Lung CT lesion and related factor correlation analysis method

Abstract

The invention provides a lung CT lesion and related factor correlation analysis method, which comprises the following steps: establishing a dynamic compensation model according to blood glucose level fluctuation data of a patient, adjusting CT scanning parameters by monitoring a blood glucose value in real time, and compensating tissue density change in a scanning process to obtain corrected original image data; performing microvascular structure enhancement on the preliminarily reconstructed CT image, extracting blood vessel features, calculating the probability that each voxel belongs to a blood vessel structure, and performing enhanced display on a microvascular network in combination with local structure tensor analysis; performing fine segmentation of the leaflet spacing and the microbronchus wall on the image after the microvessel enhancement, extracting a fine structure of the lung through multi-scale feature fusion, and performing post-processing to optimize a segmentation result; and performing anomaly detection on the whole lung area, extracting characterization characteristics of normal and abnormal tissues, and accurately positioning a suspicious focus area in combination with a multi-instance learning strategy, thereby providing auxiliary diagnosis information for early lung cancer screening.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for analyzing the association between lung CT lesions and related factors. Background Art

[0002] A key technical challenge facing CT diagnosis of interstitial lung disease is how to accurately capture and analyze the pathological changes of lung microstructures under complex physiological conditions. Fluctuations in blood sugar levels can affect the quality of CT imaging, resulting in blurred images 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 complexity and diversity of the pathological changes of interstitial lung disease, it is difficult to accurately locate suspicious lesions by relying solely on traditional image processing methods. These factors jointly restrict the application effect of CT in the diagnosis of interstitial lung disease, especially in early screening and differential diagnosis. How to improve the display ability of CT images of lung microstructures and realize accurate identification and quantitative analysis of lesions through advanced image processing and artificial intelligence technologies on the basis of considering the influence of physiological factors is the core issue in improving the level of CT diagnosis of interstitial lung disease. Summary of the invention

[0003] The present invention provides a method for analyzing the association between lung CT lesions and related factors, which mainly includes:

[0004] A dynamic compensation model is established based on the patient's blood sugar level fluctuation data. The CT scanning parameters are adjusted by real-time monitoring of blood sugar values ​​to compensate for changes in tissue density during the scanning process and obtain corrected original image data.

[0005] The corrected original image data is decomposed at multiple scales, and different frequency components are extracted by wavelet transform. The detail information is enhanced by high-frequency components, and the overall structure is retained by low-frequency components. Each frequency band is adaptively enhanced by combining nonlinear mapping functions.

[0006] Reconstruct the enhanced multi-scale image, combine the prior knowledge constraints, and introduce the regularization term based on the anatomical structure to suppress the noise and artifacts in the reconstruction process, so as to obtain the preliminary reconstructed high-quality CT image;

[0007] The microvascular structure of the initially reconstructed CT image is enhanced, the vascular features are extracted, the probability of each voxel belonging to the vascular structure is calculated, and the microvascular network is enhanced and displayed in combination with local structural tensor analysis;

[0008] Fine segmentation of lobular septa and bronchiolar walls is performed on microvascular enhanced images, and fine lung structures are extracted through multi-scale feature fusion, followed by post-processing to optimize the segmentation results.

[0009] Quantitative analysis of the segmented lung fine structure was performed to calculate the thickness of the interlobular septum and the percentage of the bronchiolar wall area. By comparing them with normal reference values, it was determined whether there were interstitial changes or airway remodeling.

[0010] Abnormalities are detected in the entire lung area, the characterization features of normal and abnormal tissues are extracted, and the suspicious lesion area is accurately located by combining multi-instance learning strategies to provide auxiliary diagnostic information for early lung cancer screening.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a method for analyzing the association between lung CT lesions and related factors. The method establishes a dynamic compensation model for blood sugar level fluctuations, adjusts 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 an iterative reconstruction algorithm is used to suppress noise and artifacts. For the reconstructed image, the present invention enhances the microvascular structure and combines a deep learning model to finely segment the interlobular septum and bronchiolar wall. By quantitatively analyzing these fine structures, it is determined whether there are interstitial changes. Finally, the present invention also uses an anomaly detection network based on an attention mechanism to accurately locate suspicious lesions. The method can significantly improve the accuracy of CT diagnosis of interstitial lung disease and provide an important basis for early lung cancer screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a method for analyzing the association between lung CT lesions and related factors.

[0014] Figure 2 It is a schematic diagram of a method for analyzing the association between lung CT lesions and related factors according to the present invention.

[0015] Figure 3 It is another schematic diagram of a method for analyzing association between lung CT lesions and related factors according to the present invention. DETAILED DESCRIPTION

[0016] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0017] like Figure 1 -3. In this embodiment, a method for analyzing the association between lung CT lesions and related factors may specifically include:

[0018] S101. A dynamic compensation model is established based on the patient's blood sugar level fluctuation data. The CT scanning parameters are adjusted by real-time monitoring of the blood sugar value to compensate for the change in tissue density during the scanning process, thereby obtaining corrected original image data.

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

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

[0021] =0.5+0.001*blood glucose value. Real-time blood glucose monitoring uses Abbott FreeStyle Libre subcutaneous implantable glucose sensor, collects data every 5 minutes, and transmits the data to the CT scanner via Bluetooth. When the blood glucose value is detected to be 8.5mmol / L, the optimal CT scanning parameters are calculated according to the mapping function as tube voltage 104.25kV, tube current 197.45mA, and scanning time 0.5085 seconds. During the CT scanning process, the Kalman filter is used to process the raw data, set the initial state estimate and error covariance matrix, and update the state estimate according to the characteristics of tissue density change. For soft tissue, the process noise covariance Q=0.01 and the measurement noise covariance R=0.1 are set; for bone tissue, Q=0.005 and R=0.05 are set. The compensation coefficient is obtained by iterative calculation to correct the raw data. Finally, the iterative reconstruction algorithm is used to generate the compensated CT image. The maximum likelihood expectation algorithm is used, the number of iterations is set to 5, and the relaxation factor is set to 0.5. The image quality was evaluated by calculating the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and spatial resolution, which were measured 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 sugar fluctuation compensation.

[0022] S102, performing multi-scale decomposition on the corrected original image data, using wavelet transform to extract different frequency components, enhancing detail information through high-frequency components, retaining the overall structure of low-frequency components, and combining nonlinear mapping functions to perform adaptive enhancement processing on each frequency band.

[0023] The corrected original image data is obtained, and a 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 the signal variance and the 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 achieve adaptive enhancement of the entire frequency band. The processed low-frequency approximation coefficients and high-frequency detail coefficients are reconstructed into an enhanced image using an 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, and Daubechies4 wavelet is used for discrete wavelet transform to decompose the image into low-frequency approximate coefficients and high-frequency detail coefficients to obtain multi-scale decomposition results. The number of decomposition layers is dynamically determined according to the image size, and 3 to 5 layers are usually selected. The high-frequency detail coefficients are enhanced by the BayesShrink method, and an adaptive threshold function is set. where σ 2 is the noise variance, σ x is the signal variance. Dynamically adjust the threshold parameters according to the coefficient distribution characteristics at different scales to extract and enhance the detail information in the image. Apply a nonlinear mapping function to the low-frequency approximation coefficients and use the Sigmoid function As an S-curve mapping function, α controls the slope of the Sigmoid curve, and β controls the center point position of the Sigmoid curve. The values ​​of these 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 to achieve adaptive enhancement of the entire frequency band. The processed low-frequency and high-frequency coefficients are reconstructed into an enhanced image using an inverse wavelet transform. The contribution of each frequency band in the reconstruction process is adaptively adjusted by determining the weight coefficient based on the local variance of the image. Weight coefficient where σ i 2 is the local variance of the ith subband, σ n 2 To estimate the noise variance, adaptive fusion of information in different frequency bands is achieved. For a 512x512 pixel rectified CT image, Daubechies4 wavelet is first used to perform a 4-layer discrete wavelet transform. In each layer of decomposition, the image is divided into approximate coefficients LL and detail coefficients in three directions, namely LH, HL, and HH. For example, after the first layer of decomposition, 256x256 LL1, LH1, HL1, and HH1 sub-bands are obtained, and the second layer of decomposition LL1 obtains 128x128 LL2, LH2, HL2, and HH2 sub-bands, and so on. The BayesShrink method is applied to enhance the high-frequency detail coefficients. Assuming the noise variance σ 2 = 20, for the 128x128 LH2 subband, the signal variance σ is calculated x =50, then the threshold T = 20 / 50 = 0.4. Set the coefficients less than 0.4 in the LH2 subband to 0, and keep the coefficients greater than 0.4 unchanged, so as to achieve noise suppression and edge enhancement. For the low-frequency approximation coefficient LL4, the image size is 32x32, that is, the width and height of the image are both 32 pixels. Apply the Sigmoid function for nonlinear mapping, set α = 0.05, β = 128, and assume that the image grayscale range is 0-255. Map each pixel value x in LL4 as follows: 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 these weights are applied in the inverse wavelet transform process to achieve adaptive fusion and obtain the final enhanced image.

[0025] S103, reconstructing the enhanced multi-scale image, combining prior knowledge constraints, and introducing anatomical structure-based regularization terms to suppress noise and artifacts in the reconstruction process, thereby obtaining 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 perform adaptive filtering on the preliminary reconstructed image to obtain a filtered image that retains important anatomical structure information; it is determined whether the filtered image meets the preset termination condition, and the termination condition is whether the difference between the images obtained by 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 observed data b, b is the projection data, i.e. the measurement result, and x is the reconstructed image. is the gradient of image x, indicating the edge information of the image, ||·|| 1 represents the L1 norm and R(x) is an additional regularization term.

[0028] S104, performing microvascular structure enhancement on the initially reconstructed CT image, extracting vascular features, calculating the probability that each voxel belongs to a vascular structure, and enhancing the display of the microvascular network in combination with local structure tensor analysis.

[0029] The CT image is convolved to extract multi-scale and multi-directional vascular characteristic responses; the probability value of each voxel belonging to the vascular structure is calculated to obtain a multi-scale vesselness response map; the structural tensor analysis method is used to determine the main direction and anisotropy information through eigenvalue decomposition; if the vesselness response is greater than the preset threshold, an adaptive anisotropic diffusion filter is constructed; according to the adaptive anisotropic diffusion filter, the microvascular network is iteratively enhanced; in each iteration, the vesselness response and the structural tensor are updated; it is determined whether the number of iterations reaches the preset value or whether the relative error between two adjacent iteration results is less than the preset threshold; if the conditions are met, the enhanced microvascular CT image is output.

[0030] Specifically, for the preliminarily reconstructed CT image, a multi-directional Gabor filter group is constructed, with 12 directions, including 0° to 165°, one every 15°, and 4 scales, including 1 to 4 pixels. By rotating and scaling the basic Gabor filter, 48 filters of different directions and scales are generated, and convolution operations are performed on the CT image to extract multi-scale and multi-directional vascular feature responses. Using the Frangiseess metric function

[0031]

[0032] V(σ) represents the Frangivesselness metric function, where S is composed of the Hessian matrix eigenvalues ​​λ 1 , 2 , 3 It is calculated that Ra and Rb are based on the Hessian matrix eigenvalue λ 1 ,λ 2 ,λ 3 The calculated curvature related quantity is R a =min(|λ 1 |,|λ 2 |,|λ 3 |)

[0033] Ra represents the minimum curvature,

[0034] R b =max(|λ 1 |,|λ 2 |,|λ 3 |)

[0035] Rb represents the absolute value of the maximum eigenvalue, σ represents the scale parameter, which is used to control the scale of the filter.

[0036]

[0037] , S represents the Hessian matrix eigenvalue λ 1 , 2 , 3 The calculated quantity, λ 1 , 2 , 3 is the eigenvalue of the Hessian matrix, S is the structural metric used to measure the anisotropy of the local structure, α, β, and c are control parameters used to adjust the behavior of the metric function; set three scale parameters σ, including 1mm, 2mm, and 3mm corresponding to blood vessels of different calibers, calculate the probability of each voxel belonging to the vascular structure, and obtain a multi-scale vesselness response map. The structural tensor analysis method is used to calculate the image gradient in the 3x3x3 local neighborhood and construct the covariance matrix Among them J ij The covariance of image intensity in the i and j directions is represented. The main direction and anisotropy information are obtained by eigenvalue decomposition. Combined with the vesselness response, an adaptive anisotropic diffusion filter is constructed. The filter coefficient is determined by the angle between the main direction and the gradient direction and the vesselness value. The microvascular network is iteratively enhanced. In each iteration, an adaptive anisotropic diffusion filter is applied. The vesselness response and the structure tensor are updated at the same time. The maximum number of iterations is set to 50 or the relative error of the results of two adjacent iterations is less than 0.1% as the convergence condition. After the iteration, the contrast window width and window level are adjusted to highlight the microvascular structure and output the enhanced microvascular CT image. For a set of 512x512x256 voxel preliminary reconstructed CT images, 48 ​​Gabor filters are first constructed, with directions from 0° to 165°, one every 15°, scales of 1, 2, 3, and 4 pixels, and filter size of 21x21. The image is convolved to obtain 48 feature response maps. Then calculate the Hessian matrix and set the scale parameter σ to 1mm, 2mm, and 3mm, corresponding to blood vessels of different calibers. Use the Frangivesselness function with parameters α=0.5, β=0.5, and c=500 to calculate the vesselness value of each voxel. For the 3x3x3 local neighborhood, calculate the image gradient, construct the structure tensor, and obtain the main direction vector v through eigenvalue decomposition. 1 and anisotropy λ 1 -λ 2 / λ 1 According to the main direction and vesselness value, an adaptive anisotropic diffusion filter is constructed, and the filter coefficient

[0038]

[0039] k represents the filter coefficient, v 1 Indicates the main direction, Represents the image gradient, κ represents the thermal conductivity, set to 0.1V(σ) represents the vesselness value, and exp is the exponential function. The filter is applied iteratively, and the vesselness and structure tensor are updated at each iteration. The maximum number of iterations is set to 50, and the relative error threshold is 0.1%. After the iteration, the image grayscale value is mapped to the range of 0-255, the window width is set to 200HU, the window level is set to 100HU, and the microvessels are highlighted. The processing takes about 5 minutes on a workstation equipped with a GPU, and the enhanced microvascular CT image is finally output. Compared with the original image, the contrast-to-noise ratio of the microvessels is improved by about 40%, and the edge clarity is improved by about 25%.

[0040] S105. Perform fine segmentation of lobular septa and bronchiolar walls on the microvascular enhanced image, extract fine lung structures through multi-scale feature fusion, and perform post-processing to optimize the segmentation results.

[0041] A microvascular enhanced CT image is obtained, and a first image is obtained according to the CT image by 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 filter 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, and the U-Net convolutional neural network model includes at least one encoder layer and at least one decoder layer. If the number of encoder layers is N, the number of 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 lobular septum and the bronchiolar wall. Morphological operations are performed on the preliminary segmentation result, including opening operations and closing operations, 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 univariate potential function and a binary potential function, wherein the univariate potential function uses the log-likelihood ratio, and the binary potential function uses the contrast-sensitive Potts model. The conditional random field model is applied to the optimized segmentation mask to obtain the final segmentation result, which includes the refined segmentation information of the lobular septa and bronchiolar walls.

[0042] Specifically, the microvascular enhanced CT images were preprocessed, and the image contrast was enhanced by contrast-limited adaptive histogram equalization (CLAHE). The block size was set to 8x8, the contrast limit parameter was set to 0.01, and bilateral filtering was used to remove noise while retaining edge information. The spatial domain standard deviation was set to 3, and the value domain standard deviation was set to 75 to obtain the preprocessed image. A U-Net convolutional neural network model was constructed, and a 4-layer encoder-decoder structure was designed. Each layer used two 3x3 convolution kernels, the activation function was ReLU, and the pooling layer was 2x2 maximum pooling. The feature information of different levels was fused through the feature pyramid network (FPN) to achieve the preliminary segmentation of the lobular septum and bronchiolar wall. Morphological operations were applied to the segmentation results output by the neural network, including the use of 3x3 circular structure elements for opening operations to remove small mis-segmented areas, the use of 5x5 circular structure elements for closing operations to fill small holes, and the use of Zhang-Suen thinning algorithm to extract the skeleton structure to obtain the optimized segmentation mask. The conditional random field was used to refine the segmentation results, and the image was divided into superpixels of size 10x10. A potential function based on pixel intensity and spatial position was constructed. The one-dimensional potential function used the log-likelihood ratio, and the binary potential function used the contrast-sensitive Potts model. The energy function was optimized by 10 iterations of the mean field inference algorithm to obtain the final segmentation results of the interlobular septum and bronchiolar wall. For a set of 512x512x256 voxel microvascular enhanced CT images, the CLAHE algorithm was first applied to divide the image into 64 8x8 blocks, and the contrast limit parameter was set to 0.01 to obtain the image after histogram equalization. Then bilateral filtering was used, and the spatial domain σ d =3, range σ r=75, the window size is 21x21, the noise of the processed image 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 contains 4 downsampling and 4 upsampling layers, each layer has two 3x3 convolutions, the starting number of channels is 64, and each layer doubles, with a maximum of 512. Using the FPN structure, the feature maps of different scales are adjusted to the same number of channels 256 through 1x1 convolution, and then upsampled and element-by-element addition fusion is performed. The model is trained on 2000 annotated images for 100 rounds, with a batch size of 8, a learning rate of 0.001, and an Adam optimizer is used. The probability map output by U-Net is thresholded with a threshold of 0.5 to obtain the preliminary segmentation result. A 3x3 circular structure element is applied for opening operation to remove areas with an area less than 10 pixels, and then a 5x5 circular structure element is used for closing operation to fill holes less than 20 pixels. The Zhang-Suen algorithm is used for refinement, and a skeleton with a width of 1 pixel is obtained after about 15 iterations. Finally, the fast SLIC algorithm was used to divide the image into about 5000 superpixels and construct a conditional random field. The unary potential used -log(p), where p is the probability value output by U-Net, and the binary potential used the contrast-sensitive Potts model with a weight of 30. The final segmentation result was obtained through 10 iterations of mean field inference optimization. The entire processing process took about 3 minutes on a workstation equipped with a GPU. Compared with using U-Net alone, the segmentation accuracy of the lobular septum and bronchiolar wall was improved by about 8%, and the Dice coefficient reached 0.92.

[0043] S106. Quantitative analysis is performed on the segmented lung fine structure to calculate the interlobular septum thickness and bronchiolar wall area percentage. By comparing with normal reference values, it is determined whether there are interstitial changes or airway remodeling.

[0044] The Zhang-Suen thinning algorithm was used to process the interlobular septum structure, and the shortest distance from each point on the skeleton to the edge was calculated by the Euclidean distance transformation algorithm to obtain the interlobular septum thickness. The bronchiolar structure was subjected to morphological expansion and corrosion operations to extract the tracheal wall area, calculate the ratio of the tracheal wall area to the corresponding tracheal total area, and obtain the percentage of the bronchiolar wall area. Based on the CT scan data of healthy volunteers, a normal reference value database was established to obtain the data of interlobular septum thickness and bronchiolar wall area percentage. The Z-score method was used to calculate the degree of deviation between the case to be tested and the normal reference value, Z-score = (x-μ) / σ, where x is the value to be tested, 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 the first preset threshold, it is judged as interstitial change; if the Z-score of the percentage of the bronchiolar wall area is greater than the second preset threshold, it is judged as airway remodeling. According to the numerical range of the Z-score, the degree of interstitial change and airway remodeling was determined.

[0045] Specifically, the segmented lobular septum structure was processed by Zhang-Suen thinning algorithm, and the shortest distance from each point on the skeleton to the edge was calculated by Euclidean distance transformation algorithm. The twice of this distance was taken as the lobular septum thickness. The thickness distribution was obtained by kernel density estimation, and the average thickness and standard deviation were calculated. For the segmented bronchiolar structure, 3x3 circular structural elements were used for morphological expansion and corrosion operations to extract the tracheal wall area, and the ratio of the tracheal wall area to the corresponding tracheal total area was calculated to obtain the percentage of bronchiolar wall area. A normal reference value database was constructed, which included CT scan data of 1000 healthy volunteers of different ages, genders, and body shapes. The same processing flow as the case to be tested was used to obtain the lobular septum thickness and bronchiolar wall area percentage data, and the Shapiro-Wilk test was used to ensure that the data was normally distributed. The Z-score method was used to calculate the degree of deviation between the case to be tested and the normal reference value, Z-score = (x-μ) / σ, where x is the value to be tested, μ and σ are the mean and standard deviation of the reference data, respectively. According to clinical research data, thresholds were set to judge interstitial changes and airway remodeling: if the Z-score of the septal thickness was greater than 2.5, it was judged as interstitial changes, 2.5 < Z-score < 3.5 was mild, 3.5 ≤ Z-score < 5 was moderate, and Z-score ≥ 5 was severe; if the Z-score of the bronchiolar wall area percentage was greater than 2.0, it was judged as airway remodeling, 2.0 < Z-score < 3 was mild, 3 ≤ Z-score < 4 was moderate, and Z-score ≥ 4 was severe, providing a quantitative basis for the diagnosis of interstitial lung disease. For a set of 512x512x256 voxel lung CT images, the Zhang-Suen thinning algorithm was first applied to the segmented septal structure, and a skeleton with a single pixel width was obtained by iteration for about 15 times. Then the Euclidean distance transform was used to calculate the distance from each point on the skeleton to the edge, and the average processing time was about 2 seconds. The distance value was multiplied by 2 to obtain the thickness, and the kernel density estimation (bandwidth 0.1 mm) was applied to obtain the thickness distribution. The calculated average thickness was 1.2 mm and the standard deviation was 0.3 mm. For the bronchiolar structure, 3x3 circular structural elements were used to perform 3 expansions and 2 corrosion operations to extract the tracheal wall area, and the area ratio was calculated to obtain the average bronchiolar wall area percentage of 23%. The normal reference database contained 1000 volunteers, aged 18-80 years, half male and half female. The Shapiro-Wilk test was performed, with p>0.05, confirming the normality of the data. The calculated mean reference value of interlobular septum thickness was 1.0 mm and the standard deviation was 0.2 mm; the mean reference value of bronchiolar wall area percentage was 20% and the standard deviation was 2%. For a case to be tested, the interlobular septum thickness was 1.8 mm and the bronchiolar wall area percentage was 28%. The calculated Z-score was 4.0 and 4.0, respectively.According to the set threshold, this case was judged to have moderate interstitial changes and moderate airway remodeling, indicating a high possibility of interstitial lung disease, and further clinical evaluation was recommended.

[0046] S107. Perform abnormality detection on the entire lung area, extract the characterization features of normal and abnormal tissues, and combine the multi-instance learning strategy to accurately locate the suspicious lesion area to provide auxiliary diagnostic information for early lung cancer screening.

[0047] Receive lung CT images and preprocess them, extract multi-scale features from the preprocessed lung CT images; perform weighted fusion of features at different spatial positions through the Transformer self-attention mechanism based on the multi-scale features to obtain a global feature representation; construct feature pairs of normal tissue features and abnormal tissue features using a contrastive learning method, and calculate the distance between feature pairs; if the feature pair is a normal-abnormal pair, maximize the distance between the feature pairs; if the feature pair is a normal-normal pair, minimize the distance between the feature pairs; input the features obtained by contrastive learning into a multi-instance learning framework, which is driven by an attention mechanism and divides the lung CT images into overlapping three-dimensional blocks; perform abnormality measurement calculation on the divided three-dimensional blocks to obtain high abnormality areas; finely locate the high abnormality areas, and use morphological operations and region growing algorithms to extract the contours of suspicious lesions.

[0048] Specifically, the input lung CT images are preprocessed, and multi-scale features are extracted using 3DResNet-50. The features at different spatial positions are weighted and fused through the Transformer self-attention mechanism to obtain a global feature representation. The contrastive learning method is used to construct feature pairs of normal and abnormal tissues. The InfoNCE loss function is used to maximize the distance between normal-abnormal pairs and minimize the distance between normal-normal pairs. The informative samples are selected using the hard case mining strategy to learn discriminative feature representations. The features obtained from contrastive learning are used as the input of multi-instance learning. Based on the attention-driven multi-instance learning framework, the lung images are divided into overlapping three-dimensional blocks of 64x64x64 size. Each block is used as an instance. The instance-level abnormality metric is calculated through the Gated-Attention mechanism, and the noisy-or pooling function is used for bag-level aggregation to calculate the overall abnormality degree. The detected high abnormality area is finely located, and the morphological opening and closing operation is performed in combination with a spherical structure element with a radius of 3. The region growing algorithm is used with a threshold of ±50HU to extract the precise contour of the suspicious lesion and calculate quantitative features such as volume, shape, and texture. According to the abnormality threshold, set to 0.8, high-risk areas are screened, and quantitative features are combined to generate early lung cancer screening recommendations, outputting abnormal detection results, positioning information, and screening recommendations. For a set of 512x512x256 voxel lung CT images, 3DResNet-50 is first used to extract features. The network contains 5 stages, and each stage outputs feature maps of 32, 64, 128, 256, and 512 channels, respectively. Then, the features are fused through the 8-head Transformer self-attention mechanism to obtain a 2048-dimensional global feature vector. In the contrastive learning stage, 100 normal patches and 100 abnormal patches are randomly sampled from each CT scan to construct 10,000 pairs of 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% of difficult samples. The learned features are input into the attention-driven multi-instance learning framework to divide the lung image into 3375 overlapping three-dimensional blocks of 64x64x64 size. A two-layer fully connected network (128-64-1) was used as a gated attention mechanism to calculate the anomaly score of each instance, and the overall anomaly was obtained by aggregating through a noisy-or pooling function (p = 0.7). Regions with an anomaly score exceeding 0.8 were finely localized, and a spherical structure element with a radius of 3 voxels was used to perform 3 morphological opening operations and 2 closing operations, and then a region growing algorithm was applied with a threshold of ±50HU and a minimum number of 100 voxels to extract the lesion contour. 18 3D radiomic features were calculated, including volume, sphericity, and gray-level co-occurrence matrix features.Finally, based on the abnormality and quantitative features, a random forest classifier was used with 100 decision trees and a maximum depth of 10 to generate an early lung cancer risk score of 0-100.

[0049] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features with similar functions disclosed in the present application (but not limited to) to form a technical solution.

Claims

1. A method for analyzing the association between lung CT lesions and related factors, characterized in that: The method comprises: A dynamic compensation model is established based on the patient's blood sugar level fluctuation data. The CT scanning parameters are adjusted by real-time monitoring of blood sugar values ​​to compensate for changes in tissue density during the scanning process and obtain corrected original image data. The corrected original image data is decomposed at multiple scales, and different frequency components are extracted by wavelet transform. The detail information is enhanced by high-frequency components, and the overall structure is retained by low-frequency components. Each frequency band is adaptively enhanced by combining nonlinear mapping functions. Reconstruct the enhanced multi-scale image, combine the prior knowledge constraints, and introduce the regularization term based on the anatomical structure to suppress the noise and artifacts in the reconstruction process, so as to obtain the preliminary reconstructed high-quality CT image; The microvascular structure of the initially reconstructed CT image is enhanced, the vascular features are extracted, the probability of each voxel belonging to the vascular structure is calculated, and the microvascular network is enhanced and displayed in combination with local structural tensor analysis; Fine segmentation of lobular septa and bronchiolar walls is performed on microvascular enhanced images, and fine lung structures are extracted through multi-scale feature fusion, followed by post-processing to optimize the segmentation results. Quantitative analysis of the segmented lung fine structure was performed to calculate the thickness of the interlobular septum and the percentage of the bronchiolar wall area. By comparing them with normal reference values, it was determined whether there were interstitial changes or airway remodeling. Abnormality detection is performed on the entire lung area, characterization features of normal and abnormal tissues are extracted, and the suspicious lesion area is accurately located by combining multi-instance learning strategy.

2. The method according to claim 1, wherein: The method of establishing a dynamic compensation model based on the patient's blood sugar level fluctuation data, adjusting the CT scanning parameters by real-time monitoring of the blood sugar value, compensating for the change in tissue density during the scanning process, and obtaining the corrected original image data includes: A dynamic compensation model is established based on the relationship between historical blood sugar data and CT image quality, and a mapping function between blood sugar value and scanning parameters is obtained; Obtaining the patient's real-time blood sugar level, and calculating the optimal CT scanning parameters corresponding to the current blood sugar level according to the mapping function; After receiving the CT scanning instruction, the raw scan data is preprocessed, including calculating the compensation coefficient according to the density change characteristics of different tissue types to correct the raw scan data; Generating a compensated CT image for the CT scan data corrected by the compensation coefficient; The CT images before and after compensation were compared, and the effect of blood glucose fluctuation compensation was evaluated by using the signal-to-noise ratio, contrast-to-noise ratio, and spatial resolution indicators.

3. The method according to claim 1, wherein: The method performs multi-scale decomposition on the corrected original image data, extracts different frequency components by wavelet transform, enhances detail information by high-frequency components, retains the overall structure of low-frequency components, and performs adaptive enhancement processing on each frequency band in combination with a nonlinear mapping function, including: Acquire the corrected original image data, and perform discrete wavelet transform on the original image data to obtain low-frequency approximation coefficients and high-frequency detail coefficients; According to the high-frequency detail coefficient, an adaptive threshold function is set, wherein parameters of the adaptive threshold function are determined by signal variance and noise variance; Applying a nonlinear mapping function to the low-frequency approximation coefficient, wherein the nonlinear mapping function is a Sigmoid function, and a parameter of the Sigmoid function is determined by the overall brightness distribution characteristics of the image; Applying the nonlinear mapping function to the high-frequency detail coefficients to achieve full-band adaptive enhancement; The processed low-frequency approximation coefficients and the high-frequency detail coefficients are reconstructed into an enhanced image using an 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 according to claim 1, wherein: The enhanced multi-scale image is reconstructed, and noise and artifacts in the reconstruction process are suppressed by introducing a regularization term based on anatomical structure in combination with prior knowledge constraints, so as to obtain a preliminarily reconstructed high-quality CT image, including: Constructing an objective function including a data fidelity term, a total variation regularization term, and an anatomical structure-based regularization term, wherein the objective function is based on a total variation minimization principle; For the objective function, an alternating direction multiplier method is used to perform optimization and solution to obtain a preliminary reconstructed image; Acquire prior knowledge of anatomical structures in a pre-trained U-Net convolutional neural network, and construct a structure-guided edge-preserving filter based on the prior knowledge of anatomical structures; Adopting the structure-guided edge-preserving filter to perform adaptive filtering on the preliminary reconstructed image to obtain a filtered image that retains important anatomical structure information; Determine whether the filtered image meets a preset termination condition, where the termination condition is whether the difference between images obtained by two adjacent iterations is less than a preset threshold; After multiple iterations of optimization, a preliminary reconstructed high-quality CT image is obtained.

5. The method according to claim 1, wherein: The method of performing microvascular structure enhancement on the initially reconstructed CT image, extracting vascular features, calculating the probability of each voxel belonging to the vascular structure, and enhancing the display of the microvascular network in combination with local structural tensor analysis includes: Perform convolution operations on CT images to extract multi-scale and multi-directional vascular feature responses; Calculate the probability value of each voxel belonging to the vascular structure and obtain a multi-scale vesselness response map; The structural tensor analysis method is used to determine the main direction and anisotropy information through eigenvalue decomposition; If the vesselness response is greater than the preset threshold, an adaptive anisotropic diffusion filter is constructed; The microvascular network is iteratively enhanced based on an adaptive anisotropic diffusion filter; In each iteration, the vesselness response and structure tensor are updated; Determine whether the number of iterations reaches a preset value or whether the relative error between two adjacent iteration results is less than a preset threshold; If the conditions are met, the enhanced microvascular CT image is output.

6. The method according to claim 1, wherein: The microvascular enhanced image is finely segmented for lobular septa and bronchiolar walls, the fine structure of the lungs is extracted by multi-scale feature fusion, and post-processing is performed to optimize the segmentation results, including: Acquire a microvascular enhanced CT image, and obtain a first image according to the CT image by using a contrast-limited adaptive histogram equalization algorithm; Processing the first image to obtain a second image, wherein the spatial domain standard deviation of the bilateral filter is a preset value A, and the range standard deviation is a preset value B; Constructing a U-Net convolutional neural network model, wherein the U-Net convolutional neural network model includes at least one encoder layer and at least one decoder layer. 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 includes segmentation information of lobular septa and bronchiolar walls; Performing morphological operations on the preliminary segmentation results, including opening operations and closing operations, to obtain optimized segmentation results; Extracting the skeleton structure of the optimized segmentation result to obtain an optimized segmentation mask; Constructing a conditional random field model, the conditional random field model includes a univariate potential function and a binary potential function, wherein the univariate 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, wherein the final segmentation result includes refined segmentation information of interlobular septa and bronchiolar walls.

7. The method according to claim 1, wherein: The method of quantitatively analyzing the fine structure of the segmented lungs, calculating the thickness of the interlobular septum and the percentage of the bronchiolar wall area, and comparing them with normal reference values ​​to determine whether there are interstitial changes or airway remodeling, includes: Extract the skeleton in the interlobular septum structure, calculate the shortest distance from each point on the skeleton to the edge, and obtain the interlobular septum thickness; The bronchiolar structure was subjected to morphological expansion and corrosion operations to extract the tracheal wall area, calculate the ratio of the tracheal wall area to the corresponding tracheal total area, and obtain the percentage of the bronchiolar wall area; A normal reference value database was established to obtain data on interlobular septum thickness and bronchiolar wall area percentage; The Z-score method is used to calculate the degree of deviation between the case to be tested and the normal reference value, Z-score = (x-μ) / σ, where x is the value to be tested, 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 the first preset threshold, it is judged as an interstitial change; If the Z-score of the bronchiolar wall area percentage is greater than the second preset threshold, it is judged as airway remodeling; The extent of interstitial changes and airway remodeling was determined based on the numerical range of the Z-score.

8. The method according to claim 1, wherein: The method of performing abnormality detection on the entire lung area, extracting the characterization features of normal and abnormal tissues, and accurately locating the suspicious lesion area by combining a multi-instance learning strategy includes: Receiving a lung CT image and preprocessing the lung CT image, and extracting multi-scale features from the preprocessed lung CT image; Based on the multi-scale features, features at different spatial positions are weightedly fused through the Transformer self-attention mechanism to obtain a global feature representation; Constructing feature pairs of normal tissue features and abnormal tissue features using a contrastive learning method, and calculating the distance between the feature pairs; If the feature pair is a normal-abnormal pair, the distance between the feature pairs is maximized; if the feature pair is a normal-normal pair, the distance between the feature pairs is minimized; Inputting the features obtained by contrastive learning into a multiple instance learning framework, wherein the multiple instance learning framework is driven by an attention mechanism and divides the lung CT image into overlapping three-dimensional blocks; Anomaly measurement calculation is performed on the divided three-dimensional blocks to obtain high anomaly areas; The high abnormality area is precisely located, and the contour of the suspicious lesion is extracted using morphological operations and region growing algorithms.

Images