A method for correlation analysis of lung CT images

Abstract

The application provides a lung CT image correlation analysis method, comprising: acquiring lung multi-modal images, including computer tomography images and positron emission tomography images, and performing denoising, enhancement and standardization pretreatment on the original images according to the imaging principles and image characteristics of different modalities; extracting a lung lesion region of interest from the registered multi-modal images, delineating and segmenting the lesion, and quantitatively analyzing the volume and shape geometric morphological characteristics of the lesion; based on the therapeutic effect, analyzing the correlation between the imaging features and the prognosis of the patient, estimating the recurrence and metastasis risk and survival period under different feature performances, combining the imaging features and the clinical indicators, constructing a therapeutic effect prediction model, and performing long-term individualized evaluation on the prognosis of the patient.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and particularly relates to a lung CT image correlation analysis method. BACKGROUND

[0002] In the diagnosis of lung diseases, multi-modal medical image fusion analysis faces a key technical problem. When the lesion features of lung CT images are correlated with other image modalities, there are challenges in the correspondence of features between modalities. At the same time, the reflected tissue characteristics of each modality are different, and the feature expression methods are different, making it difficult to establish a direct correspondence. This makes it a difficult problem to accurately match and correlate the lesion regions in multi-modal data fusion. On the other hand, when using multi-modal correlation analysis results to evaluate treatment effect, there is also a technical bottleneck. Different treatment schemes may have different effects on the lesion features in each modality image, and how to quantify and represent this dynamic changing correlation pattern remains to be solved. Especially in the long-term follow-up process, it is challenging to establish a robust longitudinal analysis model for the complex influence of patient treatment response factors on multi-modal image features. How to extract and fuse the most diagnostic feature information under the condition of limited multi-modal data to achieve accurate lesion correlation analysis and efficacy evaluation is a technical difficulty that needs to be broken through. SUMMARY

[0003] The present application provides a lung CT image correlation analysis method, mainly comprising:

[0004] Obtain lung multi-modal images, including computed tomography images and positron emission tomography images, and according to the imaging principles and image characteristics of different modalities, pre-process the original images by denoising, enhancing and standardizing;

[0005] For the pre-processed multi-modal images, use the complementary information between different modalities to optimize the registration parameters by maximizing the mutual information between images, and according to the coincidence degree of the organ tissue anatomical structure in the registered images, quantitatively evaluate the accuracy of multi-modal image registration;

[0006] Extract the lung lesion region of interest from the registered multi-modal images, delineate and segment the lesion, and quantitatively analyze the volume and shape geometric morphological features of the lesion;

[0007] For the segmented lesion region, extract multi-scale frequency domain texture features, obtain second-order statistical texture features including contrast, correlation and entropy through gray level co-occurrence matrix and run-length matrix, and combine the lesion morphological features to construct a multi-modal image omics feature vector;

[0008] The multi-modal imageomics feature vector is dimensionally reduced, a key feature subset related to curative effect is screened out, feature selection is realized by iteratively optimizing an L1 norm regularization term, the distinguishing performance of the feature subset is evaluated in combination with cross validation, and an optimal feature combination is determined;

[0009] Based on the optimal feature combination, continuous multi-modal images of a patient before and after treatment are acquired, the lesion region after registration and segmentation is quantitatively measured, the dynamic changes of the lesion volume, shape and texture features are compared and analyzed, the curative effect difference of different treatment schemes in the near future is judged, and if the lesion volume is significantly reduced, the shape tends to be regular, and the texture tends to be uniform, the curative effect is good.

[0010] Based on the curative effect result, the correlation between the imageomics features and the prognosis of the patient is analyzed, the recurrence and metastasis risk and the survival period under different feature performances are estimated, the curative effect prediction model is constructed in combination with the image features and the clinical indicators, and the long-term individualized evaluation of the prognosis of the patient is performed.

[0011] Based on the curative effect prediction model, an early multi-modal imageomics performance and a prediction model of the prognosis and curative effect of the patient are established according to multi-center clinical data, the feature importance analysis and the internal correlation mode between the image features and the prognosis are mined, and the prognosis of a new patient is individually predicted based on the same.

[0012] The technical scheme provided by the embodiment of the present application can include the following beneficial effects:

[0013] The application discloses a lung CT image correlation analysis method. The method first pre-processes and registers CT and PET images, extracts a lesion region and segments. Then multi-scale texture and shape features are extracted from the registered multi-modal images to construct an imageomics feature vector. Through feature selection and dimension reduction, a key feature subset related to curative effect is screened out. Based on the optimal feature combination, the dynamic changes of the lesion before and after treatment are analyzed to evaluate the near-term curative effect. In combination with clinical indicators, a long-term prognosis prediction model is constructed to realize individualized prediction of new patients. The present application effectively integrates the complementary information of different modal images through multi-modal imageomics analysis, improves the prediction accuracy of the curative effect and prognosis of lung cancer patients, and provides an important basis for formulating individualized treatment schemes. BRIEF DESCRIPTION OF DRAWINGS

[0014] Fig. 1 A flowchart of the lung CT image correlation analysis method of the present application.

[0015] Fig. 2 A schematic diagram of the lung CT image correlation analysis method of the present application.

[0016] Fig. 3 Another schematic diagram of the lung CT image correlation analysis method of the present application. DETAILED DESCRIPTION

[0017] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments.

[0018] As Figs. 1-3 , the lung CT image correlation analysis method of the embodiment can specifically include:

[0019] Step S101, acquiring lung multi-modal images including computed tomography (CT) images and positron emission tomography (PET) images. According to the imaging principles and image characteristics of different modalities, the original images are preprocessed by denoising, enhancing and standardizing.

[0020] The acquired lung CT images and PET images are registered by using a multi-modal image registration algorithm. According to the spatial resolution and gray scale range difference of the CT images and PET images, the mutual information maximization criterion is selected as the similarity measure, coarse registration is performed by principal component analysis, rigid registration is performed by applying affine transformation, and non-rigid registration is performed by using B-spline transformation to obtain the first registered CT image. For the first CT image, an anisotropic diffusion filtering algorithm is used for denoising processing. The anisotropic diffusion filtering algorithm adaptively adjusts the diffusion coefficient according to the local gradient of the image to obtain the second denoised CT image. The second CT image is subjected to histogram equalization processing to enhance the contrast by stretching the image gray scale dynamic range. If the image contrast enhancement ratio exceeds the preset threshold, a contrast-limited adaptive histogram equalization algorithm is used for local contrast enhancement to obtain the third CT image. According to the characteristics of the PET image, Daubechies wavelet is used for multi-scale decomposition, and the wavelet coefficients are subjected to VisuShrink soft threshold processing to remove noise. The denoised second PET image is obtained by inverse transformation reconstruction, and the second PET image is subjected to Gaussian filter smoothing processing and edge enhancement using the Laplacian operator to obtain the third PET image. The pixel values of the third CT image and the third PET image are converted to standard normal distribution by using the Z-score standardization method to complete the standardization processing.

[0021] Specifically, a multi-modal image registration algorithm is used to register the acquired lung computed tomography (CT) and positron emission tomography (PET) images. According to the differences in spatial resolution and gray scale range of the two modal images, the mutual information maximization criterion is selected as the similarity measure. Principal component analysis (PCA) is used for coarse registration, affine transformation is used for rigid registration, and finally B-spline transformation is used for fine non-rigid registration to realize image alignment. For the registered CT image, an anisotropic diffusion filtering algorithm is used for denoising. The algorithm adaptively adjusts the diffusion coefficient according to the local gradient of the image, reducing noise while preserving edge information, thus obtaining a denoised second CT image. The denoised second CT image is subjected to histogram equalization processing to enhance contrast by stretching the image gray scale dynamic range. If the contrast enhancement ratio exceeds the preset threshold, an adaptive histogram equalization algorithm with limited contrast is used for local contrast enhancement to obtain an enhanced third CT image. According to the characteristics of the PET image, Daubechies wavelet is used for multi-scale decomposition, and an appropriate threshold is selected for VisuShrink soft threshold processing of the wavelet coefficients to remove noise. The denoised second PET image is obtained by inverse transformation reconstruction. Gaussian filtering is applied to the second PET image for smoothing, and then the Laplacian operator is used for edge enhancement to obtain a third PET image. The Z-score standardization method is used to convert the pixel values of the third CT and PET images to a standard normal distribution, completing the standardization process. In the multi-modal image registration process, the CT and PET images are first preprocessed, including resampling the images to the same resolution and normalizing the gray scale range to 0-255. Then, PCA is used to extract the main features of the images, calculate the eigenvectors and eigenvalues, and select the top 3 principal components for coarse registration. Next, affine transformation matrix is applied for rigid registration, and the optimal transformation parameters are determined by minimizing the mutual information negative value. Finally, B-spline transformation is used for non-rigid registration, setting the control point grid to 16x16, and optimizing the control point positions using the gradient descent method to maximize the mutual information. Anisotropic diffusion filtering algorithm is used for denoising of the registered CT image, with a diffusion coefficient k=0.1, 20 iterations, and a time step of 0.25. The algorithm adaptively adjusts the diffusion intensity according to the local gradient, strongly denoising in smooth areas and preserving details in edge areas. Histogram equalization is performed on the denoised image, the cumulative distribution function is calculated, and the gray scale mapping is performed. If the contrast enhancement ratio exceeds the preset threshold of 1.5, the adaptive histogram equalization algorithm with limited contrast is used to divide the image into 8x8 sub-blocks, limit the contrast threshold to 0.02, and perform local enhancement.For positron emission tomography (PET) images, a three-level decomposition was performed using the Daubechies4 wavelet. VisuShrink soft thresholding was applied to the high-frequency subband coefficients, with a threshold set to σ√(2logN), where σ is the noise standard deviation and N is the number of coefficients. The denoised image was reconstructed through inverse transform. Then, a 3x3 Gaussian filter kernel was applied for smoothing, with σ = 0.5. Edge enhancement was performed using a 3x3 Laplacian operator with a convolution kernel of [[0,1,0],[1,-4,1],[0,1,0]]. Finally, Z-score normalization was performed on the two processed modal images, calculating the mean μ and standard deviation σ. The formula z = (x-μ) / σ was used to transform each pixel value x, adjusting the pixel value distribution to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0022] Step S102: For the preprocessed multimodal images, the registration parameters are optimized by maximizing the mutual information between images using complementary information between different modalities. The accuracy of multimodal image registration is quantitatively evaluated based on the degree of overlap of organ and tissue anatomical structures in the registered images.

[0023] A joint histogram is constructed based on the gray-level distribution characteristics of the preprocessed multimodal images, and the edge histogram of the joint histogram is calculated. The mutual information value between the two modalities is calculated using a normalized mutual information formula, and this normalized mutual information value serves as the optimization objective function. The registration parameters, including rotation angle, translation vector, and scaling factor, are iteratively optimized using gradient descent. The iteration stops if the change in the mutual information value is less than a preset threshold. The optimized registration parameters are used to perform a spatial transformation on the multimodal images to obtain the registered images. A voxel-based registration evaluation method is used to extract the contours of key anatomical structures from the registered images. The overlap coefficient and Hausdorff distance of the key anatomical structures are calculated, with the overlap coefficient calculated using the Dice coefficient. A weighted scoring model is established, which integrates the overlap coefficient, the Hausdorff distance, and the mutual information value. The registration results are quantitatively evaluated to obtain a registration accuracy score.

[0024] Specifically, based on the gray-level distribution characteristics of the preprocessed multimodal images, a joint histogram is constructed and an edge histogram is calculated. The mutual information value between the two modalities is then calculated using a normalized mutual information formula, and this value is used as the optimization objective function. The normalized mutual information calculation formula is as follows:

[0025] NMI(A,B) = (H(A) + H(B)) / H(A,B), where H(A) and H(B) are the entropies of images A and B, respectively, and H(A,B) is the joint entropy. Gradient descent is used to iteratively optimize the registration parameters, including rotation angle, translation vector, and scaling factor. In each iteration, the parameter values ​​are updated and the mutual information is recalculated until the maximum mutual information value or the upper limit of the number of iterations is reached. The learning rate of the gradient descent method is set to 0.01, the maximum number of iterations is 1000, and iteration stops when the change in the mutual information value is less than 0.0001. The optimized registration parameters are used to perform spatial transformations on the multimodal images, including rotation, translation, and scaling, to obtain the registered images. A voxel-based registration evaluation method is used to segment the registered multimodal images, extract the contours of the lungs, heart, and bones, and calculate the overlap coefficient and Hausdorff distance of the corresponding structures. Overlap is calculated using the DiCi coefficient, DiCi = 2|A∩B| / (|A|+|B|), where A and B represent the target regions in the two images, respectively. A weighted scoring model is established by integrating multiple evaluation metrics, including overlap coefficient, Hausdorff distance, and mutual information value, to quantitatively evaluate the registration results and obtain the final registration accuracy score. Registration accuracy score = 0.4 * overlap coefficient + 0.3 * (1 - normalized Hausdorff distance) + 0.3 * normalized mutual information value. This scoring model enables a quantitative evaluation of the registration accuracy of multimodal images. In the multimodal image registration process, a 256×256 joint histogram is first constructed from the preprocessed computed tomography and positron emission tomography images, and the edge histogram is calculated and normalized. The mutual information value was calculated using the normalized mutual information formula NMI(CT,PET) = (H(CT) + H(PET)) / H(CT,PET), where H(CT) and H(PET) are the entropies of the CT and PET images, respectively, and H(CT,PET) is the joint entropy. The initial mutual information value was 0.65. Gradient descent was used to optimize the registration parameters, with a learning rate of 0.01 and a maximum number of iterations of 1000. In each iteration, the rotation angle, translation vector, and scaling factor were updated. Optimization stopped when the change in mutual information value was less than 0.0001 or when the maximum number of iterations was reached. After 786 iterations, the mutual information value increased to 0.82, with the final rotation angle = 2.3°, translation vector = [1.5, 0.8, -0.5], and scaling factor...

[0026] =1.02. These parameters were used to perform spatial transformation on the PET images, resulting in registered PET images aligned with the CT images. Next, threshold segmentation was used to extract lung, heart, and bone structures. Thresholds for the CT images were set to -500HU, 40HU, and 300HU, respectively, while thresholds for the PET images were set to SUV1.0, SUV2.5, and SUV3.0, respectively. The Dice coefficients for these structures were calculated: 0.92 for the lungs, 0.88 for the heart, and 0.95 for the bones. The Hausdorff distances were 3.2 mm, 4.1 mm, and 2.8 mm, respectively. Finally, a weighted scoring model was used to calculate the registration accuracy: 0.4 × 0.92 + 0.3 × (1 - 3.37 / 10) + 0.3 × 0.82 = 0.8654, indicating good registration accuracy.

[0027] Step S103: Extract the region of interest of lung lesions from the registered multimodal image, delineate and segment the lesions, and quantitatively analyze the volume, shape and geometric features of the lesions.

[0028] Based on the registered multimodal images, regions of interest (ROIs) for lung lesions are extracted using a region growing algorithm to obtain preliminary lesion contours. For these preliminary contours, fine segmentation is performed using the level set method, constructing an energy function E(φ), where I is the image intensity, c1 and c2 are the average foreground and background intensities, and H(φ) is the Heaviside function. If the fine segmentation results contain noise, morphological operations are used to post-process the segmentation results, including removing small connected regions and smoothing boundaries. Multiple lesions are identified using a connected component labeling algorithm, and the volume of each lesion is calculated and its spatial coordinates are recorded. For each lesion, shape and geometric features are extracted, including maximum diameter, minor axis length, major-minor axis ratio, sphericity, compactness, and irregularity. Principal component analysis is used to calculate the major and minor axes, and sphericity is calculated using the volume-to-surface-area ratio.

[0029] Specifically, a region growing algorithm is used to extract regions of interest (ROIs) of lung lesions from the registered multimodal images. Combining computed tomography (CT) and positron emission tomography (PET) image information, the Otsu algorithm is used to automatically calculate the optimal grayscale threshold. Based on the calculated threshold and spatial connectivity conditions, the region is gradually expanded from the seed point to obtain a preliminary lesion outline. The preliminary lesion outline is then finely segmented using the level set method, and an energy function is constructed. Where I represents image intensity, c1 and c2 represent the average foreground and background intensities, and H(φ) is the Heaviside function. The level set function is iteratively updated using image gradient information and curvature constraints until convergence yields accurate lesion boundaries. Morphological operations are used to post-process the segmentation results, including removing small connected regions and smoothing boundaries. The Dice coefficient is then used to evaluate segmentation accuracy. Multiple lesions are identified using a connected region labeling algorithm, and the volume and spatial coordinates of each lesion are calculated and recorded. Geometric features of the lesions are extracted, including maximum diameter, minor axis length, major-minor axis ratio, sphericity, compactness, and irregularity. Principal component analysis is used to calculate the major and minor axes, and sphericity is calculated using the volume-to-surface-area ratio. Feature vectors are used to quantify the geometric morphological features of the lesions. In the multimodal lung image processing, a region growing algorithm is first applied to the registered computed tomography (CT) and positron emission tomography (PET) images. The Otsu algorithm is used to calculate the optimal grayscale thresholds for CT and PET images, with a threshold of -450 HU for CT and SUV2.5 for PET. Pixels with grayscale values ​​higher than -450 HU in CT images and SUV values ​​higher than 2.5 in PET images were selected as seed points. An 8-neighborhood connectivity condition was set, and the region was iteratively expanded until the threshold condition was no longer met, yielding a preliminary lesion outline. Subsequently, a level set energy function was constructed, with weight parameters λ1 = 1.0, λ2 = 1.0, and μ = 0.1. The level set function was iteratively updated 500 times or until the energy change was less than 0.001. Morphological post-processing was performed on the segmentation results. Opening operations were performed using circular structuring elements with a radius of 2 pixels to remove small connected regions, and then closing operations were performed using circular structuring elements with a radius of 1 pixel to smooth the boundaries. The Dice coefficient was calculated to evaluate the segmentation accuracy, with a threshold set to 0.85. An 8-connected region labeling algorithm was used to identify multiple lesions with a volume less than 27 mm². 3 Connected regions are treated as noise and removed. For each lesion, the volume is calculated by voxel accumulation, and the centroid coordinates are recorded. Finally, principal component analysis is used to calculate the major and minor axes, and the eigenvector corresponding to the largest eigenvalue is taken as the direction of the major axis; the sphericity is calculated as (6V√π) / (A^(3 / 2)), where V is the volume and A is the surface area; the irregularity is calculated as the ratio of the actual surface area to the surface area of ​​a sphere of equal volume. The obtained eigenvectors include volume, maximum diameter, minor axis length, major-minor axis ratio, sphericity, compactness, and irregularity, which are used to quantitatively describe the geometric morphological characteristics of the lesion.

[0030] Step S104: For the segmented lesion region, extract multi-scale frequency domain texture features, obtain second-order statistical texture features through gray-level co-occurrence matrix and run-length matrix, including contrast, correlation and entropy, and construct multimodal radiomics feature vectors by combining lesion morphology features.

[0031] Multi-scale wavelet transform is performed on the segmented lesion regions, and a three-level decomposition is performed using Daubechies4 wavelets to obtain frequency domain texture information at different scales of the lesion regions. The energy, entropy, and variance of each sub-band are calculated based on the frequency domain texture information to obtain frequency domain texture features. The mean, standard deviation, skewness, and kurtosis of the frequency domain texture features are calculated to determine the multi-scale frequency domain texture features. A gray-level co-occurrence matrix is ​​calculated for the lesion regions, and contrast, correlation, and entropy are extracted from the gray-level co-occurrence matrix to determine the second-order statistical texture features. Run matrices are constructed for computed tomography (CT) and positron emission tomography (PET) images, respectively, and run percentage, run length non-uniformity, and gray-level non-uniformity are extracted from the run matrices to obtain run features. The multi-scale frequency domain texture features, the second-order statistical texture features, and the run features are combined with lesion morphology features, and the combined features are standardized. A radiomics feature vector containing multimodal information is constructed based on the standardized features.

[0032] Specifically, multi-scale wavelet transform is performed on the segmented lesion regions, and a three-level decomposition is performed using the Daubechies4 wavelet to obtain frequency domain texture information at different scales. The energy, entropy, and variance of each sub-band are calculated as frequency domain texture features. Statistical features, including mean, standard deviation, skewness, and kurtosis, are calculated for the wavelet coefficients as multi-scale frequency domain texture features, which describe the multi-scale frequency domain texture features of the lesions. The gray-level co-occurrence matrix of the lesion region is calculated, with distance parameters set to 1, 2, and 4 pixels, and orientation angles of 0°, 45°, 90°, and 135°, for a total of 12 co-occurrence matrices. Contrast, correlation, and entropy are extracted from these co-occurrence matrices to quantitatively describe the gray-level distribution characteristics of the lesions. Run matrices are constructed for computed tomography (CT) and positron emission tomography (PET) images, respectively. The distribution of different gray levels and run lengths is statistically analyzed, and run percentage, run length non-uniformity, and gray-level non-uniformity are extracted to characterize the gray-level continuity of the lesions. Then, the features of the two modalities are concatenated. Multi-scale frequency domain texture features, gray-level co-occurrence matrix features, and run-length matrix features are combined with the previously obtained lesion morphology features. All features are standardized to the [0,1] interval and then concatenated into a long vector to construct a radiomics feature vector containing multimodal information for subsequent lesion analysis. In the process of extracting radiomics features from lung lesions, the segmented lesion regions are first decomposed into 10 sub-bands using Daubechies4 wavelets (3 levels), resulting in 10 sub-bands, including 1 low-frequency sub-band and 9 high-frequency sub-bands. Energy, entropy, and variance are calculated for each sub-band, yielding 30 frequency domain texture features. Subsequently, statistical features, including mean, standard deviation, skewness, and kurtosis, are calculated for the wavelet coefficients, resulting in 40 multi-scale frequency domain texture features. Next, the gray-level co-occurrence matrix of the lesion region is calculated, with distance parameters set to 1, 2, and 4 pixels, and orientation angles of 0°, 45°, 90°, and 135°, resulting in 12 co-occurrence matrices. Contrast, correlation, and entropy are extracted from each matrix, yielding 36 second-order statistical texture features. Run-length matrices were constructed for both CT and PET images, with 16 gray levels. The distribution of different gray levels and run lengths was statistically analyzed, and run percentage, run length non-uniformity, and gray level non-uniformity were extracted. Finally, all features were combined with the seven previously obtained morphological features, including volume, maximum diameter, minor axis length, major-minor axis ratio, sphericity, compactness, and irregularity. Min-max normalization was used to map the feature values ​​to the [0,1] interval to obtain the feature vector.

[0033] Based on the multi-scale frequency domain texture features of lesions, first-order statistical features of the lesion region are extracted, including gray-level mean, variance, skewness and kurtosis, to characterize the pixel gray-level distribution within the lesion; at the same time, second-order statistical features based on both gray-level co-occurrence matrix and run-length matrix are extracted, including contrast, correlation and entropy.

[0034] A multi-scale wavelet transform is performed on the lesion region image, and a three-level decomposition is performed using Haar wavelet basis functions to obtain the low-frequency approximation coefficients and high-frequency detail coefficients in three directions of the lesion region image. Based on the low-frequency approximation coefficients and the high-frequency detail coefficients, the energy, variance, and entropy of the wavelet coefficients are calculated to obtain the multi-scale frequency domain texture features of the lesion region image. The mean, variance, skewness, and kurtosis of the grayscale pixels in the lesion region are calculated, where the skewness is E[(X-μ)]. 3 ] / σ 3 The kurtosis is E[(X-μ)]. 4 ] / σ 4 -3, X is the gray value, μ is the mean, and σ is the standard deviation. The gray-level distribution characteristics of pixels within the lesion are determined. A gray-level co-occurrence matrix of the lesion region is constructed, with distance parameters set to 1, 2, and 4 pixels, and direction angles of 0°, 45°, 90°, and 135°. Contrast, correlation, and entropy are extracted from the gray-level co-occurrence matrix to determine the second-order statistical features of the lesion region. A run-length matrix of the lesion region is generated. By quantifying the gray level of the lesion region image, the number of consecutive pixels at each gray level is counted. The run length non-uniformity, gray-level non-uniformity, and run percentage of the run-length matrix are calculated to obtain the complete second-order statistical feature set of the lesion region.

[0035] Specifically, multi-scale wavelet transform is performed on the lesion region image, and a three-level decomposition is performed using Haar wavelet basis functions to obtain low-frequency approximation coefficients and high-frequency detail coefficients in three directions. Energy, variance, and entropy are calculated from the wavelet coefficients, where energy is the average of the sum of squares of the wavelet coefficients, variance is the dispersion of the wavelet coefficients, and entropy is -Σp*log2(p), where p is the normalized wavelet coefficient, serving as multi-scale frequency domain texture features. The mean, variance, skewness, and kurtosis of the pixels in the lesion region are calculated, where skewness = E[(X-μ)]. 3 ] / σ 3 Kurtosis = E[(X-μ)] 4 ] / σ 4 -3, X is the gray value, μ is the mean, and σ is the standard deviation. Skewness reflects the asymmetry of the gray-level distribution, and kurtosis characterizes the steepness of the gray-level distribution. These first-order statistical features characterize the pixel gray-level distribution within the lesion. A gray-level co-occurrence matrix (COM) of the lesion region is constructed, with distance parameters set to 1, 2, and 4 pixels, and orientation angles of 0°, 45°, 90°, and 135°. The COM matrix for each parameter combination is calculated. Contrast, correlation, and entropy are extracted from the COM matrix as second-order statistical features, where contrast is Σi,j|ij|. 2p(i,j) has a correlation of Σi,j[(ij)p(i,j)-μxμy] / (σxσy) and an entropy of -Σi,jp(i,j)log2p(i,j), where p(i,j) is an element of the co-occurrence matrix. To generate the run-length matrix of the lesion region, the image is first grayscaled, and then the number of consecutive pixels at each gray level is counted to form the run-length matrix. Run-length nonuniformity, gray-level nonuniformity, and run percentage are calculated. Combined with the gray-level co-occurrence matrix features, a complete second-order statistical feature set is constructed to characterize the texture complexity and gray-level distribution characteristics of the lesions. In the process of extracting radiomics features from lung lesions, the 512x512 pixel lesion region image is first decomposed into 10 sub-bands using Haar wavelets. Energy, variance, and entropy are calculated for each sub-band to obtain 30 multi-scale frequency domain texture features. Subsequently, the first-order statistical features of the lesion region were calculated: mean gray level of 120, variance of 225, skewness of 0.3, and kurtosis of 2.8. Next, gray-level co-occurrence matrices were constructed, setting four distances corresponding to pixels 1, 2, 4, and 8, and four directions corresponding to 0°, 45°, 90°, and 135°, resulting in 16 20x20 co-occurrence matrices. From each matrix, 48 second-order statistical features were extracted: contrast (0.65), correlation (0.78), and entropy (4.2). Finally, a run-length matrix was generated, quantizing the gray levels to 16 levels, resulting in a 16x512 matrix. Run-length nonuniformity (0.85), gray-level nonuniformity (0.72), and run-length percentage (0.35) were calculated.

[0036] Step S105: Dimensionality reduction is performed on the multimodal radiomics feature vectors to select key feature subsets related to efficacy. Feature selection is achieved by iteratively optimizing the L1 norm regularization term. The discriminative performance of the feature subsets is evaluated by cross-validation to determine the optimal feature combination.

[0037] Multimodal radiomics feature vectors are standardized by using the Z-score method to convert the feature values ​​into a standard normal distribution, resulting in a standardized feature matrix. Based on this standardized feature matrix, principal component analysis (PCA) is used for initial dimensionality reduction, calculating the contribution rate of each principal component until the cumulative contribution rate reaches a preset threshold, thus obtaining the dimensionality-reduced feature matrix. For this dimensionality-reduced feature matrix, a linear regression model with L1 norm regularization is constructed. The model parameters are iteratively optimized using coordinate descent to achieve feature selection. If a 5-fold cross-validation method is used to evaluate the discriminative performance of feature subsets, the optimal regularization parameter is selected from preset values ​​by adjusting the L1 regularization strength. The correlation coefficients between the features and efficacy indicators are calculated, and combined with the L1 regularization results, the optimal feature combination is determined, identifying the key feature subset related to efficacy.

[0038] Specifically, the feature vectors of multimodal image omics are standardized by using the Z-score method to convert each feature value into a standard normal distribution, eliminating the influence of different feature dimensions and obtaining a standardized feature matrix. Principal component analysis (PCA) is then used to perform initial dimensionality reduction on the standardized feature matrix. The contribution rate of each principal component is calculated and accumulated from largest to smallest until the cumulative contribution rate reaches 95%, at which point the corresponding principal component is retained, reducing redundant information in the feature space. The original feature space is reconstructed using the retained principal components, resulting in a dimensionality-reduced feature matrix. A linear regression model with L1 norm regularization is constructed within the dimensionality-reduced feature matrix, and the model parameters are iteratively optimized using coordinate descent. The weights of each feature are updated sequentially, while the weights of other features are fixed. A one-dimensional optimization problem is solved iteratively until convergence or the maximum number of iterations is reached, gradually compressing the weights of unimportant features to zero to achieve feature selection. A 5-fold cross-validation method is used to evaluate the discriminative performance of the feature subset. The L1 regularization strength is adjusted, selected from 0.01, 0.1, 1, and 10, and feature selection and performance evaluation are repeatedly performed. The correlation coefficient between each feature and the efficacy index was calculated, and combined with the L1 regularization results, until the optimal feature combination was found, resulting in a subset of key features related to efficacy. In the multimodal radiomics feature dimensionality reduction and selection process, the original feature vector containing 200 features was first Z-score standardized, adjusting the mean of each feature to 0 and the standard deviation to 1. Then, principal component analysis was applied for preliminary dimensionality reduction, calculating the eigenvalues ​​and eigenvectors of the feature covariance matrix, sorting them by eigenvalue size, and accumulating the contribution rate until it reached 95%, ultimately retaining 35 principal components. These 35 principal components were used to reconstruct the original feature space, obtaining the dimensionality-reduced feature matrix from 200 to 35. Next, a linear regression model with L1 norm regularization was constructed, initializing all feature weights to 0 and setting the maximum number of iterations to 1000. The coordinate descent method was used to optimize the model parameters, updating the weights of the 35 features sequentially in each iteration while fixing the weights of the other 34 features, solving a one-dimensional optimization problem. After 786 iterations, the model converged, and the weights of the 15 features were compressed to 0. Then, 5-fold cross-validation was performed, dividing the dataset evenly into 5 parts, with 4 parts used as the training set and 1 part as the validation set. Four L1 regularization strengths were tested: 0.01, 0.1, 1, and 10. For each strength value, feature selection and performance evaluation were repeated 5 times. Simultaneously, the Pearson correlation coefficient between each feature and the efficacy index was calculated. Finally, the best performance was achieved with a regularization strength of 0.1, and 12 key features were selected. The absolute values ​​of the correlation coefficients between these features and the efficacy index were all greater than 0.3, forming the final feature subset.

[0039] Step S106: Based on the optimal feature combination, acquire continuous multimodal images of the patient before and after treatment, and quantitatively measure the registered and segmented lesion regions. By comparing and analyzing the dynamic changes in lesion volume, morphology, and texture features, determine the short-term efficacy differences of different treatment plans. If the lesion volume significantly shrinks, the morphology becomes more regular, and the texture becomes more uniform, it indicates a good therapeutic effect.

[0040] Continuous multimodal images of the patient before and after treatment were acquired. The mutual information maximization algorithm was used to register the computed tomography (CT) and positron emission tomography (PET) images before and after treatment, resulting in registered images. The level set method was used to segment the lesion regions in the registered images, yielding segmented lesion regions. Based on the segmented lesion regions, quantitative measurements were performed on the segmented lesion regions using optimal feature combinations, including volume, maximum diameter, sphericity, contrast of the gray-level co-occurrence matrix, and correlation. Feature vectors of the lesions before and after treatment were calculated. Different types of features in the feature vectors were normalized using the Min-Max normalization method, mapping the feature values ​​to the [0, 1] interval. By comparing and analyzing the differences in the normalized lesion feature vectors before and after treatment, the volume change rate, morphological change index, and texture change index were calculated. The morphological change index is equal to half the sum of the sphericity change and the maximum diameter change, and the texture change index is equal to half the sum of the contrast change and the correlation change, thus constructing efficacy evaluation indicators. If the volume change rate exceeds a preset threshold and the morphological change index is greater than a preset threshold and the texture change index is greater than a preset threshold, then the therapeutic effect is determined to be good; if the volume change rate is less than a preset threshold and the morphological change index is less than a preset threshold and the texture change index is less than a preset threshold, then the therapeutic effect is determined to be poor.

[0041] Specifically, continuous multimodal images of the patient before and after treatment were acquired. A mutual information maximization algorithm was used to register the computed tomography (CT) and positron emission tomography (PET) images before and after treatment. Then, the level set method was used to segment the lesion region in the registered image. Based on the optimal feature combination, including volume, maximum diameter, sphericity, contrast of the gray-level co-occurrence matrix, and correlation, the segmented lesion region was quantitatively measured, and the feature vectors of the lesions before and after treatment were calculated. Different types of features were normalized using the Min-Max normalization method, mapping all feature values ​​to the [0,1] interval to ensure comparability. By comparing and analyzing the differences in lesion feature vectors before and after treatment, the volume change rate, morphological change index, and texture change index were calculated, where the morphological change index = (sphericity change + maximum diameter change) / 2, and the texture change index = (contrast change + correlation change) / 2. These were used to construct efficacy evaluation indicators and quantify the short-term efficacy of different treatment regimens. Based on clinical data statistics, efficacy evaluation thresholds were set to determine treatment effectiveness. If the lesion volume reduction rate exceeded 30%, the morphological change index was greater than 0.5, and the texture change index was greater than 0.6, the efficacy was considered good; conversely, if the volume reduction rate was less than 10%, the morphological change index was less than 0.2, and the texture change index was less than 0.3, the efficacy was considered poor. The evaluation results of different treatment plans were compared to determine the optimal treatment plan. In the evaluation of treatment efficacy for lung cancer patients, CT and PET images before and after treatment were first acquired. A mutual information maximization algorithm was used for registration, with a maximum iteration count of 200 and a convergence threshold of 0.001 to achieve accurate image alignment. Subsequently, the level set method was used to segment the lesions, with the initial contour set to a circle with a radius of 20 pixels, evolving through 500 iterations. Five key features were extracted from the segmentation results: volume, maximum diameter, sphericity, gray-level co-occurrence matrix contrast, and correlation. For example, a patient's feature vector before treatment was [5000, 25, 0.7, 0.8, 0.6], and after treatment it was [3000, 20, 0.8, 0.6, 0.7]. Normalizing the features to the [0, 1] interval, the normalized vector before treatment was [1, 1, 0.7, 1, 0.6], and after treatment it was [0.6, 0.8, 0.8, 0.75, 0.7]. Calculating the change indices, the volume reduction rate = (5000-3000) / 5000*100% = 40%, the morphological change index = (|0.8-0.7|+|20-25| / 25) / 2 = 0.15, and the texture change index = (|0.6-0.8|+|0.7-0.6|) / 2 = 0.15. The thresholds were set as follows: volume reduction >30% was considered effective, <10% was considered ineffective; morphological change >0.5 was considered effective, <0.2 was considered ineffective; texture change >0.6 was considered effective, <0.3 was considered ineffective. Based on the preset thresholds, the patient's assessment result was determined to be effective in volume reduction, but ineffective in morphological and texture changes. The overall assessment was that the treatment effect was average, and adjustment of the treatment plan was recommended.

[0042] Step S107: Based on the efficacy results, analyze the correlation between radiomics characteristics and patient prognosis, and estimate the risk of recurrence and metastasis and survival under different characteristic manifestations. Combine imaging characteristics and clinical indicators to construct an efficacy prediction model for long-term individualized assessment of patient prognosis.

[0043] Based on patient treatment outcomes and radiomics characteristics, a Cox proportional hazards regression model was used to analyze the correlation between features and patient prognosis, obtaining the hazard ratio and confidence interval for each feature. If a feature's p-value was less than a preset threshold, it was considered a significantly relevant feature and included in a feature subset. For this feature subset, Kaplan-Meier survival analysis was used for grouping, and the optimal cut-off point method was used to determine the grouping threshold for each feature, obtaining the recurrence and metastasis risk and median survival under different feature manifestations. The risk score obtained from the survival analysis was used as a new feature, combined with existing radiomics characteristics and clinical indicators, to construct a random forest prediction model. This random forest prediction model, through repeated sampling and multi-decision tree ensemble, provides long-term individualized prognosis assessment for patients. The random forest prediction model was used for prediction, and its performance was evaluated using cross-validation, calculating the C-index and the area under the time-dependent ROC curve. The C-index and the area under the ROC curve were used to determine the model's predictive accuracy and stability. Based on the prediction results of the random forest prediction model, feature importance ranking and partial dependency plots were generated. The feature importance ranking and partial dependency graph are used to interpret the model prediction results and identify the key features that have the greatest impact on prognosis.

[0044] Specifically, based on patient treatment outcomes and radiomics characteristics, a Cox proportional hazards regression model was used to analyze the correlation between features and patient prognosis. The hazard ratio and 95% confidence interval for each feature were calculated, and a p-value less than 0.05 was set as a significance threshold to select a subset of significantly correlated features. Using Kaplan-Meier survival analysis, the selected features were grouped, and the optimal cut-off point method was used to determine the grouping threshold for each feature. The risk of recurrence and metastasis and median survival under different feature manifestations were estimated, survival curves were plotted, and differences between groups were compared using the log-rank test. The risk score obtained from survival analysis was used as a new feature, combined with existing radiomics characteristics and clinical indicators, including age, gender, and tumor stage, to construct a random forest prediction model. The number of decision trees was set to 500, with a maximum depth of 10. Long-term individualized assessment of patient prognosis was performed through repeated sampling and multi-decision tree ensemble. The model performance was evaluated using 10-fold cross-validation, the C-index and area under the time-dependent ROC curve were calculated, and the 95% confidence interval was estimated using the bootstrap method to assess the model's predictive accuracy and stability. Feature importance ranking and partial dependency plots were generated to interpret model prediction results and identify key features with the greatest prognostic impact. In the prognostic analysis of lung cancer patients, Cox proportional hazards regression analysis was first performed on the radiomics features and 5-year follow-up data of 200 patients. A p-value threshold of 0.05 was set, and 10 significantly relevant features, such as tumor volume and texture heterogeneity, were selected from the initial 50 features. Subsequently, survival analysis was performed using the Kaplan-Meier method, and the X-tile algorithm was used to determine the optimal cut-off point, such as 15 cm for tumor volume. 3 Survival curves showed a 5-year survival rate of 30% for the large tumor volume group and 60% for the small tumor volume group. A random forest model was constructed by combining the risk score obtained from survival analysis with existing features, using 500 decision trees, a maximum depth of 10, and a minimum number of leaf nodes of 5. Model input included 10 radiomics features, risk score, and three clinical indicators: age, sex, and stage. Model performance was evaluated using 10-fold cross-validation, yielding an average C-index of 0.78 with a 95% confidence interval of 0.74–0.82. The time-dependent AUCs at 1, 3, and 5 years were 0.82, 0.79, and 0.75, respectively. The 95% confidence interval estimated using the bootstrap method showed good model stability. Feature importance analysis indicated that tumor volume, risk score, and texture heterogeneity were the three key factors influencing prognosis.

[0045] Step S108: Based on the efficacy prediction model, establish a predictive model for the relationship between early multimodal radiomics manifestations and patient prognosis and efficacy using multicenter clinical data. Through feature importance analysis and mining of the intrinsic correlation patterns between imaging features and prognosis, individualized predictions of the prognosis of new patients are made accordingly.

[0046] Multicenter clinical data was acquired, and missing values ​​were processed using multiple imputation. Data standardization was then performed using the Z-score method to obtain a preprocessed clinical dataset. For this preprocessed dataset, convolutional neural networks were used to extract omics features from early multimodal images. Combined with existing efficacy prediction models, a comprehensive dataset containing image features, clinical indicators, and prognostic information was constructed. If the comprehensive dataset contained multiple features, LASSO regression was used for feature selection, filtering out the most relevant feature subset. Information from different modalities was integrated using a multimodal feature fusion method to form a unified feature representation. Based on this unified feature representation, a prognostic efficacy prediction model was trained using a random forest algorithm. The parameters of the prognostic efficacy prediction model were optimized using grid search and cross-validation. Feature importance analysis was performed on the prognostic efficacy prediction model. The contribution of features to the prediction results was analyzed using SHAP values ​​to uncover the intrinsic correlation patterns between image features and prognosis, and interpretable prediction rules were constructed.

[0047] Specifically, preprocessing of multi-center clinical data included data cleaning, missing value handling, and standardization. Multiple imputation was used to handle missing values, and Z-scores were used for standardization to unify image acquisition parameters and clinical indicator standards across different centers, ensuring data quality and consistency. Convolutional neural networks were used to extract omics features from early multimodal images, and combined with existing efficacy prediction models, a comprehensive dataset containing image features, clinical indicators, and prognostic information was constructed. LASSO regression was used for feature selection, filtering out the most relevant feature subsets. A multimodal feature fusion method was employed to integrate information from different modalities, forming a unified feature representation. A random forest algorithm was used to train a prognostic efficacy prediction model based on the fused feature dataset. Model parameters were optimized through grid search and cross-validation to improve the model's generalization ability. Feature importance analysis was performed on the trained prediction model, using SHAP values ​​to analyze the contribution of features to the prediction results, uncovering the intrinsic correlation patterns between image features and prognosis, constructing interpretable prediction rules, and applying the model to new patient data to generate personalized prognostic prediction results. In the multi-center lung cancer prognostic prediction study, data from 2000 patients from 5 hospitals were first preprocessed. Multiple imputation was used to handle 15% of missing values, with 5 imputation iterations performed. All numerical features were standardized using the Z-score method, with the mean adjusted to 0 and the standard deviation adjusted to 1. CT and PET scan parameters from different centers were standardized, such as CT slice thickness being standardized to 1 mm. Subsequently, an 8-layer convolutional neural network was used to extract omics features from early multimodal images, extracting 256 features from each patient's CT and PET images. LASSO regression was used for feature selection, with a λ value set to 0.01, ultimately selecting the 50 most relevant features. An attention mechanism was used to fuse different modal features, resulting in a 100-dimensional fused feature vector. Based on the fused features, a random forest model was constructed with 500 decision trees and a maximum depth of 10. Parameters were optimized using 5-fold cross-validation and grid search, such as trying the minimum number of leaf node samples [1, 3, 5, 7, 9]. The final model achieved an AUC of 0.85 on the test set. SHAP analysis of feature importance revealed that tumor volume, SUVmax, and texture inhomogeneity were the top three prognostic factors. Construct a prediction rule if the tumor volume is >10cm 3 Furthermore, if SUVmax > 5, the 5-year survival rate decreases by 30%. The model was applied to 100 new patients to generate individualized prognostic prediction reports, including predicted 5-year survival rates and confidence intervals.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of correlation analysis of lung CT images, characterized by, The method comprises: acquiring lung multi-modal images including computed tomography images and positron emission tomography images, and performing denoising, enhancement and standardization preprocessing on the original images according to the imaging principles and image characteristics of different modalities; for the preprocessed multi-modal images, the complementary information between different modalities is utilized, the registration parameters are optimized by maximizing the mutual information between the images, and the accuracy of multi-modal image registration is quantitatively evaluated according to the coincidence degree of organ tissue anatomical structure in the registered images; the lung lesion region of interest is extracted from the registered multi-modal images, the lesion is segmented and drawn, and the volume and shape geometric morphological features of the lesion are quantitatively analyzed; for the segmented lesion region, multi-scale frequency domain texture features are extracted, second-order statistical texture features including contrast, correlation and entropy are obtained through a gray level co-occurrence matrix and a run-length matrix, and a multi-modal imaging feature vector is constructed in combination with the lesion morphological features; the multi-modal imaging feature vector is subjected to dimension reduction processing, a key feature subset related to curative effect is screened out, feature selection is realized by iteratively optimizing an L1 norm regularization term, the discrimination performance of the feature subset is evaluated in combination with cross validation, and an optimal feature combination is determined; based on the optimal feature combination, continuous multi-modal images before and after treatment of a patient are acquired, the registered and segmented lesion region is quantitatively measured, the dynamic changes of the lesion volume, shape and texture features are compared and analyzed, the curative effect difference of different treatment schemes in the near future is judged, and if the lesion volume is significantly reduced, the shape tends to be regular, and the texture tends to be uniform, the curative effect is good; based on the curative effect result, the correlation between the imaging features and the prognosis of the patient is analyzed, the recurrence and metastasis risk and the survival period under the performance of different features are estimated, the curative effect prediction model is constructed in combination of the imaging features and the clinical indexes, and the long-term individualized prognosis of the patient is evaluated; based on the curative effect prediction model, an early multi-modal imaging feature performance and patient prognosis curative effect prediction model is established according to multi-center clinical data, the feature importance analysis and the internal correlation mode between the imaging features and the prognosis are mined, and the prognosis of a new patient is individually predicted based on the same.

2. The method of claim 1, wherein, The acquisition of lung multi-modal images including computed tomography images and positron emission tomography images, and the denoising, enhancement and standardization preprocessing on the original images according to the imaging principles and image characteristics of different modalities, comprises: multi-modal image registration algorithm is adopted to register the acquired lung CT images and PET images, mutual information maximization criterion is selected as the similarity measure according to the spatial resolution and gray scale range difference of the CT images and PET images, principal component analysis method is used for coarse registration, affine transformation is applied for rigid registration, B-spline transformation is used for non-rigid registration, and the first CT image after registration is obtained; for the first CT image, anisotropic diffusion filtering algorithm is used for denoising processing, the anisotropic diffusion filtering algorithm adaptively adjusts the diffusion coefficient according to the local gradient of the image, and the second CT image after denoising is obtained; performing histogram equalization processing on the second CT image, enhancing contrast by stretching the dynamic range of image gray scale, and if the proportion of image contrast enhancement exceeds a preset threshold, performing local contrast enhancement using a contrast-limited adaptive histogram equalization algorithm to obtain a third CT image; According to the characteristics of the PET image, the Daubechies wavelet is used for multi-scale decomposition, the wavelet coefficients are processed by VisuShrink soft threshold value to remove noise, and the second PET image after denoising is obtained by inverse transformation reconstruction, Gaussian filter smoothing processing is applied to the second PET image, and edge enhancement is performed using the Laplacian operator to obtain a third PET image; The pixel values of the third CT image and the third PET image are converted into a standard normal distribution by using the Z-score standardization method to complete the standardization processing.

3. The method of claim 1, wherein, According to the preprocessed multi-modal images, the complementary information between different modalities is used to optimize the registration parameters by maximizing the mutual information between the images, and the accuracy of the multi-modal image registration is quantitatively evaluated according to the coincidence degree of the organ tissue anatomical structure in the registered images, including: Constructing a joint histogram according to the gray scale distribution characteristics of the preprocessed multi-modal images, and calculating the edge histogram of the joint histogram; The mutual information value between the two modal images is calculated by the normalized mutual information formula, and the normalized mutual information value is used as the optimization objective function; The gradient descent method is used to iteratively optimize the registration parameters, including the rotation angle, translation vector and scaling factor; If the change of the mutual information value is less than a preset threshold, the iteration is stopped; The multi-modal images are spatially transformed using the optimized registration parameters to obtain the registered images; A voxel-based registration evaluation method is used to extract the contours of the key anatomical structures from the registered images; The overlap coefficient and Hausdorff distance of the key anatomical structures are calculated, and the overlap coefficient is calculated using the Dice coefficient; A weighted scoring model is established, which integrates the overlap coefficient, Hausdorff distance and mutual information value; The registration result is quantitatively evaluated to obtain the registration accuracy score.

4. The method of claim 1, wherein, The lung lesion region of interest is extracted from the registered multi-modal images, the lesion is segmented and the volume and shape geometric features of the lesion are quantitatively analyzed, including: According to the registered multi-modal images, the region growing algorithm is used to extract the lung lesion region of interest to obtain the preliminary lesion contour; For the preliminary lesion contours, a fine segmentation is performed by a level set method, constructing an energy function where I is the image intensity, c1 and c2 are the foreground and background average intensities, is the Heaviside function; If the fine segmentation result contains noise, morphological operations are used to post-process the segmentation result, including removing small connected regions and smoothing the boundary; A connected region labeling algorithm is used to identify multiple lesions, calculate the volume of each lesion and record its spatial position coordinates; For the lesions, shape geometric features are extracted, including maximum diameter, short axis length, long-short axis ratio, sphericity, compactness and irregularity, the long-short axis is calculated using principal component analysis, and the sphericity is calculated by volume to surface area ratio.

5. The method of claim 1, wherein, The lesion region obtained by segmentation is extracted multi-scale frequency domain texture features, the second order statistical texture features are obtained by gray level co-occurrence matrix and run length matrix, including contrast, correlation and entropy, combined with lesion morphology features, a multi-modal imaging feature vector is constructed, including: The multi-scale wavelet transform is performed on the segmented lesion region, the Daubechies4 wavelet is used for 3-layer decomposition, and the frequency domain texture information of the lesion region at different scales is obtained; According to the frequency domain texture information, the energy, entropy and variance of each sub-band are calculated to obtain the frequency domain texture features; The mean, standard deviation, skewness and kurtosis of the frequency domain texture features are calculated to determine the multi-scale frequency domain texture features; The gray level co-occurrence matrix of the lesion region is calculated, and the contrast, correlation and entropy are extracted from the gray level co-occurrence matrix to determine the second order statistical texture features; The run length matrix of the computed tomography and positron emission tomography images is constructed respectively, and the run percentage, run length non-uniformity and gray non-uniformity are extracted from the run length matrix to obtain the run features; The multi-scale frequency domain texture features, the second order statistical texture features and the run features are combined with the lesion morphology features, and the combined features are standardized; An imaging feature vector containing multi-modal information is constructed according to the standardized features; Further comprising: based on the multi-scale frequency domain texture features of the lesion, the first order statistical features of the lesion region are extracted, including the mean, variance, skewness and kurtosis of the gray level, which describe the pixel gray level distribution inside the lesion; and the second order statistical features based on both the gray level co-occurrence matrix and the run length matrix are extracted, including the contrast, correlation and entropy.

6. The method of claim 5, wherein, Further comprising: based on the multi-scale frequency domain texture features of the lesion, the first order statistical features of the lesion region are extracted, including the mean, variance, skewness and kurtosis of the gray level, which describe the pixel gray level distribution inside the lesion; and the second order statistical features based on both the gray level co-occurrence matrix and the run length matrix are extracted, including the contrast, correlation and entropy. The multi-scale wavelet transform is performed on the lesion region image, the Haar wavelet basis function is used for three-layer decomposition, and the low-frequency approximation coefficient and three-direction high-frequency detail coefficient of the lesion region image are obtained; The energy, variance and entropy of the wavelet coefficient are calculated according to the low-frequency approximation coefficient and the high-frequency detail coefficient to obtain the multi-scale frequency domain texture features of the lesion region image; calculating the mean, variance, skewness and kurtosis of the gray scale of the pixels in the lesion area, wherein the skewness is E[(X-μ) 3 ] / σ 3 , the kurtosis is E[(X-μ) 4 ] / σ 4 -3, X is the gray scale, μ is the mean, and σ is the standard deviation, to determine the gray scale distribution characteristics of the pixels in the lesion The gray level co-occurrence matrix of the lesion region is constructed, the distance parameters are set to 1, 2 and 4 pixels, and the direction angles are set to 0°, 45°, 90° and 135°, the contrast, correlation and entropy are extracted from the gray level co-occurrence matrix to determine the second order statistical features of the lesion region; The run length matrix of the lesion region is generated, the continuous pixel number of each gray level is counted by gray level quantization of the lesion region image, the run length non-uniformity, gray non-uniformity and run percentage of the run length matrix are calculated to obtain the complete second order statistical feature set of the lesion region.

7. The method of claim 1, wherein, The multi-modal imageomics feature vector is subjected to dimension reduction processing, and a key feature subset related to curative effect is screened out, feature selection is realized by iteratively optimizing an L1 norm regularization term, the discrimination performance of the feature subset is evaluated in combination with cross-validation, and an optimal feature combination is determined, including: The multi-modal imageomics feature vector is subjected to standardization processing, the feature values are converted into a standard normal distribution by using a Z-score method, and a standardized feature matrix is obtained; According to the standardized feature matrix, principal component analysis is used for preliminary dimension reduction, the principal component contribution rate is calculated, and the cumulative contribution rate is accumulated until a preset threshold is reached, and a dimension-reduced feature matrix is obtained; For the dimension-reduced feature matrix, a linear regression model with an L1 norm regularization term is constructed, the model parameters are iteratively optimized by a coordinate descent method, and feature selection is realized; If a 5-fold cross-validation method is used to evaluate the discrimination performance of the feature subset, the optimal regularization parameter is selected from the preset values by adjusting the L1 regularization strength; The correlation coefficient of the feature and the curative effect index is calculated, the optimal feature combination is determined in combination with the L1 regularization result, and a key feature subset related to curative effect is determined.

8. The method of claim 1, wherein, Based on the optimal feature combination, continuous multi-modal images before and after treatment of a patient are obtained, the lesion region after registration and segmentation is quantitatively measured, the dynamic changes of the lesion volume, shape and texture features are compared and analyzed, the differences in short-term curative effect of different treatment schemes are judged, and if the lesion volume is significantly reduced, the shape tends to be regular, and the texture tends to be uniform, the curative effect is good, including: Continuous multi-modal images before and after treatment of a patient are obtained, mutual information maximization algorithm is used to register the computed tomography images and positron emission computed tomography images before and after treatment, and registered images are obtained; The level set method is used to segment the lesion region of the registered images, and the segmented lesion region is obtained; Based on the segmented lesion region, the optimal feature combination is used to quantitatively measure the segmented lesion region, the optimal feature combination includes volume, maximum diameter, sphericity, contrast and correlation of the gray level co-occurrence matrix, and the feature vectors of the lesions before and after treatment are calculated; The different types of features in the feature vectors are normalized, and the Min-Max normalization method is used to map the feature values to the interval [0, 1]; By comparing and analyzing the differences between the normalized lesion feature vectors before and after treatment, the volume change rate, the shape change index and the texture change index are calculated, the shape change index is equal to one-half of the sum of the sphericity change and the maximum diameter change, the texture change index is equal to one-half of the sum of the contrast change and the correlation change, and the curative effect evaluation index is constructed; If the volume change rate exceeds the preset threshold and the shape change index is greater than the preset threshold and the texture change index is greater than the preset threshold, the curative effect is good; If the volume change rate is less than the preset threshold and the shape change index is less than the preset threshold and the texture change index is less than the preset threshold, the curative effect is poor.

9. The method of claim 1, wherein, The correlation between the imaging features and the prognosis of the patients is analyzed based on the efficacy results, the risk of recurrence and metastasis and the survival time under different feature performances are estimated, the efficacy prediction model is constructed by combining the imaging features and the clinical indicators, and the long-term individualized evaluation of the prognosis of the patients is performed, including: The correlation between the features and the prognosis of the patients is analyzed by using the Cox proportional hazards regression model according to the efficacy results and the imaging features of the patients, and the hazard ratio and the confidence interval of each feature are obtained; If the P value of the feature is less than a preset threshold, the feature is determined as a significantly correlated feature, and is included in the feature subset; The Kaplan-Meier survival analysis method is used for grouping for the feature subset, the grouping threshold of each feature is determined by the best cut point method, and the risk of recurrence and metastasis and the median survival time under different feature performances are obtained; The risk score obtained by the survival analysis is used as a new feature, and the random forest prediction model is constructed by combining the original imaging features and the clinical indicators; The random forest prediction model performs long-term individualized evaluation of the prognosis of the patients through repeated sampling and multi-decision tree integration; The random forest prediction model is used for prediction, the model performance is evaluated by the cross-validation method, the C-index and the area under the time-dependent ROC curve are calculated; The C-index and the area under the ROC curve are used to judge the prediction accuracy and stability of the model; According to the prediction results of the random forest prediction model, the feature importance ranking and the partial dependence plot are generated; The feature importance ranking and the partial dependence plot are used to explain the prediction results of the model and determine the key features that have the greatest impact on the prognosis.

10. The method of claim 1, wherein, Based on the efficacy prediction model, the prediction model of the early multi-modal imaging features and the prognosis of the patients is established according to the multi-center clinical data, the internal correlation pattern between the imaging features and the prognosis is analyzed and mined through feature importance analysis, and the prognosis of new patients is individually predicted according to the internal correlation pattern, including: Multi-center clinical data are obtained, the missing values are processed by using the multiple imputation method according to the clinical data, and the data are standardized by using the Z-score method, so as to obtain a pre-processed clinical data set; The imaging features of the early multi-modal images are extracted by using the convolutional neural network for the pre-processed clinical data set, and a comprehensive data set containing the imaging features, the clinical indicators and the prognosis information is constructed by combining the existing efficacy prediction model; If the comprehensive data set contains multiple features, the LASSO regression is used for feature selection, the most relevant feature subset is screened out, the information of different modalities is integrated by using the multi-modal feature fusion method, and a unified feature representation is formed; The prognosis efficacy prediction model is trained by using the random forest algorithm according to the unified feature representation, and the parameters of the prognosis efficacy prediction model are optimized by using the grid search and the cross-validation; The feature importance analysis is performed on the prognosis efficacy prediction model, the contribution of the features to the prediction results is analyzed by using the SHAP value, the internal correlation pattern between the imaging features and the prognosis is mined, and the interpretable prediction rule is constructed.

Images