Cavity axis anchor point cross-modality corrected esophageal cancer radiotherapy target region generation method and system

Abstract

This invention discloses a method and system for generating esophageal cancer radiotherapy target areas using cross-modal correction of the lumen-axis anchor point, relating to the field of medical image processing technology. The proposed scheme includes acquiring esophageal images of at least two modalities from the same patient, standardizing and initially registering them to obtain a co-space esophageal image group, extracting the esophageal lumen-axis centerline from the co-space esophageal image group, using its radial neighborhood as a spatial constraint region, and setting a lumen-axis sampling segment and corresponding local cross-section along the esophageal lumen-axis centerline. This application uses the asymptotic geometric residual field of each modality as an intermediate quantity for cross-modal verification, and generates a cross-modal second-order geometric stability field based on the response co-occurrence degree within the corresponding cross-modal unit, the consistency of the Hessian principal eigenvector direction, and the correlation of local neighborhood response distribution. This solves the problem of directly using strong second-order responses from a single modality, which easily leads to modal artifacts, and achieves structural consistency access before weak boundary infiltration traces enter the target area.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, this application relates to a method and system for generating esophageal cancer radiotherapy target areas using cross-modal correction of cavity axis anchor points. Background Technology

[0002] Radiotherapy for esophageal cancer requires determining the irradiation range of the tumor based on imaging before treatment. Since a single image has limited ability to display the soft tissues in the mediastinum, multiple images of the same patient are usually acquired in clinical practice and aligned to a unified coordinate system. The tumor location is then identified through the image processing workflow, thereby determining the radiotherapy target area. The esophagus is a luminal organ located in the mediastinum and is continuously affected by respiration, heartbeat, and swallowing movements. Esophageal cancer often grows invasively along the submucosa of the esophageal wall.

[0003] In this type of multimodal target generation process, the existing mainstream approach uses the grayscale boundary between the tumor and surrounding tissues as the basic basis for identifying the tumor range, and uses the grayscale similarity between images or the distal anatomical landmarks in the mediastinum to align different modalities, supplemented by respiratory motion compensation when necessary. The common premise for the applicability of these methods is that the tumor region has a recognizable grayscale transition on the image, and that different modalities can reflect the same target in the same position after alignment.

[0004] However, submucosal infiltrates appear as progressive thickening of the esophageal wall on imaging without forming obvious gray-scale abrupt changes. When tumor extent relies on gray-scale boundary identification, these areas lacking clear boundaries are difficult to include, resulting in a lack of basis for determining the target area along the esophagus and its lateral extent. If more refined local morphological information is extracted to compensate, it is difficult to distinguish the subtle morphological changes caused by infiltration from imaging artifacts, noise, and morphological responses from adjacent normal structures such as blood vessels and trachea, easily leading to the misinclusion of irrelevant structures. Although multimodal imaging can provide complementary tissue information, conventional processing usually requires precise correspondence of each modality at the voxel level, which is difficult to achieve due to the continuous physiological movement of the esophagus. In addition, the infiltrating segment presents a weak signal in each modality, and processing by equally superimposing each modality cannot stably highlight it. Therefore, under the multiple constraints of lacking clear boundaries, unreliable single-modal morphological information, and difficulty in accurately corresponding multimodal images, how to stably identify esophageal submucosal infiltrates at the image processing level and include them in the generation of radiotherapy target areas has become a core technical problem to be solved in this scenario. Summary of the Invention

[0005] To address the aforementioned technical problems, a method and system for generating esophageal cancer radiotherapy target areas using cross-modal correction of cavity axis anchor points are provided. This technical solution resolves the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this application provides a method for generating a radiotherapy target volume for esophageal cancer with cross-modal correction of the cavity axis anchor point, the method comprising:

[0008] At least two modalities of esophageal images were acquired from the same patient, and after standardization and preliminary registration, a co-space esophageal image group was obtained;

[0009] The esophageal lumen centerline is extracted from the co-space esophageal image group, and its radial neighborhood is used as the spatial constraint area. The lumen sampling segment and corresponding local cross-section are set along the esophageal lumen centerline.

[0010] Within the spatially constrained region, gradient response suppression coefficients that decrease as the gradient magnitude increases are generated for each voxel of the esophageal image of each modality based on the first-order gradient response. Principal curvature response enhancement coefficients that increase as the structured second-order geometric response is enhanced are generated based on the combination of eigenvalues ​​of the Hessian matrix. The corresponding Hessian principal eigenvectors are recorded. The asymptotic geometric residual field of the corresponding modality is obtained by multiplying and fusing the two coefficients.

[0011] Using the cavity axis sampling segment and the corresponding local cross-section as the cross-modal corresponding unit, the spatial consistency measure of the asymptotic geometric residual field of each mode is performed to generate a cross-modal second-order geometric stability field. The spatial consistency measure includes the degree of response co-occurrence of the asymptotic geometric residual fields of different modes in the corresponding unit, and at least one of the following: the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution.

[0012] Regions where the cross-modal second-order geometric stability field satisfies the preset stability condition and the first-order gradient response does not meet the preset boundary identification condition are identified as progressive infiltration candidate segments. Regions where the first-order gradient response meets the preset boundary identification condition and satisfies the preset boundary consistency condition within the corresponding cross-modal unit are identified as strong boundary tumor regions.

[0013] Based on the continuity of progressively infiltrating candidate segments and strongly boundary tumor regions along the esophageal lumen axis, their coverage relationship within corresponding local cross sections, and the regional connectivity between them, esophageal cancer radiotherapy target data containing GTV region mask and CTV candidate region contour data is generated.

[0014] Secondly, this application provides an esophageal cancer radiotherapy target generation system with cross-modal correction of the cavity-axis anchor point, used to implement the above-mentioned method for generating an esophageal cancer radiotherapy target volume with cross-modal correction of the cavity-axis anchor point, including:

[0015] The image acquisition and registration module is used to acquire esophageal images of at least two modalities from the same patient, and obtain a co-space esophageal image group after standardization and preliminary registration.

[0016] The lumen axis extraction and segmentation module is used to extract the esophageal lumen axis centerline in the co-space esophageal image group, using its radial neighborhood as the spatial constraint area, and setting lumen axis sampling segments and corresponding local cross-sections along the esophageal lumen axis centerline;

[0017] The residual field generation module is used to generate gradient response suppression coefficients that decrease as the gradient magnitude increases for each voxel of each modality of esophageal image within the spatial constraint region, based on the first-order gradient response. It also generates principal curvature response enhancement coefficients that increase as the structured second-order geometric response increases based on the eigenvalue combination of the Hessian matrix, and records the corresponding Hessian principal eigenvectors. The product of the two coefficients is then fused to obtain the asymptotic geometric residual field of the corresponding modality.

[0018] The stability field construction module is used to measure the spatial consistency of the asymptotic geometric residual fields of each mode, taking the cavity axis sampling segment and the corresponding local cross-section as the cross-modal corresponding unit, and generate a cross-modal second-order geometric stability field. The spatial consistency measure includes the degree of response co-occurrence of the asymptotic geometric residual fields of different modes in the corresponding unit, and at least one of the following: the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution.

[0019] The region identification and discrimination module is used to identify regions where the cross-modal second-order geometric stability field satisfies the preset stability condition and the first-order gradient response does not reach the preset boundary identification condition as progressive infiltration candidate segments, and regions where the first-order gradient response reaches the preset boundary identification condition and satisfies the preset boundary consistency condition within the corresponding cross-modal unit as strong boundary tumor regions.

[0020] The target data generation module is used to generate esophageal cancer radiotherapy target data, which includes GTV region mask and CTV candidate region contour data, based on the continuity relationship between progressive infiltration candidate segments and strong boundary tumor regions along the esophageal lumen axis, the coverage relationship within the corresponding local cross-section, and the regional connectivity relationship between the two.

[0021] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for generating esophageal cancer radiotherapy target areas with cross-modal correction of cavitary axis anchor points.

[0022] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for generating esophageal cancer radiotherapy target areas with cross-modal correction of cavity axis anchor points.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This application uses the asymptotic geometric residual field of each mode as an intermediate quantity for cross-modal verification, and generates a cross-modal second-order geometric stability field by taking the response co-occurrence degree, the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution within the corresponding unit of the cross-modal. This solves the problem that the strong second-order response of the single mode is directly used and is easily mixed with modal artifacts, and realizes the structural consistency admission before the weak boundary wetting traces enter the target area.

[0025] This application addresses the problem of distinguishing between isolated low-confidence responses and true weak-boundary infiltration segments by using adjacent strong-boundary tumor regions as the starting point for inheritance, and transmitting inheritance relationships along cross-modal corresponding units that satisfy continuity, coverage, or regional connectivity relationships. It also interrupts or reduces the inheritance level at unstable second-order response regions. This transforms the selection of progressive infiltration candidate segments from point-by-point determination to dynamic inheritance determination along the esophageal lumen axis.

[0026] This application addresses the problem in the existing unidirectional process of registration followed by identification, where weak boundary infiltration segments can only rely on extrapolation from distal anatomical landmarks. By identifying regions that meet the conditions of consistent acceptance level and direction as sources of weak boundary cross-modal stability constraints and incorporating them into the update process of the preliminary registration results, this application transforms stable weak boundary structures from identification results into local registration constraints. This, in turn, drives the updating of asymptotic geometric residual field, cross-modal second-order geometric stability field, and esophageal cancer radiotherapy target area data. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0028] Figure 1 This is a flowchart of the method for generating esophageal cancer radiotherapy target area by cross-modal correction of cavity axis anchor points proposed in this invention;

[0029] Figure 2 This is a structural block diagram of the esophageal cancer radiotherapy target generation system with cross-modal correction of cavity axis anchor points proposed in this invention. Detailed Implementation

[0030] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0031] The overall process of this application includes steps S1 to S6. Figure 1The logical framework of the overall process is shown; for ease of understanding, the following explanation uses a set of thoracic esophageal images from the same patient as a consistent example: The patient acquired two modalities: planned computed tomography (CT) images and diagnostic magnetic resonance imaging (MRI) images. The scan range covered the mid-thoracic esophagus, with an intra-slice voxel spacing of approximately 1.0 mm and an inter-slice spacing of approximately 3.0 mm. The tumor was located in a segment of the esophageal wall approximately 2 to 5 cm below the carina and gradually extended downwards into the submucosa for approximately 1.5 cm.

[0032] In step S1: esophageal images of at least two modalities from the same patient are acquired, standardized, and preliminarily registered to obtain a co-space esophageal image group;

[0033] Among them, at least two modalities of esophageal imaging refer to images acquired using different imaging principles for the same patient and covering the same anatomical range of the esophagus, including but not limited to computed tomography (CT) images, magnetic resonance imaging (MRI) images, positron emission tomography (PET) images, and cone-beam computed tomography (CBCT) images. Typically, the at least two modalities are CT images and MRI images. The reason for using at least two modalities is that although the image intensity of esophageal submucosal infiltration varies under different imaging principles, it corresponds to the same histological thickened structure. Multiple modalities provide mutually corroborating inputs for subsequent cross-modal consistency measurement.

[0034] It should be noted that standardization refers to resampling and grayscale normalization of esophageal images of each modality, so that different modalities are placed in a unified voxel grid and a comparable numerical range; preliminary registration refers to aligning esophageal images of each modality to the same common space. Preliminary registration only requires achieving a coarse alignment accuracy that supports subsequent cross-modal corresponding unit level comparisons, and does not require achieving voxel-level precise correspondence.

[0035] Following the aforementioned example, the magnetic resonance images were resampled to a 1.0 mm × 1.0 mm × 3.0 mm voxel grid consistent with the computed tomography images, and preliminary registration was completed by roughly aligning the tracheal carina in the mediastinum with the anterior edge of the vertebral body. The residual misalignment of the same esophageal cross section in the resulting co-space esophageal image group between the two modalities was approximately 2 to 4 mm.

[0036] Additionally, because the esophagus is affected by breathing, heartbeat, and swallowing movements, it is difficult to achieve precise voxel-level alignment at this location. Therefore, the accuracy requirement for initial registration is proactively reduced to the corresponding unit level, removing the reliance on precise alignment from the process entry point, so that subsequent processing no longer requires voxel-level alignment.

[0037] In step S2: the esophageal lumen axis centerline is extracted from the common space esophageal image group, its radial neighborhood is used as the spatial constraint area, and the lumen axis sampling segment and corresponding local cross-section are set along the esophageal lumen axis centerline;

[0038] Among them, the esophageal lumen axis centerline refers to the curve that runs through the geometric center of the esophageal lumen and can be obtained by conventional lumen centerline extraction methods; the spatial constraint region refers to the spatial range defined by the radial neighborhood of the esophageal lumen axis centerline, which is used to limit all subsequent second-order geometric calculations; the lumen axis sampling segment and the corresponding local cross section provide geometric reference units along the lumen direction for subsequent cross-modal comparisons;

[0039] The aim is to address the problem of mistakenly including adjacent normal structures such as blood vessels and trachea when extracting fine morphological information. If the computational space is not limited and the second-order geometric morphological response is extracted across the entire map, a large number of non-esophageal tubular or sheet-like structures in the mediastinum will also produce strong responses. By using the radial neighborhood of the esophageal lumen axis centerline as an internal anchor point, the subsequent second-order geometric calculation is constrained to the space adjacent to the esophageal wall from the beginning, thus eliminating the source of responses from non-esophageal structures at the source, rather than filtering them out after the full map calculation.

[0040] In step S3: Within the spatial constraint region, for each modality of esophageal image, a gradient response suppression coefficient that decreases as the gradient magnitude increases is generated for each voxel based on the first-order gradient response. A principal curvature response enhancement coefficient that increases as the structured second-order geometric response is enhanced is generated based on the combination of eigenvalues ​​of the Hessian matrix. The corresponding Hessian principal eigenvectors are recorded. The product of the two coefficients is fused to obtain the asymptotic geometric residual field of the corresponding modality.

[0041] Among them, the progressive geometric residual field refers to the scalar field that takes values ​​for each voxel within the spatially constrained region. It takes higher values ​​at locations where the gray-level gradient is weak but there is a structured second-order geometric morphology, and is used to characterize the geometric traces left by progressive submucosal infiltration in the esophagus on the image. Since submucosal infiltration is manifested as progressive thickening of the esophageal wall rather than a sudden change in gray level, the suppression of gray-level gradient and the enhancement of second-order geometric morphology are multiplied to simultaneously satisfy the region that does not form a clear gray-level boundary and has a wall layer geometric structure.

[0042] In step S4: taking the cavity shaft sampling segment and the corresponding local cross section as the cross-modal corresponding unit, the spatial consistency measure of the asymptotic geometric residual field of each mode is performed to generate a cross-modal second-order geometric stability field. The spatial consistency measure includes the degree of response co-occurrence of different modal asymptotic geometric residual fields in the corresponding unit, and at least one of the following: the consistency of the Hessian principal eigenvector direction and the correlation of local neighborhood response distribution.

[0043] Among them, the cross-modal corresponding unit refers to the spatial unit defined by the cavity axis sampling segment and the corresponding local cross section, used for group comparisons rather than voxel comparisons between different modes; the cross-modal second-order geometric stability field refers to the scalar field that takes values ​​on each cross-modal corresponding unit, which reflects whether the asymptotic geometric residual fields of each mode present a consistent second-order geometric structure within the unit; the reason for using the corresponding unit rather than the voxel as the comparison unit is that the initial registration in the esophagus only achieves coarse alignment at the corresponding unit level, and voxel comparisons will fail due to residual misalignment, while comparisons at the corresponding unit level can accommodate this residual misalignment.

[0044] In step S5: the region where the cross-modal second-order geometric stability field satisfies the preset stability condition and the first-order gradient response does not reach the preset boundary identification condition is determined as a progressive infiltration candidate segment; the region where the first-order gradient response reaches the preset boundary identification condition and satisfies the preset boundary consistency condition within the corresponding cross-modal unit is determined as a strong boundary tumor region.

[0045] Among them, progressive infiltration candidate segments refer to regions that are geometrically stable across modal second order but have weak gray-level boundaries, corresponding to the geometric traces of progressive submucosal infiltration; strong boundary tumor regions refer to regions with clear gray-level boundaries and consistent boundaries within the corresponding cross-modal units, corresponding to the obvious solid parts of the tumor.

[0046] It should be noted that the preset stability condition is used to determine whether the cross-modal second-order geometric stability field of a cross-modal corresponding unit has reached a level where the second-order geometric structure of each modality can be considered consistent. The preset boundary recognition condition is used to determine whether the first-order gradient response of a region has reached a level where a clear gray-scale boundary can be considered to exist. The preset boundary consistency condition is used to determine whether the clear gray-scale boundary is presented by each modality within the cross-modal corresponding unit. The above three conditions serve the division of two types of regions respectively. The progressive infiltration candidate segment is characterized by a stable but weak boundary, while the strong boundary tumor region is characterized by a clear and consistent boundary. The two do not overlap.

[0047] It should also be noted that the preset stability condition, preset boundary identification condition, and preset boundary consistency condition are all given in a relative judgment manner rather than fixed absolute values, and each is expressed by a specific quantile or relative proportion: the preset stability condition takes the upper quantile of the cross-modal second-order geometric stability field distribution of the corresponding unit across each cross-modal, with a quantile value ranging from 70% to 90%, and a typical value of 75%. That is, the corresponding unit whose cross-modal second-order geometric stability field is not lower than the value corresponding to this quantile is judged to meet the preset stability condition; the preset boundary identification condition takes the upper quantile of the first-order gradient response distribution within the spatial constraint region, with a quantile value ranging from 80% to 95%, and a typical value of 90%. That is, the voxel whose first-order gradient response is not lower than the value corresponding to this quantile is judged to meet the preset boundary identification condition. Due to obvious gray-scale edges... The boundary occupies a smaller proportion than the infiltration geometric trace and has a higher gradient within the spatial constraint area, so its quantile is higher than the quantile of the preset stability condition. The preset boundary consistency condition takes that the positional deviation of the grayscale boundary of each modality within the same corresponding unit does not exceed the preset proportion of the radial scale of the corresponding unit. The proportion ranges from 1 / 5 to 1 / 2 of the radial scale, with a typical value of 1 / 4. Since there is a residual misalignment of 2 to 4 mm in the esophagus during the initial registration, corresponding to an allowable deviation of about 2.5 mm at a typical radial depth of 10 mm, this magnitude of deviation is allowed by the relative proportion of the radial scale of the corresponding unit rather than requiring complete positional overlap. Since the grayscale and gradient scale of the esophageal image vary with the patient, equipment, and modality, relative quantiles and relative proportions are used instead of absolute thresholds to keep the above conditions comparable across different data.

[0048] In step S6: Based on the continuity of the progressive infiltration candidate segment and the strong boundary tumor region along the esophageal lumen axis, the coverage relationship within the corresponding local cross section, and the regional connectivity relationship between the two, esophageal cancer radiotherapy target area data containing GTV region mask and CTV candidate region contour data is generated.

[0049] Among them, continuity relationship refers to whether the progressive infiltration candidate segment and the strong boundary tumor region are adjacent and continuous along the esophageal lumen axis; coverage relationship refers to whether the circumferential and radial ranges of the two are connected in the same local cross section; regional connectivity relationship refers to whether the two are connected to each other through adjacent cross-modal corresponding units within the spatial constraint area; GTV region mask refers to the voxel mask that identifies the gross tumor region; CTV candidate region contour data refers to the clinical target region candidate contour generated on the basis of GTV region mask according to the conventional clinical expansion rules.

[0050] It should be noted that this step only serves to connect the progressive infiltration candidate segments to the strong boundary tumor region using the above three geometric relationships and generate target area data accordingly. The outward expansion rule of the clinical target area candidate contour adopts the conventional clinical outward expansion method in this field. This solution does not limit the outward expansion rule itself, nor does it involve the determination of the planned target area based on the superposition of positioning errors and organ motion margin.

[0051] Using the aforementioned example, the strongly bordered tumor region corresponds to a significantly thickened esophageal wall approximately 2 to 5 centimeters below the carina, and the progressively infiltrated candidate segment corresponds to a progressively thickened segment approximately 1.5 centimeters below it. Since this progressively infiltrated candidate segment is adjacent to the strongly bordered tumor region along the esophageal axial direction, connects with it circumferentially along the corresponding local cross-section, and is connected to it through adjacent cross-modal corresponding units, it is included in the GTV region mask along with the strongly bordered tumor region. This causes the lower boundary of the GTV region mask along the esophageal axial direction to be shifted down by approximately 1.5 centimeters compared to when only a clearly grayscale boundary is used, and then the GTV region mask is used to generate CTV candidate region contour data according to conventional clinical outward expansion rules.

[0052] Through the above technical solution, this embodiment uses the radial neighborhood of the esophageal lumen axis centerline as the internal anchor point to define the second-order geometric calculation space, verifies the asymptotic geometric residual field with cross-modal corresponding units, and connects the three geometric relationships with continuous coverage to accept the weak boundary infiltrative segment to the clearly boundary tumor region. Unlike the existing processing method that uses gray-scale boundaries as the only entry point and pursues precise correspondence at the voxel level, this embodiment identifies and accepts the segment at the corresponding unit level of the lumen axis constraint with cross-modal structural consistency, thereby stably including the submucosal infiltrative segment under the condition of lacking clear boundaries and precise correspondence. The resulting GTV area mask and CTV candidate area contour data are output as radiotherapy target area data to the downstream radiotherapy planning stage.

[0053] In one specific embodiment of step S2, the esophageal lumen axis centerline is extracted from the co-space esophageal image group, its radial neighborhood is used as the spatial constraint region, and a lumen axis sampling segment and corresponding local cross-section are set along the esophageal lumen axis centerline, specifically including:

[0054] Sampling points along the arc length of the esophageal lumen axis centerline are set, and the arc length interval between adjacent sampling points is taken as the lumen axis sampling segment. A local tangent to the esophageal lumen axis centerline is taken at each sampling point, and the plane perpendicular to this local tangent is taken as the corresponding local cross-section. Within this cross-section, the circumferential angle zero point is determined with the lumen axis sampling point as the center. Within each local cross-section, the area around the lumen axis sampling point is divided into multiple circumferential angle zones, with the circumferential angle zero point as the starting direction. The direction from the lumen axis sampling point to the centerline of each circumferential angle zone is taken as the reference angle. The radial reference direction of the circumferential angle partition is used, and within each circumferential angle partition, the annular zone extending outward from the sampling point of the lumen axis along the corresponding radial reference direction to a preset radial depth is used as the radial neighborhood; within the radial neighborhood, the annular region that is continuous along the circumferential direction and forms a local peak of gradient magnitude relative to the adjacent annular zone is determined as the esophageal wall candidate region; the radial neighborhood in each local cross section is continuously arranged along the center line of the esophageal lumen axis to obtain the spatial constraint region, and when there is an esophageal wall candidate region, the spatial constraint region is updated with the overlapping area of ​​the esophageal wall candidate region and the radial neighborhood;

[0055] It should be noted that the construction process of the lumen axis sampling segment, local cross-section, circumferential angle partition, and radial neighborhood is as follows: Lumen axis sampling points are set at preset arc length intervals along the arc length direction of the esophageal lumen axis centerline. The arc length interval between two adjacent lumen axis sampling points constitutes a lumen axis sampling segment, serving as the basic axial unit along the lumen's direction. At each lumen axis sampling point, a local tangent to the esophageal lumen axis centerline is taken, and a plane perpendicular to this local tangent is used as the corresponding local cross-section of that lumen axis sampling point, ensuring that this corresponding local cross-section precisely intersects the lumen. Within a local cross-section, with the cavity axis sampling point as the center and a fixed circumferential reference direction as the zero point of the circumferential angle, the area around the cavity axis sampling point is uniformly divided into multiple circumferential angle partitions according to the circumferential angle. Each circumferential angle partition corresponds to a circumferential sector of the pipe wall. The direction from the cavity axis sampling point to the center line of each circumferential angle partition is taken as the radial reference direction of the circumferential angle partition. Within the circumferential angle partition, an annulus extending outward from the cavity axis sampling point along the radial reference direction to a preset radial depth is taken as the radial neighborhood of the circumferential angle partition.

[0056] The radial reference direction is set based on the fact that the esophageal wall is distributed circumferentially along the lumen in each local cross section, and the submucosal infiltration thickens radially along the wall. Therefore, the radial direction from the center of the lumen to the wall is used as the reference direction for measurement and comparison in each circumferential angle zone. This can make the subsequent direction judgment of the second-order geometry and cross-modal direction comparison have a unified reference that conforms to the anatomical direction of the wall.

[0057] It should also be noted that the preset arc length interval is the arc length distance between adjacent lumen axis sampling points along the center line of the esophageal lumen axis. This is used to determine the axial length of the lumen axis sampling segment. Too dense an interval will make it difficult to reflect axial structural changes in adjacent lumen axis sampling segments, while too sparse an interval will cause the progressive infiltration segment to be crossed axially. Therefore, its value ranges from 2 mm to 6 mm, with a typical value of 3 mm, consistent with conventional interlayer spacing. The number of circumferential angle partitions is the number of sectors divided circumferentially within each local cross-section. This is used to determine the granularity of the circumferential resolution of the tube wall. Too few partitions will result in a… The circumferential angle partitions spanning excessively large tubular wall arcs or too many partitions may result in insufficient voxels within a single circumferential angle partition to support stable statistics. Therefore, their values ​​range from 8 to 24, with a typical value of 12. The preset radial depth is the distance that the radial neighborhood extends outward from the sampling point along the radial reference direction from the luminal axis. It is used to cover the tubular wall and its adjacent submucosa without including mediastinal structures outside the tubular wall. Since too shallow a depth may miss submucosal infiltration, and too deep a depth may include paraesophageal vessels and fat, its value ranges from 6 mm to 15 mm, with a typical value of 10 mm.

[0058] The process of determining the candidate region of the esophageal wall and updating the spatial constraint region is as follows: the change of gradient modulus is examined radially within the radial neighborhood. The annular region that extends continuously along the circumference and forms a local peak of gradient modulus relative to the adjacent annular zone is determined as the candidate region of the esophageal wall. This annular region corresponds to the high gradient ring where the inner and outer interfaces of the esophageal wall are located. When the candidate region of the esophageal wall exists, the spatial constraint region is updated with the overlapping area of ​​the candidate region of the esophageal wall and the radial neighborhood, so that the spatial constraint region is further tightened to the annular zone where the esophageal wall is located. When the candidate region of the esophageal wall cannot be stably determined, the range obtained by continuously arranging the radial neighborhood in each local cross section along the center line of the esophageal lumen axis is directly used as the spatial constraint region.

[0059] It should be noted that when a certain esophageal axis sampling point cannot stably determine the annular region with a local peak of gradient magnitude in the radial neighborhood due to artifacts or interference from adjacent structures, the spatial constraint region at that esophageal axis sampling point is not updated, but its radial neighborhood is retained as the spatial constraint region. This is to ensure that the continuity of the spatial constraint region along the esophageal lumen axis centerline is not interrupted due to the lack of esophageal wall candidate regions at individual esophageal axis sampling points.

[0060] Through the above technical solution, this embodiment constructs the esophageal lumen axis sampling segment, corresponding local cross section, circumferential angle partition, radial reference direction and radial neighborhood sequentially based on the esophageal lumen axis centerline, and tightens the spatial constraint area by the overlapping area of ​​the esophageal wall candidate area and the radial neighborhood; unlike the conventional processing method of extracting morphological response on the full-view voxel grid without spatial differentiation, this embodiment organizes the calculation unit according to the circumferential angle partition in the annular space adjacent to the esophageal wall, thereby limiting the scope of the second-order geometric calculation to the esophageal wall and its submucosa.

[0061] In one specific implementation of step S3, the asymptotic geometric residual field of the corresponding mode is obtained by fusing the product of the two coefficients, specifically including:

[0062] Within the spatially constrained region, the first-order gradient response is calculated voxel-by-voxel for each modality of esophageal image. Within each circumferential angular region, the first-order gradient response is normalized according to the gradient magnitude distribution of that region, generating a gradient response suppression coefficient that decreases as the gradient magnitude increases. Within the spatially constrained region, eigenvalue combinations of the Hessian matrix are calculated voxel-by-voxel for the same modality of esophageal image at a scale adapted to the esophageal wall thickness. Voxel responses that satisfy the preset conditions for wall-like sheet structures are enhanced, while those that satisfy the preset conditions for non-wall-like structures are enhanced. The voxel response is suppressed under the structural conditions to generate principal curvature response enhancement coefficients; the eigenvectors corresponding to voxels that satisfy the preset wall-sheet structure conditions are recorded as Hessian principal eigenvectors; the deviation between the Hessian principal eigenvectors and the radial reference direction of the corresponding circumferential angle partition is compared, and the principal curvature response enhancement coefficients of voxels whose deviation exceeds the preset angle range are down-adjusted; the gradient response suppression coefficient and the principal curvature response enhancement coefficient of the same mode are multiplied and fused voxel by voxel to obtain the asymptotic geometric residual field of the mode;

[0063] It should be noted that the gradient response suppression coefficient is generated as follows: Within the spatially constrained region, the first-order gradient response is calculated voxel-by-voxel for each modality of esophageal image, and its gradient magnitude is taken; within each circumferential angular region, the gradient magnitude distribution of all voxels within that region is statistically analyzed, and this distribution is used to normalize the gradient magnitude of the voxels within the region, thus eliminating the gradient scale differences caused by varying wall thickness in different circumferential angular regions; then, the normalized gradient magnitude is used to generate the gradient response suppression coefficient through a monotonically decreasing mapping, so that the larger the gradient magnitude, the smaller the gradient response suppression coefficient.

[0064] The formula for calculating the gradient response suppression coefficient is:

[0065] ;

[0066] in, The gradient response suppression coefficient at the voxel is . This represents the normalized gradient magnitude at the voxel within the circumferential angular partition. For attenuation scale parameters, The attenuation index is used to control the gradient magnitude from the weak gradient region to the strong gradient region. Since the normalized gradient magnitude is referenced to the medium gradient level within the partition, the attenuation scale parameter is taken near the median level of the normalized gradient magnitude, so that the gradient response suppression coefficient at the medium gradient is about half. Its value ranges from 0.5 to 1.5, with a typical value of 1.0. The attenuation index is used to control the transition steepness from the weak gradient region to the strong gradient region. Its value ranges from 1 to 4, with a typical value of 2.

[0067] It should be noted that the generation process of the principal curvature response enhancement coefficient is as follows: within the spatially constrained region, the Hessian matrix is ​​calculated voxel by voxel for the same modality of esophageal image at a scale adapted to the thickness of the esophageal wall. The eigenvalues ​​of the Hessian matrix are sorted from smallest to largest to obtain the eigenvalue combination. When the eigenvalue combination satisfies the preset wall-like structure condition, that is, it presents a negative eigenvalue with a significantly large absolute value while the absolute values ​​of the other eigenvalues ​​are relatively small, corresponding to a wall-like structure that is curved in one direction and flat in the other two directions, the voxel response is enhanced. When the eigenvalue combination satisfies the preset non-wall-like structure condition, that is, all three eigenvalues ​​are significant (corresponding to point-like or mass-like structures) or present two significant negative eigenvalues ​​(corresponding to tubular structures), the voxel response is suppressed.

[0068] The formula for calculating the principal curvature response strengthening coefficient is:

[0069] ;

[0070] in, The principal curvature response enhancement coefficient at the voxel is denoted as . The eigenvalues ​​of the Hessian matrix at this voxel are sorted by absolute value from smallest to largest. To adjust the parameters in a smoother direction, The sheet-like strength adjustment parameter is used because sheet-like structures are characterized by significant bending in one direction and gentle bending in the other two directions. Therefore, the eigenvalue with the smallest absolute value represents the gentle bending direction, and the sum of the squares of the other two eigenvalues ​​represents the bending strength. The gentle bending direction adjustment parameter and the sheet-like strength adjustment parameter are taken to be of the same order of magnitude after normalization of the eigenvalue scale, so that the principal curvature response enhancement coefficient approaches one at the sheet-like structure and approaches zero at the point-like or block-like structure. This principal curvature response enhancement coefficient participates in the product fusion after the direction correction sub-step.

[0071] It should be noted that the process of determining the scale that is compatible with the thickness of the esophageal wall is as follows: the estimated thickness of the candidate region of the esophageal wall in the radial direction is used as the reference thickness, and a Hessian matrix is ​​calculated using a Gaussian scale of the same order of magnitude as the reference thickness; when the wall thickens due to infiltration, the reference thickness increases synchronously with the radial thickness of the candidate region of the esophageal wall, so that the scale is still compatible with the actual wall thickness in the infiltrated section; when a single scale is difficult to cover the range of thickness variation, several scales covering the upper and lower adjacent parts of the reference thickness are used to calculate and the larger principal curvature response enhancement coefficient is taken for each voxel as the principal curvature response enhancement coefficient of that voxel, so as to adapt to the variation of the wall thickness along the luminal axis.

[0072] The recording and direction correction process of the Hessian principal eigenvector is as follows: The eigenvector corresponding to the largest absolute value of the voxel that satisfies the preset wall sheet structure condition is recorded as the Hessian principal eigenvector. This direction corresponds to the direction in which the sheet structure bends most significantly, that is, the direction perpendicular to the wall sheet surface. The deviation between the Hessian principal eigenvector and the radial reference direction of the circumferential angle partition where the voxel is located is compared. The deviation is expressed by the omnidirectional angle between the two. When the deviation exceeds the preset angle range, it is considered that the sheet structure at the voxel deviates from the radial direction of the tube wall and is more likely to come from a non-tube wall structure. Therefore, its principal curvature response enhancement coefficient is reduced.

[0073] It should also be noted that the preset angle range is used to determine whether the orientation of the sheet structure is consistent with the radial direction of the tube wall. Since the normal of the sheet structure of the tube wall should be roughly along the radial direction of the tube wall, and the deviation caused by the local bending of the tube wall and the residual misalignment of the initial registration must be allowed, its value range is 0 degrees to 30 degrees, and the typical value is 20 degrees. That is, the principal curvature response enhancement coefficient of the voxel with a deviation of more than 20 degrees is reduced.

[0074] The product fusion process is as follows: the gradient response suppression coefficient at the same mode and the same voxel is multiplied by the principal curvature response enhancement coefficient after direction correction on a voxel-by-voxel basis. The resulting product is the asymptotic geometric residual field value of the mode at that voxel. Since the gradient response suppression coefficient takes a high value in the weak gradient region and the principal curvature response enhancement coefficient takes a high value at the sheet-like structure that conforms to the radial direction of the tube wall, the product of the two only takes a high value at the sheet-like second-order geometric structure that has neither obvious gray-level boundary nor conforms to the tube wall orientation, thereby locking the geometric traces of the submucosal progressive infiltration.

[0075] It should be noted that the first-order gradient response on which the gradient response suppression coefficient is based also serves as a criterion for determining the preset boundary identification conditions in the identification of strong boundary tumor regions. In weak gradient regions, the gradient response suppression coefficient takes a high value and the region does not meet the preset boundary identification conditions. In strong gradient regions, the gradient response suppression coefficient takes a low value and the region meets the preset boundary identification conditions. This makes the weak boundary segments identified by the asymptotic geometric residual sites naturally complementary to the strong boundary tumor regions in the gradient dimension, avoiding the same region being classified into two categories at the same time.

[0076] For example, the following provides a specific numerical implementation process for constructing asymptotic geometric residual fields: Take a voxel located in the infiltration segment of the magnetic resonance image. The normalized gradient magnitude within its circumferential angle partition is 0.4. According to the gradient response suppression coefficient calculation formula, with the attenuation scale parameter set to 1.0 and the attenuation exponent set to 2, the gradient response suppression coefficient is approximately... This indicates that the voxel is in a weak gradient region; the eigenvalues ​​of the Hessian matrix at this voxel, sorted by absolute value from smallest to largest, are approximately 0.05, 0.18, and 0.42, satisfying the pre-defined condition of a layered, sheet-like structure with one significantly large eigenvalue and the rest relatively small. The strengthening coefficient, calculated using the principal curvature response strengthening coefficient formula, is approximately 0.9; the deviation of the Hessian principal eigenvector recorded by this voxel from the radial reference direction of its circumferential angle partition is approximately 12 degrees, which does not exceed the pre-defined angle range of 20 degrees, therefore no downward adjustment is made; finally, the asymptotic geometric residual field value of this voxel is approximately... The higher value is taken; as a control, a voxel located at the esophageal blood vessel within the same circumferential angle partition is taken, and its Hessian eigenvalues ​​show two significant negative values ​​(corresponding to tubular structures). The principal curvature response enhancement coefficient is suppressed to about 0.1. Even if its gradient response suppression coefficient is 0.8, the asymptotic geometric residual field value is only about 0.08, so it is not identified as an infiltrative geometric trace.

[0077] Through the above technical solution, this embodiment constructs an asymptotic geometric residual field by fusing the voxel-by-voxel product of the three factors: the gradient response suppression coefficient normalized within the partition, the sheet-like structure enhancement coefficient adapted to the wall thickness, and the direction correction based on the radial direction of the tube wall. In this embodiment, response focusing is performed at the sheet-like structure with weak gray-scale boundaries and conforming to the radial direction of the tube wall, thereby separating the geometric traces of the submucosal progressive infiltration from non-wall structures such as blood vessels and artifacts.

[0078] In one specific embodiment of step S4, using the cavity shaft sampling segment and the corresponding local cross-section as the cross-modal corresponding unit, the spatial consistency of the asymptotic geometric residual field of each mode is measured to generate a cross-modal second-order geometric stability field, specifically including:

[0079] The combination of the cavity axis sampling segment and the circumferential angle partition is used as the cross-modal correspondence unit, and the asymptotic geometric residual field of each mode is mapped according to the cross-modal correspondence unit. Within each cross-modal correspondence unit, the number of modes in which the asymptotic geometric residual field reaches the preset residual response condition is counted to obtain the response co-occurrence degree. Within each cross-modal correspondence unit, the undirected deviation angle of the Hessian principal eigenvector of each mode record relative to the radial reference direction of the corresponding circumferential angle partition is determined, and the consistency of the Hessian principal eigenvector direction is obtained based on the difference between the undirected deviation angles of each mode. Within the local neighborhood of the cross-modal correspondence unit, the response distribution shape of the asymptotic geometric residual field of each mode is compared along the radial reference direction of the corresponding circumferential angle partition to obtain the local neighborhood response distribution correlation. The response co-occurrence degree is combined with at least one of the Hessian principal eigenvector direction consistency and the local neighborhood response distribution correlation to generate the cross-modal second-order geometric stability field.

[0080] It should be noted that the calculation process for the response co-occurrence degree is as follows: the combination of the cavity axis sampling segment and the circumferential angle partition is taken as a cross-modal correspondence unit, so that the cross-modal correspondence unit corresponds to a cavity axis sampling segment in the axial direction and a circumferential angle partition in the circumferential direction; the asymptotic geometric residual fields of each mode are mapped according to the corresponding cross-modal correspondence unit, that is, the asymptotic geometric residual field values ​​of each mode falling into the same cavity axis sampling segment and the same circumferential angle partition are assigned to the same cross-modal correspondence unit; within the cross-modal correspondence unit, it is determined whether the asymptotic geometric residual field of each mode meets the preset residual response condition, and the number of modes that meet the condition is counted. The proportion of the number of modes relative to the total number of modes participating in the measurement is used as the response co-occurrence degree.

[0081] It should also be noted that the preset residual response condition is used to determine whether the asymptotic geometric residual field of a certain mode in a certain cross-modal corresponding unit has reached the level that can be considered to have wetting geometric traces. Since the asymptotic geometric residual field is a dimensionless quantity in the range of 0 to 1, and the wetting geometric traces are higher than the normal pipe wall, its value range is 0.4 to 0.7, and the typical value is 0.5. That is, when the representative value of the asymptotic geometric residual field in the cross-modal corresponding unit is not less than 0.5, it is considered that the mode has reached the preset residual response condition. When there are only two modes involved in the measurement, the response co-occurrence degree is taken in three levels: 0, 1 / 2, and 1. At this time, the cross-modal second-order geometric stability field is gated based on the response co-occurrence degree, and then continuous values ​​are provided by the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution, thereby avoiding the overly coarse values ​​caused by counting only the number of modes.

[0082] To further clarify, the calculation process for the consistency of the Hessian principal eigenvector direction is as follows: Within each cross-modal corresponding cell, the undirected deviation angle of the Hessian principal eigenvector recorded for each modality relative to the radial reference direction of the corresponding circumferential angle partition is determined. This undirected deviation angle does not distinguish between positive and negative orientations, and only takes values ​​from 0 to... The included angle of the radian interval; when there are two modes involved in the measurement, the Hessian principal eigenvector direction consistency is obtained by the following formula based on the difference of the undirected deviation angles of the two modes. When there are more than two modes involved in the measurement, the maximum value of the pairwise difference of the undirected deviation angles of each mode is substituted into the following formula. The smaller the difference, the higher the direction consistency.

[0083] ;

[0084] in, The Hessian principal eigenvectors are consistent in direction. These are the undirected deviation angles of the Hessian principal eigenvectors of the two modes within the corresponding unit of the cross-modal region, relative to the radial reference direction, in radians; since the undirected deviation angle ranges from 0 to... Therefore, Normalization is performed in the denominator so that the Hessian principal eigenvectors are 1 when the directions are completely consistent and 0 when the directions are mutually orthogonal. The reason for using the difference of the undirected deviation angles relative to the radial reference direction instead of the direct angle between the two vectors is that there are still residual misalignments after the initial registration of each mode. Directly comparing the absolute directions of the two vectors will be affected by the misalignment, while the difference of the deviation angles of each relative to the same radial reference direction is not sensitive to the residual misalignment.

[0085] The calculation process of the local neighborhood response distribution correlation is as follows: In the local neighborhood of the cross-modal corresponding unit, take the one-dimensional response distribution of the asymptotic geometric residual field of each mode along the radial reference direction of the corresponding circumferential angle partition of the cross-modal corresponding unit, that is, the response profile from the center of the cavity outward along the radial direction; compare the shape of the one-dimensional response distribution of each mode, and use its correlation degree as the local neighborhood response distribution correlation. The higher the correlation degree, the more consistent the distribution of the wetting geometric traces of each mode along the radial direction of the pipe wall in the cross-modal corresponding unit.

[0086] The generation process of the cross-modal second-order geometric stability field is as follows: the response co-occurrence degree is used as a prerequisite gate variable. When the response co-occurrence degree is zero, the cross-modal second-order geometric stability field of the corresponding unit is zero. When the response co-occurrence degree is not zero, the combination result of at least one of the following, namely the Hessian principal eigenvector direction consistency and the correlation of local neighborhood response distribution, is weighted by the response co-occurrence degree to obtain the value of the cross-modal second-order geometric stability field of the corresponding unit. When only one of the terms is used, that term is used as the combination result. When both terms are used, the product or weighted sum of the two terms is used as the combination result.

[0087] Through the above technical solution, this embodiment uses the combination of cavity axis sampling segment and circumferential angle partition as cross-modal correspondence unit, uses the response co-occurrence degree as gating, and uses the consistency of the Hessian principal eigenvector direction relative to the radial reference direction and the correlation of the local neighborhood response distribution along the radial direction as a continuous metric for combination; this embodiment verifies the structural correspondence relationship with a relatively unified radial reference at the correspondence unit level that allows residual misalignment, thereby distinguishing between cross-modal consistent infiltration geometric traces and single-modal unique artifacts without requiring precise voxel-level correspondence.

[0088] After generating the cross-modal second-order geometric stability field in step S4, the process also includes mode-specific second-order response masking, specifically including:

[0089] The asymptotic geometric residual fields of each mode are compared with the cross-modal second-order geometric stability fields within the corresponding cross-modal units to determine the single-mode unique response region where the asymptotic geometric residual field of a certain mode meets the preset residual response condition and the corresponding cross-modal unit does not meet the preset stability condition. This single-mode unique response region is marked as an unstable second-order response region. Starting from the strong boundary tumor region adjacent to the progressive invasion candidate segment along the esophageal lumen axis, the connection relationship is passed segment by segment along the corresponding cross-modal units that satisfy continuity, coverage, or regional connectivity, thus assigning a connection to each progressive invasion candidate segment. The acceptance level decreases as the axial distance from the acceptance starting point increases; the acceptance level characterizes the reliability of the progressive invasion candidate segment continuing to the strong boundary tumor region via the acceptance relationship, and includes an effective acceptance level, a weak acceptance level, and an exclusion level; the acceptance relationship is interrupted when the acceptance relationship transmission path passes through an unstable second-order response region, and the acceptance level of the progressive invasion candidate segment after the interruption is reduced or excluded from the progressive invasion candidate segment; the effective acceptance state of the corresponding region in the cross-modal second-order geometric stability field is updated according to the excluded or downgraded acceptance level to obtain the shielded stability field.

[0090] It should be noted that the process of determining the unique single-mode response region is as follows: within each cross-modal corresponding unit, the asymptotic geometric residual field of each mode is compared with the cross-modal second-order geometric stability field of the corresponding unit; when the asymptotic geometric residual field of a certain mode reaches the preset residual response condition in the corresponding cross-modal unit, but the cross-modal second-order geometric stability field of the corresponding unit does not reach the preset stability condition, the region is determined to be a unique single-mode response region, that is, only a single mode produces a significant second-order response and does not obtain cross-modal consistency support; the unique single-mode response region is marked as an unstable second-order response region; here, the cross-modal second-order geometric stability field first determines whether the unit is cross-modal stable based on whether it reaches the preset stability condition, and then identifies the unstable second-order response region caused by the unique single-mode response in the cross-modal unstable unit. The two-level judgment is performed sequentially, the former judging whether it is stable or not, and the latter attributing it to the unique single-mode response;

[0091] It should also be noted that the acceptance level is a grading measure characterizing the reliability of a progressive invasion candidate segment continuing into a strongly bordered tumor region via an acceptance relationship. It includes three levels: effective acceptance level, weak acceptance level, and exclusion level. An effective acceptance level indicates that the progressive invasion candidate segment continues stably into a strongly bordered tumor region via a continuous, covering, or regionally connected relationship, and should be included in subsequent target volume acceptance. A weak acceptance level indicates that the acceptance relationship is interrupted or weakened, and further verification is required before inclusion. An exclusion level indicates that it fails to continue into a strongly bordered tumor region and should be excluded from the progressive invasion candidate segment.

[0092] The process of inheritance relationship transmission and interruption is as follows: taking the strong boundary tumor region adjacent to the progressive invasion candidate segment in the esophageal lumen axis as the inheritance starting point, the inheritance relationship is transmitted outward segment by segment along the esophageal lumen axis from the inheritance starting point along the cross-modal corresponding unit that satisfies the continuity relationship, coverage relationship, or regional connectivity relationship; each progressive invasion candidate segment on the transmission path is assigned an inheritance level that decreases as the axial distance from the inheritance starting point increases, so that those adjacent to the strong boundary tumor region take the effective inheritance level, and those that decrease to the weak inheritance level as the axial distance increases; when the inheritance relationship transmission path passes through the cross-modal corresponding unit marked as an unstable second-order response region, the inheritance relationship is interrupted at that point, and the inheritance level of the progressive invasion candidate segment after the interruption point is reduced or excluded from the progressive invasion candidate segment.

[0093] Regarding the uniqueness of the direction of the inheritance relationship and the handling after interruption: the inheritance relationship is transmitted unidirectionally outward from the strong boundary tumor region along the esophageal lumen axis, and does not start from the side of the progressive invasion candidate segment in reverse, so as to ensure that the inheritance starting point is unique and the inheritance level decreases monotonically with the axial interval; when the transmission path is interrupted at a certain cross-modal corresponding unit, the progressive invasion candidate segment that has obtained a valid inheritance level before the interruption point retains its inheritance level. If there are still other connecting paths with the strong boundary tumor region after the interruption point, the progressive invasion candidate segment is taken according to the highest inheritance level on the other path. If there are no other connecting paths, it is reduced to a weak inheritance level or an exclusion level, thereby avoiding the direct exclusion of all progressive invasion candidate segments after the interruption point due to the interruption of a single path;

[0094] The generation process of the shielded stability field is as follows: based on the acceptance level after exclusion or downgrading, the effective acceptance state of the corresponding region in the cross-modal second-order geometric stability field is updated. That is, the effective acceptance state of the excluded region is set to invalid, the effective acceptance state of the downgraded region is marked as the corresponding weak acceptance level, and the effective acceptance state of the region that maintains the effective acceptance level is marked as valid. The cross-modal second-order geometric stability field with effective acceptance state annotation is used as the shielded stability field. The shielded stability field is still a cross-modal second-order geometric stability field in numerical terms, but with the addition of effective acceptance state annotation for each region, for use in the third-modal cross-verification processing, registration write-back, and target area update processing.

[0095] It should be noted that the continuity, coverage, and regional connectivity relationships on which the succession relationship is based are the same set of geometric relationships used when generating radiotherapy target volume data. Succession relationship transfer is equivalent to pre-organizing these three geometric relationships in a directional and segmented manner, starting from a strong boundary tumor region, before generating radiotherapy target volume data. This ensures that the progressive infiltration candidate segments that are subsequently accepted when generating radiotherapy target volume data are all those that have passed the succession level screening. An additional effect of this pre-organization is that progressive infiltration candidate segments that rely on unstable second-order response region connections or are isolated in strong boundary tumor regions are excluded or downgraded before generating radiotherapy target volume data, thus preventing them from entering the GTV area mask.

[0096] Through the above technical solution, this embodiment uses the cross-modal second-order geometric stability field as a reference to mark the unique significant response of a single mode as an unstable second-order response region, and uses the strong boundary tumor region as the starting point to unidirectionally transfer the acceptance level along the lumen and interrupt it at the unstable second-order response region; this embodiment distinguishes between real infiltration traces and single-modal artifacts at the acceptance relationship level based on the continuity of the relatively strong boundary tumor region, thereby eliminating isolated artifacts without mistakenly deleting real weak infiltration traces.

[0097] Following the modal-specific second-order response masking process, a third-mode cross-validation process is also included, specifically:

[0098] When a third modality esophageal image exists in the shared spatial esophageal image group, the progressive geometric residual field response of the third modality esophageal image is extracted in the spatial constraint area according to the luminal axis sampling segment and circumferential angle partition, as well as the boundary auxiliary response that meets the preset boundary recognition conditions; the cross-modal corresponding units of the progressive infiltration candidate segments on both sides of the interruption of the connection relationship are respectively verified with the progressive geometric residual field response and boundary auxiliary response of the third modality esophageal image; the connection relationship is reconstructed and the connection level is improved for the progressive infiltration candidate segments that are supported by the third modality esophageal image at the interruption point; the progressive infiltration candidate segments that are not supported by the third modality esophageal image are marked as weak connection candidate segments, and their connection relationship is kept interrupted;

[0099] It should be noted that the third modality of esophageal image refers to an optional modality of image acquisition other than the at least two modalities involved in generating the cross-modal second-order geometric stability field. It is used for cross-verification rather than participating in the generation of the cross-modal second-order geometric stability field. Therefore, it is not required to calculate the complete second-order geometric shape at a scale adapted to the thickness of the esophageal wall layer. The process of extracting the response of the third modality of esophageal image is as follows: within the spatially constrained area, the image is divided into sections according to the luminal axis sampling segment and circumferential angle. At least one of the responses of the asymptotic geometric residual field or the boundary auxiliary response that meets the preset boundary recognition conditions is extracted from the third modality of esophageal image. The response of the asymptotic geometric residual field is used when the third modality itself can calculate the second-order geometric shape, and the boundary auxiliary response is used when the third modality mainly provides boundary or metabolic contour information. The two are selected or used together according to the imaging characteristics of the third modality of esophageal image.

[0100] The process of cross-verification and reconnection reconstruction is as follows: The location where the connection transmission is interrupted in the unstable second-order response region is determined. The corresponding cross-modal units of the progressive infiltration candidate segments on both sides of the interruption along the esophageal lumen axis are taken. These cross-modal units are then verified against the responses or boundary-assisted responses of the progressive geometric residual fields in the same lumen axis sampling segment and circumferential angle partition of the third-modal esophageal image. When the progressive infiltration candidate segments on both sides of the interruption obtain responses that meet the corresponding support conditions in the third-modal esophageal image, they are determined to have received support from the third-modal esophageal image. The connection is then reconstructed at the interruption point, allowing the interrupted progressive infiltration candidate segments on both sides to reconnect and improving their connection level. When no support is received from the third-modal esophageal image, the progressive infiltration candidate segment is marked as a weak connection candidate segment, and its connection remains interrupted.

[0101] When there is no third modality esophageal image in the shared space esophageal image group, the third modality cross-verification process is skipped, and the shielded stability field and its acceptance level obtained by the modality-unique second-order response shielding process are directly used in the subsequent processing. The progressive infiltration candidate segments on both sides of the interruption of the acceptance relationship maintain their acceptance level obtained in the modality-unique second-order response shielding process.

[0102] The interruption in the continuity relationship targeted by the third modal cross-verification process originates from the interruption of the continuity relationship in the unstable second-order response region in the modality-specific second-order response masking process. The linkage between the two is that the modality-specific second-order response masking process tends to interrupt the continuity at cross-modal instability to eliminate artifacts, while the third modal cross-verification process provides a way to restore the true immersion continuity by using an additional mode. The additional effect of this linkage is that it not only retains the ability to eliminate single-modal artifacts, but also reduces the risk of misjudging the true but occasionally unstable immersion segment between the original two modes as an interruption.

[0103] Through the above technical solution, this embodiment introduces a third modality of esophageal image locally at the point of interruption of the connection relationship, and uses the response of its asymptotic geometric residual field or boundary auxiliary response to cross-verify the asymptotic infiltration candidate segments on both sides of the interruption point, and reconstructs the connection relationship or marks weak connection candidate segments accordingly; this embodiment performs local on-demand verification at the interruption point of the connection relationship, thereby recovering the real infiltration connection interrupted by occasional cross-modal instability without allowing additional modal noise to enter the entire domain.

[0104] After generating esophageal cancer radiotherapy target volume data including GTV region mask and CTV candidate region contour data, the process also includes registration write-back and target volume update processing, specifically including:

[0105] In the cross-modal second-order geometric stability field after shielding stability field or third-modal cross-verification, regions with a reception level not lower than the preset write-back level are selected; within the selected regions, regions where the radial reference direction of each modality's Hessian principal feature vector relative to the corresponding circumferential angle partition satisfies the preset direction consistency condition are identified as weak boundary cross-modal stability constraint sources; the positional correspondence of the weak boundary cross-modal stability constraint sources on the cavity axis sampling segment and circumferential angle partition is written into the update processing for obtaining the preliminary registration results of the common space esophageal image group; according to the aforementioned positions... The correspondence constraint is used to define the correspondence of each modality in the arc length direction of the esophageal lumen axis centerline and the circumferential angle partition, and to maintain the consistency of the response distribution of the asymptotic geometric residual field in the radial neighborhood corresponding to each circumferential angle partition. Based on the updated registration results, the asymptotic geometric residual field and the cross-modal second-order geometric stability field of each modality are regenerated in the spatial constraint area, and the inheritance relationship and inheritance level are retransmitted. The asymptotic infiltration candidate segment is updated according to the retransmitted inheritance level, and when the preset convergence condition is met or the preset number of updates is reached, the regenerated or corrected esophageal cancer radiotherapy target area data is output.

[0106] It should be noted that the screening process for weak boundary cross-modal stability constraint sources is as follows: In the cross-modal second-order geometric stability field after shielding stability field or cross-verification by the third mode, regions with a bearing level not lower than the preset write-back level are first screened to ensure that only regions with stable bearing to strong boundary tumor regions are used; then, within the screened regions, regions whose Hessian principal eigenvectors of each mode satisfy the preset direction consistency condition relative to the radial reference direction of the corresponding circumferential angle partition are selected and determined as weak boundary cross-modal stability constraint sources, so that the regions used as constraint sources have both bearing reliability and cross-modal direction consistency;

[0107] It should also be noted that the preset write-back level is the lowest acceptance level allowed as a weak boundary cross-modal stability constraint source. It is used to limit the constraint source to a region with reliable acceptance. Since the acceptance relationship of weak acceptance candidate segments may be interrupted or weakened, and is not suitable for reverse constraint registration, the preset write-back level is the effective acceptance level. That is, only regions with an effective acceptance level can participate in write-back. The preset direction consistency condition is the maximum undirected deviation angle of the Hessian principal eigenvector of each mode as a constraint source relative to the radial reference direction. It is used to exclude regions with discrete directions. Since the structure used as a registration constraint should be directionally stable and consistent across modes, its value ranges from 0 degrees to 15 degrees, with a typical value of 10 degrees.

[0108] To further explain, the registration write-back and regeneration process is as follows: The positional correspondence of the weak boundary cross-modal stability constraint source on the luminal axis sampling segment and circumferential angle partition is written into the update processing used to obtain the preliminary registration results of the common space esophageal image group. That is, the position of each modality at the weak boundary cross-modal stability constraint source along the arc length direction of the esophageal luminal axis centerline and on the circumferential angle partition is added to the registration as a new correspondence; the correspondence of each modality on the arc length direction of the esophageal luminal axis centerline and on the circumferential angle partition is constrained according to the positional correspondence, and the response distribution pattern of the asymptotic geometric residual field is kept consistent in the radial neighborhood corresponding to each circumferential angle partition, so that the registration obtains the originally missing local correspondence anchor point in the infiltration segment of the weak gray-scale boundary; based on the updated registration results, the asymptotic geometric residual field and cross-modal second-order geometric stability field of each modality are regenerated in the spatial constraint area, and the inheritance relationship and inheritance level are retransmitted with the updated strong boundary tumor region as the inheritance starting point;

[0109] The process of terminating the iteration and outputting the target area data is as follows: the retransmitted acceptance level is used to update the progressive invasion candidate segment, and the radiotherapy target area data is regenerated or corrected based on the updated progressive invasion candidate segment and the strong boundary tumor region; after each round of regeneration, it is determined whether the preset convergence condition or the preset number of updates is met. If either condition is met, the iteration stops and the regenerated or corrected esophageal cancer radiotherapy target area data is output. If neither condition is met, the registration result updated in this round and the cross-modal second-order geometric stability field are used as inputs to continue the next round of registration write-back and target area update processing.

[0110] The preset convergence condition and preset update count are used to ensure that the registration write-back and target area update processing terminate within a limited number of rounds: The preset convergence condition is that the change between the GTV area masks obtained in two adjacent rounds is lower than the change threshold. Since the continuous decrease in change indicates that the registration and recognition have become stable, the voxel difference ratio of the GTV area masks in two adjacent rounds is lower than the change threshold as the convergence criterion. The change threshold ranges from 1 / 100 to 5 / 100, with a typical value of 2 / 100. The preset update count is the maximum number of rounds that registration write-back and target area update processing are allowed to be performed. It is used to force termination when the convergence is slow. Since the corresponding anchor points of the registration in the immersion section have been basically filled after one or two rounds of write-back, its value ranges from 1 to 5 times, with a typical value of 2 times.

[0111] It should be noted that the esophageal cancer radiotherapy target area data output in this step includes the following information fields depending on the termination scenario: In the case of termination due to meeting the preset convergence conditions, the output fields include the voxel identifier set of the GTV region mask, the contour point set of the CTV candidate region contour data, each progressive infiltration candidate segment and its final acceptance level, and the actual number of update rounds performed; In the case of termination due to reaching the preset number of updates, the output fields are supplemented with a non-convergence mark to indicate that the result has not met the preset convergence conditions within the preset number of updates.

[0112] For example, following the aforementioned through example, after the second round of updates, the voxel difference ratio between adjacent rounds of GTV region masks drops to approximately 1.5 / 100, which is lower than the change threshold of 2 / 100, thus satisfying the preset convergence condition and terminating. The output esophageal cancer radiotherapy target area data includes a set of voxel identifiers of the GTV region mask covering a tumor region with strong boundaries of approximately 2 to 5 cm below the carina and a progressive infiltration candidate segment approximately 1.5 cm below it, a set of CTV candidate region contour points obtained by expanding outward based on the GTV region mask, effective acceptance level labels corresponding to each progressive infiltration candidate segment, and the actual number of update rounds performed being 2.

[0113] It should be noted that the registration write-back and target update processing rely on the availability of each modal image when regenerating the asymptotic geometric residual field of each modality. When a modal image has a large range of missing data or artifact contamination in the spatial constraint region under the updated registration result, causing its asymptotic geometric residual field to be unable to be stably generated in most cross-modal corresponding units, the registration write-back and target update processing is terminated, and the previous round of converged or the most recent round of effective esophageal cancer radiotherapy target data is used as the output to avoid iterative divergence caused by modal data abnormalities.

[0114] Through the above technical solution, this embodiment uses the dual screening of weak boundary cross-modal stable constraint sources based on the consistency of acceptance level and direction, writes their positional correspondence back to the initial registration, and regenerates the asymptotic geometric residual field and the cross-modal second-order geometric stability field based on the updated registration results. When the preset convergence condition is met or the preset number of updates is reached, the target area data is output. This embodiment uses the identified weak boundary stable structure to perform reverse constraint registration, forming a closed loop that drives each other at the corresponding unit level, thereby supplementing the missing local corresponding anchor points for the infiltration segment of the weak gray-scale boundary. The regenerated or corrected esophageal cancer radiotherapy target area data produced is provided as the final output of this method to the downstream radiotherapy planning stage.

[0115] See Figure 2 As shown, this scheme proposes an esophageal cancer radiotherapy target volume generation system with cross-modal correction of the cavity axis anchor point, which is used to realize the above-mentioned method for generating esophageal cancer radiotherapy target volume with cross-modal correction of the cavity axis anchor point. It includes: an image acquisition and registration module, which is used to acquire esophageal images of at least two modalities of the same patient, and obtain a co-space esophageal image group after standardization and preliminary registration.

[0116] The lumen axis extraction and segmentation module is used to extract the esophageal lumen axis centerline in the co-space esophageal image group, using its radial neighborhood as the spatial constraint area, and setting lumen axis sampling segments and corresponding local cross-sections along the esophageal lumen axis centerline;

[0117] The residual field generation module is used to generate gradient response suppression coefficients that decrease as the gradient magnitude increases for each voxel of each modality of esophageal image within the spatial constraint region, based on the first-order gradient response. It also generates principal curvature response enhancement coefficients that increase as the structured second-order geometric response increases based on the eigenvalue combination of the Hessian matrix, and records the corresponding Hessian principal eigenvectors. The product of the two coefficients is then fused to obtain the asymptotic geometric residual field of the corresponding modality.

[0118] The stability field construction module is used to measure the spatial consistency of the asymptotic geometric residual fields of each mode, taking the cavity axis sampling segment and the corresponding local cross-section as the cross-modal corresponding unit, and generate a cross-modal second-order geometric stability field. The spatial consistency measure includes the degree of response co-occurrence of the asymptotic geometric residual fields of different modes in the corresponding unit, and at least one of the following: the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution.

[0119] The region identification and discrimination module is used to identify regions where the cross-modal second-order geometric stability field satisfies the preset stability condition and the first-order gradient response does not reach the preset boundary identification condition as progressive infiltration candidate segments, and regions where the first-order gradient response reaches the preset boundary identification condition and satisfies the preset boundary consistency condition within the corresponding cross-modal unit as strong boundary tumor regions.

[0120] The target data generation module is used to generate esophageal cancer radiotherapy target data, which includes GTV region mask and CTV candidate region contour data, based on the continuity relationship between progressive infiltration candidate segments and strong boundary tumor regions along the esophageal lumen axis, the coverage relationship within the corresponding local cross-section, and the regional connectivity relationship between the two.

[0121] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0122] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0123] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0124] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for generating esophageal cancer radiotherapy target volume with cross-modal correction of cavity axis anchor points, characterized in that, The method includes: At least two modalities of esophageal images were acquired from the same patient, and after standardization and preliminary registration, a co-space esophageal image group was obtained; The esophageal lumen centerline is extracted from the co-space esophageal image group, and its radial neighborhood is used as the spatial constraint area. The lumen sampling segment and corresponding local cross-section are set along the esophageal lumen centerline. Within the spatially constrained region, for each modality of esophageal image, a gradient response suppression coefficient that decreases as the gradient magnitude increases is generated for each voxel based on the first-order gradient response. A principal curvature response enhancement coefficient that increases as the structured second-order geometric response is enhanced is generated based on the combination of eigenvalues ​​of the Hessian matrix. The corresponding Hessian principal eigenvectors are recorded. The asymptotic geometric residual field of the corresponding modality is obtained by multiplying and fusing the two coefficients. Using the cavity axis sampling segment and the corresponding local cross-section as the cross-modal corresponding unit, the spatial consistency measure of the asymptotic geometric residual field of each mode is performed to generate a cross-modal second-order geometric stability field. The spatial consistency measure includes the degree of response co-occurrence of the asymptotic geometric residual fields of different modes in the corresponding unit, and at least one of the following: the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution. Regions where the cross-modal second-order geometric stability field satisfies the preset stability condition and the first-order gradient response does not meet the preset boundary identification condition are identified as progressive infiltration candidate segments. Regions where the first-order gradient response meets the preset boundary identification condition and satisfies the preset boundary consistency condition within the corresponding cross-modal unit are identified as strong boundary tumor regions. Based on the continuity of progressively infiltrating candidate segments and strongly boundary tumor regions along the esophageal lumen axis, their coverage relationship within corresponding local cross sections, and the regional connectivity between them, esophageal cancer radiotherapy target data containing GTV region mask and CTV candidate region contour data is generated.

2. The method according to claim 1, characterized in that, The esophageal lumen centerline is extracted from the shared-space esophageal image set. Its radial neighborhood is used as the spatial constraint region, and a lumen centerline sampling segment and corresponding local cross-section are set along the esophageal lumen centerline. Specifically, this includes: Sampling points along the arc length of the esophageal lumen axis are set, and the arc length interval between adjacent lumen axis sampling points is taken as the lumen axis sampling segment. At each cavity axis sampling point, a local tangent of the esophageal cavity axis centerline is taken, and a plane perpendicular to the local tangent is taken as the corresponding local cross section. Within this cross section, the circumferential angle zero position is determined with the cavity axis sampling point as the center. Within each local cross section, taking the zero position of the circumferential angle as the starting direction, the area around the cavity axis sampling point is divided into multiple circumferential angle partitions. The direction from the cavity axis sampling point to the center line of each circumferential angle partition is taken as the radial reference direction of the corresponding circumferential angle partition. Within each circumferential angle partition, the ring extending outward from the cavity axis sampling point along the corresponding radial reference direction to a preset radial depth is taken as the radial neighborhood. Within the radial neighborhood, a ring-shaped region that is continuous along the circumferential direction and forms a local peak of gradient modulus relative to the adjacent ring zone is identified as a candidate region of the esophageal wall. The spatial constraint region is obtained by continuously arranging the radial neighborhoods within each local cross-section along the esophageal lumen axis centerline, and when there are candidate regions of the esophageal wall, the spatial constraint region is updated by the overlapping area of ​​the candidate region of the esophageal wall and the radial neighborhood.

3. The method according to claim 2, characterized in that, The asymptotic geometric residual field of the corresponding mode is obtained by multiplying and fusing the product of the two coefficients, specifically including: Within the spatially constrained region, the first-order gradient response is calculated voxel by voxel for each modality of esophageal image. The first-order gradient response is normalized according to the gradient magnitude distribution of each circumferential angle partition, generating a gradient response suppression coefficient that decreases as the gradient magnitude increases. Within the spatial constraint region, the eigenvalue combination of the Hessian matrix is ​​calculated voxel by voxel for the same modality of esophageal images according to a scale adapted to the thickness of the esophageal wall. The voxel response that satisfies the preset conditions of the wall sheet structure is enhanced, and the voxel response that satisfies the preset conditions of the non-wall structure is suppressed, generating the principal curvature response enhancement coefficient. The feature vectors corresponding to voxels that meet the preset wall-layer sheet structure conditions are recorded as Hessian principal feature vectors. Compare the degree of deviation between the Hessian principal eigenvector and the radial reference direction of the corresponding circumferential angle partition, and adjust the principal curvature response enhancement coefficient of the voxel whose deviation exceeds the preset angle range. By fusing the gradient response suppression coefficient and principal curvature response enhancement coefficient of the same mode on a voxel-by-voxel basis, the asymptotic geometric residual field of that mode is obtained.

4. The method according to claim 3, characterized in that, Using the cavity shaft sampling segment and its corresponding local cross-section as the cross-modal corresponding unit, the spatial consistency of the asymptotic geometric residual field of each mode is measured to generate a cross-modal second-order geometric stability field, specifically including: The combination of cavity axis sampling segment and circumferential angle partition is used as the cross-modal corresponding unit, and the asymptotic geometric residual field of each mode is mapped according to the cross-modal corresponding unit. The number of modes in each cross-modal corresponding unit that achieve the preset residual response condition is counted to obtain the response co-occurrence degree; Within each cross-modal corresponding unit, the undirected deviation angle of the Hessian principal eigenvector of each modality record relative to the radial reference direction of the corresponding circumferential angle partition is determined, and the consistency of the Hessian principal eigenvector direction is obtained based on the difference between the undirected deviation angles of each modality. Within the local neighborhood of the corresponding unit across modes, the response distribution of each mode's asymptotic geometric residual field is compared along the radial reference direction of the corresponding circumferential angle partition to obtain the correlation of the local neighborhood response distribution. By combining at least one of the following: the degree of response co-occurrence with the consistency of the Hessian principal eigenvector direction and the correlation of local neighborhood response distribution, a cross-modal second-order geometrically stable field is generated.

5. The method according to claim 4, characterized in that, After generating the cross-modal second-order geometrically stable field, the process also includes mode-specific second-order response masking, specifically including: The asymptotic geometric residual fields of each mode are compared with the cross-modal second-order geometric stability fields in the corresponding cross-modal units to determine the single-mode unique response region where the asymptotic geometric residual field of a certain mode reaches the preset residual response condition and the cross-modal corresponding unit does not reach the preset stability condition. The unique response region of a single mode is marked as an unstable second-order response region; Starting from a strongly bordered tumor region adjacent to the progressive invasion candidate segment along the esophageal axial direction, the inheritance relationship is passed segment by segment along cross-modal corresponding units that satisfy continuity, coverage, or regional connectivity. Each progressive invasion candidate segment is assigned an inheritance level that decreases as its axial distance from the inheritance starting point increases. The inheritance level characterizes the reliability of the progressive invasion candidate segment continuing to the strongly bordered tumor region via the inheritance relationship, and includes an effective inheritance level, a weak inheritance level, and an exclusion level. The succession relationship is interrupted when the succession relationship transmission path passes through the unstable second-order response region, and the succession level of the asymptotic infiltration candidate segment after the interruption point is reduced or excluded from the asymptotic infiltration candidate segment. The effective acceptance state of the corresponding region in the cross-modal second-order geometric stability field is updated based on the acceptance level after exclusion or downgrading, thus obtaining the shielded stability field.

6. The method according to claim 5, characterized in that, Following the modal-specific second-order response masking process, a third-mode cross-validation process is also included, specifically: When a third modality of esophageal image exists in the shared space esophageal image group, the progressive geometric residual field response of the third modality of esophageal image is extracted in the spatial constraint area according to the luminal axis sampling segment and circumferential angle partition, as well as the boundary auxiliary response that meets the preset boundary recognition conditions. The cross-modal corresponding units of the progressive infiltration candidate segments on both sides of the interruption of the connection relationship are respectively verified with the progressive geometric residual field response and boundary auxiliary response of the third modality esophageal image. Reconstruct the suture relationship and improve the suture level for progressive infiltration candidate segments that have obtained third-modality esophageal imaging support at the interruption site; Progressive infiltration candidate segments that do not receive support from third-modality esophageal imaging are marked as weakly successive candidate segments, and their succession relationship is maintained as interrupted.

7. The method according to claim 6, characterized in that, After generating esophageal cancer radiotherapy target volume data including GTV region mask and CTV candidate region contour data, the process also includes registration write-back and target volume update processing, specifically including: In a shielded stability field or a cross-modal second-order geometric stability field after third-mode cross-verification, regions with a reception level not lower than the preset write-back level are selected. Within the selected region, the region where the Hessian principal eigenvector of each mode satisfies the preset direction consistency condition with respect to the radial reference direction of the corresponding circumferential angle partition is selected and determined as the weak boundary cross-modal stability constraint source. The positional correspondence of the weak boundary cross-modal stability constraint source on the cavity axis sampling segment and circumferential angle partition is written into the update processing used to obtain the preliminary registration results of the common space esophageal image group; The correspondence between the modes is constrained according to the positional correspondence in the arc length direction of the esophageal lumen axis centerline and the circumferential angle partition, and the response distribution pattern of the asymptotic geometric residual field is kept consistent in the radial neighborhood corresponding to each circumferential angle partition. Based on the updated registration results, the asymptotic geometric residual field and cross-modal second-order geometric stability field of each mode are regenerated within the spatial constraint region, and the inheritance relationship and inheritance level are retransmitted. The progressive infiltration candidate segments are updated based on the retransmitted acceptance level, and the regenerated or corrected esophageal cancer radiotherapy target area data is output when the preset convergence condition is met or the preset number of updates is reached.

8. A cross-modal correction system for esophageal cancer radiotherapy target generation based on cavity axis anchor points, characterized in that, A method for generating esophageal cancer radiotherapy target volume to achieve cross-modal correction of the cavity axis anchor point as described in any one of claims 1-7, comprising: The image acquisition and registration module is used to acquire esophageal images of at least two modalities from the same patient, and obtain a co-space esophageal image group after standardization and preliminary registration. The lumen axis extraction and segmentation module is used to extract the esophageal lumen axis centerline in the co-space esophageal image group, using its radial neighborhood as the spatial constraint area, and setting lumen axis sampling segments and corresponding local cross-sections along the esophageal lumen axis centerline; The residual field generation module is used to generate gradient response suppression coefficients that decrease as the gradient magnitude increases for each voxel of each modality of esophageal image within the spatial constraint region, based on the first-order gradient response. It also generates principal curvature response enhancement coefficients that increase as the structured second-order geometric response increases based on the eigenvalue combination of the Hessian matrix, and records the corresponding Hessian principal eigenvectors. The product of the two coefficients is then fused to obtain the asymptotic geometric residual field of the corresponding modality. The stability field construction module is used to measure the spatial consistency of the asymptotic geometric residual fields of each mode, taking the cavity axis sampling segment and the corresponding local cross-section as the cross-modal corresponding unit, and generate a cross-modal second-order geometric stability field. The spatial consistency measure includes the degree of response co-occurrence of the asymptotic geometric residual fields of different modes in the corresponding unit, and at least one of the following: the consistency of the Hessian principal eigenvector direction and the correlation of the local neighborhood response distribution. The region identification and discrimination module is used to identify regions where the cross-modal second-order geometric stability field satisfies the preset stability condition and the first-order gradient response does not reach the preset boundary identification condition as progressive infiltration candidate segments, and regions where the first-order gradient response reaches the preset boundary identification condition and satisfies the preset boundary consistency condition within the corresponding cross-modal unit as strong boundary tumor regions. The target data generation module is used to generate esophageal cancer radiotherapy target data, which includes GTV region mask and CTV candidate region contour data, based on the continuity relationship between progressive infiltration candidate segments and strong boundary tumor regions along the esophageal lumen axis, the coverage relationship within the corresponding local cross-section, and the regional connectivity relationship between the two.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images