A method for analyzing correlation between lung CT lesions and related factors

Abstract

The application provides a lung CT lesion and related factor correlation analysis method, comprising the following steps: establishing a dynamic compensation model according to patient blood glucose level fluctuation data, adjusting CT scanning parameters through real-time monitoring of blood glucose value, compensating for tissue density changes in the scanning process, and obtaining corrected original image data; enhancing microvessel structure for the preliminary reconstructed CT image, extracting blood vessel features, calculating the probability of each voxel belonging to the blood vessel structure, and combining local structure tensor analysis to enhance and display the microvessel network; finely segmenting the interlobular septum and bronchiole wall of the microvessel enhanced image, extracting the fine structure of the lung through multi-scale feature fusion, and optimizing the segmentation result through post-processing; and performing abnormal detection on the entire lung region, extracting the characteristic features of normal and abnormal tissues, and combining a multi-instance learning strategy to accurately locate the suspicious lesion area, thereby providing auxiliary diagnosis information for early lung cancer screening.

Description

TECHNICAL FIELD

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

[0002] A key technical challenge in the CT diagnosis of interstitial lung disease is how to accurately capture and analyze the pathological changes of fine lung structures under complex physiological conditions. Fluctuations in blood glucose levels can affect the quality of CT imaging, resulting in image blurring and artifacts. At the same time, early lesions of interstitial lung disease often manifest as abnormalities in fine structures such as microvessels, interlobular septa, and bronchiolar walls, which are difficult to clearly display and quantitatively analyze in conventional CT images. In addition, due to the complex and diverse pathological changes of interstitial lung disease, it is difficult to achieve accurate positioning of suspicious lesions solely relying on traditional image processing methods. These factors collectively constrain the application effect of CT in the diagnosis of interstitial lung disease, particularly in early screening and differential diagnosis. How to improve the display capability of CT images for fine lung structures based on consideration of physiological factors, and achieve accurate identification and quantitative analysis of lesions through advanced image processing and artificial intelligence technology, is the core problem of improving the CT diagnosis level of interstitial lung disease. SUMMARY

[0003] The present application provides a lung CT lesion and related factor correlation analysis method, mainly comprising:

[0004] A dynamic compensation model is established according to patient blood glucose level fluctuation data, CT scanning parameters are adjusted by real-time monitoring of blood glucose values, and the density changes of tissues during scanning are compensated to obtain corrected original image data;

[0005] The corrected original image data is subjected to multi-scale decomposition, wavelet transform is used to extract different frequency components, high-frequency components are used to enhance detailed information, low-frequency components are used to retain overall structure, and nonlinear mapping functions are used to perform adaptive enhancement processing on each frequency band;

[0006] The enhanced multi-scale images are reconstructed, prior knowledge constraints are combined, noise and artifacts in the reconstruction process are suppressed by introducing a regularization term based on anatomical structure, and high-quality CT images are obtained after preliminary reconstruction;

[0007] The preliminary reconstructed CT images are subjected to microvessel structure enhancement, vessel features are extracted, the probability of each voxel belonging to the vessel structure is calculated, and the microvessel network is enhanced and displayed by combining local structure tensor analysis;

[0008] The microvessel-enhanced images are subjected to fine segmentation of interlobular septa and bronchiolar walls, fine lung structures are extracted by multi-scale feature fusion, and the segmentation results are optimized by post-processing.

[0009] The quantitative analysis is carried out on the fine structure of the segmented lung, the interlobular septal thickness and the percentage of bronchiole wall area are calculated, and the interstitial changes or airway remodeling are judged by comparing with the normal reference value;

[0010] The abnormal detection is carried out on the whole lung region, the characteristic features of normal and abnormal tissues are extracted, the suspicious lesion region is accurately positioned by combining the multi-instance learning strategy, and the auxiliary diagnosis information is provided for early lung cancer screening.

[0011] The technical scheme provided by the embodiment of the present application can include the following beneficial effects:

[0012] The present application discloses a lung CT lesion and related factor correlation analysis method. The method establishes a dynamic compensation model of blood glucose level fluctuation, adjusts the CT scanning parameters in real time, and corrects the original image. Multi-scale decomposition and adaptive enhancement processing are used to improve image quality, and iterative reconstruction algorithm is used to suppress noise and artifacts. For the reconstructed image, the present application enhances the microvascular structure and combines a deep learning model to finely segment the interlobular septum and bronchiole wall. By quantitatively analyzing these fine structures, it is determined whether there is interstitial change. Finally, the present application also uses an abnormal detection network based on attention mechanism to accurately position the suspicious lesion. This method can significantly improve the CT diagnostic accuracy of interstitial lung disease and provide important basis for early lung cancer screening. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The flowchart of the lung CT lesion and related factor correlation analysis method of the present application.

[0014] Figure 2 The schematic diagram of the lung CT lesion and related factor correlation analysis method of the present application.

[0015] Figure 3 The schematic diagram of the lung CT lesion and related factor correlation analysis method of the present application. DETAILED DESCRIPTION

[0016] The technical scheme in the embodiments of the present application will be described clearly and in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application.

[0017] As Figure 1 -3, the lung CT lesion and related factor correlation analysis method of the present application can specifically include:

[0018] S101, a dynamic compensation model is established according to the blood glucose level fluctuation data of the patient, the CT scanning parameters are adjusted by real-time monitoring of the blood glucose value, the tissue density change in the scanning process is compensated, and the corrected original image data is obtained.

[0019] A dynamic compensation model is established according to the relationship between historical blood glucose data and CT image quality, and a mapping function of blood glucose value and scanning parameter is obtained. The real-time blood glucose value of the patient is obtained, and the optimal CT scanning parameter corresponding to the current blood glucose level is calculated according to the mapping function. After receiving the CT scanning instruction, the original scanning data is preprocessed, including calculating the compensation coefficient according to the density change characteristics of different tissue types to correct the original scanning data. For the CT scanning data corrected by the compensation coefficient, the compensated CT image is generated. The effects of blood glucose fluctuation compensation are evaluated by comparing the CT images before and after compensation through signal-to-noise ratio, contrast-to-noise ratio and spatial resolution indicators.

[0020] Specifically, a dynamic compensation model is established according to the blood glucose level fluctuation data of the patient, the relationship between historical blood glucose data and CT image quality is analyzed by multiple linear regression, and a mapping function of blood glucose value and scanning parameter is obtained, which includes parameters such as tube voltage, tube current and scanning time. The tube voltage refers to the potential difference between the cathode and anode in the X-ray tube, and the tube current refers to the current intensity flowing through the X-ray tube. The subcutaneous implantable glucose sensor is used to monitor the blood glucose value of the patient in real time, and the data is collected every 5 minutes. According to the mapping function, the optimal CT scanning parameter corresponding to the current blood glucose level is calculated, and the parameters such as tube voltage, tube current and scanning time are dynamically adjusted. During the CT scanning process, the Kalman filter is used to process the original scanning data in real time, and the corresponding compensation coefficient is calculated according to the density change characteristics of different tissue types to correct the original data. For the corrected CT scanning data, the iterative reconstruction algorithm is used to generate the compensated CT image, and the effects of blood glucose fluctuation compensation are evaluated by comparing the image quality before and after compensation, wherein the image quality evaluation indicators include signal-to-noise ratio, contrast-to-noise ratio and spatial resolution. When establishing the dynamic compensation model, first, the historical blood glucose data and the corresponding CT image quality scores of 100 patients are collected to form a data set. Using multiple linear regression analysis, the blood glucose value is used as the independent variable, and the CT scanning parameters, including tube voltage, tube current and scanning time, are used as the dependent variable, to establish the mapping function. For example, the obtained mapping function may be in the form of tube voltage = 100 + 0.5 * blood glucose value, tube current = 200 - 0.3 * blood glucose value, and scanning time

[0021] = 0.5 + 0.001 * blood glucose value. Real-time blood glucose monitoring was performed using an Abbott FreeStyle Libre subcutaneous implantable glucose sensor, which collected data every 5 minutes. The data were transmitted to the CT scanner via Bluetooth. When the blood glucose value was detected to be 8.5 mmol / L, the optimal CT scanning parameters were calculated according to the mapping function as tube voltage 104.25 kV, tube current 197.45 mA, and scanning time 0.5085 seconds. During the CT scanning process, the original data were processed using a Kalman filter, the initial state estimate and error covariance matrix were set, and the state estimate was updated according to the tissue density variation characteristics. For soft tissue, the process noise covariance Q was set to 0.01, and the measurement noise covariance R was set to 0.1; for bone tissue, Q was set to 0.005, and R was set to 0.05. The compensation coefficient was obtained by iterative calculation, and the original data were corrected. Finally, the compensated CT image was generated using an iterative reconstruction algorithm, the maximum likelihood expectation algorithm was used, and the iteration number was set to 5 times and the relaxation factor was set to 0.5. The image quality was evaluated in terms of signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and spatial resolution, which was expressed in line pairs per centimeter. The indicators before and after compensation were compared, such as SNR increased from 15 to 18, CNR increased from 10 to 12, and spatial resolution increased from 10 line pairs per centimeter to 12 line pairs per centimeter, thereby verifying the effectiveness of blood glucose fluctuation compensation.

[0022] In S102, multi-scale decomposition is performed on the corrected original image data, wavelet transform is used to extract different frequency components, high-frequency components are used to enhance detailed information, low-frequency components are used to retain overall structure, and adaptive enhancement processing is performed on each frequency band by combining a nonlinear mapping function.

[0023] The corrected original image data is obtained, discrete wavelet transform is performed on the original image data to obtain low-frequency approximation coefficients and high-frequency detail coefficients. According to the high-frequency detail coefficients, an adaptive threshold function is set, wherein the parameters of the adaptive threshold function are determined by signal variance and noise variance. A nonlinear mapping function is applied to the low-frequency approximation coefficients, wherein the nonlinear mapping function is a Sigmoid function, and the parameters of the Sigmoid function are determined by the overall brightness distribution characteristics of the image. The nonlinear mapping function is applied to the high-frequency detail coefficients to realize adaptive enhancement of all frequency bands. The processed low-frequency approximation coefficients and high-frequency detail coefficients are reconstructed into enhanced images by inverse wavelet transform, wherein the contribution of each frequency band in the reconstruction process is determined by a weight coefficient based on the local variance of the image.

[0024] Specifically, wavelet transform is performed on the corrected original image data, Daubechies4 wavelet is used for discrete wavelet transform, the image is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients, and a multi-scale decomposition result is obtained. The number of decomposition layers is dynamically determined according to the image size, and is usually selected to be 3 to 5 layers. The high-frequency detail coefficients are enhanced by the BayesShrink method, and an adaptive threshold function is set where σ 2 is a noise variance, and σ x is a signal variance. The threshold parameter is dynamically adjusted according to the coefficient distribution characteristics under different scales, and the detail information in the image is extracted and enhanced. The low-frequency approximation coefficients are applied to a nonlinear mapping function, and the Sigmoid function is used as an S-shaped curve mapping function, where α controls the Sigmoid curve slope, and β controls the center point position of the Sigmoid curve. The values of the two parameters are determined according to the overall brightness distribution characteristics of the image. The nonlinear mapping is extended to each frequency band, and the corresponding nonlinear mapping is also applied to the high-frequency detail coefficients, so as to realize adaptive enhancement of the full frequency band. The processed low-frequency and high-frequency coefficients are reconstructed into an enhanced image by using inverse wavelet transform, and the contribution of each frequency band in the reconstruction process is adaptively adjusted by using a weight coefficient determination method based on the local variance of the image. The weight coefficient where σ i 2 is the local variance of the i-th subband, σ n 2 is the estimated noise variance, and adaptive fusion of information in different frequency bands is realized. For a corrected CT image of 512x512 pixels, Daubechies4 wavelet is first used for 4-layer discrete wavelet transform. In each layer of decomposition, the image is divided into approximation coefficients LL and three direction detail coefficients, that is, LH, HL and HH. For example, after the first layer of decomposition, 256x256 LL1, LH1, HL1 and HH1 subbands are obtained, and after the second layer of decomposition of LL1, 128x128 LL2, LH2, HL2 and HH2 subbands are obtained, and so on. The BayesShrink method is applied to enhance the high-frequency detail coefficients, assuming that the noise variance σ 2 = 20, for the 128x128 LH2 subband, the signal variance σ x = 50, then the threshold T = 20 / 50 = 0.4. The coefficients less than 0.4 in the LH2 subband are set to 0, and the coefficients greater than 0.4 are kept unchanged, so as to realize noise suppression and edge enhancement. For the low-frequency approximation coefficient LL4, the size of the image is 32x32, that is, the width and height of the image are both 32 pixels, a nonlinear mapping is performed on the low-frequency approximation coefficient LL4 by using the Sigmoid function, α = 0.05 and β = 128 are set, and assuming that the image grayscale range is 0-255, each pixel value x in LL4 is mapped to e is the base of the natural logarithm, which enhances the overall contrast of the image. Similar nonlinear mapping is extended to each high-frequency subband, but different α and β values ​​are used, such as setting α = 0.1 and β = 0 for the LH1 subband. During the reconstruction process, the weight coefficient is calculated based on the local variance of each subband. Assume that the local variance σ of the LL4 subband is LL4 2 =1000, estimated noise variance σ n 2 =20, then the weight coefficient w of LL4 LL4 =1000 / (1000+20)≈0.98. Similarly, the weight coefficients of other sub-bands are calculated and applied in the inverse wavelet transform process to achieve adaptive fusion and obtain the final enhanced image.

[0025] S103. Reconstruct the enhanced multi-scale image, combine the prior knowledge constraints, and introduce the anatomical structure-based regularization term to suppress the noise and artifacts in the reconstruction process, so as to obtain a preliminary reconstructed high-quality CT image.

[0026] An objective function including a data fidelity term, a total variation regularization term, and an anatomical structure-based regularization term is constructed, and the objective function is based on the principle of minimizing the total variation. The alternating direction multiplier method is used to optimize and solve the objective function to obtain a preliminary reconstructed image. The anatomical structure prior knowledge in the pre-trained U-Net convolutional neural network is obtained, and a structure-guided edge-preserving filter is constructed based on the anatomical structure prior knowledge. The structure-guided edge-preserving filter is used to adaptively filter the preliminary reconstructed image to obtain a filtered image that retains important anatomical structure information. It is judged whether the filtered image meets the preset termination condition, and the termination condition is whether the difference between the images obtained in two adjacent iterations is less than a preset threshold. After multiple iterative optimizations, a preliminary reconstructed high-quality CT image is obtained.

[0027] Specifically, based on the enhanced multi-scale image data, an objective function based on minimizing the total variation is constructed, and the function form is Where A is the system matrix, which describes the relationship between the image x and the observation data b, b is the projection data, that is, the measurement result, and x is the reconstructed image. is the gradient of image x, representing the edge information of the image, ||·||1 represents the L1 norm, and R(x) is an additional regularization term.

[0028] S104. Perform microvascular structure enhancement on the preliminarily reconstructed CT image, extract vascular features, calculate the probability of each voxel belonging to a vascular structure, and enhance the display of the microvascular network in combination with local structure tensor analysis.

[0029] The CT image is subjected to convolution operation, and multi-scale and multi-directional blood vessel feature responses are extracted; the probability value of each voxel belonging to the blood vessel structure is calculated, and a multi-scale vesselness response map is obtained; a structural tensor analysis method is adopted, and the principal direction and anisotropy information are determined through eigenvalue decomposition; if the vesselness response is greater than a preset threshold value, an adaptive anisotropic diffusion filter is constructed; according to the adaptive anisotropic diffusion filter, the microvessel network is subjected to iterative enhancement processing; in each iteration, the vesselness response and the structural tensor are updated; it is judged whether the iteration number reaches a preset value or the relative error of adjacent two iterations is less than a preset threshold value; if the conditions are met, the enhanced microvessel CT image is output.

[0030] Specifically, for the preliminarily reconstructed CT image, a multi-directional Gabor filter bank is constructed, 12 directions are set, including 0° to 165°, every 15°, and 4 scales, including 1 to 4 pixels, 48 filters of different directions and scales are generated by rotating and scaling the basic Gabor filter, the CT image is subjected to convolution operation, and multi-scale and multi-directional blood vessel feature responses are extracted. The Frangivesselness metric function

[0031]

[0032] V(σ) represents the Frangivesselness metric function, wherein S is calculated from the Hessian matrix eigenvalues λ1, λ2, λ3, Ra and Rb are curvature-related quantities calculated based on the Hessian matrix eigenvalues λ1, λ2, λ3, R a = min(|λ1|, |λ2|, |λ3|)

[0033] Ra represents the minimum curvature,

[0034] R b = max(|λ1|, |λ2|, |λ3|)

[0035] Rb represents the absolute value of the maximum eigenvalue, and σ represents a scale parameter for controlling the scale of the filter.

[0036]

[0037] , S represents a quantity calculated from the eigenvalues λ1, λ2, λ3 of the Hessian matrix, λ1, λ2, λ3 are eigenvalues of the Hessian matrix, S is a structure measure for measuring the anisotropy of local structure, and α, β, c are control parameters for adjusting the behavior of the measure function; three scale parameters σ are set, including 1 mm, 2 mm, and 3 mm corresponding to blood vessels of different diameters, the probability of each voxel belonging to a blood vessel structure is calculated, and a multi-scale vesselness response map is obtained. The image gradient in a 3x3x3 local neighborhood is calculated using the structure tensor analysis method, and a covariance matrix where J ij represents the covariance of image intensity in i and j directions, the principal direction and anisotropy information are obtained by eigenvalue decomposition, an adaptive anisotropic diffusion filter is constructed in combination with the vesselness response, and the filter coefficient is determined by the angle between the principal direction and the gradient direction and the vesselness value. The microvascular network is iteratively enhanced, the adaptive anisotropic diffusion filter is applied in each iteration, and the vesselness response and structure tensor are updated. The maximum number of iterations is set to 50 or the relative error between adjacent two iterations is less than 0.1% as the convergence condition. After the iteration, the contrast window width and window position are adjusted to highlight the microvascular structure, and the enhanced microvascular CT image is output. For a set of 512x512x256 voxel preliminary reconstruction CT images, 48 Gabor filters are first constructed, the direction is from 0° to 165° every 15°, the scale is 1, 2, 3, and 4 pixels, and the filter size is 21x21. The image is convolved to obtain 48 feature response maps. Then the Hessian matrix is calculated, and the scale parameter σ is set to 1 mm, 2 mm, and 3 mm corresponding to blood vessels of different diameters. The Frangivesselness function is used, the parameters α=0.5, β=0.5, and c=500, and the vesselness value of each voxel is calculated. For a 3x3x3 local neighborhood, the image gradient is calculated, the structure tensor is constructed, the principal direction vector v1 and the anisotropy λ1-λ2 / λ1 are obtained by eigenvalue decomposition. According to the principal direction and the vesselness value, an adaptive anisotropic diffusion filter is constructed, and the filter coefficient

[0038]

[0039] k represents the filter coefficient, v1 represents the principal direction, where represents image gradient, represents thermal conductivity, is set to 0.1 V, represents vesselness value, and exp is exponential function. The filter is iteratively applied, and vesselness and structure tensor are updated in each iteration. The maximum iteration number is set to 50, and the relative error threshold is set to 0.1%. After the iteration, the image grayscale value is mapped to the range of 0-255, the window width is set to 200 HU, and the window level is set to 100 HU, so as to highlight the microvessels. The processing process takes about 5 minutes on a workstation equipped with a GPU, and the final output is an enhanced microvessel CT image. Compared with the original image, the contrast noise ratio of the microvessels is increased by about 40%, and the edge definition is improved by about 25%.

[0040] S105, fine segmentation of interlobular septa and bronchiole walls is performed on the microvessel enhanced image, fine structures of the lung are extracted through multi-scale feature fusion, and the segmentation result is optimized through post-processing.

[0041] The CT image after microvessel enhancement is obtained, and a first image is obtained from the CT image through a contrast limited adaptive histogram equalization algorithm. The first image is processed to obtain a second image, wherein the spatial domain standard deviation of the bilateral filtering is a preset value A, and the value domain standard deviation is a preset value B. A U-Net convolutional neural network model is constructed, the U-Net convolutional neural network model includes at least one encoder layer and at least one decoder layer, and if the number of the encoder layers is N, the number of the decoder layers is N. The U-Net convolutional neural network model is applied to the second image to obtain a preliminary segmentation result, and the preliminary segmentation result includes segmentation information of the interlobular septa and the bronchiole walls. Morphological operations including opening operation and closing operation are performed on the preliminary segmentation result to obtain an optimized segmentation result. The skeleton structure of the optimized segmentation result is extracted to obtain an optimized segmentation mask. A conditional random field model is constructed, and the conditional random field model includes a unary potential function and a binary potential function, wherein the unary potential function adopts a log-likelihood ratio, and the binary potential function adopts a contrast-sensitive Potts model. The conditional random field model is applied to the optimized segmentation mask to obtain a final segmentation result, and the final segmentation result includes fine segmentation information of the interlobular septa and the bronchiole walls.

[0042] Specifically, the microvascular enhanced CT image is preprocessed, the image contrast is enhanced by contrast limited adaptive histogram equalization (CLAHE), the block size is set to 8x8, the contrast limit parameter is set to 0.01, the bilateral filter is used to remove noise while preserving edge information, the spatial domain standard deviation is set to 3, and the value domain standard deviation is set to 75, to obtain the preprocessed image. A U-Net convolutional neural network model is constructed, a 4-layer encoder-decoder structure is designed, two 3x3 convolution kernels are used in each layer, the activation function is ReLU, the pooling layer is 2x2 max pooling, the feature information of different levels is fused through a feature pyramid network (FPN), and the preliminary segmentation of the interlobular septa and the bronchiovascular wall is realized. Morphological operations are applied to the segmentation results output by the neural network, including open operation using a 3x3 circular structural element to remove small missegmentation regions, close operation using a 5x5 circular structural element to fill small holes, and Zhang-Suen thinning algorithm to extract the skeleton structure, to obtain the optimized segmentation mask. Conditional random fields are used to refine the segmentation results, the image is divided into superpixels with a size of 10x10, a potential function based on pixel intensity and spatial position is constructed, the unary potential function uses a log-likelihood ratio, and the binary potential function uses a contrast-sensitive Potts model, the energy function is optimized by mean field inference algorithm for 10 iterations to obtain the final interlobular septa and bronchiovascular wall segmentation results. For a group of 512x512x256 voxel microvascular enhanced CT images, first, the CLAHE algorithm is applied, the image is divided into 64 8x8 blocks, the contrast limit parameter is set to 0.01, and the histogram equalized image is obtained. Then, the bilateral filter is used, the spatial domain σ d = 3, the value domain σ r= 75, window size is 21x21, and the processed image noise is reduced by about 30%, and the edge sharpness is maintained above 95%. Then the processed image is input into the U-Net model, which includes 4 down-sampling and 4 up-sampling layers, each layer has two 3x3 convolutions, the initial channel number is 64, and each layer is doubled to a maximum of 512. Using the FPN structure, the feature maps of different scales are adjusted to the same channel number 256 through 1x1 convolution, and then up-sampling and element-wise addition fusion are performed. The model is trained on 2000 labeled images for 100 rounds, with a batch size of 8 and a learning rate of 0.001, using the Adam optimizer. The probability map output by the U-Net is thresholded, with a threshold of 0.5, to obtain the preliminary segmentation result. A 3x3 circular structural element is applied to the open operation to remove areas smaller than 10 pixels, and then a 5x5 circular structural element is used for the close operation to fill holes smaller than 20 pixels. The Zhang-Suen algorithm is used for thinning, and about 15 iterations are performed to obtain a 1-pixel-wide skeleton. Finally, the image is divided into about 5000 superpixels using the fast SLIC algorithm, and a conditional random field is constructed. The unary potential uses -log(p), and p is the probability value output by the U-Net. The binary potential uses the contrast-sensitive Potts model with a weight of 30. Through 10 iterations of mean field inference optimization, the final segmentation result is obtained. The entire processing process takes about 3 minutes on a workstation equipped with a GPU. Compared with using U-Net alone, the segmentation accuracy of interlobular septa and bronchiole walls is improved by about 8%, and the Dice coefficient reaches 0.92.

[0043] S106, quantitative analysis is performed on the segmented fine lung structures to calculate the interlobular septum thickness and bronchiole wall area percentage indicators, and to determine whether there is interstitial change or airway remodeling by comparing with normal reference values.

[0044] The interlobular septum structure is processed using the Zhang-Suen thinning algorithm, and the shortest distance from each point on the skeleton to the edge is calculated using the Euclidean distance transform algorithm to obtain the interlobular septum thickness. Morphological dilation and erosion operations are performed on the bronchiole structure to extract the bronchial wall area, and the bronchiole wall area percentage is calculated by calculating the ratio of the bronchial wall area to the corresponding total bronchial area. According to the CT scan data of healthy volunteers, a normal reference value database is established to obtain interlobular septum thickness and bronchiole wall area percentage data. The Z-score method is used to calculate the deviation of the test case from the normal reference value, where x is the test value, μ and σ are the mean and standard deviation of the reference data, respectively. If the Z-score of the interlobular septum thickness is greater than a first preset threshold, it is determined to be interstitial change; if the Z-score of the bronchiole wall area percentage is greater than a second preset threshold, it is determined to be airway remodeling. According to the numerical range of the Z-score, the degree of interstitial change and airway remodeling is determined.

[0045] Specifically, the interlobular septa structure obtained by segmentation is processed by Zhang-Suen thinning algorithm, and the shortest distance from each point on the skeleton to the edge is calculated by Euclidean distance transform algorithm, and twice the distance is taken as the thickness of the interlobular septa, the thickness distribution is obtained by kernel density estimation, and the average thickness and standard deviation are calculated. For the segmented bronchiole structure, morphological dilation and erosion operations are performed using a 3x3 circular structural element to extract the bronchial wall area, and the ratio of the bronchial wall area to the corresponding total bronchial area is calculated to obtain the bronchiole wall area percentage. A normal reference database is constructed, containing CT scan data of 1000 healthy volunteers of different ages, genders and body types, the interlobular septa thickness and bronchiole wall area percentage data are obtained using the same processing flow as the test case, and Shapiro-Wilk test is used to ensure that the data is normally distributed. The deviation of the test case from the normal reference value is calculated using the Z-score method, Z-score=(x-μ) / σ, where x is the test value, μ and σ are the mean and standard deviation of the reference data, respectively. According to the clinical research data, the threshold is set to judge the interstitial changes and airway remodeling: if the interlobular septa thickness Z-score is greater than 2.5, it is judged as interstitial changes, 2.5Z-score<3.5 is mild, 3.5≤Z-score<5 is moderate, and Z-score≥5 is severe; if the bronchiole wall area percentage Z-score is greater than 2.0, it is judged as airway remodeling, 2.0Z-score<4 is moderate, and Z-score≥4 is severe, providing quantitative basis for the diagnosis of interstitial lung disease. For a set of 512x512x256 voxel lung CT images, first apply the Zhang-Suen thinning algorithm to the interlobular septa structure obtained by segmentation, and iterate about 15 times to get a single-pixel width skeleton. Then use Euclidean distance transform to calculate the distance from each point on the skeleton to the edge, with an average processing time of about 2 seconds. Multiply the distance value by 2 to get the thickness, apply kernel density estimation (bandwidth 0.1mm) to obtain the thickness distribution, calculate the average thickness 1.2mm, and the standard deviation 0.3mm. For the bronchiole structure, use a 3x3 circular structural element for 3 times of dilation and 2 times of erosion operation to extract the bronchial wall area, calculate the area ratio, and obtain the average bronchiole wall area percentage 23%. The normal reference database contains 1000 volunteers, aged 18-80, half male and half female, and Shapiro-Wilk test shows that p>0.05, confirming the normality of the data. The reference value of the interlobular septa thickness is 1.0mm, and the standard deviation is 0.2mm; the reference value of the bronchiole wall area percentage is 20%, and the standard deviation is 2%. For a test case, the interlobular septa thickness is 1.8mm, and the bronchiole wall area percentage is 28%, the Z-score is 4.0 and 4.0 respectively.According to the set threshold, the case is judged as moderate interstitial change and moderate airway remodeling, suggesting a higher possibility of interstitial lung disease, and further clinical evaluation is recommended.

[0046] S107, abnormality detection is performed on the whole lung region, and the characteristic features of normal and abnormal tissues are extracted, and a multi-instance learning strategy is combined to accurately locate the suspicious lesion region, thereby providing auxiliary diagnostic information for early lung cancer screening.

[0047] A lung CT image is received and preprocessed, and multi-scale features are extracted from the preprocessed lung CT image; the features of different spatial positions are weighted and fused through a Transformer self-attention mechanism according to the multi-scale features to obtain a global feature representation; a contrast learning method is used to construct a feature pair of normal tissue features and abnormal tissue features, and the distance between the feature pair is calculated; if the feature pair is a normal-abnormal pair, the distance between the feature pair is maximized, and if the feature pair is a normal-normal pair, the distance between the feature pair is minimized; the features obtained by contrast learning are input into a multi-instance learning framework, which is driven based on an attention mechanism, and the lung CT image is divided into overlapping three-dimensional blocks; abnormality measurement calculation is performed on the three-dimensional blocks obtained by division to obtain a high abnormality degree region; fine positioning is performed on the high abnormality degree region, and a morphological operation and a region growing algorithm are used to extract the outline of the suspicious lesion.

[0048] Specifically, the input lung CT image is preprocessed, and 3DResNet-50 is used to extract multi-scale features. The features of different spatial positions are weighted and fused through the Transformer self-attention mechanism to obtain the global feature representation. By using the contrast learning method, the feature pairs of normal and abnormal tissues are constructed, the InfoNCE loss function is used to maximize the distance between normal-abnormal pairs and minimize the distance between normal-normal pairs, and the informative samples are selected by the hard example mining strategy to learn the discriminative feature representation. The features obtained by contrast learning are used as the input of multi-instance learning, and the lung image is divided into 64x64x64 overlapping three-dimensional blocks based on the attention-driven multi-instance learning framework. Each block is taken as an instance, the Gated-Attention mechanism is used to calculate the abnormality measurement at the instance level, the noisy-or pooling function is used for bag-level aggregation, and the overall abnormality degree is calculated. The high abnormality region detected is finely positioned, the morphological opening and closing operation is performed combined with the spherical structure element with a radius of 3, the region growing algorithm is used with a threshold of ±50HU to extract the accurate contour of the suspicious lesion, and the quantitative features such as volume, shape and texture are calculated. According to the abnormality threshold, the high-risk area is selected with an abnormality threshold of 0.8, and the quantitative features are combined to generate early lung cancer screening suggestions, and the abnormality detection results, positioning information and screening suggestions are output. For a set of 512x512x256 voxels of lung CT image, 3DResNet-50 is used to extract features, and the network contains 5 stages, each stage outputs 32, 64, 128, 256 and 512 channel feature maps respectively. Then the features are fused through 8 heads of Transformer self-attention mechanism to obtain a 2048-dimensional global feature vector. In the contrast learning stage, 100 normal patches and 100 abnormal patches are randomly sampled from each CT scan to construct 10000 feature pairs. The InfoNCE loss function with a temperature parameter of 0.07 is used for optimization, and the hard negative mining strategy is used to select the top 20% difficult samples. The learned features are input into the attention-driven multi-instance learning framework, and the lung image is divided into 3375 overlapping three-dimensional blocks with a size of 64x64x64. A two-layer fully connected network (128-64-1) is used as the Gated-Attention mechanism to calculate the abnormality score of each instance, and the noisy-or pooling function (p=0.7) is used for aggregation to obtain the overall abnormality degree. The regions with abnormality degree exceeding 0.8 are finely positioned, the spherical structure element with a radius of 3 voxels is used for 3 times of morphological opening operation and 2 times of closing operation, and then the region growing algorithm is applied with a threshold of ±50HU and a minimum voxel number of 100 to extract the lesion contour. 18 3D image features are calculated, including volume, sphericity, gray level co-occurrence matrix features, etc.Finally, a random forest classifier with 100 decision trees and a maximum depth of 10 was used to generate a risk score for early stage lung cancer between 0 and 100 based on the abnormality scores and quantitative features.

[0049] The above description is merely that of the preferred embodiments of the application and of the principles thereof. It is to be understood that the present application is not limited to the specific technical features described above and that many modifications, changes and substitutions in accordance with the spirit of the application can be made by those skilled in the art without departing from the application. For example, the features described above can be replaced by other technical features disclosed in the present application (but not limited to) having similar functions to form other technical solutions.

Claims

1. A method for analyzing the correlation between lesions of lung CT and related factors, characterized in that, The method comprises: establishing a dynamic compensation model according to patient blood glucose level fluctuation data, adjusting CT scan parameters by real-time monitoring of blood glucose values, compensating for tissue density changes during scanning to obtain corrected original image data; multi-scale decomposition is performed on the corrected original image data, different frequency components are extracted using wavelet transform, detail information is enhanced through high-frequency components, overall structure is retained through low-frequency components, and each frequency band is adaptively enhanced by combining a nonlinear mapping function; the enhanced multi-scale image is reconstructed, prior knowledge constraints are combined, and a regularization term based on anatomical structure is introduced to suppress noise and artifacts during reconstruction to obtain a preliminary reconstructed high-quality CT image; the preliminary reconstructed CT image is enhanced for microvessel structure, the blood vessel features are extracted, the probability of each voxel belonging to the blood vessel structure is calculated, and the microvessel network is enhanced and displayed by combining local structure tensor analysis; the microvessel enhanced image is finely segmented for interlobular septa and bronchiole walls, the lung fine structure is extracted by multi-scale feature fusion, and the segmentation result is optimized by post-processing; the lung fine structure obtained by segmentation is quantitatively analyzed, the indexes of interlobular septal thickness and bronchiole wall area percentage are calculated, and whether there is interstitial change or airway remodeling is judged by comparing with the normal reference value; abnormal detection is performed on the whole lung region, the characteristic features of normal and abnormal tissues are extracted, and the suspicious lesion area is accurately positioned by combining a multi-instance learning strategy.

2. The method of claim 1, wherein, The method comprises: establishing a dynamic compensation model according to patient blood glucose level fluctuation data, adjusting CT scan parameters by real-time monitoring of blood glucose values, compensating for tissue density changes during scanning to obtain corrected original image data; multi-scale decomposition is performed on the corrected original image data, different frequency components are extracted using wavelet transform, detail information is enhanced through high-frequency components, overall structure is retained through low-frequency components, and each frequency band is adaptively enhanced by combining a nonlinear mapping function; the enhanced multi-scale image is reconstructed, prior knowledge constraints are combined, and a regularization term based on anatomical structure is introduced to suppress noise and artifacts during reconstruction to obtain a preliminary reconstructed high-quality CT image; the preliminary reconstructed CT image is enhanced for microvessel structure, the blood vessel features are extracted, the probability of each voxel belonging to the blood vessel structure is calculated, and the microvessel network is enhanced and displayed by combining local structure tensor analysis; the microvessel enhanced image is finely segmented for interlobular septa and bronchiole walls, the lung fine structure is extracted by multi-scale feature fusion, and the segmentation result is optimized by post-processing; the lung fine structure obtained by segmentation is quantitatively analyzed, the indexes of interlobular septal thickness and bronchiole wall area percentage are calculated, and whether there is interstitial change or airway remodeling is judged by comparing with the normal reference value; abnormal detection is performed on the whole lung region, the characteristic features of normal and abnormal tissues are extracted, and the suspicious lesion area is accurately positioned by combining a multi-instance learning strategy.

3. The method of claim 1, wherein, The method comprises the following steps: obtaining the corrected original image data, performing discrete wavelet transform on the original image data to obtain low-frequency approximation coefficients and high-frequency detail coefficients; setting an adaptive threshold function according to the high-frequency detail coefficients, wherein the parameters of the adaptive threshold function are determined by signal variance and noise variance; applying a nonlinear mapping function to the low-frequency approximation coefficients, wherein the nonlinear mapping function is a Sigmoid function, and the parameters of the Sigmoid function are determined by the overall brightness distribution characteristics of the image; applying the nonlinear mapping function to the high-frequency detail coefficients to realize adaptive enhancement of all frequency bands; and reconstructing the processed low-frequency approximation coefficients and the high-frequency detail coefficients into an enhanced image by inverse wavelet transform, wherein the contribution of each frequency band in the reconstruction process is determined by a weight coefficient based on the local variance of the image.

4. The method of claim 1, wherein, The method comprises the following steps: constructing a target function including a data fidelity term, a total variation regularization term and an anatomical structure regularization term, wherein the target function is based on the total variation minimization principle; performing optimization and solving of the target function by using an alternating direction multiplier method to obtain a preliminary reconstructed image; obtaining anatomical structure prior knowledge in a pre-trained U-Net convolutional neural network, and constructing a structure-guided edge-preserving filter according to the anatomical structure prior knowledge; performing adaptive filtering processing on the preliminary reconstructed image by using the structure-guided edge-preserving filter to obtain a filtered image that retains important anatomical structure information; judging whether the filtered image satisfies a preset termination condition, wherein the termination condition is whether the difference between images obtained by two adjacent iterations is less than a preset threshold; and obtaining a preliminary reconstructed high-quality CT image after multiple iterations of optimization.

5. The method according to claim 1, wherein The method comprises the following steps: performing convolution operation on the CT image to extract multi-scale and multi-directional blood vessel feature response; calculating the probability value of each voxel belonging to the blood vessel structure to obtain a multi-scale vesselness response map; using a structure tensor analysis method, determining the principal direction and anisotropy information through eigenvalue decomposition; if the vesselness response is greater than a preset threshold, an adaptive anisotropic diffusion filter is constructed; the microvascular network is iteratively enhanced according to the adaptive anisotropic diffusion filter; in each iteration, the vesselness response and the structure tensor are updated; it is judged whether the iteration number reaches a preset value or the relative error of adjacent two iterations is less than a preset threshold; if the conditions are met, the enhanced microvascular CT image is output.

6. The method of claim 1, wherein, The method comprises the following steps: obtaining the microvascular enhanced CT image, and obtaining a first image by using a contrast limited adaptive histogram equalization algorithm on the CT image; performing bilateral filtering on the first image to obtain a second image, wherein the spatial domain standard deviation of the bilateral filtering is a preset value A, and the value domain standard deviation is a preset value B; constructing a U-Net convolutional neural network model, wherein the U-Net convolutional neural network model comprises at least one encoder layer and at least one decoder layer, and if the number of the encoder layers is N, the number of the decoder layers is N; applying the U-Net convolutional neural network model to the second image to obtain a preliminary segmentation result, wherein the preliminary segmentation result comprises the segmentation information of the interlobular septa and the bronchiole walls; performing morphological operations on the preliminary segmentation result, including opening operation and closing operation, to obtain an optimized segmentation result; extracting the skeleton structure of the optimized segmentation result to obtain an optimized segmentation mask; constructing a conditional random field model, wherein the conditional random field model comprises a unary potential function and a binary potential function, the unary potential function adopts a log-likelihood ratio, and the binary potential function adopts a contrast-sensitive Potts model; applying the conditional random field model to the optimized segmentation mask to obtain a final segmentation result, wherein the final segmentation result comprises the fine segmentation information of the interlobular septa and the bronchiole walls.

7. The method of claim 1, wherein, The fine structure of the lung obtained by segmentation is quantitatively analyzed, the interlobular septal thickness and the percentage of bronchiole wall area are calculated, and by comparing with the normal reference value, it is judged whether there is interstitial change or airway remodeling, including: extracting the skeleton of the interlobular septal structure, calculating the shortest distance from each point on the skeleton to the edge to obtain the interlobular septal thickness; morphological dilation and erosion operation is carried out on the bronchiole structure, the bronchial wall area is extracted, the ratio of bronchial wall area to corresponding total bronchial area is calculated, and the percentage of bronchiole wall area is obtained; a normal reference value database is established, and the interlobular septal thickness and the percentage of bronchiole wall area data are obtained; the deviation degree of the case to be detected and the normal reference value is calculated by using Z-score method, Z-score=(x-μ) / σ, wherein x is the value to be detected, μ and σ are the mean and standard deviation of the reference data respectively; if the Z-score of the interlobular septal thickness is greater than the first preset threshold, it is judged that there is interstitial change; if the Z-score of the percentage of bronchiole wall area is greater than the second preset threshold, it is judged that there is airway remodeling; according to the numerical range of Z-score, the degree of interstitial change and airway remodeling is determined.

8. The method of claim 1, wherein, The whole lung region is detected for abnormalities, the characteristic features of normal and abnormal tissues are extracted, and a multi-instance learning strategy is combined to accurately locate the suspicious lesion region, including: receiving a lung CT image and pre-processing the lung CT image, and extracting multi-scale features from the pre-processed lung CT image; according to the multi-scale features, the features of different spatial positions are weighted and fused by the Transformer self-attention mechanism to obtain global feature representation; a feature pair of normal tissue features and abnormal tissue features is constructed by using a contrast learning method, and the distance between the feature pair is calculated; if the feature pair is a normal-abnormal pair, the distance between the feature pair is maximized, and if the feature pair is a normal-normal pair, the distance between the feature pair is minimized; the features obtained by contrast learning are input into a multi-instance learning framework, and the multi-instance learning framework is driven based on an attention mechanism to divide the lung CT image into overlapping three-dimensional blocks; the three-dimensional blocks obtained by division are calculated for abnormality measurement to obtain a high abnormality degree region; the high abnormality degree region is finely located, and a morphological operation and a region growing algorithm are used to extract a suspicious lesion contour.

Images