Computer vision segmentation system for epileptogenic zone nuclear magnetic medical images

Abstract

The application discloses a computer vision segmentation system for epilepsy lesion in nuclear magnetic medical images, and relates to the technical field of medical image computer processing. The system comprises a data acquisition module, a seed determination module, an interlayer deformation calculation module, an interlayer index calculation module, a graph construction module, an optimal path generation module, a cross-slice filling module and a reconstruction output module. The data acquisition module acquires MRI tomographic scanning slice sequences and obtains single-slice lesion candidate contours through segmentation. The seed determination module calculates contour definition indexes and determines seed slices. The interlayer deformation calculation module calculates cross-slice contour displacement vectors, interlayer deformation amplitudes and thin plate spline deformation fields based on anatomical landmark points. The interlayer index calculation module calculates interlayer contour similarity and interlayer emergence intensity. The graph construction module constructs an interlayer emergence consistency graph with the seed slices as root nodes. The optimal path generation module generates an optimal backtracking path tree covering all the slices. The cross-slice filling module maps and fuses parent node contours to the current slice. The reconstruction output module reconstructs a three-dimensional consistent lesion body and outputs the same. The application can make the lesion edge keep a real anatomical shape and spatial continuity in a three-dimensional space.

Description

Technical Field

[0001] This invention relates to the field of medical image computer processing technology, specifically to a computer vision segmentation system for MRI medical images of epilepsy lesions. Background Technology

[0002] Epilepsy is a common neurological disorder. Hippocampal sclerosis is the main pathological basis for drug-resistant epilepsy. In preoperative evaluation, magnetic resonance imaging is the core means to locate the epileptogenic focus. Physicians identify the lesion by analyzing tomographic scan sections and determine the location and extent of the sclerotic area.

[0003] However, the slice thickness of routine clinical MRI is usually 3 to 5 millimeters. Hippocampal sclerosis lesions often appear as blurred boundaries and mixed signals on a single slice. Some slices may even only show faint abnormalities. Current processing usually performs segmentation on each slice independently and then simply stacks the results into three-dimensional volume data. This processing method ignores the spatial continuity of lesions between adjacent slices, resulting in false connections or breaks between slices. The edges of lesions are difficult to maintain the true anatomical shape in three-dimensional space.

[0004] Because hippocampal sclerosis lesions have the pathological characteristic of infiltrating and growing along the gray-white matter boundary, their true three-dimensional morphology often cannot be fully displayed on a single slice. When the segmentation results of a single slice are blurred or missing, the clear information from adjacent slices should provide supplementary information. However, existing technologies lack a mechanism to effectively integrate interlayer spatial correlation into segmentation decisions.

[0005] Therefore, how to effectively complete the blurred or broken lesion boundaries in a single slice by utilizing the spatial correlation information between adjacent slices in a conventional MRI tomographic scan sequence, so as to obtain a three-dimensional spatially continuous epileptogenic focus localization result that is consistent with the pathological morphology, has become an urgent problem to be solved. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a computer vision segmentation system for MRI medical images of epileptiform lesions. This system can effectively complete the blurred or broken lesion boundaries in a single slice by utilizing the spatial correlation information between adjacent slices in a conventional MRI tomographic scan sequence, thereby obtaining a three-dimensional spatially continuous epileptiform lesion localization result that is consistent with the pathological morphology.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention discloses a computer vision segmentation system for magnetic resonance imaging of epilepsy lesions, comprising:

[0009] The data acquisition module is used to acquire MRI tomographic scan slice sequences, extract the region of interest in the hippocampus from each slice in the MRI tomographic scan slice sequence and perform lesion segmentation to obtain single-slice lesion candidate contours;

[0010] The seed determination module is used to calculate the contour clarity index of candidate lesion contours for each single slice, and to determine the slice with the highest contour clarity index as the seed slice.

[0011] The interlayer deformation calculation module is used to calculate the cross-slice contour displacement vector, interlayer deformation amplitude, and thin plate strip deformation field based on the corresponding anatomical landmarks in the region of interest of the hippocampus of adjacent slices.

[0012] The inter-layer index calculation module is used to calculate the inter-layer contour similarity based on the cross-slice contour displacement vector and the inter-layer emergence intensity based on the contour sharpness index.

[0013] The graph construction module is used to construct a set of candidate parent nodes for each non-root node with the seed slice as the root node and calculate the corresponding edge weights. The edge weights include adjacent layer edge weights and cross-layer edge weights. Based on the edge weights, an inter-layer emergent consistency graph covering all slice nodes is constructed.

[0014] The optimal path generation module is used to generate an optimal backtracking path tree covering all slice nodes, starting from the seed slice and selecting the node with the largest edge weight from its set of candidate parent nodes for each non-root node based on the inter-layer emergent consistency graph.

[0015] The cross-slice filling module is used to perform cross-slice evidence filling along the optimal backtracking path tree. Based on the thin plate strip deformation field, the lesion candidate contour of the parent node slice is mapped to the slice corresponding to the current non-root node and fused to obtain the filled cross-slice contour.

[0016] The reconstruction output module is used to perform three-dimensional consistency verification on the filled cross-slice contours, remove isolated artifact contours, reconstruct them into three-dimensional consistent lesions, and output them.

[0017] Secondly, this invention discloses a computer vision segmentation method for MRI medical images of epilepsy lesions, including:

[0018] MRI tomographic scan slice sequences are obtained. For each slice in the MRI tomographic scan slice sequence, the region of interest in the hippocampus is extracted and lesion segmentation is performed to obtain candidate lesion contours for a single slice.

[0019] Calculate the contour clarity index of each single slice candidate lesion contour, and determine the slice with the highest contour clarity index as the seed slice;

[0020] Based on the corresponding anatomical landmarks in the regions of interest of the hippocampus in adjacent slices, calculate the cross-slice contour displacement vector, inter-layer deformation amplitude, and thin-plate strip deformation field;

[0021] Inter-layer contour similarity is calculated based on cross-slice contour displacement vectors, and inter-layer emergence intensity is calculated based on contour sharpness index.

[0022] Using the seed slice as the root node, a set of candidate parent nodes is constructed for each non-root node and the corresponding edge weights are calculated. The edge weights include the weights of adjacent layers and the weights of cross layers. Based on the edge weights, an inter-layer emergent consistency graph covering all slice nodes is constructed.

[0023] Based on the inter-layer emergent consistency graph, starting from the seed slice, for each non-root node, the node with the largest edge weight is selected from its set of candidate parent nodes as the parent node, generating the optimal backtracking path tree covering all slice nodes.

[0024] Cross-slice evidence filling is performed along the optimal backtracking path tree. Based on the thin plate strip deformation field, the lesion candidate contour of the parent node slice is mapped to the slice corresponding to the current non-root node and fused to obtain the filled cross-slice contour.

[0025] Perform 3D consistency verification on the filled cross-slice contour, remove isolated artifact contours, reconstruct a 3D consistent lesion, and output it.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. An optimal backtracking path tree covering all slices is established by the interlayer emergence consistency map construction module and the optimal path generation module. Then, the cross-slice filling module follows the path tree to map the clear lesion candidate contours of the parent node slices through thin plate strip deformation field mapping and fuse them into the low-resolution child node slices. This effectively integrates the interlayer spatial correlation into the segmentation decision. The clear contour information of adjacent slices is used to fill in the evidence of the blurred slices, so that the edge of the lesion maintains the true anatomical shape in three-dimensional space and avoids false connections or breaks between slices.

[0028] 2. The slice with the highest contour clarity index is selected as the backtracking benchmark through the seed determination module, and cross-layer skip connections are supported by the inter-layer emergent consistency map. When the contour correspondence between adjacent layers fails due to motion artifacts or anatomical deformation, the low-reliability local path is automatically bypassed to avoid the accumulation and propagation of segmentation errors along the low-resolution slices. At the same time, the weighted fusion mechanism based on contour filling confidence suppresses the interference of unreliable mapping on the contour, thereby significantly improving the global consistency and robustness of the three-dimensional lesion reconstruction results between layers. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is an overall block diagram of the system according to Embodiment 1 of the present invention;

[0031] Figure 2 This is a schematic diagram illustrating the principle of constructing an interlayer emergent consistency map in this invention.

[0032] Figure 3 This is an overall block diagram of the method in Embodiment 2 of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1:

[0035] like Figures 1-2 As shown, the computer vision segmentation system for MRI medical images of epilepsy lesions includes:

[0036] The data acquisition module is used to acquire MRI tomographic scan slice sequences, extract the region of interest in the hippocampus from each slice in the MRI tomographic scan slice sequence and perform lesion segmentation to obtain single-slice lesion candidate contours.

[0037] For ease of understanding, we will use a hippocampal MRI scan of a patient with drug-resistant epilepsy as an example. This patient underwent T2-weighted sequence scanning, acquiring a total of 30 tomographic slices, with the slice spatial index denoted as... ,in Total number of slices. Layer thickness. pixel pitch The original size of each slice is Pixel.

[0038] It should be noted that the present invention provides a medical image computer vision segmentation system, which outputs image segmentation results of three-dimensional consistent lesions. These results are only for clinical physicians as an auxiliary diagnostic reference and cannot be directly used as a basis for disease diagnosis or treatment. The final diagnostic conclusion must be made by qualified medical personnel in combination with clinical information.

[0039] This module first receives the tomographic scan slice sequence output by the MRI equipment and establishes an ordered queue according to the slice spatial index. Simultaneously, it extracts slice metadata, including slice thickness. Pixel pitch echo time This method establishes a mapping relationship between slice metadata and slice spatial indexes. The applicable imaging conditions for this method are: the MRI sequence is a T1-weighted or T2-weighted sequence, and the slice thickness meets the following requirements. The pixel spacing is isotropic or near isotropic, the minimum detectable size of the lesion is not less than the area of ​​a single pixel, and the image signal-to-noise ratio meets the basic requirements for clinical diagnosis.

[0040] Subsequently, the region of interest in the hippocampus was extracted from each slice in the MRI tomographic scan sequence. For the first... Zhang slice Based on the prior knowledge of the hippocampus's anatomical location within the skull, the following processing was performed: Calculation of slices The grayscale histogram was fitted using a bimodal Gaussian mixture model, where gray matter corresponds to the intermediate grayscale peak, white matter to the high grayscale peak, and cerebrospinal fluid to the low grayscale peak. Grayscale intervals corresponding to the grayscale peaks were extracted, and pixels within these intervals were used as grayscale candidate regions. Within these candidate regions, based on the prior knowledge of the hippocampus's curved tubular anatomical shape, an 8-connected region labeling algorithm was used to segment the grayscale candidate regions into several connected branches. Connected branches satisfying spatial connectivity and curvature constraints were selected. Specifically, the curvature constraint was achieved by calculating the mean curvature and variance of the curvature of each connected branch contour, retaining branches with a mean curvature greater than a certain value. And the curvature variance is less than The branches are selected; those whose perimeter-to-area ratio satisfies the curved tubular shape constraint are retained. Specifically, the perimeter-to-area ratio constraint is achieved by calculating the ratio of the perimeter of each branch's profile to the area of ​​its circumscribed convex hull, and branches with a ratio within the specified range are retained. to Within the range of branches; if multiple branches simultaneously satisfy the morphological constraints, sort them by area from largest to smallest, retain the top three branches with the largest areas, and then select the branch closest to the expected distance to the hippocampal anatomical coordinates. Specifically, using the hippocampal center coordinates in the standard MNI space as a reference, rigid registration is used to map the patient image to the standard space, with the expected hippocampal center located at the standard coordinates. , , Within the range (right hippocampus), retain the branch whose spatial location is closest to the expected coordinates; if none of the branches satisfy the perimeter-area ratio constraint, relax the curvature constraint threshold to the original threshold. Double-check the selection to ensure no hippocampal region is missed. Expand outwards from the bounding rectangle that retains the branches. After cropping the image to a pixel boundary, a sub-image containing only the hippocampus and its adjacent anatomical structures is obtained, denoted as the hippocampal region of interest slice. ,in and Slices of the region of interest in the hippocampus direction and Pixel size in the direction.

[0041] Next, slices of the region of interest in each hippocampus were prepared. Perform lesion segmentation. Calculate. After smoothing the grayscale histogram with a Gaussian kernel, local maxima are identified by finding the zero-crossing points of the first derivative. The local maximum with the highest grayscale value is selected as the high-signal peak, and the peak position of the peak is recorded. With half-peak width Based on the peak position and half-width at half-maximum (WHM) of the high-signal peak, the grayscale threshold of the lesion is determined. ,in This is the peak offset coefficient. The value is adaptively determined based on the half-width at half-maximum (WHM) of the high-signal peak and the overall dispersion of the histogram; a smaller WHM is used when the WHM is large. The threshold value is relaxed. If multi-peak analysis of the grayscale histogram fails to identify significant high-signal peaks (the difference between the peak value and the adjacent valley value is less than the total dynamic range of the histogram), the threshold value is relaxed. If the dynamic threshold is not met, an adaptive threshold based on the Otsu algorithm is used instead for lesion candidate region segmentation to ensure that lesion candidate contours can still be extracted in low-contrast cases. Gray values ​​higher than this threshold are considered acceptable. The pixels marked are used as candidate pixels for lesions, forming a lesion candidate region mask. A radius of... A circular structuring element of a pixel performs a morphological closing operation on the lesion candidate region mask, connecting broken areas caused by grayscale fluctuations. The size of the structuring element is adaptively adjusted according to the minimum detectable size of the lesion and the pixel spacing to ensure that the details of the real lesion are not eliminated, resulting in connected lesion candidate regions. The set of pixel coordinates of the outer boundary of the connected lesion candidate regions is extracted as the lesion candidate contour for a single slice. .

[0042] Following the previous example, in the first In the slice, the lesion exhibits a high signal abnormality at the gray-white matter boundary. After the above segmentation process, the candidate contour of the lesion in a single slice is obtained. , containing approximately Each boundary pixel.

[0043] The seed determination module is used to calculate the contour clarity index of candidate lesion contours for each single slice, and the slice with the highest contour clarity index is determined as the seed slice.

[0044] Single-slice lesion candidate contours for each slice Calculate its outline sharpness index First, regarding the outline... For each boundary pixel, calculate its gradient magnitude. ,in The gradient magnitude represents the degree of grayscale change in the region of interest slice of the Hammerhead model at that pixel location. Slices of the region of interest in the hippocampus. and These are pixel coordinates. The average gradient magnitude of all boundary pixels is taken as the average gradient magnitude. Secondly, the statistical variance of the gradient magnitudes of all pixels on the contour is calculated as the boundary gradient variance. ,in The boundary gradient variance reflects the dispersion of the boundary gradient distribution and is related to the pathological heterogeneity of the lesion boundary. For the number of boundary pixels, For the first The gradient magnitude of each boundary pixel. This represents the average gradient magnitude. Next, the statistical variance of all pixel grayscale values ​​along the contour is calculated as the boundary pixel grayscale variance. .

[0045] Outline sharpness index The weighted combination of the above three components:

[0046] ;

[0047] in, The outline sharpness index characterizes the sharpness of the candidate lesion outline boundary and the consistency of grayscale in a single slice. , , The weight coefficients for each component are determined through data-driven calibration on the training set: the contour quality ranking annotated by experts is used as the supervision signal, and grid search optimization is employed. , , To improve the outline sharpness index Maximize the Spearman correlation coefficient with expert ranking, where , , , Negative values ​​are allowed to reflect the physical reality that contour sharpness may decrease as boundary gradient variance increases. The average gradient magnitude. For the boundary gradient variance, The grayscale variance of the boundary pixels.

[0048] Before identifying seed slices, a minimum threshold for contour sharpness is preset. If the maximum value of the outline sharpness index for all slices If this occurs, a warning is triggered and automatic segmentation is terminated, prompting the physician to conduct a manual review to avoid using noisy slices as a retrospective benchmark. The outline sharpness index of all slices is then iterated. The slice with the largest index value is identified as the seed slice. Let the spatial index of the seed slice be . ,satisfy: ;

[0049] in, Spatial index for seed slices, This is an indicator of outline sharpness.

[0050] Following the previous example, after calculation, the first... The outline sharpness index of the slice For all The maximum value in the slice, and Therefore, the first The slide was identified as a seed slide, and its lesion candidate outline was determined. This provides a baseline profile for subsequent analysis.

[0051] The interlayer deformation calculation module is used to calculate the cross-slice contour displacement vector, interlayer deformation amplitude, and thin-plate spline deformation field based on the corresponding anatomical landmarks in the region of interest of the hippocampus of adjacent slices.

[0052] For each slice of the region of interest in the hippocampus The hippocampal contour was extracted. Seed slices were used. Based on this, the following calculations are performed pairwise on adjacent slices: the hippocampal contour is parameterized to obtain the contour parameterization curve. Calculate its curvature: ;

[0053] in, Parametric curves for contours in parameters curvature at that point and The first derivative, and This is the second derivative. For discrete pixel contours, the pixels of the hippocampal contour boundary are arranged in clockwise order, with arc length as the parameter. By performing uniform parameterization, a discrete point sequence is obtained. The central difference method is used to calculate the derivatives of each order: , , and Similarly, substitute the values ​​into the curvature formula above to calculate the curvature value of each pixel. .

[0054] Identify the pixel coordinates where the curvature reaches a local maximum or minimum as the set of anatomical landmarks. For adjacent slices and Based on the set of anatomical landmarks and By establishing corresponding marker point pairs through nearest neighbor search, and using Euclidean distance as the distance metric, bidirectional nearest neighbor matching is performed: only if the marker point... The nearest neighbor is ,and The nearest neighbor is also Only then are corresponding marker pairs established. This process eliminates one-to-many false matches. For each pair of corresponding marker points, the pixel coordinate difference is calculated, and the arithmetic mean of all marker point displacement vectors is taken as the cross-slice contour displacement vector of that adjacent slice pair. , The average spatial offset of the corresponding marker points of the hippocampal contour between adjacent slices.

[0055] Set of anatomical landmarks based on adjacent slices and The deformation field of a thin plate spline is calculated by interpolation. Thin-plate spline interpolation is based on radial basis functions. ,in Let be the Euclidean distance between the marker points. Let the sets of corresponding marker points of adjacent slices be the source point sets. With the target point set Thin plate spline deformation field By solving the following system of linear equations, we obtain the minimum bending energy function. While satisfying interpolation constraints The obtained deformation field parameters include affine components and non-rigid components based on radial basis functions. The superposition of these two components constitutes the complete deformation field of the thin plate spline. The thin-plate stripe deformation field Characterizes the non-rigid mapping relationship of pixel coordinates between adjacent slices. For the strip deformation field of thin plate samples... Covered pixel positions Calculate the offset magnitude of the pixel coordinates at that location after non-rigid mapping. The mean of the offset magnitudes at all pixel locations is calculated to obtain the interlayer deformation amplitude. :

[0056] ;

[0057] in, The interlayer deformation amplitude characterizes the degree of overall non-rigid deformation of the hippocampal contour between adjacent slices. For position The deformation vector at that point, Slicing the region of interest in the hippocampus The total number of pixels.

[0058] Continuing with the previous example, the interlayer deformation amplitude between most adjacent slice pairs... Approximately to This indicates that the hippocampal outline maintains a stable anatomical correspondence between layers. However, in the first... Zhang Yudi Between slices, due to the patient's head movement during scanning, the degree of inter-slice deformation varies. It is significantly higher than other layer pairs.

[0059] The inter-layer index calculation module is used to calculate the inter-layer contour similarity based on the cross-slice contour displacement vector and the inter-layer emergence intensity based on the contour sharpness index.

[0060] Interlayer contour similarity Based on cross-slice contour displacement vector The modulus length is calculated based on exponential decay:

[0061] ;

[0062] in, Inter-slice contour similarity characterizes the degree of spatial correspondence between lesion contours in adjacent slices. The displacement attenuation coefficient, Transverse slice contour displacement vector The modulus length. This exponential decay is used to quantify the effect of spatial offset of lesion contours between adjacent slices on the degree of contour correspondence.

[0063] Interlayer emergence intensity The calculation process is as follows: For slices with adjacent slices on both sides ( The arithmetic mean of the outline sharpness indices of the two adjacent slices is calculated as the joint measure of outline sharpness. :

[0064] ;

[0065] in, This is a joint measure of contour sharpness, reflecting the expected contour sharpness of the current slice under joint analysis of adjacent slices. and These are the outline sharpness indices for adjacent slices.

[0066] For slices with only one side adjacent slices ( or ), calculate the outline sharpness index of its adjacent slices as a joint measure of outline sharpness: ;

[0067] in, For the joint metric of outline sharpness, This is an indicator of the sharpness of the outlines of adjacent slices.

[0068] For the current slice Calculate the outline sharpness index Combined with outline sharpness measurement The absolute difference is used to obtain the interlayer emergence intensity. : ;

[0069] in, Interlayer emergence intensity, characterizing a single slice The intensity of the blurred contour information due to lesion projection truncation is revealed by joint analysis of adjacent slices.

[0070] Following the previous example, the first The outline sharpness index of the slice , No. The outline sharpness index of the slice Then the first Joint metric of the outline sharpness of the slice Interlayer emergence intensity . No. The outline sharpness index of the slice its adjacent slices , ,but Interlayer emergence intensity This indicates that there are significant differences in interlayer information in the slice, and the lesion outline may be blurred due to projection truncation.

[0071] The graph construction module is used to construct a set of candidate parent nodes for each non-root node, with the seed slice as the root node, and calculate the corresponding edge weights. The edge weights include adjacent layer edge weights and cross-layer edge weights. Based on the edge weights, an inter-layer emergent consistency graph covering all slice nodes is constructed.

[0072] Indexed by seed slice space The root node is defined as [node name], and all non-root nodes are defined as [node name]. Construct a set of candidate parent nodes. For the current non-root node, determine the inter-layer deformation magnitude between it and its neighboring nodes. Does it exceed the preset interlayer deformation tolerance? .

[0073] If the number of adjacent nodes does not exceed the limit, then the adjacent nodes are included in the candidate parent node set.

[0074] If it exceeds the limit, then search in the preset cross-layer search window. The function internally calculates the cumulative cross-slice profile displacement vector and cumulative inter-layer deformation amplitude between the current non-root node and each candidate cross-layer node. For non-root nodes in the decreasing direction, the function is applied to the window according to a symmetry rule. The cumulative cross-slice profile displacement vector and cumulative inter-layer deformation amplitude are calculated internally. This cumulative displacement vector is the vector sum of the displacement vectors of each adjacent layer within the window in spatial order. For example, for the ... Zhang slice and candidate parent node Slices, cumulative cross-slice contour displacement vectors For from the first Zhang Daodi Zhang, Di Zhang Daodi The vector sum of the cross-slice contour displacement vectors of Zhang, i.e. Cumulative interlayer deformation amplitude This is the arithmetic sum of the corresponding interlayer deformation amplitudes, i.e. .in, To accumulate the cross-slice contour displacement vector, and This represents the cross-slice contour displacement vector between adjacent slices. To accumulate the interlayer deformation amplitude, and This represents the interlayer deformation amplitude between adjacent slices. The cross-layer search window size. According to the layer thickness The expected cross-layer continuity length of the lesion is adaptively determined. If the cross-layer edge weights of all candidate cross-layer nodes within the cross-layer search window are lower than the minimum admission threshold... Then expand the window to Re-search, maximum not exceeding If no node meets the conditions, the adjacent node will be included in the candidate parent node set by default to ensure that the path tree is not interrupted.

[0075] Calculate cross-layer contour similarity based on the magnitude of the cumulative cross-slice contour displacement vector:

[0076] ;

[0077] in, For cross-layer contour similarity, The displacement attenuation coefficient, This is the cumulative cross-slice contour displacement vector.

[0078] The cross-layer edge weights are obtained by weighting the monotonically decreasing mapping of cross-layer contour similarity and cumulative inter-layer deformation amplitude with the inter-layer emergence intensity of candidate cross-layer nodes:

[0079] ;

[0080] in, For cross-layer edge weights, , , These are the weighted combination coefficients. This is a deformation mapping function, specifically in the form of: , These are the deformation mapping coefficients. The inter-layer emergence intensity of candidate cross-layer nodes. This is to accumulate the interlayer deformation amplitude.

[0081] The cross-layer edge weights must meet the preset minimum admission threshold. Candidate cross-layer nodes are included in the candidate parent node set. If the edge weights of all nodes in the candidate parent node set are lower than the minimum admission threshold... If so, then by default, adjacent nodes will be included in the candidate parent node set.

[0082] For each pair of adjacent slices Calculate the edge weights of adjacent layers :

[0083] ;

[0084] in, The weights of adjacent layers. For interlayer contour similarity, This represents the interlayer deformation amplitude. For interlayer emergence intensity, Interlayer deformation amplitude The monotonically decreasing mapping.

[0085] Using the seed slice as the root node, each non-root node It has a set of candidate parent nodes, forming an inter-layer emergent consistency graph that covers all slice nodes.

[0086] Following the previous example, interlayer deformation tolerance Cross-layer search window Minimum admission threshold Displacement attenuation coefficient Deformation mapping coefficient Weighted combination coefficient , , Expanding in the decreasing direction to the first When the slice is cut, it is related to the adjacent slice. Interlayer deformation amplitude of a slice Exceeding the inter-layer deformation tolerance triggers a cross-layer search. (In the window...) Within, calculate the first according to the symmetry rule of decreasing direction. Zhang Daodi Zhang's cumulative cross-slice contour displacement vector With cumulative interlayer deformation amplitude ,Right now and Due to the first Zhang Zhidi Zhang's displacement vector With the Zhang Zhidi Zhang's displacement vector The directions are opposite, resulting in a smaller cumulative modulus, which is then used to calculate the cross-layer contour similarity and cross-layer edge weights. The calculated cross-layer edge weights... It satisfies the minimum admission threshold and is greater than the weight of the adjacent layer edge. Therefore, the first Zhang slice included in the first The set of candidate parent nodes for each slice is used as the parent node in subsequent optimal path selection.

[0087] The optimal path generation module is used to generate an optimal backtracking path tree covering all slice nodes based on the inter-layer emergent consistency graph, starting from the seed slice and selecting the node with the largest edge weight from its candidate parent node set for each non-root node.

[0088] From seed slices Starting from the beginning, the process expands layer by layer towards both ends along the sequence of MRI tomographic slices. In each expansion step, the current non-root node selects the node with the largest edge weight in the inter-layer emergent consistency graph from its set of candidate parent nodes as its parent node, establishing a parent-child connection. The set of candidate parent nodes includes adjacent nodes and nodes across layers.

[0089] For non-root nodes located in the direction of seed slice growth ( ), select the node with the largest edge weight from its candidate parent node set as the parent node; for non-root nodes located in the decreasing direction of the seed slice ( Following the symmetry rule, the tree expands layer by layer until all slice nodes are included in the spanning tree, forming an optimal backtracking path tree that covers the entire slice space index range. .

[0090] This spanning tree allows for jumping across levels while ensuring that no nodes are missed.

[0091] Following the previous example, from seed slices Starting from the beginning and expanding in the direction of increase, the first... Slice number 1 is selected from the set of candidate parent nodes. Zhang slice is the parent node; the first Slice selection number The slice is used as the parent node. When expanding in the decreasing direction, the first slice... Slice selection number Zhang slice is the parent node; the first Slice selection number The slice is used as the parent node. Expanding in the decreasing direction to the [number]th slice... When slicing a slice, since its candidate parent node set contains the first slice... Slice (adjacent layer edge weights) ) and the Slice (cross-layer edge weight) The weight of cross-layer edges is larger, therefore the first layer edge has a higher weight. Slice selection number A slice serves as the parent node, forming a cross-layer jump connection. This process expands layer by layer until it eventually covers the entire area. Slice nodes to form the optimal backtracking path tree. .

[0092] The cross-slice filling module is used to perform cross-slice evidence filling along the optimal backtracking path tree. Based on the thin plate strip deformation field, the lesion candidate contour of the parent node slice is mapped to the slice corresponding to the current non-root node and fused to obtain the filled cross-slice contour.

[0093] Along the optimal backtracking path tree The determined parent-child connection order is advanced layer by layer from the seed slice to the child nodes. For each edge in the optimal backtracking path tree, the contour clarity index of the slice corresponding to the current non-root node is compared with the contour clarity index of the parent node's slice. If the contour clarity index of the slice corresponding to the current non-root node is lower than that of the parent node's slice, cross-slice evidence filling is performed on the slice corresponding to the current non-root node.

[0094] If the parent node is an adjacent node, the inverse mapping relationship is directly solved numerically through the deformation field of the thin-plate sample. This maps the lesion candidate contour of the parent node slice to the coordinate space of the slice corresponding to the current non-root node, thus obtaining the mapped contour. Specifically, for the pixels on the parent node contour... Solve for the condition that satisfies of As the mapped coordinates in the current slice, the fixed-point iteration method is used to solve the problem, with the initial value set to... until the Euclidean distance between two adjacent iterations is less than 1. Pixels or reaching the maximum number of iterations The signal converges at this time.

[0095] If the parent node is a cross-layer node, then the thin-plate strip deformation fields of each adjacent layer between the parent node slice and the slice corresponding to the current non-root node are obtained. The pixel coordinates of the lesion candidate contour of the parent node slice are sequentially transferred to the coordinate space of the slice corresponding to the current non-root node through the inverse mapping of the thin-plate strip deformation fields of each adjacent layer, thus obtaining the mapped contour. Specifically, let the parent node slice be... The current child node slice is ,and (Increasing direction), the deformation field of the adjacent thin-plate sample between the two is as follows: For any pixel on the outline of the parent node. First, calculate its position. Layer mapping coordinates That is, through The inverse mapping is passed to Layer, then calculate exist Layer mapping coordinates And so on, until it is passed on to The final mapped coordinates are obtained by traversing the layers. The accuracy of this composite mapping decreases as the number of skip layers increases, and the cross-layer search window... It is designed to keep accumulated errors within an acceptable range.

[0096] Calculate the Hausdorff distance between the mapped contour and the candidate lesion contour of the slice corresponding to the current non-root node. If Hausdorf is far Greater than the preset contour filling threshold If the pixel coordinate distance between the mapped contour and the candidate contour of the lesion in the slice corresponding to the current non-root node is greater than the preset spatial tolerance, then... The pixel values ​​are used as the normalized edge weights of the current parent-child connection as the confidence score for contour filling. The candidate lesion contours are supplemented to the slice corresponding to the current non-root node according to the nearest neighbor principle. Specifically, the contour filling confidence level is... ,in The maximum value of all edge weights in the inter-layer emergent consistency graph represents the reliability of cross-layer mapping; the coordinates of the supplementary points... Its nearest neighbor in the current contour Coordinates after fusion This is to reflect the reliability of cross-layer mapping and avoid excessive interference of low-confidence mapping on the contour.

[0097] Boundary smoothing based on B-spline curve fitting is performed on the supplemented contour to obtain the filled cross-slice contour. Specifically, adopt A B-spline curve is used, with the supplemented contour pixels as data points, and a control point sequence is obtained by least squares fitting. The number of control points is set to the number of contour pixels. to and not less than To ensure a smooth finish while preserving the true pathological morphology of the lesion boundaries, the fitted B-spline curve was resampled into a pixel sequence using a uniform sampling method to obtain the filled cross-slice contour. .

[0098] Following the previous example, proceed in the decreasing direction to the first... When slicing, its outline sharpness index Lower than the parent node The outline sharpness index of the slice This triggers cross-slice evidence filling. Because the parent node's... Zhang slice and the current number The slices are in a cross-layer skip relationship, obtaining the first... Zhang Yudi Zhang, Di Zhang Yudi The thin-plate stripe deformation field between the two plates will be the first The candidate lesion contours of the slices are sequentially transmitted to the first slice via the inverse mapping of the above two thin-plate strip deformation fields. The coordinate space of the slice is used to obtain the mapped contour. The mapping contour is then calculated to the coordinate space of the slice. Hausdorff distance between candidate contours of original lesions on slices pixels larger than the contour filling threshold Pixel. Map the contour to the first pixel. The pixel coordinate distance of the candidate contour of the lesion in the slice is greater than the spatial tolerance. The confidence level is filled with the outline of each pixel. To supplement the weights, B-spline curve fitting and boundary smoothing are performed to eliminate jagged artifacts introduced by cross-layer mapping, thus obtaining the first... The cross-slice outline after the fill of the slice .

[0099] Continue along the optimal backtracking path tree in the decreasing direction until the th When slicing, its outline sharpness index Lower than the parent node The outline sharpness index of the slice This triggers cross-slice evidence filling. The parent node is an adjacent node, directly based on the thin-plate strip deformation field. The inverse mapping relationship is solved by numerical iteration, and the first... The candidate lesion contours of the slice are mapped to the first... The mapped contour is obtained by slicing the coordinate space of the first slice. The mapped contour is then compared with the first slice. Hausdorff distance between candidate contours of original lesions on slices pixels larger than the contour filling threshold Pixel, will map the contour to the first The pixel coordinate distance of the candidate contour of the lesion in the slice is greater than the spatial tolerance. The confidence level is filled with the outline of each pixel. To supplement the weights, perform B-spline curve fitting and boundary smoothing to obtain the first... The cross-slice outline after the fill of the slice .

[0100] The reconstruction output module is used to perform three-dimensional consistency verification on the filled cross-slice contours, remove isolated artifact contours, reconstruct them into three-dimensional consistent lesions, and output them.

[0101] For the filled cross-slice contour set Perform the following processing: fill the cross-slice outline of each slice. Convert to contour-bound region mask Specifically, a scan-line filling algorithm is used: after filling each slice, the cross-slice contour is... ,exist The scanning process proceeds line by line, calculating the intersection points between the scan lines and the contour boundaries. Pixels between adjacent intersection points are marked as true (inside) and pixels outside the contour are marked as false, generating a two-dimensional binary mask.

[0102] For each pair of adjacent slices and Based on the mask of the region enclosed by its contour and Calculate the overlap area ratio of the regions: ;

[0103] in, This represents the number of pixels in the intersection region of the two masks where the true value is true. The number of pixels in the union region of the two masks that are true.

[0104] Based on the deformation field of thin plate spline Calculate the deformation uniformity correction factor. The deformation uniformity correction factor is an exponential decay function of the average amplitude of the deformation field of the thin plate spline:

[0105] ;

[0106] in, This is the deformation consistency correction coefficient, which characterizes the effect of the degree of deformation between adjacent slices on the continuity of the contour. The deformation uniformity attenuation coefficient, For thin plate strip deformation field The mean of the offset magnitude at all pixel locations, i.e. .

[0107] Multiplying the area overlap ratio by the deformation consistency correction factor yields the contour continuity measure. :

[0108] ;

[0109] in, It is a measure of contour continuity, characterizing the degree of continuity of lesion contours between adjacent slices under the dual constraints of space and deformation.

[0110] If the continuity measure of the cross-slice contour after filling is lower than the continuity threshold of both the cross-slice contours after filling of the adjacent slices on both sides, then... If so, the filled cross-slice contour is determined to be an isolated artifact and removed. For edge slices ( or Only the contour continuity metric of its only adjacent side slice is checked. If the contour continuity metric of that single side slice is lower than the continuity threshold, the subsequent slice is considered closed. of If the value is multiplied by a factor of 1, the edge slice contour is determined to be an isolated artifact and removed; The coefficient is used to compensate for the confidence loss of unilateral examination compared to bilateral examination, avoiding excessive rejection of true lesions at the margins due to the lack of contralateral reference. Continuity threshold. The determination is based on the type of lesion and the thickness of the slice. For early lesions with a smaller expected volume, Downgraded to Only when both sides are lower than Elimination is triggered only when the lesion is below a certain threshold, and elimination is not triggered if the lesion is on one side, in order to avoid the accidental deletion of small lesions.

[0111] Obtain the slice thickness of MRI computed tomography (CT) slice sequences With pixel spacing The cross-slice contours, after removing isolated artifacts, are filled and retained according to the slice thickness. With pixel spacing Voxel stacking and inter-slice interpolation are performed along the arrangement direction of the MRI tomographic slice sequence to reconstruct a three-dimensional volume mask, resulting in a three-dimensionally consistent lesion. During the voxel stacking stage, based on the pixel pitch... Sure - In-plane voxel resolution depends on layer thickness Sure The interlayer spacing along the axial direction forms a mask that encloses the two-dimensional contours of each slice. along Axis according to layer thickness The layers are stacked at equal intervals to form the initial three-dimensional volume data. Due to the layer thickness... Typically larger than the pixel pitch , Axis resolution lower than - For planar areas, interlayer interpolation is required: for regions with gentle contour changes between adjacent slices, linear interpolation is used to generate an intermediate layer contour between the two layers; for regions with drastic contour changes or large deformations, a thin-plate spline deformation field-based interpolation method is used. Deformation-guided interpolation is used to deform the upper layer contour along the deformation field direction, which is then used as the middle layer contour. Voxels are filled and lesion areas are marked to obtain a three-dimensional volume mask. .

[0112] Continuing with the previous example, the continuity threshold Deformation uniformity attenuation coefficient In the first At the point of the slice, the filled cross-slice outline is consistent with the first slice. Zhang, Di The contour continuity measures of the slice contours are respectively and All are below the continuity threshold. Determine the first The contours of the selected slices are isolated artifacts and are removed. The contour continuity metrics of the remaining slices are all above a threshold and are retained. The retained slices... The outline of the slice is based on the layer thickness. With pixel spacing Voxel stacking and interlayer interpolation are performed along the slice arrangement direction to reconstruct a three-dimensional volume mask. The size is The voxels belonging to the lesion area are marked as true, and the rest are marked as false. The three-dimensional consistent lesion body is output.

[0113] Through the collaborative processing of the above modules, this system focuses computational resources on potential lesion regions by extracting the region of interest in the hippocampus and performing adaptive lesion segmentation; it provides a highly reliable baseline contour for inter-layer backtracking by selecting the seed slice with the highest contour clarity index; it establishes a non-rigid mapping relationship between pixel coordinates of adjacent slices by calculating cross-slice contour displacement vectors, inter-layer deformation amplitude, and thin-plate strip deformation fields based on anatomical landmarks, overcoming the limitation of rigid registration in adapting to soft tissue deformation; it explicitly encodes the spatial offset of lesion contours and inter-layer information differences into measurable indicators by using inter-layer contour similarity and inter-layer emergence intensity, enabling lesion information hidden in blurred slices due to projection truncation to be revealed under inter-layer joint analysis; and it constructs an inter-layer emergence consistency map and allows cross-layer jump connections, in order to achieve the desired effect. When the deformation of adjacent layers is too large, a more reliable alternative path is automatically sought to ensure that the optimal backtracking path tree fully covers all slice nodes. By performing cross-slice evidence filling along the optimal backtracking path tree, the lesion contour information of the high-resolution parent node slice is mapped and fused to the low-resolution child node slice, filling in the contour loss caused by partial volume effect or boundary blurring. At the same time, weighted fusion based on contour filling confidence and B-spline boundary smoothing suppress cross-layer mapping artifacts. By performing three-dimensional consistency verification on the filled cross-slice contour, isolated artifacts are removed based on contour continuity measurement, and voxel stacking and inter-layer interpolation are performed according to slice thickness and pixel spacing to reconstruct a three-dimensional consistent lesion in three-dimensional space that maintains the true anatomical morphology and spatial continuity, providing physicians with three-dimensional visualization auxiliary information consistent with pathological morphology.

[0114] Example 2:

[0115] like Figure 3 As shown, the computer vision segmentation method for epileptic lesions on MRI images includes the following steps:

[0116] MRI tomographic scan slice sequences are obtained. For each slice in the MRI tomographic scan slice sequence, the region of interest in the hippocampus is extracted and lesion segmentation is performed to obtain candidate lesion contours for a single slice.

[0117] Calculate the contour clarity index of each single slice candidate lesion contour, and determine the slice with the highest contour clarity index as the seed slice;

[0118] Based on the corresponding anatomical landmarks in the regions of interest of the hippocampus in adjacent slices, calculate the cross-slice contour displacement vector, inter-layer deformation amplitude, and thin-plate strip deformation field;

[0119] Inter-layer contour similarity is calculated based on cross-slice contour displacement vectors, and inter-layer emergence intensity is calculated based on contour sharpness index.

[0120] Using the seed slice as the root node, a set of candidate parent nodes is constructed for each non-root node and the corresponding edge weights are calculated. The edge weights include the weights of adjacent layers and the weights of cross layers. Based on the edge weights, an inter-layer emergent consistency graph covering all slice nodes is constructed.

[0121] Based on the inter-layer emergent consistency graph, starting from the seed slice, for each non-root node, the node with the largest edge weight is selected from its set of candidate parent nodes as the parent node, generating the optimal backtracking path tree covering all slice nodes.

[0122] Cross-slice evidence filling is performed along the optimal backtracking path tree. Based on the thin plate strip deformation field, the lesion candidate contour of the parent node slice is mapped to the slice corresponding to the current non-root node and fused to obtain the filled cross-slice contour.

[0123] Perform 3D consistency verification on the filled cross-slice contour, remove isolated artifact contours, reconstruct a 3D consistent lesion, and output it.

[0124] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0125] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0126] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A computer vision segmentation system for MRI medical images of epilepsy lesions, characterized in that, include: The data acquisition module is used to acquire MRI tomographic scan slice sequences, extract the region of interest in the hippocampus from each slice in the MRI tomographic scan slice sequence and perform lesion segmentation to obtain single-slice lesion candidate contours. The seed determination module is used to calculate the contour clarity index of candidate lesion contours for each single slice, and to determine the slice with the highest contour clarity index as the seed slice. The interlayer deformation calculation module is used to calculate the cross-slice contour displacement vector, interlayer deformation amplitude, and thin plate strip deformation field based on the corresponding anatomical landmarks in the region of interest of the hippocampus of adjacent slices. The inter-layer index calculation module is used to calculate the inter-layer contour similarity based on the cross-slice contour displacement vector and to calculate the inter-layer emergence intensity based on the contour sharpness index. The graph construction module is used to construct a set of candidate parent nodes for each non-root node with the seed slice as the root node and calculate the corresponding edge weights. The edge weights include adjacent layer edge weights and cross-layer edge weights. Based on the edge weights, an inter-layer emergent consistency graph covering all slice nodes is constructed. The optimal path generation module is used to generate an optimal backtracking path tree covering all slice nodes by starting from the seed slice based on the inter-layer emergent consistency graph and selecting the node with the largest edge weight from its candidate parent node set for each non-root node. The cross-slice filling module is used to perform cross-slice evidence filling along the optimal backtracking path tree. Based on the thin plate strip deformation field, the lesion candidate contour of the parent node slice is mapped to the slice corresponding to the current non-root node and fused to obtain the filled cross-slice contour. The reconstruction output module is used to perform three-dimensional consistency verification on the filled cross-slice contour, remove isolated artifact contours, reconstruct a three-dimensional consistent lesion, and output it.

2. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The calculation process for the outline sharpness index includes: Calculate the mean gradient magnitude, boundary gradient variance, and boundary pixel grayscale variance of each pixel in the candidate contour of the single-slice lesion; The contour sharpness index is obtained by weighting and combining the mean gradient magnitude, the boundary gradient variance, and the reciprocal of the boundary pixel grayscale variance.

3. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The calculation process of the cross-slice contour displacement vector includes: parameterizing the hippocampal contour of the region of interest of the hippocampus, calculating the curvature of the contour parameterization curve, identifying the position where the curvature reaches a local extremum as an anatomical landmark, performing nearest neighbor matching on the anatomical landmarks of adjacent slices, and obtaining the cross-slice contour displacement vector based on the pixel coordinate difference of the matched landmarks. The calculation process of the interlayer deformation amplitude includes: based on the set of anatomical landmarks of adjacent slices, calculating the thin plate spline deformation field through thin plate spline interpolation. The thin plate spline deformation field characterizes the non-rigid mapping relationship of pixel coordinates between adjacent slices. For each pixel position covered by the thin plate spline deformation field, the offset magnitude of the pixel coordinates at that position after non-rigid mapping is calculated. The average value of the offset magnitude of all pixel positions is obtained to obtain the interlayer deformation amplitude.

4. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The weights of adjacent layer edges are calculated in the following way: The inter-layer contour similarity is calculated based on the magnitude of the cross-slice contour displacement vector with exponential decay. The exponential decay is used to quantify the influence of the spatial offset of the lesion contour between adjacent slices on the degree of contour correspondence. The adjacent layer edge weight is a weighted combination of the interlayer contour similarity, the monotonically decreasing mapping of the interlayer deformation amplitude, and the interlayer emergence intensity. The weighted combination is used to quantify the reliability of lesion contour transmission between adjacent slices.

5. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The calculation process for the interlayer emergence intensity includes: For a slice with two adjacent slices on both sides, the arithmetic mean of the outline sharpness indices of the two adjacent slices is calculated as a joint measure of outline sharpness. For slices with only one-sided adjacent slices, the outline sharpness index of its adjacent slices is calculated as a joint measure of outline sharpness. For the current slice, the absolute difference between the contour sharpness index and the joint contour sharpness measure is calculated to obtain the interlayer emergence intensity.

6. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The process of generating the optimal backtracking path tree includes: Starting from the seed slice, the sequence extends layer by layer towards both ends along the arrangement order of the MRI tomographic scan slice sequence; In each expansion step, the current non-root node selects the node with the largest edge weight in the inter-layer emergent consistency graph from its set of candidate parent nodes as the parent node. The set of candidate parent nodes includes adjacent nodes and cross-layer nodes. This process is repeated layer by layer until all slice nodes of the MRI tomographic slice sequence are covered.

7. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The process of filling in evidence across slices includes: Compare the outline sharpness index of the slice corresponding to the current non-root node with the outline sharpness index of the slice corresponding to the parent node. If the outline sharpness index of the slice corresponding to the current non-root node is lower than that of the slice corresponding to the parent node, then perform cross-slice evidence filling on the slice corresponding to the current non-root node. If the parent node is an adjacent node, the lesion candidate contour of the parent node slice is directly mapped to the coordinate space of the slice corresponding to the current non-root node based on the thin plate strip deformation field to obtain the mapped contour. If the parent node is a cross-layer node, the thin-plate strip deformation field of each adjacent layer between the parent node slice and the slice corresponding to the current non-root node is obtained. The pixel coordinates of the lesion candidate contour of the parent node slice are sequentially mapped through the thin-plate strip deformation field of each adjacent layer and transferred to the coordinate space of the slice corresponding to the current non-root node to obtain the mapped contour. Calculate the Hausdorff distance between the mapped contour and the candidate contour of the lesion in the slice corresponding to the current non-root node. If the Hausdorff distance is greater than the preset contour filling threshold, then add the pixels in the mapped contour whose pixel coordinate distance from the candidate contour of the lesion in the slice corresponding to the current non-root node is greater than the preset spatial tolerance to the candidate contour of the lesion in the slice corresponding to the current non-root node according to the nearest neighbor principle. Boundary smoothing based on B-spline curve fitting is performed on the supplemented contour to obtain the filled cross-slice contour.

8. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The process of three-dimensional conformance verification includes: The filled cross-slice contour is converted into a contour-enclosed region mask; For each pair of adjacent slices, calculate the region overlap ratio based on the mask of the region enclosed by its contour; The deformation consistency correction coefficient is calculated based on the deformation field of the thin plate spline, and the deformation consistency correction coefficient is an exponential decay function of the average amplitude of the deformation field of the thin plate spline. Multiply the region overlap ratio by the deformation consistency correction coefficient to obtain the contour continuity measure; If the contour continuity measure of a filled cross-slice contour and the filled cross-slice contours of its two adjacent slices are both lower than the continuity threshold, then the filled cross-slice contour is determined to be an isolated artifact and is removed. The slice thickness and pixel spacing of the MRI tomographic scan slice sequence are obtained. The cross-slice contour, which is filled after removing isolated artifacts, is reconstructed into a three-dimensional volume mask by performing voxel stacking and inter-slice interpolation along the arrangement direction of the MRI tomographic scan slice sequence based on the slice thickness and pixel spacing, thereby obtaining the three-dimensional consistent lesion.

9. The computer vision segmentation system for epilepsy lesion MRI medical images according to claim 1, characterized in that: The process of constructing a set of candidate parent nodes for each non-root node includes: For the current non-root node, determine whether the inter-layer deformation amplitude between it and its adjacent nodes exceeds the preset inter-layer deformation tolerance. If the number of adjacent nodes does not exceed the limit, then the adjacent nodes are included in the candidate parent node set. If the number of nodes exceeds the limit, the cumulative cross-slice contour displacement vector and cumulative inter-layer deformation amplitude between the current non-root node and each candidate cross-layer node are calculated within the preset cross-layer search window. The cross-layer edge weight is calculated based on the cumulative cross-slice contour displacement vector and the cumulative inter-layer deformation amplitude. Candidate cross-layer nodes whose cross-layer edge weights meet the preset minimum admission threshold are included in the candidate parent node set. If the edge weights of all nodes in the candidate parent node set are lower than the minimum admission threshold, then the adjacent node is included in the candidate parent node set by default. The cross-layer edge weights are calculated as follows: the cross-layer contour similarity is calculated based on the magnitude of the cumulative cross-slice contour displacement vector. The cross-layer contour similarity decreases as the magnitude of the cumulative cross-slice contour displacement vector increases. The cross-layer contour similarity, the monotonically decreasing mapping of the cumulative inter-layer deformation amplitude, and the inter-layer emergence intensity of the candidate cross-layer nodes are weighted and combined to obtain the cross-layer edge weights. The cumulative cross-slice contour displacement vector is the layer-by-layer summation of the cross-slice contour displacement vectors between adjacent slices within the cross-layer search window; The cumulative interlayer deformation amplitude is the sum of the interlayer deformation amplitudes between adjacent slices within the cross-layer search window.

10. A computer vision segmentation method for epileptic lesions based on magnetic resonance imaging, characterized in that, Includes the following steps: MRI tomographic scan slice sequences are obtained, and the region of interest in the hippocampus is extracted from each slice in the MRI tomographic scan slice sequence and lesion segmentation is performed to obtain single-slice lesion candidate contours. Calculate the contour clarity index of each single slice candidate lesion contour, and determine the slice with the highest contour clarity index as the seed slice; Based on the corresponding anatomical landmarks in the regions of interest of the hippocampus in adjacent slices, calculate the cross-slice contour displacement vector, inter-layer deformation amplitude, and thin-plate strip deformation field; The inter-layer contour similarity is calculated based on the cross-slice contour displacement vector, and the inter-layer emergence intensity is calculated based on the contour sharpness index. Using the seed slice as the root node, a set of candidate parent nodes is constructed for each non-root node and the corresponding edge weights are calculated. The edge weights include adjacent layer edge weights and cross-layer edge weights. Based on the edge weights, an inter-layer emergent consistency graph covering all slice nodes is constructed. Based on the inter-layer emergent consistency graph, starting from the seed slice, for each non-root node, the node with the largest edge weight is selected from its set of candidate parent nodes as the parent node, generating an optimal backtracking path tree covering all slice nodes. Cross-slice evidence filling is performed along the optimal backtracking path tree. Based on the thin plate strip deformation field, the lesion candidate contour of the parent node slice is mapped to the slice corresponding to the current non-root node and fused to obtain the filled cross-slice contour. Perform 3D consistency verification on the filled cross-slice contour, remove isolated artifact contours, reconstruct a 3D consistent lesion, and output it.

Images