Method for automatic detection and counting of peritoneal metastasis of ovarian cancer based on CT image

Abstract

The present application relates to the technical field of image processing algorithm, and particularly relates to an ovarian cancer peritoneal metastasis automatic detection and counting method based on CT image.The present application acquires a local lesion area in full abdominal enhanced CT image data, wherein the local lesion area comprises a plurality of lesion voxels; a lesion center voxel is determined based on distance distribution of the lesion voxels and the region edge; gray level continuity is analyzed according to the difference of Hounsfield values between adjacent voxels to obtain a tissue attribution probability; the voxels are mapped to graph nodes, the distance distribution and the tissue attribution probability are fused to calculate a connected cost, and a voxel connected cost graph is constructed; all lesion center voxels are taken as source points, multi-source shortest path expansion is performed on the cost graph, a handover boundary is divided according to the attribution relationship, and morphological correction is performed on the local lesion area to obtain an independent lesion detection result.The present application effectively distinguishes the external adhesion gap and the internal necrotic cavity by fusing the spatial depth and the gray level continuity, avoids over-segmentation, and realizes accurate counting.

Description

Technical Field

[0001] This invention relates to the field of image processing algorithm technology, specifically to an automatic detection and counting method for peritoneal metastases of ovarian cancer based on CT images. Background Technology

[0002] In the clinical diagnosis and treatment of ovarian cancer, assessing the number of peritoneal metastases is crucial for developing surgical plans and scoring the Peritoneal Cancer Index (PCI). Enhanced CT imaging of the entire abdomen is currently the primary method for observing the distribution of metastases. However, due to the crowding of organs within the abdominal cavity, multiple ovarian cancer metastatic nodules often compress and physically adhere to each other, appearing as adhesions on CT images. Simultaneously, larger single ovarian cancer nodules commonly exhibit deep tissue liquefaction and necrosis, forming internal necrotic cavities.

[0003] In existing medical image post-processing and morphological segmentation algorithms (such as conventional watershed algorithms or graph cut algorithms), CT grayscale jumps or grayscale minima are typically used to determine and cut lesion boundaries. However, external adhesion sutures (often containing trace amounts of fat or ascites) and internal necrotic cavities both appear as low-grayscale features on CT images. Existing algorithms rely solely on single local grayscale features and cannot effectively distinguish between these two types of low-grayscale regions with different origins but similar grayscale appearances. This leads to the algorithm incorrectly fragmenting the interior of necrotic tumors with a single, intact shell while attempting to cut open external adhesion nodules. This oversegmentation phenomenon results in the system outputting a significantly higher lesion count than the actual number of independent lesions. Summary of the Invention

[0004] To address the technical problem that existing algorithms cannot distinguish between external adhesion sutures and internal necrotic cavities, easily leading to over-segmentation of necrotic cavities, the present invention aims to provide an automatic detection and counting method for peritoneal metastases of ovarian cancer based on CT images. The specific technical solution adopted is as follows: This invention proposes an automated detection and counting method for peritoneal metastases of ovarian cancer based on CT images, the method comprising: A local lesion region is acquired from the enhanced CT image data of the whole abdomen, the local lesion region including several lesion voxels; the central voxel of the local lesion region is determined based on the distance distribution between the lesion voxels and the edge of the local lesion region. Based on the difference in Henle values ​​between every two adjacent voxels within the local lesion region, the gray-level coherence of each voxel is analyzed to obtain the tissue attribution probability of each voxel. Each voxel is mapped as a graph node. Based on the distance distribution between each graph node and the edge of the local lesion region, as well as the tissue attribution probability, the connectivity cost between adjacent graph nodes is calculated to construct the voxel connectivity cost graph of the local lesion region. Using all lesion center voxels within the local lesion region as source points, multi-source shortest path expansion is performed on the voxel connectivity cost graph. The affiliation of each graph node is determined based on the expansion results. The junction boundary is defined according to the location of the lesion voxel corresponding to the graph node in the affiliation relationship. Morphological correction is performed on the local lesion region to obtain the detection results of independent lesions.

[0005] Furthermore, the method for determining the central voxel of the local lesion region includes: For any lesion voxel within a local lesion region, identify marginal lesion voxels within the neighborhood space of the lesion voxel; calculate the Euclidean distance between the lesion voxel and each marginal lesion voxel, and take the smallest Euclidean distance as the boundary distance of the lesion voxel; determine the central lesion voxel based on the boundary distances of all lesion voxels within the neighborhood space of the lesion voxel.

[0006] Further, determining the central voxel of the lesion based on the boundary distances of all lesion voxels within the neighborhood space of the lesion voxel includes: Extract lesion voxels whose boundary distance is greater than or equal to the boundary distance of all lesion voxels within the neighborhood space as candidate extreme value voxels; Connectivity analysis is performed on local lesion areas. Based on the average level of the locations of all candidate extreme voxels in the same connected region in the analysis results, the lesion center voxel of each connected region is determined.

[0007] Furthermore, the method for obtaining the tissue attribution probability of each lesion voxel includes: Based on the negative correlation mapping results of the differences in Henle values ​​between each lesion voxel and its neighboring lesion voxels, the gray-level similarity between each lesion voxel and its neighboring lesion voxels is determined. A coherent probability matrix is ​​constructed based on the gray-level similarity between each lesion voxel and its neighboring lesion voxels; a feature vector is constructed based on the preset weights of each lesion voxel, and the feature vector is adjusted in combination with the coherent probability matrix to obtain the tissue attribution probability of each lesion voxel.

[0008] Furthermore, the step of adjusting the eigenvector by combining the coherent probability matrix to obtain the tissue attribution probability of each lesion voxel includes: The product of the coherent probability matrix and the eigenvector is subjected to a predetermined fixed number of diffusion iterations to obtain the final eigenvector. Extract the value of each lesion voxel in the final feature vector to obtain the tissue attribution probability of each lesion voxel.

[0009] Furthermore, the method for determining the grayscale similarity between each lesion voxel and adjacent lesion voxels includes: The Heinz values ​​of the lesion voxels within the local lesion area are mapped to obtain the gray values ​​of the lesion voxels; By performing negative correlation mapping on the absolute difference of gray values ​​between each lesion voxel and its neighboring lesion voxels, the gray similarity between each lesion voxel and its neighboring lesion voxels can be determined.

[0010] Furthermore, the method for constructing the voxel connectivity cost map of the local lesion region includes: Using the exponential decay model, the voxel connectivity value between each graph node and its adjacent graph nodes is obtained based on the tissue assignment probability and boundary distance of the voxel corresponding to each graph node. By utilizing all graph nodes within the local lesion region, the connected edges between each graph node and its adjacent graph nodes, and the voxel connectivity cost, a voxel connectivity cost graph of the local lesion region is constructed.

[0011] Furthermore, determining the affiliation of each graph node based on the expansion result includes: Obtain the shortest path to the local lesion area, and there is a relationship between each lesion voxel on the shortest path and the corresponding lesion center voxel.

[0012] Further, the step of defining the boundary based on the location of the lesion voxel corresponding to the graph node in the attribution relationship, performing morphological correction on the local lesion region, and obtaining the detection result of the independent lesion includes: For any lesion center voxel in a connected domain, obtain the lesion voxels that are adjacent to the lesion voxels in the neighborhood space and belong to other lesion center voxels, and form the boundary. Based on the boundary, the lesion voxels in the local lesion area are corrected to obtain the detection results of independent lesions.

[0013] Furthermore, the method for obtaining detection results of independent lesions includes: A three-dimensional morphological dilation operation is performed on the junction boundary to obtain the foreground boundary region; the voxel values ​​of the lesion voxels in the foreground boundary region are modified; the number of independent connected regions after modification is counted to obtain the detection results of independent lesions.

[0014] The present invention has the following beneficial effects: This invention acquires local lesion regions from enhanced abdominal CT images and determines the central voxel of the lesion based on the distance distribution between the lesion voxel and the region edge. It utilizes spatial depth information to accurately locate the core of each independent lesion, avoiding the problem of misidentifying necrotic cavities as the center due to similar gray levels, a problem inherent in traditional methods. By analyzing the differences in Heinz values ​​between adjacent voxels to obtain tissue attribution probabilities, and then fusing distance distribution and probability to calculate connectivity costs, a voxel connectivity cost map is constructed. This results in high resistance at superficial real gaps while maintaining low resistance in deep necrotic cavity regions, fundamentally solving the technical challenge of traditional algorithms failing to distinguish between external adhesion gaps and internal necrotic cavities. Furthermore, multi-source shortest path expansion is performed using the central voxel of the lesion as the source point. The boundary is delineated based on attribution relationships and morphological corrections are applied, achieving precise separation and independent counting of adhesion lesions. This effectively overcomes oversegmentation defects and significantly improves the clinical reliability and accuracy of automatic detection and counting of peritoneal metastases in ovarian cancer. Attached Figure Description

[0015] Figure 1 The flowchart illustrates an automatic detection and counting method for peritoneal metastases of ovarian cancer based on CT images, as provided in one embodiment of the present invention. Figure 2 This is a flowchart of the steps of a method for determining the tissue attribution probability of each lesion voxel according to an embodiment of the present invention. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details the specific scheme of the automatic detection and counting method for peritoneal metastases of ovarian cancer based on CT images provided by the present invention.

[0017] Please see Figure 1 The diagram illustrates a flowchart of an automatic detection and counting method for peritoneal metastases of ovarian cancer based on CT images, according to an embodiment of the present invention. The method includes: Step S1: Obtain the local lesion region in the enhanced CT image data of the whole abdomen. The local lesion region includes several lesion voxels. Based on the distance distribution between the lesion voxels in the local lesion region and the edge of the local lesion region, determine the lesion center voxel in the local lesion region.

[0018] To achieve precise localization and independent counting of multiple peritoneal metastases of ovarian cancer in enhanced abdominal CT images, it is first necessary to extract local lesion regions from global data containing multiple lesion voxels, and then locate the core location of each potential independent lesion within each local region. Traditional methods rely solely on CT grayscale thresholding or simple connected component analysis. For multiple lesions that are physically adhered together, they may be incorrectly classified as the same connected region. Furthermore, liquefied necrotic cavities commonly found within tumors are often misjudged as boundaries due to their extremely low grayscale values, leading to central localization shifts or oversegmentation.

[0019] In this embodiment of the invention, the local lesion area is a spatially connected set of lesion voxels obtained by dividing the area into regions. The distance distribution of each lesion voxel to the edge (surface) of the region can quantify the depth of the lesion voxel inside the lesion. The greater the distance, the closer the lesion voxel is to the core. Based on this distance distribution, the central voxels of the lesion are extracted. These central voxels represent the deepest core position of each independent lesion and are not sensitive to superficial low-grayscale noise such as necrotic cavities and calcification points.

[0020] Therefore, to overcome the limitation of traditional grayscale analysis in distinguishing between adhesive lesions and necrotic cavities, this invention first acquires local lesion regions (including several lesion voxels) from enhanced abdominal CT images. Then, based on the distance distribution between the lesion voxels and the edge of the local lesion region, the central voxel of the local lesion region is determined. In adhesive scenarios, this provides robust and accurate starting anchor points for each potential independent lesion, laying a reliable foundation for subsequent multi-source segmentation and counting based on connected cost graphs.

[0021] Furthermore, the method for obtaining the local lesion area includes: The enhanced abdominal CT image data is organized into a three-dimensional voxel matrix according to the spatial location of voxels. This matrix serves as the basic input source for the entire algorithm, providing spatial structure and density feature input to the preset three-dimensional convolutional neural network. The three-dimensional voxel matrix is ​​then input into the preset three-dimensional convolutional neural network. Through layer-by-layer extraction of image features, a three-dimensional probability tensor with the same size as the enhanced abdominal CT image is output. Each value in this tensor represents the predicted probability that the corresponding voxel belongs to a tumor region. The three-dimensional probability tensor is then compared with a preset binarization threshold. Voxel values ​​corresponding to voxels whose three-dimensional probability tensor is greater than the preset binarization threshold are set to 1, denoted as foreground voxels, while the voxel values ​​of other voxels are set to 0, denoted as background voxels. The connected regions formed by all foreground voxels constitute the initial three-dimensional connected region. Within the initial three-dimensional connected region, all voxel values ​​are 1, representing all suspected lesion locations initially identified by the three-dimensional convolutional neural network. It should be noted that in this embodiment of the invention, the preset three-dimensional convolutional neural network is the 3DU-Net network commonly used in the field of medical image segmentation. In other embodiments, the implementer can choose a suitable three-dimensional convolutional neural network according to the specific implementation scenario. The preset binarization threshold is 0.5, which represents the minimum confidence boundary of the algorithm in identifying voxels as tumor tissue.

[0022] Furthermore, the spatial position of each foreground voxel within its 26-neighborhood is analyzed within the initial 3D connected region. If other foreground voxels exist within the 26-neighborhood of each foreground voxel, then the foreground voxel is considered spatially connected to these other foreground voxels within its 26-neighborhood, and they are assigned the same region label. After checking all foreground voxels within the initial 3D connected region, the set of foreground voxels with the same region label is considered as an independent local lesion region. It should be noted that subsequent steps are performed independently within a single local lesion region, and the voxels within the local lesion region are recorded as lesion voxels.

[0023] It should be noted that during the initial data acquisition phase, the output of a full abdominal enhanced CT scan consists of multiple layers of two-dimensional cross-sectional images arranged continuously along the head-to-toe direction (Z-axis). Each image is composed of several two-dimensional pixels, and each pixel corresponds to a specific spatial location and a Henle value (HU). However, due to the definite slice thickness and interslice spacing of CT scans, these continuous two-dimensional images are not isolated in space, but are tightly stacked according to their respective Z-axis coordinates, forming a regular three-dimensional data raster. Therefore, after reading all the two-dimensional images, based on the Z-axis coordinates of each layer and the X and Y coordinates of the pixels within the image, each two-dimensional pixel is expanded into a three-dimensional unit with depth information, namely a voxel. Specifically, the pixel with coordinates (x, y) in the k-th layer image is mapped to the voxel position in the three-dimensional matrix with Z-axis index k, X-axis index x, and Y-axis index y. The value of this voxel is the Henle value of the original pixel. Through this coordinate mapping and stacking operation, the originally discrete two-dimensional CT image sequence can be reconstructed into a three-dimensional enhanced CT voxel matrix with dimensions (Width × Height × Depth), where Width and Height are the width and height dimensions of a single CT image, and Depth is the total number of CT image slices. This three-dimensional voxel matrix completely preserves the spatial geometric relationship of the lesion in the X, Y, and Z directions, and serves as the basic input for all subsequent three-dimensional image processing steps.

[0024] Preferably, in some implementations of the present invention, the method for determining the lesion center voxel of the local lesion region includes: For any lesion voxel within a local lesion region, identify marginal lesion voxels within the neighborhood space of the lesion voxel; calculate the Euclidean distance between the lesion voxel and each marginal lesion voxel, and take the smallest Euclidean distance as the boundary distance of the lesion voxel; determine the central lesion voxel based on the boundary distances of all lesion voxels within the neighborhood space of the lesion voxel.

[0025] In local lesion regions, lesion voxels refer to foreground voxels in enhanced abdominal CT images, serving as the basic unit for subsequent depth analysis and feature voxel localization. This embodiment uses any one lesion voxel as an example to illustrate the process, examining the distribution of other foreground or background voxels within the lesion voxel's neighborhood space to determine whether the corresponding lesion voxel is located on the surface or inside the lesion. In this embodiment, the neighborhood space range of each lesion voxel is used as the neighborhood space range. If at least one background voxel exists within the 26 neighborhoods of a lesion voxel, then that lesion voxel is identified as an edge lesion voxel, representing that the lesion voxel is located on the boundary surface of the lesion. All edge lesion voxels are collected, and using the three-dimensional Euclidean distance formula, for any lesion voxel, the Euclidean distance (spatial distance) between the lesion voxel and each edge lesion voxel is calculated. The minimum Euclidean distance is taken as the boundary distance of that lesion voxel. The minimum Euclidean distance represents the closest distance from a single lesion voxel to the lesion surface (edge), while the boundary distance quantifies the depth of the foreground voxel within the lesion. The farther the lesion voxel is from the lesion surface (edge), the deeper it is embedded within the lesion, and the larger the corresponding boundary distance.

[0026] For each lesion voxel within the local lesion region, its boundary distance is compared one by one with the boundary distances of all lesion voxels within its 26 neighborhoods. Based on the comparison results, the lesion center voxel is obtained. The lesion center voxel represents the deepest core location of a potential independent lesion, serving as the starting point for subsequent multi-source pathway expansion.

[0027] Preferably, in some implementations of the present invention, determining the central voxel of the lesion based on the boundary distances of all lesion voxels within the neighborhood space of the lesion voxel includes: Extract lesion voxels whose boundary distance is greater than or equal to the boundary distance of all lesion voxels within the neighborhood space as candidate extreme value voxels; perform connected component analysis on the local lesion region, and determine the lesion center voxel for each connected component based on the average level of the location of all candidate extreme value voxels in the same connected component in the analysis results.

[0028] Specifically, by extracting local maxima based on the boundary distance of lesion voxels, the potential local deepest point inside the lesion is reflected. As a specific example, taking any lesion voxel in the local lesion region as an example, if the boundary distance of the lesion voxel is greater than or equal to the boundary distance of all other lesion voxels within the 26-neighborhood of the lesion voxel, then the lesion voxel is regarded as a candidate extreme voxel.

[0029] Candidate extreme voxels represent potential local deepest points within a lesion. However, a flat region (an independent lesion) may contain multiple local deepest points. Using the local deepest point as the source point for subsequent path expansion would cause an independent lesion to split into multiple lesions, interfering with lesion counting. Therefore, it is necessary to further determine a central voxel to represent an independent lesion.

[0030] Furthermore, based on connected component labeling, all candidate extremum voxels are scanned, and spatially connected candidate extremum voxels (i.e., candidate extremum voxels are adjacent to each other within a 26-neighborhood) are grouped together and labeled with the same label. All candidate extremum voxels in the deepest region within the same independent lesion are clustered into an independent connected component. A connected component represents an independent lesion; multiple independent lesions result in multiple independent connected components. The positions of candidate extremum voxels within each connected component are averaged to obtain the candidate extremum voxel corresponding to the geometric center of the connected component, which is then used as the lesion center voxel. It should be noted that if no corresponding candidate extremum voxel exists at the geometric center of a connected component, the candidate extremum voxel with the closest Euclidean distance to the geometric center of that connected component is used as the lesion center voxel for that connected component.

[0031] The lesion center voxel reflects the unique geometric center of the deepest core region of each potential independent lesion. Even if the deepest region of the lesion is a flat plateau (multiple voxel distance values ​​are equal), it can be merged into a precise center point, avoiding redundant centers or positioning deviations caused by multiple candidate extreme points. This provides a robust and unique starting anchor point for subsequent multi-source path expansion. It should be noted that the method of averaging the three-dimensional coordinates of each candidate extreme voxel and the method of calculating Euclidean distance are well-known techniques and will not be elaborated further.

[0032] The coordinates of all candidate extremum voxels within the same connected domain collectively describe the spatial extent of the deepest region of the independent lesion. The coordinates of the lesion center voxel are obtained by averaging all candidate extremum voxels within the same connected domain, thus aggregating multiple discrete candidate extremum voxels into a single lesion center voxel. This lesion center voxel may not be located at the same position as the candidate extremum voxels, but it represents the geometric center of the lesion's core region. At this point, each lesion center voxel represents an independent lesion.

[0033] Step S2: Based on the difference in Henle values ​​between every two adjacent voxels in the local lesion area, analyze the gray-level coherence of each voxel and obtain the tissue attribution probability of each voxel; map each voxel to a graph node, calculate the connectivity cost between adjacent graph nodes based on the distance distribution between each graph node and the edge of the local lesion area, and the tissue attribution probability, and construct the voxel connectivity cost graph of the local lesion area.

[0034] To achieve accurate segmentation between adherent lesions within a local lesion area, while avoiding incorrect cutting of internal necrotic cavities, a passage resistance map is needed to distinguish between superficial true sutures and deep necrotic cavities. Traditional methods rely solely on the grayscale differences (Henness differences) between adjacent lesion voxels to determine whether to cut. However, grayscale differences exhibit low similarity at both external sutures and internal necrotic cavities, making it impossible to distinguish between the two. This results in segmentation algorithms either failing to cut through true adhesions or mistakenly cutting necrotic cavities as boundaries.

[0035] In this embodiment of the invention, the difference in Henle values ​​between two adjacent voxels within a local lesion region can be used to analyze the grayscale coherence of each voxel, thereby obtaining the tissue attribution probability of each voxel. This probability can indicate whether the voxel belongs to continuous solid tissue (the probability is low in real gaps and necrotic cavities, both of which exhibit low grayscale). However, this tissue attribution probability alone cannot distinguish between the two. After mapping each lesion voxel to a graph node, the distance distribution between each graph node and the edge of the local lesion region (i.e., boundary distance, reflecting spatial depth) and the tissue attribution probability are fused to construct a depth-adaptive connectivity cost model. In this model, the boundary distance acts as an exponential suppression factor, which can suppress the connectivity cost of deep necrotic cavities to a low value (keeping them unobstructed), while amplifying the connectivity cost of shallow real gaps to a high value (forming a barrier).

[0036] Therefore, to overcome the limitation of traditional grayscale segmentation in distinguishing between external gaps and internal necrotic cavities, this invention first analyzes the grayscale coherence of each lesion voxel based on the difference in Heinz values ​​between every two adjacent voxels within the local lesion region, obtaining the tissue attribution probability of each voxel. Then, each voxel is mapped as a graph node. Based on the distance distribution between each graph node and the edge of the local lesion region, and the tissue attribution probability, the connectivity cost between adjacent graph nodes is calculated, constructing a voxel connectivity cost map of the local lesion region. Nonlinear correction of the grayscale probability using spatial depth information ensures that only shallow, real gaps exhibit high traffic resistance in the final cost map, while deep necrotic cavity regions maintain low resistance. This provides an accurate map basis for subsequent multi-source path expansion to achieve "cutting only gaps, not necrotic cavities."

[0037] Preferably, such as Figure 2 As shown, in some implementations of the present invention, the method for obtaining the tissue attribution probability of each lesion voxel can be implemented by steps S201 to S203.

[0038] Step S201: Based on the negative correlation mapping results of the differences in Heinz values ​​between each lesion voxel and adjacent lesion voxels, determine the gray-level similarity between each lesion voxel and adjacent lesion voxels.

[0039] As a specific example, other lesion voxels within a 26-neighborhood of the lesion voxel are considered as adjacent lesion voxels of the lesion voxel.

[0040] The greater the difference in Henle value between each lesion voxel and its neighboring lesions, the greater the difference in tissue density between the two lesions (e.g., the junction of tumor and fat), and the more likely they belong to different entities. Negative correlation mapping converts the difference into gray-level similarity between neighboring lesion voxels using an inverse proportional function; the smaller the difference, the closer the gray-level similarity between each lesion voxel and its neighboring lesions is to 1; the larger the difference, the closer the gray-level similarity is to 0. This mapping quantifies the gray-level relationship between each lesion voxel and its neighboring lesions into a similarity value between 0 and 1, serving as the basis for subsequent adjustments (diffusion iteration).

[0041] Specifically, the normalization process is illustrated using any lesion voxel as an example. Any adjacent lesion voxel of the lesion voxel is designated as the selected adjacent lesion voxel. The gray-level similarity between the lesion voxel and the selected adjacent lesion voxel is normalized. The sum of the gray-level similarities between the selected adjacent lesion voxel and all adjacent lesion voxels is calculated. Then, the gray-level similarity between the lesion voxel and the selected adjacent lesion voxel is divided by this sum to obtain the normalized gray-level similarity, thus achieving the normalization operation.

[0042] Step S202: Construct a coherent probability matrix based on the gray-level similarity between each lesion voxel and its neighboring lesion voxels.

[0043] Following the global index order of lesion voxels, the normalized gray-level similarity is filled into the coherent probability matrix. The number of rows and columns of the matrix equals the total number of lesion voxels within the local lesion region. Each row of the matrix corresponds to one lesion voxel, and each column corresponds to one lesion voxel. The element at the i-th row and j-th column of the matrix represents the gray-level similarity between the i-th and j-th lesion voxels. It should be understood that the sum of all values ​​in the same column of the coherent probability matrix is ​​1.

[0044] Specifically, the grayscale similarity between each lesion voxel and other non-adjacent lesion voxels is 0. It should be understood that if a lesion voxel is outside the 26-neighborhood, it is not adjacent.

[0045] For example, if the i-th lesion voxel and the j-th lesion voxel are adjacent lesion voxels, the element value is the gray-level similarity between the i-th lesion voxel and the j-th lesion voxel; if the i-th lesion voxel and the j-th lesion voxel are non-adjacent lesion voxels, the element value is 0. This characterizes the local probability transfer relationship between lesion voxels based on gray level, while facilitating subsequent processing and preventing memory overflow caused by large-scale data.

[0046] Step S203: Construct a feature vector based on the preset weight of each lesion voxel, and adjust the feature vector in combination with the coherent probability matrix to obtain the tissue attribution probability of each lesion voxel.

[0047] Because independent lesions may contain calcifications or contrast agent residues in microvessels, the local grayscale abrupt changes exhibited by these internal tissues are often equal to or even greater than the grayscale tomography at the external compression gaps. To smooth out the interference of such internal noise, it is necessary to adjust the feature vector by combining it with a coherent probability matrix. Initially, each lesion voxel is assigned a preset weight, forming a feature vector. Multiplying the feature vector by the coherent probability matrix is ​​equivalent to performing a "probability diffusion." For each lesion voxel, a new weight is calculated based on the current weights of its neighboring lesion voxels (i.e., other lesion voxels in direct contact within a 26-neighborhood) and the grayscale similarity between them, thus obtaining a new feature vector. The value of each voxel in the new feature vector is the tissue affiliation probability, reflecting the confidence that the voxel belongs to a continuous solid tissue.

[0048] It should be noted that, in the embodiments of the present invention, the preset weight of each lesion voxel is the reciprocal of the total number of lesion voxels in the local lesion area, which means that the weight of all lesion voxels is the same in the initial state.

[0049] Preferably, in some implementations of the present invention, the step of adjusting the feature vector by combining the coherent probability matrix to obtain the tissue attribution probability of each lesion voxel includes: The product of the coherent probability matrix and the eigenvector is subjected to a predetermined number of diffusion iterations to obtain the final eigenvector; the value of each lesion voxel in the final eigenvector is extracted to obtain the tissue attribution probability of each lesion voxel.

[0050] Multiplying the coherent probability matrix by the eigenvector on the left is equivalent to performing one probabilistic iterative diffusion, outputting the eigenvector of the first iteration. Multiplying the eigenvector of the first iteration by the coherent probability matrix again yields the eigenvector of the second iteration. After a predetermined fixed number of iterative product operations, the final eigenvector is obtained. Each iteration further diffuses the weights in regions with strong connectivity, while gradually losing weights in regions with weak connectivity (such as gaps). The final eigenvector stably reflects the strength of each lesion voxel's affiliation in the globally connected structure.

[0051] If calcifications or blood vessels are present within an isolated lesion, localized grayscale abrupt changes will occur, initially appearing as a fault line. However, it is surrounded by continuous healthy tissue. Through multiple iterations, the numerous high-probability weights of the surrounding tissue will "feed back" and submerge this noise, proving that it is merely internal impurity rather than a true boundary. At true adhesion gaps, the sharp drop in grayscale effectively severs the "bridge" of tissue attribution probability transmission. Tissue attribution probabilities cannot cross the gap to complement each other, causing the weights of edge voxels to continuously decrease during iterations, ultimately forming a true low-probability cliff at the gap. It should be understood that through multiple such iterative diffusions, the weight of each lesion voxel will gradually reflect the overall grayscale continuity of its region. Voxels located within continuous solid tissue will maintain a high weight level because their surrounding adjacent voxels have similar grayscale values ​​(high grayscale similarity); while voxels located in true gaps or necrotic cavities will gradually lose weight due to large grayscale differences with adjacent voxels (low grayscale similarity), ultimately resulting in a low tissue attribution probability.

[0052] After iteration, the value corresponding to each lesion voxel in the feature vector is the tissue assignment probability of that lesion voxel. A high tissue assignment probability (close to 1) indicates that the lesion voxel is surrounded by a large amount of tissue with similar gray levels, belonging to the interior of continuous solid tissue; a low tissue assignment probability (close to 0) indicates that the lesion voxel is located at a gray-level discontinuity (such as a crevice or necrotic cavity). This tissue assignment probability provides a coherent evaluation of the gray-level attribution level for subsequent fusion with boundary distance.

[0053] It should be noted that, in this embodiment of the invention, the preset fixed number of times is in the range of [10, 50], so as to ensure that the diffusion range covers the local noise area while avoiding excessive smoothing. Preferably, in this embodiment, the preset fixed number of times is 20, which is an empirical value obtained through a large number of experiments.

[0054] Preferably, in some implementations of the present invention, the method for obtaining the grayscale similarity of adjacent lesion voxels includes: The Heinz values ​​of the lesion voxels within the local lesion area are mapped to obtain the gray values ​​of the lesion voxels; the absolute difference of the gray values ​​of adjacent lesion voxels is negatively correlated to obtain the gray similarity between adjacent lesion voxels.

[0055] As a specific example, for all lesion voxels within a local lesion area, when the Heinz value of a lesion voxel is greater than the upper grayscale bound of a preset soft tissue window, the Heinz value of the lesion voxel is replaced with the upper grayscale bound; when the Heinz value is less than the lower grayscale bound of the preset soft tissue window, the Heinz value of the lesion voxel is replaced with the lower grayscale bound, thus completing the interval truncation. Using a linear interpolation algorithm, the truncated Heinz values ​​between the lower and upper bounds are proportionally mapped to the [0, 255] interval to generate the grayscale value of each lesion voxel. This process effectively eliminates extreme value interference from surrounding irrelevant tissues, ensuring that subsequent steps have a unified and stable data scale when calculating the grayscale similarity of adjacent voxels. It should be noted that, in this embodiment of the invention, the upper and lower grayscale bounds of the preset soft tissue window are 215HU and -135HU, respectively; this process is a well-known technical means and will not be elaborated here.

[0056] It should be noted that in this embodiment of the invention, a preset soft tissue window and a linear interpolation algorithm are used to map the Henness values. In other embodiments, the implementer may choose a suitable mapping method according to the specific implementation scenario.

[0057] The absolute difference in grayscale values ​​between adjacent lesion voxels visually quantifies the magnitude of density variation between adjacent lesion voxels (tissues). In CT imaging physics, when adjacent tissues belong to the same entity (such as within a continuous tumor), the density change is gradual, and the absolute difference in grayscale is minimal. However, when tissues undergo physical discontinuity or encounter substances of different densities (such as adhesions with fat or ascites), the density jumps, and the absolute difference in grayscale increases sharply. Adding a positive integer 1 to the absolute difference and taking the reciprocal transforms this difference into a probability of continuity. The smaller the difference, the larger the reciprocal, and the higher the probability of classifying them as belonging to the same tissue; the larger the difference, the smaller the reciprocal, and the higher the probability of classifying them as physically discontinuous.

[0058] Preferably, in some implementations of the present invention, the method for constructing the voxel connectivity cost map of the local lesion region includes: Using the exponential decay model, the voxel connectivity cost between each graph node and its adjacent graph nodes is obtained based on the tissue affiliation probability and boundary distance of the voxel corresponding to each graph node. Using all graph nodes in the local lesion region, the connectivity edges between each graph node and its adjacent graph nodes, and the voxel connectivity cost, a voxel connectivity cost graph of the local lesion region is established.

[0059] Since the liquefied necrotic tissue deep within independent lesions has lost the structural morphology of normal healthy cells, the tissue attribution probability calculated in the aforementioned steps will also be extremely small. If the division boundary is determined solely by this extremely small tissue attribution probability, it will inevitably lead to the incorrect fragmentation of the internal necrotic region. Therefore, the boundary distance, reflecting the depth of voxel space, is introduced into the exponential decay model to forcibly modify the connectivity costs at different depths, achieving selective control by preserving deep connectivity and blocking superficial connectivity.

[0060] Optionally, before performing the preset exponential decay model solution, in order to prevent the tissue attribution probability of all graph nodes from being too small and causing subsequent multiplication calculations to lose mathematical precision, a preliminary probability relativization operation is first performed. This involves traversing all lesion voxels within the local lesion region and obtaining the maximum value among all tissue attribution probabilities. Using this maximum value, a division operation is performed on the tissue attribution probability of each lesion voxel within the local lesion region, thereby obtaining a relatively normalized probability with values ​​distributed in the interval (0,1].

[0061] As a specific example, the graph nodes corresponding to other voxels within the 26 neighborhoods of the graph node corresponding to the lesion voxel are taken as the adjacent graph nodes of the graph node. A connection relationship is established between the graph node and the adjacent graph nodes, and there is no connection relationship between the graph node and the non-adjacent graph nodes.

[0062] Substituting the relative normalized probability of the graph node and the normalized proportion of the boundary distance into the exponential decay model yields the voxel connectivity value between adjacent graph nodes. A higher voxel connectivity value indicates greater graph structural resistance for the algorithm to cross adjacent lesion voxels at that location, and the algorithm is more likely to stop path growth and generate a boundary at that position. For example, taking any graph node and any adjacent graph node as an example, the exponential decay model is as follows: in, The voxel connectivity value between graph node i and its adjacent graph node j is represented. Indicates the value of basic space. This represents the preset cost amplification factor. , This represents the relative normalized probability of graph node i and its adjacent graph node j. This represents the ratio of the boundary distance of node i in the graph to the maximum boundary distance within the local lesion region. This represents the ratio of the boundary distance between adjacent graph node j to the maximum boundary distance within the local lesion area.

[0063] This formula uses spatial depth to nonlinearly suppress grayscale probability, resulting in high passage cost in shallow gaps and low passage cost in deep necrotic cavities. This guides subsequent multi-source shortest path expansion to converge and cut off only at the actual adhesion boundary, thus achieving lesion segmentation constraints.

[0064] It should be noted that the basic spatial cost is used to compensate for the difference in the actual spatial step size in the discrete 3D mesh. In this embodiment, the value is 1.0 when two adjacent lesion voxels share a face; 1.414 when two adjacent lesion voxels share an edge; and 1.732 when two adjacent lesion voxels share a vertex. The preset cost amplification factor is used to control the limit of the jump in resistance value when encountering a real gap. In this embodiment, this constant is set in the range of [50, 200], preferably 100.

[0065] Using each voxel within the local lesion region as a graph node, connected edges are established for nodes with existing connectivity, while no edges are established for nodes without connectivity. The calculated voxel connectivity cost is assigned to the corresponding connected edges, thus constructing a complete weighted graph structure. This graph is a voxel connectivity cost graph, representing a three-dimensional access resistance map of the local lesion region. Edges with low cost values ​​represent easily accessible areas (inside the lesion or deep necrotic cavities), while edges with high cost values ​​represent difficult-to-access barriers (shallow real gaps). Subsequent multi-source shortest path expansion will be performed on this graph to determine the segmentation boundaries of different lesion center voxels.

[0066] Step S3: Using all the central voxels of the lesions in the local lesion area as the source points, perform multi-source shortest path expansion on the voxel connectivity cost graph, and determine the attribution relationship of each graph node based on the expansion results; divide the junction boundary according to the location of the lesion voxel corresponding to the graph node in the attribution relationship, perform morphological correction on the local lesion area, and obtain the detection results of independent lesions.

[0067] To completely separate and accurately count multiple interconnected independent tumors within a local lesion region, it is necessary to utilize a pre-constructed voxel connectivity cost map. Starting from the central voxel of each lesion, competitive expansion is used to determine the boundaries of different lesions. Traditional methods directly apply thresholding or watershed algorithms to grayscale images, failing to utilize the depth adaptive information in the cost map. This results in incorrect boundaries at necrotic cavities or failure to cut at actual gaps.

[0068] This invention considers using all lesion center voxels within a local lesion area as source points and performing multi-source shortest path expansion on the voxel connectivity cost graph. This allows each source point to occupy voxels outward along a low-cost path, and the expansions of different source points will naturally meet at high-cost junction boundaries. Based on the expansion results, the affiliation of each graph node (i.e., which source point each voxel ultimately belongs to) is determined, thereby obtaining the affiliation division of the entire region. Based on the location of the lesion voxels corresponding to the graph nodes in the affiliation relationship, the interface between adjacent voxels with different affiliations is extracted as the junction boundary. Morphological correction (such as dilation and widening) is applied to this junction boundary to completely sever the adhesion region, ultimately obtaining the detection result of independent lesions.

[0069] Therefore, to overcome the limitation of traditional segmentation in not being able to distinguish between gaps and necrotic cavities using depth information, this embodiment of the invention first uses all lesion center voxels within the local lesion area as source points and performs multi-source shortest path expansion on the voxel connectivity cost graph. Then, based on the expansion results, the attribution relationship of each graph node is determined. The boundary is then defined according to the location of the lesion voxel corresponding to the graph node in the attribution relationship. Finally, morphological correction is performed on the local lesion area to obtain the detection results of independent lesions. Utilizing the characteristic of high cost for shallow gaps and low cost for deep necrotic cavities in the cost graph, multi-source expansion automatically forms the boundary at the actual gap, avoiding incorrect cutting of necrotic cavities and achieving accurate separation and counting of adherent lesions.

[0070] Preferably, in some implementations of the present invention, determining the affiliation of each graph node based on the expansion result includes: Obtain the shortest path to the local lesion area, and there is a relationship between each lesion voxel on the shortest path and the corresponding lesion center voxel.

[0071] Using all lesion center voxels within the local lesion region as the starting point (i.e., source point) for path expansion, the algorithm employs a multi-source shortest path expansion mechanism. A custom voxel connectivity cost, integrating lesion depth features (boundary distance) and tissue affiliation probability, is used as the edge weight to obtain the shortest path for the local lesion region. Each lesion center voxel is assigned a unique identifier (number), with an initial connectivity cost of 0, ensuring that subsequent expansion starts simultaneously from all lesion center voxels. Therefore, each source point will alternately spread outwards, macroscopically appearing as simultaneous outward expansion from multiple centers until a collision occurs at a gap. If an adjacent lesion voxel is reached by the current lesion voxel with a smaller connectivity cost, the affiliation number of the adjacent lesion voxel is updated to the number of the lesion center voxel to which the current lesion voxel belongs, achieving dynamic allocation of region affiliation. Once all lesion voxels have been processed and each lesion voxel has been assigned to a source point (lesion center voxel), the attribution relationship of each lesion represents the partitioning result of all lesion voxels in the local lesion area, and the boundary between different numbered areas is the lesion segmentation surface.

[0072] Preferably, in some implementations of the present invention, the step of dividing the boundary according to the location of the lesion voxel corresponding to the graph node in the attribution relationship, performing morphological correction on the local lesion region, and obtaining the detection result of the independent lesion includes: For any lesion center voxel in a connected domain, obtain the lesion voxels that are adjacent to the lesion voxel in the neighborhood space and belong to other lesion center voxels, forming the boundary; based on the boundary, correct the lesion voxels in the local lesion region to obtain the detection results of independent lesions.

[0073] If the lesion voxel of the connected domain exists within the 26-neighborhood of the lesion center voxel, it means that the lesion center voxel and the lesion voxel of the connected domain are adjacent in the neighborhood space.

[0074] In a local lesion region, each connected domain corresponds to a unique lesion center voxel, which represents the deepest core location of an independent lesion within that domain. For any lesion center voxel in a connected domain, its affiliation is determined by the multi-source shortest path expansion result. All lesion voxels ultimately dominated by this lesion center voxel (i.e., lesion voxels on the same path as the lesion center voxel in the multi-source shortest path expansion result) constitute the voxel set occupied by an independent lesion within the local lesion region. We then obtain lesion voxels that are spatially adjacent to the lesion voxels in this voxel set but belong to other lesion center voxels. These lesion voxels are spatially adjacent to the lesion represented by the lesion center voxel but belong to different independent individuals. The spatial interface between them is precisely where the expansion paths intersect, corresponding to real physical adhesion gaps (lacking continuous tissue transitions, with tiny compression gaps causing abrupt grayscale changes). We extract these lesion voxels located at the affiliation boundary and combine them to form the junction boundary.

[0075] In a localized lesion region, the voxel value of the lesion voxel is 1. The voxel value of the lesion voxel at the boundary is changed from 1 to 0, thus modifying the voxel value of the lesion voxels located at the boundary within the localized lesion region, thereby severing the connectivity between regions representing different attribution relationships. In the corrected localized lesion region, the originally adhered independent lesions are separated, forming multiple unconnected lesion voxel clusters, each corresponding to an independent lesion. By counting the number of these lesion voxel clusters, the detection result of the independent lesion can be obtained.

[0076] Preferably, in some implementations of the present invention, the method for obtaining the detection results of independent lesions includes: A three-dimensional morphological dilation operation is performed on the junction boundary to obtain the foreground boundary region; the voxel values ​​of the lesion voxels in the foreground boundary region are modified; the number of independent connected regions after modification is counted to obtain the detection results of independent lesions.

[0077] To prevent severance failure, morphological reinforcement is introduced. In a discrete 3D mesh space, simply changing the voxel value of a single-layer lesion voxel to zero may still allow adjacent lesion voxels to penetrate and connect along the diagonal direction of the 26-neighborhood structure. Therefore, 3D structural elements (such as 3×3×3 cubes) are used to expand the lesion voxels at the junction boundary, thickening the cutting interface from a single layer to multiple layers. This eliminates the diagonal connectivity remaining due to only vertex or edge contact in the discrete mesh, ensuring that the adhered area is completely severed, resulting in the foreground boundary region. The voxel values ​​of all lesion voxels covered by the foreground boundary region in the local lesion region are changed from 1 to 0, spatially isolating voxel clusters belonging to different lesion centers and achieving physical independence. Connectivity labeling is performed on the modified local lesion region, and the number of non-connected lesion voxel clusters is counted. Each cluster corresponds to an independent lesion, and this statistical result is the detection result of the independent lesion. The morphological expansion operation described above ensures that the cutting interface has sufficient voxel thickness, avoids residual adhesion caused by diagonal connectivity in the discrete mesh, and guarantees the thoroughness of lesion separation.

[0078] Furthermore, all steps S1 to S3 above are continuously executed in a loop until all local lesion areas in the enhanced abdominal CT image have been processed. The detection results of individual lesions in each processed local lesion area are summed up to obtain the global final detection result. This global final detection result is output to the medical terminal display to complete the automatic detection and counting of peritoneal metastases of ovarian cancer.

[0079] In summary, this invention first obtains local lesion regions containing several lesion voxels from enhanced abdominal CT images. The central voxel of the lesion is determined based on the distance distribution between the voxels and the region edges. Then, within the local lesion region, grayscale coherence is analyzed based on the difference in Henle values ​​between adjacent voxels to obtain the tissue attribution probability of each voxel. The voxels are then mapped as graph nodes, and the connectivity cost between adjacent nodes is calculated by fusing the distance distribution and tissue attribution probability to construct a voxel connectivity cost graph. Finally, using all central voxels as source points, multi-source shortest path expansion is performed on the cost graph. Boundary boundaries are delineated based on attribution relationships, and morphological corrections are applied to the local lesion regions to obtain the detection results of independent lesions. This invention effectively distinguishes between external adhesion gaps and internal necrotic cavities by fusing spatial depth features and grayscale coherence, avoiding over-segmentation and achieving accurate independent lesion counting.

[0080] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0081] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. An automated detection and counting method for peritoneal metastases of ovarian cancer based on CT images, characterized in that, The method includes: A local lesion region is acquired from the enhanced CT image data of the whole abdomen, the local lesion region including several lesion voxels; the central voxel of the local lesion region is determined based on the distance distribution between the lesion voxels and the edge of the local lesion region. Based on the difference in Henle values ​​between every two adjacent voxels within the local lesion region, the gray-level coherence of each voxel is analyzed to obtain the tissue attribution probability of each voxel. Each voxel is mapped as a graph node. Based on the distance distribution between each graph node and the edge of the local lesion region, as well as the tissue attribution probability, the connectivity cost between adjacent graph nodes is calculated to construct the voxel connectivity cost graph of the local lesion region. Using all lesion center voxels within the local lesion region as source points, multi-source shortest path expansion is performed on the voxel connectivity cost graph. The affiliation of each graph node is determined based on the expansion results. The junction boundary is defined according to the location of the lesion voxel corresponding to the graph node in the affiliation relationship. Morphological correction is performed on the local lesion region to obtain the detection results of independent lesions. The method for constructing the voxel connectivity cost map of the local lesion region includes: Using the exponential decay model, the voxel connectivity value between each graph node and its adjacent graph nodes is obtained based on the tissue assignment probability and boundary distance of the voxel corresponding to each graph node. A voxel connectivity cost graph of the local lesion region is constructed by utilizing all graph nodes within the local lesion region, the connected edges between each graph node and its adjacent graph nodes, and the voxel connectivity cost. The exponential decay model is as follows: in, The voxel connectivity value between graph node i and its adjacent graph node j is represented. Indicates the value of basic space. This represents the preset cost amplification factor. , This represents the relative normalized probability of graph node i and its adjacent graph node j. This represents the ratio of the boundary distance of node i in the graph to the maximum boundary distance within the local lesion region. This represents the ratio of the boundary distance between adjacent graph node j to the maximum boundary distance within the local lesion area.

2. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 1, characterized in that, The method for determining the central voxel of a local lesion region includes: For any lesion voxel within a local lesion region, identify marginal lesion voxels within the neighborhood space of the lesion voxel; calculate the Euclidean distance between the lesion voxel and each marginal lesion voxel, and take the smallest Euclidean distance as the boundary distance of the lesion voxel; determine the central lesion voxel based on the boundary distances of all lesion voxels within the neighborhood space of the lesion voxel.

3. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 2, characterized in that, The step of determining the central voxel of the lesion based on the boundary distances of all lesion voxels within the neighborhood space of the lesion voxel includes: Extract lesion voxels whose boundary distance is greater than or equal to the boundary distance of all lesion voxels within the neighborhood space as candidate extreme value voxels; Connectivity analysis is performed on local lesion areas. Based on the average level of the locations of all candidate extreme voxels in the same connected region in the analysis results, the lesion center voxel of each connected region is determined.

4. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 1, characterized in that, The method for obtaining the tissue attribution probability of each lesion voxel includes: Based on the negative correlation mapping results of the differences in Henle values ​​between each lesion voxel and its neighboring lesion voxels, the gray-level similarity between each lesion voxel and its neighboring lesion voxels is determined. A coherent probability matrix is ​​constructed based on the gray-level similarity between each lesion voxel and its neighboring lesion voxels; a feature vector is constructed based on the preset weights of each lesion voxel, and the feature vector is adjusted in combination with the coherent probability matrix to obtain the tissue attribution probability of each lesion voxel.

5. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 4, characterized in that, The process of adjusting the eigenvector by combining the coherent probability matrix to obtain the tissue attribution probability of each lesion voxel includes: The product of the coherent probability matrix and the eigenvector is subjected to a predetermined fixed number of diffusion iterations to obtain the final eigenvector. Extract the value of each lesion voxel in the final feature vector to obtain the tissue attribution probability of each lesion voxel.

6. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 4, characterized in that, The method for determining the grayscale similarity between each lesion voxel and adjacent lesion voxels includes: The Heinz values ​​of the lesion voxels within the local lesion area are mapped to obtain the gray values ​​of the lesion voxels; By performing negative correlation mapping on the absolute difference of gray values ​​between each lesion voxel and its neighboring lesion voxels, the gray similarity between each lesion voxel and its neighboring lesion voxels can be determined.

7. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 3, characterized in that, The process of determining the affiliation of each graph node based on the expansion results includes: Obtain the shortest path to the local lesion area, and there is a relationship between each lesion voxel on the shortest path and the corresponding lesion center voxel.

8. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 7, characterized in that, The step of defining the boundary based on the location of the lesion voxel corresponding to the node in the attribution relationship, performing morphological correction on the local lesion region, and obtaining the detection result of the independent lesion includes: For any lesion center voxel in a connected domain, obtain the lesion voxels that are adjacent to the lesion voxels in the neighborhood space and belong to other lesion center voxels, and form the boundary. Based on the boundary, the lesion voxels in the local lesion area are corrected to obtain the detection results of independent lesions.

9. The method for automatic detection and counting of peritoneal metastases of ovarian cancer based on CT images according to claim 8, characterized in that, The methods for obtaining detection results for independent lesions include: A three-dimensional morphological dilation operation is performed on the junction boundary to obtain the foreground boundary region; the voxel values ​​of the lesion voxels in the foreground boundary region are modified; the number of independent connected regions after modification is counted to obtain the detection results of independent lesions.

Images