Unsupervised liver tumor CT image segmentation method based on manual features

Abstract

The invention proposes an unsupervised liver tumor CT image segmentation method based on manual features, and relates to the technical field of deep learning, and the method comprises the steps: firstly collecting a normal liver CT image as a data set, and selecting a tumor position in combination with clinical knowledge; a texture similar to a real image is generated through a three-dimensional simple noise binary mask, and morphological modeling is carried out by using histological characteristics. And then, superposing the tumor texture and the liver CT image at a selected position to synthesize a new liver tumor CT image. And training a segmentation model by using the CT image with the annotation, and performing medical image segmentation by using the synthesized image. And finally, evaluating a segmentation result by adopting a Dice coefficient, a Hausdorff distance, a standardized surface distance and a surface distance index. The method is helpful for artificially synthesizing tumor focus images and generating voxel-level annotations, and provides a new perspective for solving medical image annotation challenges and promoting liver tumor diagnosis and treatment.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an unsupervised liver tumor CT image segmentation method based on manual features. Background Art

[0002] Three-dimensional medical imaging technology can accurately visualize and quantitatively analyze internal structures, thus revolutionizing clinical practice. Accurate segmentation of three-dimensional images is essential for early detection and effective diagnosis of liver tumors. Tumor segmentation has become a basic task in medical image analysis, which is of great value for accurate diagnosis of lesions and treatment decisions, and has always been the focus of researchers and clinicians.

[0003] However, existing tumor CT image segmentation methods have the following problems: first, the volumetric characteristics of three-dimensional imaging make manual segmentation costly and subjective; second, conventional medical image segmentation methods usually require a large amount of manually segmented training data, while the scale of well-annotated tumor data is limited; third, the complexity and morphological differences of tumors make tumor image segmentation more difficult.

[0004] Therefore, it is urgent to provide a solution to improve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide an unsupervised liver tumor CT image segmentation method based on manual features to improve the problems of the above-mentioned background technology.

[0006] The present invention provides an unsupervised liver tumor CT image segmentation method based on manual features, which adopts the following technical solutions:

[0007] CT scan images of normal liver are collected as a data set, and the potential tumor location of the liver is verified by volume and gradient to determine the tumor location of the liver;

[0008] Acquiring a final tumor texture based on the tumor position of the liver, and performing morphological modeling of the liver tumor based on the final tumor texture to synthesize a new liver tumor CT image;

[0009] Annotated CT scan images are used to train the segmentation model, synthetic tumor images are used as datasets for medical image segmentation, and the segmentation performance is verified based on evaluation metrics.

[0010] The beneficial effect of the unsupervised liver tumor CT image segmentation method based on manual features provided by the present invention is that the present invention first collects normal liver CT images as a data set, and selects the tumor location in combination with clinical knowledge. A texture similar to real imaging is generated by a three-dimensional minimalist noise binary mask, and morphological modeling is performed using histological features. Subsequently, the tumor texture is superimposed on the liver CT image at a selected position to synthesize a new liver tumor CT image. The segmentation model is trained using annotated CT images, and the synthesized images are used for medical image segmentation. Finally, the Dice coefficient, Hausdorff distance, standardized surface distance and surface distance indicators are used to evaluate the segmentation results. The present invention helps to artificially synthesize tumor lesion images and generate voxel-level annotations, providing a new perspective for solving the challenges of medical image annotation and promoting the diagnosis and treatment of liver tumors.

[0011] Optional, volumetric and gradient verification of potential tumor locations in the liver. The process of determining the tumor location in the liver includes:

[0012] Generating spherical coordinates within the liver Convert spherical coordinates to Cartesian coordinates x=(x, y, z), and convert Cartesian coordinates and spherical coordinates based on the ellipse axis and half-price;

[0013] Iterate and verify each potential tumor location x to determine whether the point is located in the liver until the tumor location x is determined. t =(x t ,y t ,z t ) to generate a tumor position of the liver, wherein the tumor position is located inside the liver and has a sufficient distance from the liver boundary.

[0014] Optional, a procedure to generate the tumor location in the liver, including:

[0015] Obtaining the number of liver masks in a neighborhood volume around the predicted position based on the neighborhood volume inspection, and obtaining a first tumor position based on the number of liver masks and a first threshold;

[0016] A gradient value of the scan volume at the predicted position is calculated based on the gradient amplitude evaluation, and a second tumor position is acquired based on the gradient value and a second threshold, and a tumor position of the liver is generated based on the first tumor position and the second tumor position.

[0017] Optionally, the unsupervised liver tumor CT image segmentation method based on manual features is characterized in that the process of obtaining the final tumor texture includes:

[0018] Acquire three-dimensional minimalist noise based on predefined parameters, and use nonlinear diffusion and wavelet transform to filter artifacts and retain high-frequency details of the noise respectively to acquire a texture to be processed;

[0019] Normalizing the texture to be processed to obtain a normalized texture, and performing smoothing on the normalized texture based on a first Gaussian filter to obtain a smoothed texture;

[0020] Proportionally scaling and moving the smoothed texture to obtain a first texture;

[0021] A first binary mask is obtained based on random sampling, the first binary mask is smoothed based on a second Gaussian filter to obtain a second binary mask, and the first texture is scaled and cropped based on the second binary mask to generate a final tumor texture.

[0022] Optionally, a procedure for morphological modeling of liver tumors may be performed, including:

[0023] The synthetic tumors were classified into four size categories: tiny, small, medium, and large, and the number of synthetic samples was drawn from a discrete uniform distribution;

[0024] The tumor shape is simulated based on an ellipsoid defined by parameters, where the semi-major axes of the ellipsoid in the x, y, and z coordinates are taken from a uniform distribution U(0.75r, 1.25r) and the coordinates x in the scan volume f(x) are t =(x t ,y t ,z t ) is the center, and the initial tumor mask is defined within this range;

[0025] The initial tumor mask is subjected to nonlinear transformation processing based on nonlinear elastic deformation to obtain a deformed tumor mask:

[0026] The deformed tumor mask is convolved based on a third Gaussian filter to obtain a synthetic mask.

[0027] Optionally, the process of synthesizing a new liver tumor CT image includes:

[0028] Quantify the density of the liver: divide the liver into three density intervals and standardize the intensity values ​​within each interval;

[0029] Adjust tumor density and reduce the density of high-density tissue;

[0030] The tumor is integrated into the scan volume f(x) and the corresponding liver mask l(x).

[0031] Optionally, the process of quantifying the density of the liver comprises:

[0032] Get the lower limit L and upper limit U of the liver region P and calculate the density scale Δ:

[0033] Δ=(UL) / 3;

[0034] For each three-dimensional coordinate point P within the liver region ijk perform the following operations:

[0035] If P ijk < L + Δ, let P ijk = L; if L + Δ ≤ P ijk < L + 2Δ, then let P ijk = L + Δ; if L + Δ ≤ P ijk < U, then let P ijk = L + 2Δ; if P ijk ≥ U, then let P ijk = U;

[0036] Obtain the density distribution map D according to the above operations. For each D on the density distribution map D ijk perform processing: If D ijk = U and P ijk ≠ U, then let D ijk = L + 2Δ.

[0037] Optionally, the process of adjusting the tumor density includes:

[0038] Obtain the lower limit L and upper limit U of the liver region P, and calculate the density scale Δ:

[0039] Δ = (U - L) / 3;

[0040] For each coordinate point F in the scanning volume domain F ijk perform the following processing:

[0041] When the coordinate point F ijk coincides with the generated tumor position, if D ijk = U, then let If D ijk = L + 2Δ then let Otherwise, retain F ijk ;

[0042] where D ijk represents the coordinate point on the density distribution map D, U represents the upper limit of the liver region, and L represents the lower limit of the liver region.

[0043] Optionally, the process of integrating the tumor into the scanning volume domain f(x) and the corresponding liver mask l(x) includes:

[0044] Integrate the tumor with the scanning volume:

[0045] f′(x) = [1 - t″(x)] ⊙ f(x) + t″(x) ⊙ T″(x);

[0046] Wherein, ⊙ represents element-wise multiplication, t″(x) represents the tumor mask array, and T″(x) represents the tumor texture array. Through the above transformation, a new scan volume f′(x) can be obtained, and a new liver mask containing the tumor can be synthesized in the manner of l′(x)=l(x)+t″(x);

[0047] Determine the border area of ​​the tumor:

[0048]

[0049] where λ l represents the lower intensity threshold for defining the tumor boundary area, λ u represents the upper intensity threshold that defines the tumor boundary region;

[0050] Use the Gaussian filter G(x; σ3) to convolve with b(x) to generate a smooth boundary image b′(x):

[0051] b′(x)=b(x)*G(x;σ3);

[0052] Among them, σ3 determines the standard deviation of the Gaussian kernel and controls the smoothness of the boundary transition;

[0053] The boundary map b′(x) enhances the voxel intensity at the periphery of the tumor by adding a fixed intensity offset δ, which simulates the increase of Hounsfield units (HU) in the cyst:

[0054] f″′(x)=f″(x)+δ·b′(x);

[0055] where δ represents the HU intensity difference between the tumor core and the fibrous capsule, and f″′(x) represents the final volume scan result.

[0056] Optionally, when the synthesized tumor image is used as a data set for medical image segmentation and the segmentation performance is verified based on the evaluation index, the image segmentation is performed by training a segmentation model using 118 CT scans with liver and liver tumor annotations, the data set includes 281 three-dimensional abdominal CTs, each data set has a resolution of 512×512 pixels, spiral data is reconstructed in the axial plane, the slice thickness is 1.25 mm, and the interval is 1.0 mm. The types of the data sets can be divided into liver tumor segmentation data sets and normal liver CT data sets, wherein,

[0057] The liver tumor segmentation dataset contains detailed per-pixel annotations of liver tumors and is part of the LiTS challenge;

[0058] The normal liver CT dataset is a normal liver dataset consisting of 116 CT scans compiled from healthy subjects in the CHAOS, Pancreas-CT, BTCV, and LiTS datasets, which serves as a baseline for generating synthetic tumor images. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The present invention provides a flowchart of an unsupervised liver tumor CT image segmentation method based on manual features. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be the usual meanings understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0061] The embodiment of the present invention provides an unsupervised liver tumor CT image segmentation method based on manual features, comprising:

[0062] S1. Collect CT scan images of normal liver as a data set, and perform volume and gradient verification on the potential tumor location of the liver to determine the tumor location of the liver;

[0063] S2. acquiring a final tumor texture based on the tumor position of the liver, and performing morphological modeling of the liver tumor based on the final tumor texture to synthesize a new liver tumor CT image;

[0064] S3. Use annotated CT scan images to train the segmentation model, use the synthesized tumor images as a dataset for medical image segmentation, and verify the segmentation performance based on evaluation indicators.

[0065] In some embodiments, the process of performing volume and gradient verification on the potential tumor location of the liver to determine the tumor location of the liver in step S1 specifically includes:

[0066] S1-1. Generating spherical coordinates in the liver Convert spherical coordinates to Cartesian coordinates x=(x, y, z), and convert Cartesian coordinates and spherical coordinates based on the ellipse axis and half-price;

[0067] S1-2, iteratively verify each potential tumor location x to determine whether the point is located in the liver until the tumor location x is determined. t =(x t ,y t ,z t ) to generate a tumor position of the liver, wherein the tumor position is located inside the liver and has a sufficient distance from the liver boundary.

[0068] Specifically, when executing step S1-1, a spherical coordinate is generated in the liver The spherical coordinates are converted to Cartesian coordinates x = (x, y, z), which are further transformed according to the axis and radius of the ellipsoid to ensure that the tumor location is spatially uniform within the defined ellipsoid.

[0069] Furthermore, the process of generating the tumor location of the liver in step S1-2 includes:

[0070] S1-2-1. Obtaining the number of liver masks in a neighborhood volume around the predicted position based on a neighborhood volume inspection, and obtaining a first tumor position based on the number of liver masks and a first threshold;

[0071] S1-2-2. Calculate the gradient value of the scan volume at the predicted position based on the gradient amplitude evaluation, obtain the second tumor position based on the gradient value and a second threshold, and generate the tumor position of the liver based on the first tumor position and the second tumor position.

[0072] Specifically, when executing step S1-2-1, the mathematical expression of the neighborhood volume is:

[0073]

[0074] Among them, τ is the first threshold, l(x i ) is the liver mask value, V(x) represents the ellipsoidal volume around the predicted position x, x i Represents every point or pixel in the domain volume V(x).

[0075] Furthermore, when executing step S1-2-2, the mathematical expression of the gradient value is:

[0076]

[0077] in, Represents the gradient value, x, y, z represent the three-dimensional coordinate point.

[0078] Furthermore, the gradient value and compared with the second threshold k. Exceeding the second threshold indicates that the gradient value The corresponding point is close to the liver boundary.

[0079] Furthermore, the combination of neighborhood volume inspection and gradient amplitude evaluation can ensure that the selected tumor location is located within the liver tissue and away from the boundary. The expression is:

[0080]

[0081] Among them, ∧ represents the logical AND operator, which means that the two conditions must be met at the same time, x i Represents every point or pixel in the domain volume V(x).

[0082] In fact, This is an iterative process, which continues until the optimal tumor position x is determined. t =(x t ,y t ,z t ), the optimal tumor position x t Located inside the liver, with a sufficient distance from the liver border.

[0083] In some embodiments, the process of obtaining the final tumor texture in step S2 includes:

[0084] S2-1, obtaining three-dimensional simple noise based on predefined parameters, and using nonlinear diffusion and wavelet transform to filter artifacts and retain high-frequency details of the noise to obtain a texture to be processed;

[0085] S2-2, normalizing the texture to be processed to obtain a normalized texture, and smoothing the normalized texture based on a first Gaussian filter to obtain a smoothed texture;

[0086] S2-3, proportionally scaling and moving the smoothed texture to obtain a first texture;

[0087] S2-4. Obtain a first binary mask based on random sampling, smooth the first binary mask based on a second Gaussian filter to obtain a second binary mask, and scale and crop the first texture based on the second binary mask to generate a final tumor texture.

[0088] Specifically, in the process of executing step S2-2 to smooth the normalized texture to obtain the smoothed texture, the standard deviation is σ b The Gaussian filter G(x;σ b ) performs smoothing processing on the normalized texture to obtain M′(x). The standard deviation of the first Gaussian filter G(x; σ1) is σ1.

[0089] Further, when executing step S2-3 to obtain the first texture, the mathematical expression of the first texture is:

[0090] T(x)=α·(S″(x)*G(x;σ1))+β;

[0091] Wherein, * represents a convolution operation, α represents a scaling ratio, β represents a displacement, and T(x) represents the first texture. Further, the process of executing step S2-4 to generate the final tumor texture includes:

[0092] T″(x)=T(x)·M′(x);

[0093] Where T″(x) represents the final tumor texture.

[0094] Furthermore, when executing step S2-4, the second Gaussian filter G(x; σ b ) is the standard deviation of b .

[0095] In some embodiments, the process of performing morphological modeling of liver tumors in step S2 includes:

[0096] S2-5, classify the synthetic tumors into four size categories: tiny, small, medium, and large, and draw the number of synthetic samples n from a discrete uniform distribution;

[0097] S2-6. The tumor shape is simulated based on an ellipsoid defined by parameters. The semi-principal axes of the ellipsoid in the x, y and z coordinates are taken from the uniform distribution U(0.75r, 1.25r) and the coordinates x in the scanning volume f(x) are t =(x t ,y t ,z t ) is the center, and the initial tumor mask is defined within this range;

[0098] S2-7, based on nonlinear elastic deformation The initial tumor mask is subjected to nonlinear transformation processing to obtain a deformed tumor mask:

[0099] S2-8. Perform convolution processing on the deformed tumor mask based on a third Gaussian filter to obtain a synthetic mask.

[0100] Specifically, when executing step S2-5, each tumor category has a specific radius r and a number n of synthetic samples drawn from the discrete uniform distribution U[a,b];

[0101] Specifically, when executing step S2-7, the deformed tumor mask is closer to the anisotropic growth pattern observed in vivo, and its expression is:

[0102]

[0103] Among them, t′(x) represents the deformed tumor mask, and t(x) represents the tumor mask before deformation.

[0104] Specifically, when performing step S2-8, the mathematical expression of the synthesized mask is:

[0105] t″(x) = t′(x) * G(x; σ2);

[0106] Among them, t″(x) represents the synthesized mask, and G(x; σ2) represents the Gaussian filter with a deviation of σ2.

[0107] In some embodiments, when performing step S2, the process of synthesizing the new liver tumor CT image includes:

[0108] S2-9. Quantify the density of the liver: Divide the liver into three density intervals and standardize the intensity values within each interval;

[0109] S2-10. Adjust the tumor density to reduce the density of high-density tissues;

[0110] S2-11. Integrate the tumor into the scan volume domain f(x) and the corresponding liver mask l(x).

[0111] In some embodiments, the specific process of quantifying the liver density when performing step S2-9 is as follows:

[0112] S2-9-1. Obtain the lower limit L and upper limit U of the liver region P, and calculate the density scale Δ:

[0113] Δ = (U - L) / 3;

[0114] S2-9-2. For each three-dimensional coordinate point P within the liver region ijk Perform the following operations:

[0115] If P ijk < L + Δ, let P ijk = L; if L + Δ ≤ P ijk < L + 2Δ, then let P ijk = L + Δ; if L + 2Δ ≤ P ijk < U, then let P ijk = L + 2Δ; if P ijk ≥ U, then let P ijk = U.

[0116] S2-9-3. According to the above operations, obtain the density distribution map D, and process each D on the density distribution map D ijk Perform processing: If D ijk = U and Pijk ≠U, then let D ijk =L+2Δ.

[0117] Specifically, the process of adjusting the tumor density in step S2-10 includes:

[0118] S2-10-1. Obtain the lower limit L and upper limit U of the liver region P, and calculate the density scale Δ:

[0119] Δ=(UL) / 3;

[0120] S2-10-2. For each coordinate point F in the scan volume F ijk Perform the following processing:

[0121] When the coordinate point F ijk When the position of the generated tumor coincides with that of the generated tumor, if D ijk =U, then let If D ijk =L+2Δ then let Otherwise, keep F ijk ;

[0122] Among them, D ijk represents the coordinate point on the density distribution map D, U represents the upper limit of the liver area, and L represents the lower limit of the liver area.

[0123] In some embodiments, the specific steps of integrating the tumor into the scan volume f(x) and the corresponding liver mask l(x) in step S2-11 are as follows:

[0124] S2-11-1. Integrate the tumor with the scan volume:

[0125] f′(x)=[1-t″(x)]⊙f(x)+t″(x)⊙T″(x);

[0126] Among them, ⊙ represents element-by-element multiplication, t″(x) represents the tumor mask array, and T″(x) represents the tumor texture array. The new scanning volume f′(x) can be obtained through the above transformation, and the new liver mask containing the tumor can be synthesized in the manner of l′(x)=l(x)+t″(x).

[0127] In reality, the capsule is formed due to the fibrotic reaction caused by chronic compression of the surrounding liver tissue. This capsule is characterized by a hyperechoic rim around the tumor, representing a layer of dense connective tissue. The formation of the capsule can be simulated by amplifying the intensity of the tumor border, and the border area of ​​the tumor b(x) is usually identified as:

[0128]

[0129] Among them, λ lrepresents the lower intensity threshold for defining the tumor boundary area, λ u represents the upper intensity threshold that defines the tumor boundary region;

[0130] S2-11-2. Use the Gaussian filter G(x; σ3) to convolve with b(x) to generate a smooth boundary image b′(x):

[0131] b′(x)=b(x)*G(x;σ3);

[0132] Among them, σ3 determines the standard deviation of the Gaussian kernel and controls the smoothness of the boundary transition;

[0133] In practice, the boundary image b′(x) is enhanced by adding a fixed intensity offset δ to the voxels at the periphery of the tumor, which simulates the increase of Hounsfield units (HU) in the cyst:

[0134] f″′(x)=f″(x)+δ·b′(x);

[0135] Among them, δ represents the HU intensity difference between the tumor core and the fibrous capsule. The final volume scan result f″′(x) reflects both the mass effect and the formation of the capsule.

[0136] In some embodiments, when executing step S3 to perform medical image segmentation using the synthesized tumor image as a data set and verifying the segmentation performance based on evaluation indicators, the image segmentation is performed using 118 CT scans with liver and liver tumor annotations to train a segmentation model, the data set includes 281 three-dimensional abdominal CTs, each data set has a resolution of 512×512 pixels, spiral data is reconstructed in the axial plane, the slice thickness is 1.25 mm, and the interval is 1.0 mm.

[0137] Among them, the types of datasets can be divided into liver tumor segmentation datasets and normal liver CT datasets. The liver tumor segmentation dataset contains detailed per-pixel annotations of liver tumors and is part of the LiTS challenge. The normal liver CT dataset is a normal liver dataset consisting of 116 CT scans compiled from healthy subjects in CHAOS, Pancreas-CT, BTCV and LiTS datasets. This dataset is a baseline for generating synthetic tumor images.

[0138] Specifically, synthetic tumor images are crucial to creating diverse image-label pairs. These synthetic tumors are systematically inserted into normal CT scans to increase the training data for the segmentation model.

[0139] Furthermore, when executing step S3, the evaluation index includes: Dice coefficient (DSC index), Hausdorff distance index (HD), normalized surface distance (NSD) and surface distance (SD) evaluation index.

[0140] Furthermore, to calculate the DSC index, we should first calculate the four indicators of the model segmentation results: true positive (TP), true negative (TN), false positive (FP) and false negative (FN). The calculation method is as follows:

[0141]

[0142] Among them, p i ,q i ∈{0,1}, 0 represents negative voxels, 1 represents positive voxels, p i is the i-th voxel result predicted by the model, q i To label the synthetic tumor on the liver, n represents the number of all pixels in the image. The Dice calculation method is as follows:

[0143]

[0144] Furthermore, the calculation method of Hausdorff distance index (HD) is as follows:

[0145]

[0146] Where h(A,B)=max a∈A {min b∈B ||ab||},h(B,A)=max b∈B {min a∈A ||ba||}, a and b represent points in point set A and point set B respectively. Specifically, a is a point in point set A and b is a point in point set B.

[0147] Among them, ||·|| is the distance normal form between the A and B point sets.

[0148] Furthermore, the normalized surface distance (NSD) is calculated as follows:

[0149]

[0150] Among them, S ref represents the reference surface, S seg represents the surface point set of the segmentation result, d(x,S) represents the minimum distance from point x to point set S, |·| represents the number of points in the point set, τ is the set distance threshold, x represents S ref or S seg The pixels in .

[0151] Furthermore, the surface distance (SD) is calculated as follows:

[0152]

[0153] Among them, S ref represents the reference surface, S seg The surface point set representing the segmentation result, d(x,S ref ) represents the point x to the point set S ref The minimum distance.

[0154] Furthermore, Table 1 is a comparison table of segmentation results of the present invention and other existing medical image segmentation methods.

[0155] Table 1 Comparison of the segmentation results of the present invention and other existing medical image segmentation methods

[0156]

[0157] The cells with horizontal lines in the table indicate that the corresponding method does not use the evaluation index. From the evaluation results, it can be seen that the method of the present invention achieves better segmentation results on four different evaluation indexes.

[0158] Although the embodiments of the present invention are described in detail above, it is obvious to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as described in the claims. Moreover, the present invention described herein may have other embodiments and may be implemented or realized in a variety of ways.

Claims

1. An unsupervised liver tumor CT image segmentation method based on manual features, characterized in that: include: CT scan images of normal liver are collected as a data set, and the potential tumor location of the liver is verified by volume and gradient to determine the tumor location of the liver; Acquire a final tumor texture based on the tumor position of the liver, perform morphological modeling on the liver tumor based on the final tumor texture, and synthesize a new liver tumor CT image; Annotated CT scan images are used to train the segmentation model, synthetic tumor images are used as datasets for medical image segmentation, and the segmentation performance is verified based on evaluation metrics.

2. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 1 is characterized in that: Volumetric and gradient verification of potential tumor locations in the liver The process of determining the tumor location in the liver includes: Generating spherical coordinates within the liver Convert spherical coordinates to Cartesian coordinates x=(x, y, z), and convert Cartesian coordinates and spherical coordinates based on the ellipse axis and half-price; Iterate and verify each potential tumor location x to determine whether the point is located in the liver until the tumor location x is determined. t =(x t ,y t ,z t ) to generate a tumor position of the liver, wherein the tumor position is located inside the liver and has a sufficient distance from the liver boundary.

3. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 2 is characterized in that: The process of generating a tumor location in the liver includes: Obtaining the number of liver masks in a neighborhood volume around the predicted position based on the neighborhood volume inspection, and obtaining a first tumor position based on the number of liver masks and a first threshold; A gradient value of the scan volume at the predicted position is calculated based on the gradient amplitude evaluation, and a second tumor position is acquired based on the gradient value and a second threshold, and a tumor position of the liver is generated based on the first tumor position and the second tumor position.

4. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 1, characterized in that: The process of obtaining the final tumor texture includes: Acquire three-dimensional minimalist noise based on predefined parameters, and use nonlinear diffusion and wavelet transform to filter artifacts and retain high-frequency details of the noise respectively to acquire a texture to be processed; Normalizing the texture to be processed to obtain a normalized texture, and performing smoothing processing on the normalized texture based on a first Gaussian filter to obtain a smoothed texture; Proportionally scaling and moving the smoothed texture to obtain a first texture; A first binary mask is obtained based on random sampling, the first binary mask is smoothed based on a second Gaussian filter to obtain a second binary mask, and the first texture is scaled and cropped based on the second binary mask to generate a final tumor texture.

5. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 1, characterized in that: The process of morphological modeling of liver tumors includes: The synthetic tumors were classified into four size categories: tiny, small, medium, and large, and the number of synthetic samples was drawn from a discrete uniform distribution; The tumor shape is simulated based on an ellipsoid defined by parameters, where the semi-major axes of the ellipsoid in the x, y, and z coordinates are taken from a uniform distribution U(0.75r, 1.25r) and the coordinates x in the scan volume f(x) are t =(x t ,y t ,z t ) is the center, and the initial tumor mask is defined within this range; The initial tumor mask is subjected to nonlinear transformation processing based on nonlinear elastic deformation to obtain a deformed tumor mask: The deformed tumor mask is convolved based on a third Gaussian filter to obtain a synthetic mask.

6. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 1, characterized in that: The process of synthesizing a new liver tumor CT image includes: Quantify the density of the liver: divide the liver into three density intervals and standardize the intensity values ​​within each interval; Adjust tumor density and reduce the density of high-density tissue; The tumor is integrated into the scan volume f(x) and the corresponding liver mask l(x).

7. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 6, characterized in that: The process of quantifying the density of the liver comprises: Get the lower limit L and upper limit U of the liver region P and calculate the density scale Δ: Δ=(UL) / 3; For each 3D coordinate point Pijk in the liver area, the following operations are performed: If P ijk <L + Δ, let P ijk = L; if L + Δ ≤ P ijk < L + 2Δ, then let P ijk = L + Δ; if L + Δ ≤ P ijk <U, then let P ijk = L + 2Δ; if P ijk ≥ U, then let P ijk = U; According to the above operation, we can get the density distribution map D. For each D on the density distribution map D, ijk Processing: If D ijk =U and P ijk ≠U, then let D ijk =L+2Δ.

8. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 6, characterized in that: The process of adjusting for tumor density includes: Get the lower limit L and upper limit U of the liver region P and calculate the density scale Δ: Δ=(UL) / 3; For each coordinate point F in the scan volume F ijk Perform the following processing: When the coordinate point F ijk When the position of the generated tumor coincides with that of the generated tumor, if D ijk =U, then let If D ijk =L+2Δ then let Otherwise, keep F ijk ; Among them, D ijk represents the coordinate point on the density distribution map D, U represents the upper limit of the liver area, and L represents the lower limit of the liver area.

9. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 6, characterized in that: The process of integrating the tumor into the scan volume f(x) and the corresponding liver mask l(x) includes: Integrate the tumor with the scan volume: f′(x)=[1-t″(x)]⊙f(x)+t″(x)⊙T″(x); Wherein, ⊙ represents element-wise multiplication, t″(x) represents the tumor mask array, and T″(x) represents the tumor texture array. Through the above transformation, a new scan volume f′(x) can be obtained, and a new liver mask containing the tumor can be synthesized in the manner of l′(x)=l(x)+t″(x); Determine the border area of ​​the tumor: where λ l represents the lower intensity threshold for defining the tumor boundary area, λ u Represents the upper intensity threshold that defines the tumor boundary area: Use the Gaussian filter G(x; σ3) to convolve with b(x) to generate a smooth boundary image b′(x): b′(x)=b(x)*G(x;σ3); Among them, σ3 determines the standard deviation of the Gaussian kernel and controls the smoothness of the boundary transition; The boundary map b′(x) enhances the voxel intensity at the periphery of the tumor by adding a fixed intensity offset δ, which simulates the increase of Hounsfield units (HU) in the cyst: f″′(x)=f″(x)+δ·b′(x); where δ represents the HU intensity difference between the tumor core and the fibrous capsule, and f″′(x) represents the final volume scan result.

10. The unsupervised liver tumor CT image segmentation method based on manual features according to claim 1, characterized in that: When the synthesized tumor image is used as a data set for medical image segmentation and the segmentation performance is verified based on an evaluation index, the image segmentation is performed by training a segmentation model using 118 CT scans with liver and liver tumor annotations, the data set includes 281 three-dimensional abdominal CTs, each data set has a resolution of 512×512 pixels, spiral data is reconstructed in an axial plane, the slice thickness is 1.25 mm, and the interval is 1.0 mm. The types of the data sets can be divided into liver tumor segmentation data sets and normal liver CT data sets, wherein, The liver tumor segmentation dataset contains detailed per-pixel annotations of liver tumors and is part of the LiTS challenge; The normal liver CT dataset is a normal liver dataset consisting of 116 CT scans compiled from healthy subjects in the CHAOS, Pancreas-CT, BTCV, and LiTS datasets, which serves as a baseline for generating synthetic tumor images.

Images