Lung tumor CT image 3D segmentation method and system based on multi-modal image fusion

Abstract

The invention relates to the technical field of medical image processing, in particular to a lung tumor CT image 3D segmentation method and system based on multi-modal image fusion. The method comprises the following steps: acquiring a lung tumor CT image; determining a first texture feature based on the lung tumor CT image; identifying a lung tumor boundary by using the first texture feature; detecting a nodule protrusion area in the boundary of the lung tumor; obtaining tissue infiltration data from the nodule protrusion area; determining a second texture feature according to the tissue infiltration data; determining a tumor heterogeneity feature according to the first texture feature and the second texture feature; evaluating the potential malignancy degree by utilizing tumor heterogeneity characteristics; and dividing a tumor risk area of the lung tumor CT image based on the potential malignancy degree. According to the invention, accurate heterogeneity identification and risk region division of the lung tumor CT image are realized based on a medical image processing technology, and the accuracy of lung tumor 3D segmentation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a lung tumor CT image 3D segmentation method and system based on multi-modal image fusion. BACKGROUND

[0002] Currently, the precise diagnosis and treatment of lung tumors rely on high-quality medical image analysis, especially the accurate segmentation of three-dimensional CT images plays an important role in clinical practice. Traditional lung tumor CT image segmentation methods are mostly based on single modal data, which is difficult to fully reflect the complex structure and heterogeneity of tumors, resulting in insufficient accuracy of segmentation results. In addition, existing technologies often rely on empirical parameters or simple algorithms in key links such as tumor boundary identification, nodule protrusion detection and tissue invasion analysis, lack effective fusion of multiple image information, and cannot fully capture the subtle changes of tumor internal and peripheral tissues, affecting the accurate assessment of potential malignancy and the division of risk areas. The existing segmentation system cannot dynamically respond to tumor risk characteristics in three-dimensional reconstruction and segmentation coefficient adjustment, resulting in insufficient coherence and accuracy of three-dimensional segmentation results, which limits its application effect in clinical auxiliary diagnosis. SUMMARY

[0003] Therefore, it is necessary to provide a lung tumor CT image 3D segmentation method based on multi-modal image fusion to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a lung tumor CT image 3D segmentation method based on multi-modal image fusion comprises the following steps:

[0005] Step S1: obtaining a lung tumor CT image; determining a first texture feature based on the lung tumor CT image; identifying a lung tumor boundary using the first texture feature;

[0006] Step S2: detecting a nodule protrusion region in the lung tumor boundary; obtaining tissue invasion data from the nodule protrusion region; determining a second texture feature according to the tissue invasion data;

[0007] Step S3: determining a tumor heterogeneity feature according to the first texture feature and the second texture feature; evaluating the potential malignancy using the tumor heterogeneity feature; dividing the tumor risk area of the lung tumor CT image based on the potential malignancy;

[0008] Step S4: adjusting the image segmentation coefficient based on the tumor risk area; performing 3D segmentation on the lung tumor CT image using the image segmentation coefficient to generate a 3D segmentation image.

[0009] The specification also provides a lung tumor CT image 3D segmentation system based on multi-modal image fusion, for executing the lung tumor CT image 3D segmentation method based on multi-modal image fusion as described above, the lung tumor CT image 3D segmentation system based on multi-modal image fusion comprises:

[0010] A lung tumor boundary recognition module is configured to acquire a lung tumor CT image, determine a first texture feature based on the lung tumor CT image, and recognize a lung tumor boundary using the first texture feature.

[0011] A nodule protrusion region recognition module is configured to detect a nodule protrusion region in the lung tumor boundary, obtain tissue invasion data from the nodule protrusion region, and determine a second texture feature based on the tissue invasion data.

[0012] A tumor risk region division module is configured to determine a tumor heterogeneity feature based on the first texture feature and the second texture feature, evaluate a potential malignancy degree using the tumor heterogeneity feature, and divide a tumor risk region of the lung tumor CT image based on the potential malignancy degree.

[0013] A 3D segmented image generation module is configured to adjust an image segmentation coefficient based on the tumor risk region, and perform 3D segmentation on the lung tumor CT image using the image segmentation coefficient to generate a 3D segmented image.

[0014] The present application has the following advantages:

[0015] On the one hand, by fusing lung CT image data of different modalities, the complementary advantages of each modality in terms of tissue contrast, density distribution, and edge sharpness can be comprehensively utilized, thereby achieving information enhancement and redundancy suppression in lung tumor region identification, significantly improving the positioning accuracy of tumor boundaries and the coherence of overall segmentation, overcoming the problems of boundary blurring and easy missed detection of tissue adhesion regions in traditional single-modality methods, and effectively enhancing the structural integrity and diagnostic reliability of three-dimensional segmentation results.

[0016] On the other hand, by joint extraction and comparison based on texture features and gray scale distribution features, the tissue heterogeneity features within the tumor region can be accurately identified, and the risk region preliminary screening and invasion boundary judgment can be performed in combination with the voxel-level analysis method, thereby improving the recognition ability of complex morphological tumors (such as ground glass nodules and irregular infiltrating lesions), effectively supporting preoperative risk grading and radiotherapy target area setting, and overcoming the limitations of existing methods in terms of slow response to lesion shape changes and weak feature recognition ability.

[0017] On the other hand, by introducing a spatial connectivity constraint mechanism in the three-dimensional image data, and combining the dynamic sliding window to adjust the boundary response coefficient of the segmentation region, the false segmentation and artifact interference can be inhibited, the stability of the overall structure of the image is enhanced while the edge details are maintained, and the robustness of the segmentation algorithm in complex lung structures (such as blood vessel intersection and chest wall adjacent area) is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the drawings:

[0019] Fig. 1 A step flowchart of a lung tumor CT image 3D segmentation method based on multi-modal image fusion according to the present application is shown in the figure;

[0020] Fig. 2 A module schematic diagram of a lung tumor CT image 3D segmentation system based on multi-modal image fusion according to the present application is shown in the figure;

[0021] Fig. 3 A lung tumor CT image according to the present application is shown in the figure;

[0022] The implementation of the object, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0023] The technical method of the present application will be described clearly and completely below in combination with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] In addition, the accompanying drawings are only schematic illustrations of the present application, and are not necessarily drawn to scale. The same reference signs in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0025] It should be understood that, although the terms "first", "second", etc. can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element without departing from the scope of the example embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the associated associated items.

[0026] To achieve the above object, there is provided Figs. 1-3 The application provides a lung tumor CT image 3D segmentation method based on multi-modal image fusion, which comprises the following steps:

[0027] Step S1: acquiring lung tumor CT images; determining first texture features based on the lung tumor CT images; and identifying lung tumor boundaries using the first texture features;

[0028] In an embodiment, a multi-slice spiral CT scanner is used to scan the chest of a patient to obtain three-dimensional CT image data with a resolution of 512x512x300, a layer thickness of 1 mm, and a reconstruction interval of 0.5 mm. A gray level co-occurrence matrix (GLCM) is used to extract first texture features of the lung region in the image, including energy, contrast, correlation, and entropy indicators. The texture features of each block are calculated in the suspected tumor region by a sliding window method, and a support vector machine (SVM) is used for preliminary boundary classification to identify lung tumor boundary regions with obvious gray scale structure changes.

[0029] In another embodiment, it is assumed that a set of input lung CT images has a total of 240 layers, each layer having an image size of 512x512, and the lung suspected lesion area appears in 45 layers. Texture features are calculated in units of 32x32 pixel blocks, and the average value of the extracted GLCM contrast is set as a threshold of 0.25 to determine the boundary region. According to statistics, a total of 5200 texture abnormal regions are identified, and an edge detection algorithm (such as Canny) is further used to accurately outline the tumor boundary profile.

[0030] Step S2: detecting nodule protrusion regions in the lung tumor boundary; obtaining tissue invasion data from the nodule protrusion regions; and determining second texture features according to the tissue invasion data;

[0031] In an embodiment, for the identified tumor boundary, a three-dimensional morphological analysis method is used to extract the boundary curvature and convex hull profile, and identify the local protrusion region. The region with a voxel curvature radius less than a set threshold (e.g., 3 mm) is used as a nodule protrusion criterion, and the gray scale gradient direction and distribution of its neighborhood are extracted to form the tissue invasion data. Based on the data, the Laws texture energy filter output is further calculated to obtain the second type of texture feature, which reflects the abnormal growth morphology of the tissue structure.

[0032] In another embodiment, it is assumed that 76 nodule protrusions are detected in the tumor boundary region, and the protrusion voxel volume range is between 25 and 150 mm3. The gradient direction variance of the 5x5x5 voxels in the neighborhood of these regions is calculated, and an average variance value greater than 0.8 is considered as a significant tissue invasion region. Subsequently, the texture response energy distribution (E5L5, L5S5, etc. convolution template output) of these regions is extracted as a second texture feature set, which is used to supplement the complexity description of the tumor structure.

[0033] Step S3: determining a tumor heterogeneity feature according to the first texture feature and the second texture feature; evaluating a potential malignancy degree using the tumor heterogeneity feature; and dividing a tumor risk region of the lung tumor CT image based on the potential malignancy degree;

[0034] In an embodiment, the first texture feature and the second texture feature are registered in the spatial domain to obtain a texture overlap region. Based on these regions, the dispersion (such as variance, coefficient of variation) of the texture response intensity and the spatial distribution density are calculated. These two statistics are fused to construct a tumor heterogeneity feature index. Then, a trained logistic regression model is used to classify and predict the heterogeneity index, output a malignancy probability score, and divide the tumor region into three categories of high-risk (malignancy probability ≥ 0.7), medium-risk (0.4-0.7), and low-risk (<0.4) regions according to the score results.

[0035] In another embodiment, it is assumed that 38 overlapping regions are obtained after texture feature matching in the lung tumor image, and each region has a size of about 10x10 pixel blocks. Statistics show that 15 of these regions have a heterogeneity coefficient of variation >0.6, belonging to the structure significantly uneven region. These regions are divided into high-risk regions (9 regions, accounting for 31% of the total tumor volume), medium-risk regions (12 regions, 40%), and low-risk regions (17 regions, 29%) according to their corresponding malignancy degree scores, and are visualized and labeled in the 3D model with different colors.

[0036] Step S4: adjusting an image segmentation coefficient based on the tumor risk region; and performing 3D segmentation on the lung tumor CT image using the image segmentation coefficient to generate a 3D segmentation image.

[0037] In an embodiment, different image segmentation super parameters are set for the risk area, for example: a small smoothing parameter (λ=0.1) is used in the high-risk area to retain boundary details; a large regional consistency coefficient (λ=0.5) is used in the low-risk area to enhance segmentation stability. Based on the improved 3D U-Net model input adjusted feature map, voxel-level lung tumor segmentation is performed. The final output includes a three-dimensional tumor structure graph with high, medium and low risk labels, realizing the risk-aware segmentation strategy.

[0038] In another embodiment, it is assumed that the entire lung tumor volume is 48000 voxels, which is divided into 15000 voxels in the high-risk area, 19000 voxels in the medium-risk area, and 14000 voxels in the low-risk area. The segmentation parameters set are: high-risk area boundary retention weight α=0.7, medium-risk α=0.5, and low-risk α=0.3. The model after training performs an average Dice coefficient of 0.89 on an independent test set, which is about 4.3% higher than the unified parameter segmentation method, which can effectively improve the segmentation accuracy and the reliability of the region identification.

[0039] Preferably, the first texture feature determined based on the lung tumor CT image in step S1 comprises:

[0040] For any one pixel point in the lung tumor CT image, a fixed size neighborhood region centered on the pixel point is extracted; in the neighborhood region, the gray difference between each pixel point and the center pixel point is calculated respectively;

[0041] In an embodiment, the input lung CT image is traversed pixel by pixel, and any one pixel point in the image is selected as the center point, for example, a CT slice image with a size of 512x512 is selected as the center point (256, 256). A 7x7 neighborhood region is extracted centered on the point, which contains the center point and its surrounding 49 pixels. For each pixel in the neighborhood, the difference between its gray value and the gray value of the center pixel is calculated respectively. For example, if the gray value of the center pixel is 125 and the gray value of the adjacent pixel is =1…48), then the difference , a 7x7 gray difference matrix is formed. The difference matrix reflects the degree of gray fluctuation within the neighborhood and is the basis for subsequent texture feature extraction.

[0042] ​In another embodiment, assuming that the grayscale value of pixel (300, 300) in the CT image is 110, a 5×5 neighborhood region is extracted. This region contains 25 pixels with grayscale values: G={108,112,109,106,115,...,111}. The difference between the grayscale value and the center pixel is calculated: D={∣108–110∣,∣112–110∣,...,∣111–110∣}={2,2,1,4,5,...,1}. A 5×5 grayscale difference matrix is ​​obtained, with values ​​ranging from [0,7], providing raw data for subsequent normalization processing and orientation extraction.

[0043] The grayscale difference is used as the texture response value and normalized; a texture consistency metric map is constructed based on the normalized texture response value.

[0044] In one embodiment, the gray-level difference matrix is ​​normalized so that its values ​​are distributed between [0,1], serving as the texture response values. Subsequently, the same processing is performed on each pixel across the entire image, ultimately constructing a texture consistency metric map to represent the gray-level consistency variations in different regions of the image.

[0045] In another embodiment, after normalizing the aforementioned 5×5 gray-level difference matrix, the maximum response value is 1.0, the minimum response value is 0.0, and the remaining values ​​are linearly distributed. This operation is repeated to perform a sliding window scan on the entire lung CT image, generating a texture response map. This response map exhibits a characteristic pattern of dramatic gray-level differences at the tumor edges and uniform texture in the lung field region, providing a basis for subsequent orientation extraction.

[0046] In the texture consistency metric map, the direction of the line connecting the pixel with the maximum response value and the pixel with the minimum response value is determined, and this direction of the line is used as the first texture feature.

[0047] In one embodiment, in the texture consistency metric map, the pixel with the largest and smallest response value in each processing window is found, and the direction of the line connecting these two points is determined based on their coordinates. This direction is defined as the first texture feature of the center pixel of that window. The first texture feature is used to characterize the dominant direction of texture grayscale changes in that region.

[0048] In another embodiment, assuming that in a 5×5 neighborhood, the maximum texture response point is located at (x+2, y–1) and the minimum texture response point is located at (x–1, y+1), the line connecting these two points forms a direction vector. After angle conversion, the direction is approximately 45°, which is used as the first texture feature direction for the central pixel. In the lung tumor boundary region, statistical analysis shows that most texture directions are concentrated in the range of 40° to 70°, while in the lung field background, texture directions are randomly distributed. This feature has a significant distinguishing effect in tumor boundary detection.

[0049] Preferably, step S1, which involves identifying the lung tumor boundary using the first texture feature, includes:

[0050] A first texture feature image of lung tumor CT images is constructed based on the first texture feature;

[0051] In one embodiment, the input lung CT image sequence is first preprocessed, including standardization and noise suppression. Then, the gray-level co-occurrence matrix (GLCM) method is used to extract texture features of each CT image frame, such as energy, entropy, contrast, and correlation. Based on the texture feature dimensions, the energy and entropy features that best reflect tissue structure differences are selected. Using a sliding window (11×11 window size), local statistics are performed on the entire image to generate the corresponding first texture feature image. The pixel gray values ​​of this feature image represent the texture complexity of the corresponding region, providing a basis for subsequent boundary extraction.

[0052] In another embodiment, assuming the input CT image size is 512×512 and the sliding window step size is set to 5 pixels, energy features are extracted to form a first texture feature image, and the energy values ​​are normalized to the range of [0,1]. Assuming that the texture energy of the central tumor region is concentrated between 0.6 and 0.9, while the background energy value of the lung tissue is mainly distributed between 0.2 and 0.4, this feature image can highlight the texture difference between the tumor region and the background.

[0053] The first texture feature image is subjected to edge enhancement processing; in the edge-enhanced texture feature image, the texture direction change rate of each pixel is calculated;

[0054] In one embodiment, the Sobel or Laplacian operator is applied to the first texture feature image for edge enhancement to highlight regions of abrupt texture changes. Then, the texture direction change rate for each pixel is calculated in the enhanced image. Using structural tensor analysis, the intensity of abrupt texture direction changes is evaluated by analyzing the degree of change in the local gradient direction field, and the output is a direction change rate map.

[0055] In another embodiment, assuming the enhancement process uses a 3×3 Sobel operator to obtain a texture gradient map, the standard deviation of the principal direction within the neighborhood (radius of 3 pixels) of each pixel is then calculated as the texture direction change rate. If the direction change rate of a pixel is 0.2, it indicates that its local texture direction is relatively consistent; if it is 0.9, it indicates a drastic texture mutation, which usually occurs at the tumor boundary or tissue transition zone.

[0056] Based on a preset texture change rate threshold, pixels with a texture direction change rate greater than the texture change rate threshold are extracted to form an initial set of boundary candidate points.

[0057] In one embodiment, a preset texture direction change rate threshold (e.g., 0.7) is set, and all pixels with a change rate greater than this threshold are extracted as candidate points for boundary abrupt change regions, forming an initial boundary candidate point set. This point set typically exhibits a distribution pattern around the tumor edge and has strong boundary discrimination capability.

[0058] In another embodiment, it is assumed that the extracted direction change rate map contains 512 × 512 = 262,144 pixels, of which 15,840 pixels have a change rate greater than 0.7, accounting for 6.04% of the total number of pixels. These pixels are distributed in a ring shape on the image, surrounding the high-energy region, which constitutes the initial boundary candidate point set.

[0059] Morphological closing operations are performed on the initial set of candidate boundary points to fill the boundary gaps and smooth the edge contours, resulting in candidate boundary regions. Connectivity analysis is then performed on the candidate boundary regions to remove isolated pixels and retain the boundary regions that satisfy the area and shape constraints. The remaining regions are used as the boundaries of the lung tumor.

[0060] In one embodiment, a closing operation with a 5×5 circular kernel as the structuring element is used to perform edge closure processing on the initial set of candidate boundary points, effectively connecting the boundary interruption regions and smoothing the edge shape. Subsequently, a connected component analysis algorithm is used to extract continuous regions. Regions with an area threshold of 500 pixels and a roundness factor greater than 0.4 are retained, while other isolated pixels or noise regions are removed. The final retained boundary region is the lung tumor boundary.

[0061] In another embodiment, assuming the initial boundary candidate points form multiple small regions, a closing operation yields 7 boundary candidate regions. After connectivity analysis, 4 regions have an area less than 300 pixels, and 2 regions have a roundness less than 0.4, so they are all excluded. Finally, one main boundary region with an area of ​​1380 pixels and a roundness of 0.61 is retained. This region is located slightly to the left of the center of the image, with an overlap rate of 89.7% with the manually labeled boundary, and its visual outline is clear.

[0062] Preferably, step S2 includes the following steps:

[0063] Step S21: Differentiate the lung tumor boundary and calculate the curvature value of each point; use the curvature value to mark the high curvature points of the lung tumor boundary to form candidate protrusion segments; calculate the convexity value of the selected protrusion segments, and if it is greater than the preset convexity threshold, it is determined to be a nodule protrusion region.

[0064] In one embodiment, Gaussian smoothing combined with the Sobolev gradient operator was used to perform first and second derivative operations on the lung tumor boundary in the 3D CT image to obtain the principal curvature and mean curvature at each point on the boundary. Then, a threshold κ_thr = 0.35 was set on the boundary curvature map, and all points with curvature values ​​greater than κ_thr were clustered to select spatially continuous regions as candidate protrusion segments. Next, the minimum convex hull of the enclosing shell was calculated for each protrusion segment, and the ratio of the actual surface area to the convex hull area was determined as the convexity index C. If C ≥ 1.2, the segment was marked as a nodular protrusion region. In the experiment, an average of 2-5 effective nodular regions could be identified per case, mainly concentrated at the tumor edge or near the pleura.

[0065] In another embodiment, assuming a lung tumor CT image has 5600 boundary points, after curvature calculation, approximately 780 points satisfy κ>0.4; cluster analysis yields 6 protruding segments. The convexity of these 6 segments is calculated, resulting in a maximum convexity value of 1.48, a minimum of 1.07, and an average of 1.23. Setting a convexity threshold of 1.2, 4 segments are identified as nodular protrusion regions, denoted as regions N1 to N4, providing spatial targets for subsequent tissue invasion analysis.

[0066] Step S22: Obtain tissue infiltration data from the nodule protrusion area;

[0067] In one embodiment, the changes in grayscale gradient between the tumor region and adjacent tissues in enhanced CT images are used to extract the region of origin (ROI) of areas identified as nodular protrusions, and a comparative analysis of regional growth and local texture differences is performed. The gray-level co-occurrence matrix (GLCM) is used to extract parameters such as energy, contrast, and homogeneity. If the texture gradient within a region changes abruptly and the homogeneity is below a set threshold (e.g., 0.45), the region is determined to have signs of tissue invasion. Simultaneously, the CT value transition rate is calculated as a supplementary reference indicator, combining the grayscale differences before and after image enhancement.

[0068] In another embodiment, assuming that the local texture homogeneity values ​​of nodule regions N1 to N4 are calculated as follows: N1=0.62, N2=0.39, N3=0.47, and N4=0.33, respectively, where N2 and N4 are below the threshold of 0.45, initially indicating the presence of tissue infiltration. Further CT value transition analysis was performed on regions N2 and N4, and the results showed that the transition rate of N2 was 8.5% and that of N4 was 12.1%, both significantly higher than the set standard of 5%, thus confirming them as suspicious infiltration regions, denoted as R2 and R4.

[0069] Step S23: Extract vascular invasion labels based on tissue invasion data, and use the vascular invasion labels to determine invasion candidate regions; in the invasion candidate regions, set sliding windows, and calculate the main direction deviation angle of each sliding window;

[0070] In one embodiment, for confirmed infiltration areas (such as R2 and R4), vascular structures are identified through multimodal image fusion (CT and enhanced angiography) to determine whether there is a structural discontinuity in the area where the tumor edge contacts the blood vessel. If the blood vessel contour is found to be disturbed or abnormally curved, this area is marked as a "vascular invasion label". Based on this, a three-dimensional sliding window with a size of 7×7×5 (pixels) is set inside the infiltration area, the principal direction vector of each window is extracted, and compared with the overall principal axis direction of the tumor to calculate the deviation angle.

[0071] In another embodiment, it is assumed that 45 sliding windows are set in region R2, of which 18 windows deviate from their main direction by an angle exceeding 30°; while 40 windows are set in region R4, of which 22 windows deviate from their main direction by an angle greater than 35°. The statistical results of the deviation angles preliminarily indicate that the tissue arrangement in both regions exhibits an abnormal bending and expansion trend, providing input for subsequent assessment of the degree of structural disorder.

[0072] Step S24: Calculate the variance of each sliding window based on the deviation angle of the main direction, and evaluate the degree of tissue structure disorder based on the variance; use the degree of tissue structure disorder and the deviation angle of the main direction as the second texture feature.

[0073] In one embodiment, the distribution sequence of the main direction deviation angles of each sliding window in step S23 is constructed, and its mean and variance are statistically analyzed. A higher variance value indicates more drastic changes in tissue orientation and a more irregular structure within the region; therefore, this variance value is defined as "structural disorder." Simultaneously, the mean of the main direction deviation angles is retained as a supplementary indicator. Finally, the "structural disorder" and the "mean deviation angle" are combined to form a second texture feature, used to distinguish tumor tissues in different infiltration states.

[0074] In another embodiment, it is assumed that the mean of the deviation angle sequence in region R2 is 27° and the variance is 64.3; while the mean of region R4 is 32° and the variance is 89.7. Based on the set disorder threshold of 60, both R2 and R4 satisfy the characteristic condition of highly disordered structure. Finally, the second texture features of the two regions are represented as vectors. , , serving as an important input parameter for heterogeneity analysis and lesion classification.

[0075] Preferably, step S22 includes the following steps:

[0076] Step S221: Detect the gray-scale discontinuities and locations of significant changes in direction in the nodule protrusion region, and identify the blurred boundary region based on the gray-scale discontinuities and locations of significant changes in direction.

[0077] In one embodiment, the Sobel operator and the Canny edge detection algorithm are used to perform multi-directional gradient calculations on suspected nodule regions in CT images, extracting grayscale change gradient angle maps. Abrupt grayscale direction changes are identified in the angle maps, and regions where the grayscale direction difference between adjacent pixels is greater than 45 degrees are determined as grayscale direction discontinuities. Simultaneously, the variance of direction change within a 3×3 neighborhood is calculated; regions with variance values ​​exceeding a preset threshold (e.g., 20°²) are identified as locations with significant direction changes. These two regions are combined to form a boundary-blurred region for subsequent processing.

[0078] In another embodiment, it is assumed that 1053 points with discontinuous gray-scale direction and 982 points with significant directional changes are detected within a certain lung nodule region, with the spatial overlap between the two accounting for approximately 67%. This overlapping area is considered as a region with blurred boundaries, with a total area of ​​approximately 38 mm², distributed in a band-shaped area between 1.5 mm and 3.8 mm around the nodule, mainly located on the dorsal side of the right upper lung apex segment.

[0079] Step S222: Divide the blurred boundary region into equal-angle sector areas, and count the edge point density in each equal-angle sector area; calculate the edge density change rate in the equal-angle sector areas, extract the angle intervals where the edge density change rate is significant, and generate the edge perturbation response curve.

[0080] In one embodiment, the blurred boundary region is divided into 36 equiangular sector regions, each 10°, with the nodule center as the origin. The number of edge points within each sector region is counted and standardized to a density per unit area. A one-dimensional perturbation function is constructed using the edge density change rate (difference ratio) between adjacent sectors, and angle segments on this function whose density change rate exceeds twice the average standard deviation are identified as significant perturbation intervals. Edge perturbation response maps are plotted based on the density change rate curves of all sector regions for subsequent structure identification.

[0081] In another embodiment, it is assumed that each sector within the blurred boundary region contains an average of 42 edge points with a standard deviation of 15. In sectors 5, 6, 11, and 19, the rate of change in edge point density reaches 1.9 to 2.3 times the standard deviation, corresponding to angle ranges of 40°–60°, 100°–120°, and 180°–200°. The response curve shows three main peaks, indicating the presence of typical burr-like outward protrusions and disturbances in this region.

[0082] Step S223: Identify the peak response point in the edge perturbation response curve, and perform directional fitting on the corresponding edge point trajectory to extract the burr structure line clusters that diverge outward in strip shape;

[0083] In one embodiment, edge point tracking is performed on the main peak region of the edge disturbance response curve, and the Hough line fitting method is used to linearly approximate each set of edge trajectories. If the angle between the fitted line and the line connecting the nodule center is within ±20°, and its length is greater than 2.5 mm and its thickness ratio is greater than 3:1, it is determined to be a burr segment. Burr segments with the same direction and an adjacent distance of less than 1 mm are classified into the same burr structure cluster.

[0084] In another embodiment, a total of 58 spur lines were identified within the three main response peak angle intervals, of which 42 met the morphological requirements. Finally, eight clusters of spur structure lines were obtained, with the largest cluster containing 12 spur lines with an average divergence length of 3.2 mm. They generally pointed towards the edge of the lung lobe and were morphologically characterized as slender radial protrusions.

[0085] Step S224: Calculate the spatial density of the burr structure clusters and mark the burr-like infiltration candidate regions with high spatial density;

[0086] In one embodiment, the burr structure line clusters are projected onto the outer contour surface of the nodules in 3D space, and the directional distribution density of each cluster is calculated using a unit spherical division method. If the average directional density of a certain cluster exceeds 1.5 times the overall density average, and the concentration at the end of the line segment exceeds a set threshold (e.g., the proportion of convergence points within the maximum radius > 70%), it is marked as a high-density burr infiltration candidate area.

[0087] In another embodiment, the directional density threshold was set to 0.12 (unit: stripes / mm² / sr), and the terminal concentration threshold was set to 70%. Three structural clusters were detected that met the requirements. The largest infiltrated cluster was distributed on the right anterior lateral side of the nodule, with a density of 0.18 stripes / mm² / sr and a terminal concentration of 83%. This cluster had a spatial span of 4.8 mm and a volume of approximately 12.6 mm³, making it a highly suspected burr infiltrated area.

[0088] Step S225: Merge the spiculated invasion candidate area and the blurred boundary area to obtain the tissue invasion area; evaluate the tissue penetration degree of the tissue invasion area to obtain tissue invasion data.

[0089] In one embodiment, a morphological dilation algorithm is used to merge the burr infiltration candidate area with the boundary blurred area to form a complete tissue infiltration area. The texture consistency, gray-level gradient continuity and spatial expansion direction of the area are analyzed. Combined with the infiltrative feature definition with expert annotation, tissue penetration indexes are extracted, including average gray-level difference, texture direction difference and local expansion rate.

[0090] In another embodiment, the volume of the fused tissue infiltration area was 27.3 mm³, of which the spiculated area accounted for 46% and the blurred boundary area accounted for 54%; the average grayscale difference was 23.1, the directional difference was 0.68, and the local expansion rate was 14.5%. According to empirical rules, the tissue infiltration level was rated as "moderate to severe infiltration", indicating that the tumor tissue has an extensional growth trend.

[0091] Preferably, step S225 includes:

[0092] Pixel-level alignment is performed on the candidate areas of spiky infiltration and the areas with blurred boundaries to obtain a mask comparison map. In the mask comparison map, each corresponding pixel position is traversed. If both the spiky mask and the blurred boundary mask have positive values ​​at that position, they are counted as overlapping pixels. All overlapping pixels are merged to determine the tissue infiltration area.

[0093] In one embodiment, based on preprocessed lung CT images, a U-Net model is used to segment and obtain mask images of candidate spiculated infiltration regions and mask images of regions with blurred boundaries. Registration between the two images is achieved through sub-pixel-level bilinear interpolation, and the image sizes are standardized (e.g., adjusted to 512×512 resolution). Subsequently, a pixel-by-pixel traversal is performed on the mask comparison image. If a pixel's value is greater than 0.5 (normalized probability map) in both the spiculated and blurred boundary mask images, it is identified as an overlapping pixel. Morphological closing operations are used to fuse and fill in all overlapping pixels, generating the final tissue infiltration region mask for subsequent analysis.

[0094] In another embodiment, assuming that in a CT slice, the spiky mask contains 2350 positive pixels and the mask of the blurred boundary region contains 1980 positive pixels, after pixel-level alignment and overlap judgment, 1275 pixels satisfy the condition that "both masks are positive", and after fusion, they form a connected tissue infiltration region. This region will be used as a whole for tissue anomaly analysis and quantification in subsequent processing.

[0095] For the connected regions of the tissue invasion area, the cross-sectional area in different slice directions is extracted; equally spaced sections are constructed along the main axis of the connected region, and the contour complexity of the equally spaced sections is calculated; abnormal tissue fragments are identified using the cross-sectional area and contour complexity; the proportion of abnormal tissue fragments in the tissue invasion area is statistically analyzed and defined as the tissue invasion degree, thus obtaining tissue invasion data.

[0096] In one embodiment, connectivity analysis is first performed on the three-dimensional tissue infiltration region to extract all 3D connected components. For each connected component, its cross-section is extracted along the coronal, sagittal, and axial planes, and the cross-sectional area on each slice is calculated. Then, the principal axis direction is determined based on a principal axis extraction algorithm (such as PCA), and a profile patch is constructed every 2 mm along the principal axis direction. After contour extraction on the profile, its contour complexity index (such as the contour perimeter to area ratio, fractal dimension, etc.) is calculated. When the contour complexity of some profiles exceeds a threshold (e.g., set to the average plus 1.5 times the standard deviation) and the cross-sectional area changes abruptly (the rate of change exceeds 25%), the profile is determined to be an abnormal tissue fragment. Finally, the proportion of the total area of ​​all abnormal profiles to the total area of ​​the infiltrated region is calculated as the tissue infiltration degree, and the infiltration data is output for subsequent grading or prognostic assessment.

[0097] In another embodiment, suppose a certain 3D connected tissue infiltration region comprises 30 CT slices, with an average cross-sectional area of ​​450 mm² per slice. Seven of these slices have an area ranging from 680 to 710 mm², and their contour complexity is greater than 0.85 (higher than the average of 0.62). These seven slices are identified as abnormal tissue fragments. Their total area is approximately 4930 mm², accounting for 36.5% of the total infiltration region of 13500 mm². Based on this, the tissue infiltration degree of this region is determined to be 36.5%, which is higher than a set threshold (30%), indicating the presence of high-level tissue abnormalities requiring further investigation.

[0098] Preferably, determining tumor heterogeneity characteristics in step S3 includes:

[0099] The first texture feature and the second texture feature are matched according to their spatial location to obtain the matching spatial region;

[0100] In one embodiment, firstly, first texture features (e.g., gray-level co-occurrence matrix features) and second texture features (e.g., run length matrix features) from multimodal image sources (e.g., CT and MRI) are acquired, corresponding to the spatial coordinates of each pixel. The two feature maps are then spatially aligned using image registration techniques (e.g., affine transformation-based registration methods). Next, the image region is divided into regions with a fixed window size (e.g., 16×16 pixels). Within each window, it is determined whether the texture responses at corresponding positions in the first and second texture feature maps are consistent (e.g., texture direction consistency is higher than 0.8). Regions meeting the consistency condition are designated as matching spatial regions for subsequent analysis.

[0101] In another embodiment, assuming an image size of 512×512 pixels, a 32×32 pixel sliding window is used for scanning, scanning a total of 256 regions. The first texture feature is defined as a local entropy feature, and the second texture feature as a texture contrast feature. After aligning the two texture feature maps using an affine matrix, the cosine similarity between the two feature maps within each window is calculated. If the similarity is greater than 0.85, a match is considered successful. Ultimately, 87 matching spatial regions are obtained, covering the main part of the entire tumor region.

[0102] Calculate the texture dispersion value of each pixel within the matching space region; identify high dispersion regions within the matching space region based on the texture dispersion value; count the number of pixel blocks within the matching space region whose texture dispersion value is greater than a preset dispersion threshold;

[0103] In one embodiment, for each matching spatial region, the texture dispersion value of a pixel is calculated in both the first texture feature map and the second texture feature map, defined as the standard deviation of the texture values ​​of that pixel and its neighborhood (5×5 window). The average of the two dispersion maps is taken as the final dispersion map. A dispersion threshold is set to the mean of the overall image dispersion plus twice the standard deviation. Pixel blocks in each matching region with dispersion values ​​higher than this threshold are identified and their number is recorded to measure the degree of local texture variation.

[0104] In another embodiment, the dispersion value range of the matching region is set to [0,1], and the preset dispersion threshold is 0.65. The number of pixel blocks with a dispersion value greater than 0.65 is counted within each matching region. On average, 12 highly dispersed blocks are detected per region, with the largest block containing 64 pixels and the smallest block containing 9 pixels. These highly dispersed regions are mostly concentrated at the tumor margin or in necrotic areas within the tumor, suggesting an increased potential for tissue heterogeneity.

[0105] The density information is obtained by calculating the percentage of pixel blocks in the matching spatial region; the distribution information of pixel blocks in the matching spatial region is calculated; and the density information and distribution information are fused as tumor heterogeneity features.

[0106] In one embodiment, for each matching spatial region, the percentage of the total area of ​​discrete pixel blocks to the entire region area is first calculated to obtain density information. Then, cluster analysis (such as DBSCAN clustering) is performed on the spatial distribution of these discrete blocks to evaluate their dispersion and central tendency within the region, and distribution information is extracted. Finally, the density information and distribution information are normalized separately and then weighted and fused (with a weight ratio of 0.6:0.4) to serve as a tumor heterogeneity characteristic index for the region, which is used by downstream models for tumor subtyping or efficacy prediction.

[0107] In another embodiment, assuming a matching spatial region has a total of 1024 pixels and the detected highly discrete pixel blocks have a total of 256 pixels, the density is 25%. Spatial coordinate clustering analysis is performed on these 256 pixels, revealing that they can cluster into 4 densely distributed clusters and several discrete pixels. The average distance standard deviation is calculated to be 18 pixels, and the distribution index is 0.72. Finally, the density value of 0.25 and the distribution index of 0.72 (weighted at a ratio of 0.5:0.5) are merged to obtain a heterogeneity index of 0.485, indicating that the region has a moderate degree of structural heterogeneity.

[0108] Preferably, step S3, which utilizes tumor heterogeneity characteristics to assess potential malignancy, includes:

[0109] The tumor heterogeneity characteristics are used to segment the tumor heterogeneous regions; the tumor heterogeneous regions are divided into grid cells, and the texture orientation variance of the grid cells is calculated; the structurally disordered regions are identified based on the texture orientation variance.

[0110] In one embodiment, firstly, based on the fused CT image, first texture features (such as the energy and contrast of the gray-level co-occurrence matrix) and second texture features (such as histogram of oriented gradients or Local Binary Patterns - LBP) are extracted within the tumor region. A clustering algorithm (such as K-means or DBSCAN) is used to perform feature consistency analysis on the different texture features, thereby dividing the region into multiple areas with high heterogeneity. Then, each heterogeneous region is divided into 5×5×5 voxel cubic grid cells. The principal texture direction of all pixels within each grid is calculated, and the variance of each direction is taken as the texture direction variance of that grid. If the texture direction variance within a certain grid exceeds the mean plus one standard deviation of all grids, it is marked as a structurally disordered region.

[0111] In another embodiment, the tumor region is assumed to be divided into 8 heterogeneous regions, each of which is further divided into 125 5×5×5 voxel grid cells. After calculating the principal texture direction of each grid, its texture direction variance range is found to be [0.12, 0.95]. The average texture direction variance of all grids is calculated to be 0.42, and the standard deviation is 0.18. Therefore, the identification threshold is 0.42 + 0.18 = 0.60. Finally, a total of 84 grids in heterogeneous regions 3, 5, and 7 are identified as structurally disordered regions.

[0112] Identify the infiltration path direction in structurally disordered regions; construct an infiltration path vector set using the infiltration path direction; obtain a lung field spatial segmentation mask; project the infiltration path vector set onto the lung field spatial segmentation mask to obtain spatially overlapping regions;

[0113] In one embodiment, for the structurally disordered regions identified in the previous step, local gradient tracing is first performed on the CT images. Based on the gradient direction and continuity, the paths spreading outward from the tumor are extracted as infiltration paths. Each path can be represented as a vector using three-dimensional coordinates from the start point to the end point, constructing an infiltration path vector set. Then, an existing lung field segmentation model (such as a UNet-based segmentation network) is used to extract the lung field spatial mask. The infiltration path vectors are projected one by one onto the lung field mask coordinate system to determine whether the end of the path is located in the lung field space. If it is located in the lung field space, it is considered that there is a spatially overlapping region.

[0114] In another embodiment, it is assumed that 84 structurally disordered mesh regions were identified, from which 210 infiltration path vectors were extracted. The lung field mask is a three-dimensional voxel mask with a size of 512×512×256. After path projection analysis, it was found that the endpoint coordinates of 124 paths fall within the lung field mask, with the intersecting regions mainly concentrated between the left middle lobe and the right lower lobe. This spatially overlapping region will be used for subsequent directional analysis.

[0115] The angle between the infiltration path vector and the lung field boundary normal vector in the spatially overlapping region is calculated. If the angle is less than a preset threshold, it is recorded as the direction toward the lung field. The proportion of the path vector set in the direction toward the lung field is counted, and the potential malignancy is assessed.

[0116] In one embodiment, for each infiltration path vector identified in the spatially overlapping region in the previous step, the boundary normal vector at the path's endpoint is first calculated using a lung field boundary mask, and then the angle between the normal vector and the path vector is calculated. If the angle is less than a set threshold (e.g., 30°), the path is considered to be oriented towards the lung field. Finally, the proportion of paths meeting the criteria is counted out of all paths to assess whether the tumor has a strong tendency to infiltrate the lung field. If the proportion exceeds 50%, it suggests a high probability of malignancy.

[0117] In another embodiment, an angle threshold of 30° was set. Angle calculations were performed on 124 paths, revealing 68 paths with angles less than 30°, representing approximately 54.8%. This suggests a strong tendency for directional infiltration and certain malignant characteristics of the tumor. Further analysis, considering factors such as tumor location and structural disorder, can provide auxiliary references for radiotherapy planning and surgical boundary design.

[0118] Preferably, in step S3, the tumor risk region in lung tumor CT images is divided based on the degree of potential malignancy, including:

[0119] Set low-risk, medium-risk, and high-risk thresholds based on the degree of potential malignancy;

[0120] In one embodiment, statistical analysis is performed on the CT image data of existing lung tumor patients to extract biomarker texture features related to malignancy (such as CT value, gray variance, texture roughness, etc.). Through expert experience and retrospective analysis, the classification boundaries of three risk levels are determined respectively: the low-risk threshold is set to T1 = 120, mainly corresponding to fibrous nodules or calcified foci; the medium-risk threshold is set to T2 = 180, corresponding to solid nodules with blurred edges but no significant blood supply; the high-risk threshold is set to T3 = 250, corresponding to high-density nodules or solid tumor regions with irregular edges and rich blood supply. All thresholds are based on the HU (Hounsfield Unit) gray scale range.

[0121] In another embodiment, assume that a training statistical analysis is performed on a batch of CT images to extract the gray mean μ and standard deviation σ of each tumor region. Assume that the gray mean range is 80–300 and the standard deviation range is 5–60. Based on the results of K-means clustering analysis, set the low-risk threshold , the medium-risk threshold , and the high-risk threshold . The distribution intervals corresponding to the cluster centers are defined as: T ≤ 100: may be a benign nodule; 100 < T ≤ 200: suspected early malignant transformation area; T > 200: high-risk tissue with potential invasiveness.

[0122] Use the low-risk threshold, medium-risk threshold, and high-risk threshold to divide the lung tumor CT images to obtain low-risk regions, medium-risk regions, and high-risk regions; smooth the boundaries of the low-risk regions, medium-risk regions, and high-risk regions to form tumor risk regions.

[0123] In one embodiment, for each three-dimensional CT image, scan its gray value voxel by voxel in a sliding window manner, classify and label it according to its所属区间 (low, medium, high-risk thresholds) to construct three types of initial mask regions. Subsequently, use the 3D region growing algorithm and conditional connectivity judgment to剔除 small isolated mislabeled regions (such as those with a volume less than 5 voxels) and保留 the coherent main tumor region. Then, perform boundary smoothing on each type of region using Gaussian filtering and 3D morphological operations (dilation - erosion) to消除 "zigzag" pseudo-boundaries and fragmented mutation regions and form a regular tumor risk partition structure.

[0124] In another embodiment, assuming a 64×64×64 lung CT image block is processed, an average of 2300 low-risk voxels, 5100 medium-risk voxels, and 1600 high-risk voxels are detected, with the original distribution exhibiting interlaced boundaries. After applying a boundary smoothing algorithm, the average boundary change rate of each type of region decreases by 25%, with the medium-risk region showing the greatest improvement in boundary smoothness, where the boundary curvature coefficient decreases from 0.65 to 0.41. This ultimately results in three clearly defined tumor risk sub-blocks, labeled R_low, R_mid, and R_high, which can provide regional guidance for subsequent tumor heterogeneity analysis or 3D segmentation modeling.

[0125] Please see Fig. 2 This specification also provides a 3D segmentation system for lung tumor CT images based on multimodal image fusion, used to perform the 3D segmentation method for lung tumor CT images based on multimodal image fusion as described above. The 3D segmentation system for lung tumor CT images based on multimodal image fusion includes:

[0126] Lung tumor boundary recognition module 101 is used to acquire lung tumor CT images; determine first texture features based on lung tumor CT images; and recognize lung tumor boundaries using the first texture features.

[0127] The nodule protrusion region identification module 102 is used to detect nodule protrusion regions in the boundary of lung tumors; obtain tissue infiltration data from the nodule protrusion regions; and determine second texture features based on the tissue infiltration data.

[0128] The tumor risk region segmentation module 103 is used to determine tumor heterogeneity features based on the first texture feature and the second texture feature; to assess the potential malignancy using the tumor heterogeneity features; and to segment the tumor risk region of lung tumor CT images based on the potential malignancy.

[0129] The 3D segmentation image generation module 104 is used to adjust the image segmentation coefficients based on the tumor risk region; and to perform 3D segmentation on lung tumor CT images using the image segmentation coefficients to generate 3D segmented images.

[0130] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0131] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A 3D segmentation method for lung tumor CT images based on multimodal image fusion, characterized in that, Includes the following steps: Step S1: Acquire CT images of lung tumors; determine the first texture feature based on the CT images of lung tumors; Lung tumor boundaries were identified using first texture features; Step S2: Detect nodular protrusion regions within the lung tumor boundary; obtain tissue infiltration data from the nodular protrusion regions; determine the second texture feature based on the tissue infiltration data; Step S3: Determine tumor heterogeneity features based on the first and second texture features; assess the potential malignancy using the tumor heterogeneity features; delineate the tumor risk region in lung tumor CT images based on the potential malignancy. Step S4: Adjust the image segmentation coefficients based on the tumor risk region; use the image segmentation coefficients to perform 3D segmentation of the lung tumor CT image to generate a 3D segmented image.

2. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 1, characterized in that, Step S1, which determines the first texture feature based on lung tumor CT images, includes: For any pixel in a lung tumor CT image, extract a neighborhood region of a fixed size centered on that pixel; within the neighborhood region, calculate the grayscale difference between each pixel and the central pixel. The grayscale difference is used as the texture response value and normalized; a texture consistency metric map is constructed based on the normalized texture response value. In the texture consistency metric map, the direction of the line connecting the pixel with the maximum response value and the pixel with the minimum response value is determined, and this direction of the line is used as the first texture feature.

3. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 1, characterized in that, Step S1, which uses the first texture feature to identify the lung tumor boundary, includes: A first texture feature image of lung tumor CT images is constructed based on the first texture feature; The first texture feature image is subjected to edge enhancement processing; in the edge-enhanced texture feature image, the texture direction change rate of each pixel is calculated; Based on a preset texture change rate threshold, pixels with texture direction change rate greater than the texture change rate threshold are extracted to form an initial boundary candidate point set. Morphological closing operations are performed on the initial set of candidate boundary points to fill the boundary gaps and smooth the edge contours, resulting in candidate boundary regions. Connectivity analysis is then performed on the candidate boundary regions to remove isolated pixels and retain the boundary regions that satisfy the area and shape constraints. The remaining regions are used as the boundaries of the lung tumor.

4. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Differentiate the lung tumor boundary and calculate the curvature value of each point; use the curvature value to mark the high curvature points of the lung tumor boundary to form candidate protrusion segments; calculate the convexity value of the selected protrusion segments, and if it is greater than the preset convexity threshold, it is determined to be a nodule protrusion region. Step S22: Obtain tissue infiltration data from the nodule protrusion area; Step S23: Extract vascular invasion labels based on tissue invasion data, and use the vascular invasion labels to determine invasion candidate regions; in the invasion candidate regions, set sliding windows, and calculate the main direction deviation angle of each sliding window; Step S24: Calculate the variance of each sliding window based on the deviation angle of the main direction, and evaluate the degree of tissue structure disorder based on the variance; use the degree of tissue structure disorder and the deviation angle of the main direction as the second texture feature.

5. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 4, characterized in that, Step S22 includes the following steps: Step S221: Detect the gray-scale discontinuities and locations of significant changes in direction in the nodule protrusion region, and identify the blurred boundary region based on the gray-scale discontinuities and locations of significant changes in direction. Step S222: Divide the blurred boundary region into equal-angle sector areas, and count the edge point density in each equal-angle sector area; calculate the edge density change rate in the equal-angle sector areas, extract the angle intervals where the edge density change rate is significant, and generate the edge perturbation response curve. Step S223: Identify the peak response point in the edge perturbation response curve, and perform directional fitting on the corresponding edge point trajectory to extract the burr structure line clusters that diverge outward in strip shape; Step S224: Calculate the spatial density of the burr structure clusters and mark the burr-like infiltration candidate regions with high spatial density; Step S225: Merge the spiculated invasion candidate area and the blurred boundary area to obtain the tissue invasion area; evaluate the tissue penetration degree of the tissue invasion area to obtain tissue invasion data.

6. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 5, characterized in that, Step S225 includes: Pixel-level alignment is performed on the candidate areas of spiky infiltration and the areas with blurred boundaries to obtain a mask comparison map. In the mask comparison map, each corresponding pixel position is traversed. If both the spiky mask and the blurred boundary mask have positive values ​​at that position, they are counted as overlapping pixels. All overlapping pixels are merged to determine the tissue infiltration area. For the connected regions of the tissue invasion area, the cross-sectional area in different slice directions is extracted; equally spaced sections are constructed along the main axis of the connected region, and the contour complexity of the equally spaced sections is calculated; abnormal tissue fragments are identified using the cross-sectional area and contour complexity; the proportion of abnormal tissue fragments in the tissue invasion area is statistically analyzed and defined as the tissue invasion degree, thus obtaining tissue invasion data.

7. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 1, characterized in that, Step S3, which identifies tumor heterogeneity characteristics, includes: The first texture feature and the second texture feature are matched according to their spatial location to obtain the matching spatial region; Calculate the texture dispersion value of each pixel within the matching space region; identify high dispersion regions within the matching space region based on the texture dispersion value; count the number of pixel blocks within the matching space region whose texture dispersion value is greater than a preset dispersion threshold; The density information is obtained by calculating the percentage of pixel blocks in the matching spatial region; the distribution information of pixel blocks in the matching spatial region is calculated; and the density information and distribution information are fused as tumor heterogeneity features.

8. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 1, characterized in that, Step S3, which utilizes tumor heterogeneity characteristics to assess potential malignancy, includes: The tumor heterogeneity characteristics are used to segment the tumor heterogeneous regions; the tumor heterogeneous regions are divided into grid cells, and the texture orientation variance of the grid cells is calculated; the structurally disordered regions are identified based on the texture orientation variance. Identify the infiltration path direction in structurally disordered regions; construct an infiltration path vector set using the infiltration path direction; obtain a lung field spatial segmentation mask; project the infiltration path vector set onto the lung field spatial segmentation mask to obtain spatially overlapping regions; The angle between the infiltration path vector and the lung field boundary normal vector in the spatially overlapping region is calculated. If the angle is less than a preset threshold, it is recorded as the direction toward the lung field. The proportion of the path vector set in the direction toward the lung field is counted, and the potential malignancy is assessed.

9. The method for 3D segmentation of lung tumor CT images based on multimodal image fusion according to claim 8, characterized in that, Step S3, which involves classifying the tumor risk regions on lung tumor CT images based on their potential malignancy, includes: Set low-risk, medium-risk, and high-risk thresholds based on the degree of potential malignancy; Lung tumor CT images are divided into low-risk, medium-risk, and high-risk regions using low-risk, medium-risk, and high-risk thresholds. The boundaries of these regions are then smoothed to form tumor risk areas.

10. A 3D segmentation system for lung tumor CT images based on multimodal image fusion, characterized in that, For performing the 3D segmentation method for lung tumor CT images based on multimodal image fusion as described in claim 1, the 3D segmentation system for lung tumor CT images based on multimodal image fusion includes: The lung tumor boundary recognition module is used to acquire lung tumor CT images; determine the first texture feature based on the lung tumor CT images; and identify the lung tumor boundary using the first texture feature. The nodule protrusion region identification module is used to detect nodule protrusion regions in the boundary of lung tumors; obtain tissue infiltration data from the nodule protrusion regions; and determine the second texture feature based on the tissue infiltration data. The tumor risk region segmentation module is used to determine tumor heterogeneity features based on the first texture feature and the second texture feature; to assess the potential malignancy using the tumor heterogeneity features; and to segment the tumor risk region of lung tumor CT images based on the potential malignancy. The 3D segmentation image generation module is used to adjust the image segmentation coefficients based on the tumor risk region; and to perform 3D segmentation on lung tumor CT images using the image segmentation coefficients to generate 3D segmented images.

Images