the fourth vector and the fifth vector are 0. b' is the input vector composed of the fourth vector and the fifth vector when they are 0. Extract the fifth vector from the original feature vector, and obtain the position information of the feature points constituting the target contour of the target organ based on the fifth vector.

[0060] Specifically, the present invention estimates the contour boundary of the target organ in the second region based on a boundary prediction method of linear programming. Its principle can be divided into the following steps: First, pixel information is extracted from the input image to generate feature points or key points. These feature points are usually located at the edges or corners of the object and can effectively represent the appearance of the object. The extracted feature points are matched with the pre-acquired feature space to determine the appearance information of the object. The matching process utilizes the geometric relationship and appearance features between the feature points. By using an iterative optimization method (such as gradient descent or projection-based method), the pose parameters of the target are adjusted in the feature space to make the target features match the features of the input image best, that is, to select the best feature space and efficiently estimate the pose information of the object. During the optimization process, the contour position of the object in the image is inferred by combining the position and pose information of the target features. This process only depends on the pixel information of the input image, so no additional external data is required.

[0061] The above method is described below in combination with the first formula. The gray values of all pixels are extracted from the second image and denoted as the fourth vector V4. The vector matrix generated by all the main feature vectors in the optimal feature space is denoted as R. Where the dimension of B is the sum of the pixel value information dimension (assumed to be N) and the contour information dimension (assumed to be Q). Since the initial target contour of the second region is unknown, the last Q-dimensional data on the diagonal of the B matrix is set to 0. b' = {v 4 , 0, 0, 0, 0, …, 0, 0}. The target contour of the second region in the input vector is also unknown, so the Q-dimensional data part in b' is set to 0.

[0062] The meaning of the first formula is as follows: The original data b' is projected into the optimal feature space R, and the projected vector is R T b'. The purpose of this step is to convert the high-dimensional data into a low-dimensional representation. (R T BR) -1 This step is for the back-projection operation, and finally the original feature vector b is obtained. The fifth vector is separated from the original combined vector b, the fifth vector is analyzed, the coordinate information is extracted, and a point set representing the target contour is generated. This method is based on a clear mathematical model and optimization process, and the result has high interpretability. Each step and decision can be traced back to specific formulas and constraints, and its result has high stability and is not easily affected by data noise and outliers.

[0063] In a specific embodiment, determining whether the organ type of the target organ in the second image is correct based on the verification model specifically includes the following steps:

[0064] The verification model is a deep convolutional neural network. The third image containing multiple target organs is input into the verification model for training to obtain the organ type and location information of each target organ. The second image is input into the trained verification model to obtain the organ type of each target organ in the second image, and it is determined whether it is the same as the organ type obtained from the optimal feature space. If so, it is determined that the organ type identified by the optimal feature space is correct, and the verification model outputs the organ type of the target organ. Otherwise, repeat step S3 and continue to search for the optimal feature space until the generated organ type is the same as the organ type output by the verification model.

[0065] Specifically, a third image containing multiple target organs is obtained from a medical image database or a PACS system. The third image is input into the CNN for forward propagation to obtain a prediction result. The second image to be processed is input into the verification model to obtain the organ type of each target organ in the second image. The obtained organ type is compared with the organ type obtained through the optimal feature space. If the two are the same, it is determined that the organ type identified by the optimal feature space is correct, and the CNN verification model outputs the organ type of the target organ. If the two are different, repeat step S3 and continue to search for the optimal feature space until the requirements are met and the organ type is confirmed by the verification model, which can reduce segmentation errors.

[0066] If the optimal feature space misidentifies the organ type, the verification model can identify this error, thereby improving the accuracy of segmentation. Finally, the target contour obtained through the optimal feature space is used for the segmentation of the target organ. This segmentation method usually relies on less training data because the model is constructed based on explicit mathematical and statistical methods rather than relying on a large amount of data to learn features. Its segmentation results usually have high stability and are not easily affected by data noise and outliers. At the same time, the verification method integrating deep learning greatly enhances the segmentation efficiency of the present invention, and can also identify multiple target organs for synchronous segmentation.

[0067] The above describes a method for segmenting chest CT images based on deep learning in an embodiment of the present application. Next, a system for segmenting chest CT images based on deep learning in an embodiment of the present application will be described. Please refer to Figure 4 An embodiment of a system for segmenting chest CT images based on deep learning in an embodiment of the present application includes:

[0068] An acquisition module, configured to acquire a chest CT image labeled with contour information of multiple target organs, defined as the first image. Based on the contour information, the first image is divided into multiple first regions, and a first label is set for each first region to label the organ type corresponding to the first region. An image set composed of first regions with the same first label is defined as the target data set;

[0069] An analysis module, configured to obtain first pixel information and contour information of a target organ in each target dataset, perform principal component analysis on each target dataset based on the first pixel information and the contour information, and generate a feature space corresponding to each target dataset;

[0070] A contour generation module, configured to obtain a chest CT image to be segmented, defined as a second image, extract a rectangular region of the target organ from the second image, defined as a second region, obtain second pixel information of the second region, and select an optimal feature space from all the feature spaces based on the second pixel information, to obtain a target contour and an organ type of the target organ in the second image;

[0071] A verification module, configured to establish a verification model based on deep learning, determine whether the organ type of the target organ in the second image is correct based on the verification model, and if so, output the organ type of the target organ and segment the target organ based on the target contour.

[0072] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0073] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc that can store program codes.

[0074] The above is the case. The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A chest CT image segmentation method based on deep learning, characterized in that: The method comprises: Step S1: obtaining a chest CT image annotated with contour information of a plurality of target organs, defining the image as a first image, dividing the first image into a plurality of first regions based on the contour information, and setting a first label for each first region to annotate the organ type corresponding to the first region, and defining an image set consisting of first regions having the same first label as a target data set; Step S2: acquiring first pixel information and contour information of the target organ in each target data set, performing principal component analysis on each target data set based on the first pixel information and the contour information, and generating a feature space corresponding to each target data set; Step S3: obtaining a chest CT image to be segmented, which is defined as a second image; extracting a rectangular region of a target organ from the second image, which is defined as a second region; obtaining second pixel information of the second region; selecting an optimal feature space from all the feature spaces based on the second pixel information; and obtaining a target contour and an organ type of the target organ in the second image; Step S4: Establish a verification model based on deep learning, and determine whether the organ type of the target organ in the second image is correct based on the verification model. If so, output the organ type of the target organ and segment the target organ based on the target contour.

2. The method according to claim 1, characterized in that Performing principal component analysis on each target data set based on the first pixel information and contour information comprises the following steps: The first pixel information in the first area is arranged in a preset order to generate a first vector, a plurality of feature points are taken on the contour information, and the position coordinates of all the feature points are combined to generate a second vector, the first vector and the second vector are fused to generate a third vector, principal component analysis is performed on the third vector, the eigenvector and the corresponding eigenvalue corresponding to each target data set are obtained, the cumulative explained variance of each eigenvector is calculated, and the first M eigenvectors when the cumulative explained variance is greater than a first threshold are defined as main eigenvectors, and a corresponding feature space is constructed based on the main eigenvector and the mean vector of each target organ, and each feature space matches the first label of the corresponding target organ.

3. The method according to claim 1, characterized in that Extracting a rectangular region of the target organ from the second image comprises the following steps: A pixel value histogram of the second image is generated based on second pixel information of the second image, a plurality of peak points in the pixel value histogram are identified, the position coordinates of the pixels corresponding to each peak point in the second image are obtained, a rectangular range is set to expand outward with the position coordinates as the center node, and the edge of the target organ is identified based on an edge detection algorithm at the same time, until all pixels of the target organ are within the rectangular range, and the formed rectangular range is defined as the second region of the target organ.

4. The method according to claim 3, characterized in that Selecting the optimal feature space from all the feature spaces based on the second pixel information comprises the following steps: The second pixel information of each second area in the second image is converted into a fourth vector, and is projected into all the feature spaces respectively to obtain the projected first data, where the first data is a combination of multiple feature vectors of the target organ in the second area, and the first data is back-projected into the original image space to obtain the second data, and the mean square error between the fourth vector and each second data is calculated, and the feature space whose mean square error is less than a second threshold is defined as the optimal feature space of the second area.

5. The method according to claim 4, characterized in that Obtaining the target contour and organ type of the target organ in the second image comprises the following steps: The first label corresponding to the optimal feature space is used as the organ type of the target organ in the second image, and a vector composed of multiple feature points in the target contour of the target organ in the second image is defined as a fifth vector. The optimal feature space is used to perform matrix calculation based on the first formula to obtain the original feature vector b of the target organ in the second image. The first formula is: b=R(R T BR) -1 R T b', wherein b comprises a combination of a fourth vector representing the second pixel information and a fifth vector representing the target contour, R is a vector matrix generated by all main eigenvectors in the optimal feature space, B is a diagonal matrix formed when the fourth vector and the fifth vector are 0, b' is an input vector composed when the fourth vector and the fifth vector are 0, the fifth vector is extracted from the original eigenvector, and the position information of the feature points constituting the target contour of the target organ is obtained based on the fifth vector.

6. The method according to claim 1, characterized in that Judging whether the organ type of the target organ in the second image is correct based on the verification model comprises the following steps: The verification model is a deep convolutional neural network. A third image containing multiple target organs is input into the verification model for training to obtain the organ type and position information of each target organ. The second image is input into the trained verification model to obtain the organ type of each target organ in the second image, and it is determined whether the organ type is the same as the organ type obtained by the optimal feature space. If so, it is determined that the organ type identified by the optimal feature space is correct, and the verification model outputs the organ type of the target organ. Otherwise, step S3 is repeated to continue searching for the optimal feature space until the generated organ type is the same as the organ type output by the verification model.

7. A deep learning-based chest CT image segmentation system, used to implement the deep learning-based chest CT image segmentation method according to any one of claims 1 to 6, characterized in that: The system comprises: an acquisition module, configured to acquire a chest CT image annotated with contour information of a plurality of target organs, defined as a first image, divide the first image into a plurality of first regions based on the contour information, and set a first label for each first region to annotate the organ type corresponding to the first region, and define an image set consisting of first regions with the same first label as a target data set; An analysis module, used for obtaining first pixel information and contour information of the target organ in each target data set, performing principal component analysis on each target data set based on the first pixel information and the contour information, and generating a feature space corresponding to each target data set; A contour generation module is used to obtain a chest CT image to be segmented, which is defined as a second image, extract a rectangular area of ​​a target organ from the second image, which is defined as a second area, obtain second pixel information of the second area, select an optimal feature space from all the feature spaces based on the second pixel information, and obtain a target contour and an organ type of the target organ in the second image; A verification module is used to establish a verification model based on deep learning, and to determine whether the organ type of the target organ in the second image is correct based on the verification model. If so, the organ type of the target organ is output, and the target organ is segmented based on the target contour.

Images