Inner ear sub-region segmentation control method for semicircular canal elongated structure fracture correction

Abstract

The application discloses an inner ear sub-region segmentation control method for semicircular canal elongated structure fracture correction, which is developed around the existing inner ear sub-region segmentation result, does not take retraining the model as a necessary condition, and forms a closed-loop processing flow of "detection-scoring-correction-revision-output". The method can conditionally correct the local fracture and abnormal connection of the semicircular canal on the basis of preserving the overall morphology of the original segmentation, so that the segmentation result is not only acceptable in the voxel overlap layer, but also more stable in the semicircular canal structure continuity and anatomical rationality.

Description

Technical Field

[0001] This embodiment relates to the field of ear subregion segmentation quality control technology, specifically to an inner ear subregion segmentation control method for correcting fractures in the slender semicircular canal structure. Background Technology

[0002] Inner ear subregion segmentation is an important aspect of medical image post-processing, commonly used for detailed anatomical observation of the inner ear, auxiliary disease assessment, preoperative planning, research data analysis, and automated annotation quality control. Current techniques typically rely on three-dimensional magnetic resonance imaging (MRI) images, such as 3D T2-weighted images, CISS sequences, and FIESTA sequences, which can display the fluid-filled structures of the inner ear. Preprocessing steps such as image resampling, registration, cropping, and intensity normalization are used to input the regions to be segmented into an automatic segmentation algorithm. Common segmentation algorithms include 3D U-Net networks, 3D Transformer segmentation networks, and cascaded convolutional neural networks, as well as traditional image segmentation methods such as thresholding, region growing, atlas registration, and morphological processing. The front-end segmentation module generally outputs two types of results: a probability map corresponding to each inner ear subregion and a label map with a defined category for each voxel. Label categories can be classified in a coarse-grained manner, such as basal cochlear turns, mid-apical cochlear turns, vestibule, and semicircular canals, or further differentiated between the anterior semicircular canals, posterior semicircular canals, lateral semicircular canals, vestibule, and different turns of the cochlea. The basic principle is to utilize image grayscale, spatial context, and annotation patterns in training samples to classify three-dimensional voxels point-by-point or block-by-block, and then generate the final label map through the maximum probability principle, threshold filtering, or post-processing rules. After the automatic segmentation results are output, existing technologies often use methods such as small connected component deletion, maximum connected component preservation, hole filling, dilation, erosion, and smoothing to improve the label appearance, and evaluate the similarity between the segmentation results and manual annotations through indicators such as Dice coefficient, intersection-union ratio, Hausdorff distance, and volume difference. For relatively large regions such as the overall structure of the inner ear or the cochlea, the above techniques can improve segmentation efficiency to a certain extent, reduce the workload of manual layer-by-layer delineation, and provide a basis for automation in batch case analysis.

[0003] Current technologies still have significant limitations when dealing with slender, curved structures like the semicircular canals, which have small diameters and complex spatial orientations. In 3D images, the semicircular canals typically appear as thin, tubular, annular, or arc-shaped structures with small local cross-sectional areas. Their boundaries are significantly affected by spatial resolution, slice thickness, partial volumetric effects, noise, motion artifacts, and low contrast. Current segmentation models primarily focus on voxel classification, prioritizing whether each voxel belongs to a particular category rather than ensuring the overall topological continuity of the semicircular canal. When the front-end model experiences a decrease in confidence at a small segment of the semicircular canal, that local voxel is easily mistaken for background, leading to breaks, missing segments, or isolated distal segments in the label image. Since the hard labels at the break points have disappeared, simply relying on maximum connectivity preservation may also treat real distal semicircular canal segments as noise and delete them, making it impossible to recover the original structure later. If ordinary expansion, closure, or hole filling is used for remediation, it may forcibly connect adjacent structures in the absence of directional, radii, and imaging evidence, causing abnormal adhesion between the semicircular canals and the vestibule, cochlea, other branches of the semicircular canals, or background noise.

[0004] Most existing post-processing methods are general morphological processing methods, lacking structural constraints specific to the anatomical morphology of the semicircular canals. The semicircular canals are characterized by their elongated shape, continuity, relatively stable local diameter, and smooth, tortuous path. The two ends of a fracture typically exhibit certain correspondences in spatial distance, endpoint orientation, local diameter, and image grayscale. Conventional post-processing operates only based on voxel adjacency, connected region size, or local morphology, failing to determine whether the two endpoints belong to the same semicircular canal fracture, or whether there is a weak probability map response or original MRI grayscale support between the two endpoints. Especially in images with anisotropic voxel spacing, a single voxel interval in the slice thickness direction may correspond to a large actual physical distance. If connectivity is still judged based on ordinary voxel distances, the fracture distance is easily underestimated or oblique connections are misjudged. When the semicircular canals and vestibule are often combined into the same label, the vestibule body is relatively large, while the semicircular canal branches are relatively slender. The lack of a clear boundary between them within the same label makes it difficult for existing techniques to reliably extract the truly quality-controlled candidate semicircular canal regions from the combined area, easily treating vestibular edges, noisy bridges, or local false positive branches as semicircular canals.

[0005] Existing evaluation and quality control methods are insufficient to fully reflect structural errors in semicircular canal segmentation. The Dice coefficient and cross-union ratio primarily reflect voxel overlap and are not sensitive to topological errors such as small-scale breaks, short-bridge misconnections, and missing local canals. A segmentation result with high overall overlap may still fail to meet the requirements for subsequent anatomical analysis or clinical examination regarding semicircular canal continuity. While data augmentation, label preservation strategies, or network structure optimization during model training can improve average segmentation performance, they cannot guarantee that each output result is free of local breaks or anomalous connections. Existing technologies typically lack an independent post-output quality control layer, failing to automatically identify break endpoints, assess the degree of completion, limit completion paths, remove low-confidence anomalous connections, and verify corrected volume changes, connectivity, diameter changes, and label conflicts. Due to the lack of correction marker maps and structural quality reports, physicians or annotators also find it difficult to quickly identify which locations the algorithm modified, which regions had low confidence, and which breaks were only indicated but not automatically corrected. Therefore, while existing automatic segmentation techniques for inner ear subregions can provide initial labeling results, they still lack specific technical means to address the continuity, anatomical rationality, and controllability of corrections for the slender structures of the semicircular canals. Summary of the Invention

[0006] The main objective of this invention is to provide an inner ear subregion segmentation control system for correcting fractures in the slender structure of the semicircular canal, thereby addressing the shortcomings of related technologies.

[0007] To achieve the above objectives, according to a first aspect of the present invention, a method for inner ear subregion segmentation control oriented towards fracture correction of slender semicircular canal structures is provided, comprising: Acquire a three-dimensional inner ear magnetic resonance image, the corresponding triaxial voxel spacing, the subregion probability map output by the front-end segmentation, and an initial subregion label map; extract candidate semicircular canal regions to be quality controlled from the initial subregion label map; perform centerline extraction on the candidate semicircular canal regions to obtain a corresponding skeleton set; identify a skeleton endpoint set based on the skeleton set; filter endpoint pairs that need to be automatically completed from the skeleton endpoint set; generate a center completion path for the endpoint pairs that need to be automatically completed; and generate a tubular completion mask based on the local diameter of the endpoints after obtaining the center completion path; merge the tubular completion mask into the corresponding label region of the semicircular canal in the initial subregion label map to obtain a merged subregion label map; re-perform centerline extraction on the corresponding label region of the semicircular canal in the merged subregion label map to obtain a skeleton set to be reviewed; perform abnormal connection removal based on the skeleton set to be reviewed to obtain a temporary corrected label map; perform a post-correction review on the temporary corrected label map; after the review is passed, output the corrected subregion label map and the corresponding correction marker map.

[0008] According to a second aspect of this embodiment, an inner ear subregion segmentation control system for correcting fractures in the slender structure of the semicircular canal is provided, comprising an endpoint pair determination unit, acquiring a three-dimensional inner ear magnetic resonance image, the triaxial voxel spacing corresponding to the three-dimensional inner ear magnetic resonance image, a subregion probability map and an initial subregion label map output by the front-end segmentation; extracting candidate regions of the semicircular canal to be controlled from the initial subregion label map; performing centerline extraction on the candidate regions of the semicircular canal to obtain a corresponding skeleton set; identifying a skeleton endpoint set based on the skeleton set; filtering endpoint pairs that need to be automatically completed from the skeleton endpoint set; and a completion unit for automatically completing endpoint pairs. The automatic completion of endpoint pairs generates a center completion path; and after obtaining the center completion path, a tubular completion mask is generated based on the local diameter of the endpoints; the tubular completion mask is merged into the corresponding label region of the semicircular tube in the initial subregion label map to obtain a merged subregion label map; the centerline extraction is re-performed on the corresponding label region of the semicircular tube in the merged subregion label map to obtain a skeleton set to be reviewed; abnormal connections are removed based on the skeleton set to be reviewed to obtain a temporary corrected label map; a review unit is used to perform a review on the temporary corrected label map after correction; after the review is passed, the corrected subregion label map and the corresponding corrected label map are output.

[0009] This embodiment presents an inner ear subregion segmentation control method for correcting fractures in the slender structure of the semicircular canal. The method includes acquiring a three-dimensional inner ear magnetic resonance image, the triaxial voxel spacing corresponding to the three-dimensional inner ear magnetic resonance image, a subregion probability map output by the front-end segmentation, and an initial subregion label map; extracting candidate regions of the semicircular canal to be controlled from the initial subregion label map; performing centerline extraction on the candidate regions of the semicircular canal to obtain a corresponding skeleton set; identifying a skeleton endpoint set based on the skeleton set; filtering endpoint pairs that need to be automatically completed from the skeleton endpoint set; and generating... A central completion path is generated; and after obtaining the central completion path, a tubular completion mask is generated based on the local diameter of the endpoints; the tubular completion mask is merged into the corresponding label region of the semicircular canal in the initial subregion label map to obtain a merged subregion label map; the centerline extraction is re-performed on the corresponding label region of the semicircular canal in the merged subregion label map to obtain a skeleton set to be verified; abnormal connections are removed based on the skeleton set to be verified to obtain a temporary corrected label map; the temporary corrected label map is verified after correction; after verification, the corrected subregion label map and the corresponding corrected label map are output. Based on the existing inner ear subregion segmentation results, retraining the model is not a necessary condition, but rather a closed-loop processing flow of "detection—scoring—correction—verification—output" is formed. It can conditionally correct local breaks and abnormal connections of the semicircular canal while preserving the overall shape of the original segmentation, making the segmentation results not only acceptable at the voxel overlap level, but also more stable in terms of the continuity and anatomical rationality of the semicircular canal structure. Attached Figure Description

[0010] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the method according to an embodiment of the present invention; Figure 2- Figure 3 This is a schematic diagram illustrating the application of the method in an embodiment of the present invention; Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the solutions of this embodiment, the technical solutions of this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this embodiment, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this embodiment.

[0013] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this embodiment are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in this embodiment can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0015] According to this embodiment, a method for inner ear subregion segmentation control to correct fractures in the slender structure of the semicircular canals is provided, such as... Figure 1 As shown, steps 101 to 103 are included below: Step 101: Acquire a three-dimensional inner ear magnetic resonance image, the triaxial voxel spacing corresponding to the three-dimensional inner ear magnetic resonance image, the subregion probability map and the initial subregion label map output by the front-end segmentation; extract the semicircular canal candidate regions to be quality controlled from the initial subregion label map; perform centerline extraction on the semicircular canal candidate regions to obtain the corresponding skeleton set; identify the skeleton endpoint set based on the skeleton set; filter out the fracture candidate endpoint pairs from the skeleton endpoint set; determine the endpoint pairs that need to be automatically completed from the fracture candidate endpoint pairs.

[0016] This method, applied to the automatic segmentation of medical images, performs structural consistency checks, local break completion, abnormal connection removal, and result verification on the obtained inner ear subregion segmentation results. This method does not require the front-end to use a fixed segmentation model; the front-end segmentation results can come from 3D Transformer segmentation networks, 3D U-Net networks, cascaded convolutional networks, or traditional image segmentation algorithms, or from manual coarse annotation or semi-automatic segmentation results. As long as the front-end can output label maps of the inner ear subregions, or probability maps corresponding to each subregion, this embodiment can be executed as a subsequent quality control module. Existing technologies can already achieve automatic segmentation of inner ear subregions based on 3DT2-weighted MRI, dividing the inner ear into the basal cochlear turn, the mid-parietal cochlear turn, and regions containing the vestibule and semicircular canals. However, the main improvements are still focused on segmentation model training and label-preserving data augmentation. This embodiment further addresses the problem of local discontinuities, missing segments, or abnormal adhesions in the segmentation results of slender and curved structures such as the semicircular canals by establishing an executable structural quality control and correction process after model output.

[0017] The first type of input received is a three-dimensional inner ear magnetic resonance image, denoted as... The image can be a 3DT2-weighted image, a heavily T2-weighted image, a CISS sequence, a FIESTA sequence, or other three-dimensional images capable of displaying the fluid-filled structures of the inner ear. The image can be at its original acquisition resolution or can be an image that has undergone resampling, registration, cropping, or intensity normalization. To ensure that subsequent spatial distance calculations have practical physical meaning, this embodiment simultaneously reads the voxel spacing during image acquisition, denoted as . ,in , , These represent the voxel dimensions of the image in the three coordinate directions, typically in millimeters. If the image resolution is significantly lower in the layer thickness direction than the in-plane resolution, for example... Greater than or If the value is more than twice that of the image, this embodiment can first perform isotropic resampling on the local processing area, or it can remain unchanged by introducing voxel spacing weights in distance, path length, curvature, and neighborhood search. This can avoid mistaking a voxel adjacent in the layer thickness direction as having the same spatial distance as an adjacent voxel in the plane, thereby reducing false breakage detection and incorrect completion.

[0018] The second type of input received in this embodiment is the sub-region probability map output by the front-end segmentation module, denoted as... ,in Represents the background probability map. to These represent probability diagrams for different inner ear subregions. This represents the number of foreground categories. If the front-end model outputs a three-category classification based on the basal cochlea, mid-parietal cochlea, and vestibular and semicircular canal categories, then... It can be 3; if the front-end model further distinguishes between different turns of the anterior semicircular canal, posterior semicircular canal, lateral semicircular canal, vestibule, and cochlea, then The value can be greater than 3. The voxel values ​​in each probability map typically range from 0 to 1, representing the confidence level of the voxel belonging to the corresponding category. If the front-end algorithm cannot output a probability map but only a hard label map, this embodiment can assign a value of 1 to the inside of the corresponding label region and a value of 0 to the outside, forming an approximate probability map; alternatively, it can use only the label map and the original image grayscale to perform subsequent judgments. When a probability map is available, it is preferred to use it for fracture path scoring because although the semicircular canal fracture may not be selected by the hard label, there may still be a low but continuous foreground response in the probability map. This information is of practical value in determining whether completion is necessary.

[0019] The third type of input received in this embodiment is the initial subregion label map, denoted as... .exist In this model, each voxel has a specific label, such as background, basal cochlear torsion, mid-apical cochlear torsion, vestibule, and semicircular canals, or even finer semicircular canal labels. If the front-end segmentation model only outputs a probability map without hard labels, this embodiment can obtain an initial label map using the maximum probability principle, i.e., selecting the category with the highest probability value for each voxel as its label; alternatively, different thresholds can be set for different categories, such as using a lower threshold for the slender semicircular canal category to reduce the direct deletion of weakly responding structures. For cases where there is significant small noise in the initial label map, this embodiment can first perform lightweight noise removal, but this removal is not considered a final quality control conclusion; it is only used to reduce computational interference during subsequent scaffolding. The scope of noise removal should be limited, for example, only deleting isolated voxel clusters with a volume of less than 3 to 20 voxels and a distance from the main inner ear structures exceeding a preset range, to avoid mistakenly deleting actual semicircular canal terminal segments.

[0020] After obtaining the input data, this embodiment first determines the candidate region of the semicircular canal from the initial label map, denoted as... If the front-end label image already contains independent labels for the anterior, posterior, and lateral semicircular canals, then the corresponding label area is directly used as the candidate area for quality control of the semicircular canals. The three semicircular canals can be processed separately before overall consistency verification. If the front-end label image only contains the merged area of ​​the vestibule and semicircular canals, this embodiment first distinguishes the vestibular body and the slender branches of the semicircular canals within this merged area. Specifically, a three-dimensional distance transformation can be performed on the merged area to obtain the shortest distance from each foreground voxel to the background boundary, denoted as... .in To represent a voxel position, A higher value indicates that the voxel is closer to the center of a larger structure; a lower value indicates that the voxel is more likely to be located near a thin wall, a narrow tube, or a boundary. Semicircular canals typically exhibit a small diameter, a long centerline, and a curved extension along space, while the vestibular body typically exhibits a relatively enlarged cavity structure. Therefore, this embodiment can satisfy... Furthermore, the voxels located on the peripheral extension branches of the vestibular main body were identified as candidate voxels for the semicircular canals, among which... The tube diameter screening threshold can be set from 0.3 mm to 1.5 mm, preferably from 0.4 mm to 1.0 mm. If the image resolution is low or the layer thickness is large, this threshold can be appropriately relaxed, but subsequent grayscale support or probability support requirements should be increased at the same time to prevent noise from being mistaken for semicircular tubes.

[0021] In another implementation, this embodiment can first extract the centerline of the merging region of the vestibule and semicircular canals, and then filter candidate branches of the semicircular canals based on the geometric properties of the centerline branches. After centerline extraction, the vestibular body typically corresponds to a short and complex central region, while the semicircular canals correspond to slender branches extending from the vicinity of the vestibule. For each branch, this embodiment calculates its centerline length, average local radius, curvature change, and connection position with the vestibular body. If the centerline length of a branch is greater than a preset length threshold... If the average local radius of the branch is less than a preset proportion of the average radius of the vestibular body, for example, less than 0.6 or 0.7 times, then the branch and its neighboring voxels are identified as candidate regions for the semicircular canals. The length threshold... The value can be set from 2mm to 10mm, preferably from 3mm to 6mm; the average local radius can be estimated by the distance transformation value or by the cross-sectional area perpendicular to the centerline. If the image space is small and only covers part of the inner ear region, this embodiment allows the use of a relative threshold, that is, using the maximum path length of the ipsilateral inner ear foreground structure or the main radius of the vestibular body as a reference to automatically adjust the semicircular canal candidate determination conditions.

[0022] If the candidate region for the semicircular canals is empty, or the volume of the candidate region is significantly lower than a reasonable range, this embodiment does not directly generate the semicircular canal structure. Instead, it outputs a quality flag indicating "insufficient candidate region for the semicircular canals" and enters a conservative processing mode. The reasonable range can be set based on the overall foreground volume of the ipsilateral inner ear. For example, when the volume of the candidate region for the semicircular canals is smaller than the foreground volume of the ipsilateral inner ear... to If the front-end segmentation result is deemed insufficient to support the semicircular canal, this embodiment can search for weak-response elongated regions in the probability map. However, local completion is only performed when the weak-response region simultaneously satisfies continuity, grayscale support, and spatial location constraints; otherwise, the original result is retained and manual review is prompted. This setting aims to prevent the algorithm from fabricating non-existent or unverifiable structures based on a single morphological prior in cases of severe low contrast, motion artifacts, or insufficient scanning range.

[0023] Identifying candidate regions of the semicircular canals Subsequently, this embodiment performs three-dimensional connected component analysis. Connected components can be defined using 6-neighborhood, 18-neighborhood, or 26-neighborhood. For slender structures like the semicircular canal with complex spatial orientation and potentially oblique traversal, 26-neighborhood is preferred as the initial connected component definition to reduce the possibility of incorrectly splitting true oblique connections. The connected component analysis yields several connected components, denoted as... This embodiment does not employ a simple strategy of retaining only the largest connected component, because the actual semicircular canal segment after a break may be small in volume. If only the largest connected component is retained, the distal segment of the break may be mistakenly deleted, making subsequent reconstruction impossible. Instead, this embodiment calculates the volume, centerline length, average local radius, distance to the vestibular main body, closest distance to the main semicircular canal region, and endpoint orientation for each connected component, and then determines whether it belongs to the main structure, a suspected broken segment, isolated noise, or part of an abnormal connection.

[0024] For volume smaller And the length of the centerline is less than A connected component is labeled as an isolated false positive fragment if the distance between this component and any endpoint of the main structure is greater than the maximum connectable distance, or if its image grayscale and probability support are both below a preset threshold. It can be set to 3 to 50 voxels, or it can be set to the volume of the ipsilateral semicircular canal candidate region. to ; The value can be set from 0.3mm to 2mm. If a small component, although small in size, is located near the endpoint of the main path and its local direction is consistent with the extension direction of the main path, this embodiment does not immediately delete the component, but retains it as a suspected distal fragment of a fracture, and proceeds to the subsequent endpoint pairing process. This process can suppress noise and avoid mistakenly deleting genuine semicircular canal fragments.

[0025] Subsequently, in this embodiment, centerline extraction is performed on the candidate regions of the semicircular canals to obtain the framework set. Centerline extraction can employ 3D topology refinement algorithms, distance-transform ridge extraction algorithms, shortest path-based centerline extraction algorithms, or graph search-based centerline extraction algorithms. The purpose of centerline extraction is not to alter the original segmentation, but rather to simplify the tubular structure into easily identifiable spatial paths, making fracture endpoints, anomalous bifurcations, and pseudo-connection bridges easier to identify. Before scaffolding, a small-scale hole-filling or surface smoothing process can be performed on the candidate region, for example, filling small holes with a diameter of no more than 1 to 2 voxels. However, large-scale dilation should be avoided to prevent obscuring the true fracture location.

[0026] After obtaining the frame, this embodiment identifies the set of frame endpoints, denoted as... An endpoint is defined as a scaffold point in a scaffolding graph that has only one adjacent scaffold point. If multiple adjacent endpoints appear in a local area, for example, if scaffolding noise causes several short, spiky endpoints near a fracture end, spatial clustering of the endpoints can be performed first. Endpoints with a distance of less than 1 to 3 voxels or 0.2 mm to 0.8 mm can be merged into an endpoint cluster, and the geometric center or the point with the highest probability of the endpoint cluster can be used as the representative endpoint. This can reduce the repeated counting of the same fracture location and also reduce the number of subsequent endpoint pairing combinations.

[0027] As an optional implementation of this embodiment, the process of selecting endpoint pairs to be automatically completed from the skeleton endpoint set includes: for any two endpoints in the skeleton endpoint set, determining them as the first endpoint and the second endpoint respectively, and determining the endpoint pair composed of the first endpoint and the second endpoint as a candidate endpoint pair for breakage; extracting a continuous skeleton segment of a preset real spatial arc length from each endpoint along its respective skeleton branch into the candidate region of the semicircular canal as the endpoint neighborhood of the corresponding endpoint; determining the local direction vector of the corresponding endpoint based on the spatial coordinates of the skeleton voxels in the neighborhood of each endpoint, and generating a scoring candidate path connecting the two endpoints using the first endpoint and the second endpoint as path endpoints and the local direction vectors of the two endpoints as direction constraints. Distance fit is determined based on the actual spatial distance between the two endpoints. The consistency of direction is determined based on the angle between the local direction vectors of the two endpoints and the direction of the line connecting the two endpoints. ; along the scoring candidate path Perform path sampling and calculate path image support according to the following formula. ,in, For the scoring candidate path The number of path sampling points on the path voxels in the subregion probability map The probability value of belonging to the semicircular canal or vestibular semicircular canal category. For voxels The fluid structure support value is obtained by normalizing the grayscale of the original magnetic resonance image. To predetermine the fusion weights, a three-dimensional Euclidean distance transformation considering the three-axis voxel spacing is performed on the binary mask of the semicircular canal candidate region to obtain a distance transformation map. The distance transformation value of any foreground voxel in the distance transformation map is the shortest true spatial distance from that foreground voxel to the boundary of the semicircular canal candidate region. The median of the distance transformation values ​​corresponding to the skeleton voxels in the endpoint neighborhoods of the first and second endpoints is taken to obtain the local radius of the first endpoint. Local radius of the second endpoint And calculate local radius consistency According to the scoring candidate path The number of path sampling points falling into the preset uncoverable area, accounting for a certain percentage of the scoring candidate paths. The proportion of the total number of path sampling points determines the boundary violation penalty. The preset non-coverable regions include the cochlear high-confidence label region, the contralateral inner ear label region, the background high-confidence region, and other prohibited completion regions; the broken connectivity score of the candidate broken endpoint pairs is calculated. ,in, to Preset non-negative weights; score broken connections. Candidate endpoint pairs with broken connections that are not lower than a preset connection threshold are identified as endpoint pairs that need to be automatically completed.

[0028] Once the set of endpoints is determined, for any two endpoints and If the two belong to different connected components, or belong to the same connected component but are located on opposite sides of an abnormally interrupted region, then the actual spatial distance between them is calculated. The distance calculation formula is as follows: ,in, and Representing the endpoints respectively and endpoints Coordinates in a voxel coordinate system , , These represent the voxel spacing in the three coordinate directions. The function of this formula is to convert the voxel coordinate differences into actual physical distances. For isotropic images, this formula is approximately equivalent to ordinary three-dimensional Euclidean distance; for images with large slice thickness, this formula ensures that the spacing in the slice thickness direction is calculated as actual millimeter distances, thus avoiding an underestimation of the fracture distance. Greater than the maximum completion distance If the endpoint does not enter the completion process, then the endpoint will not be included. It can be set from 0.5mm to 5mm, preferably 1mm to... For low-resolution images, the distance can be relaxed to 6mm, but higher probability support or grayscale support conditions must be met simultaneously. If the endpoint distance is less than the minimum distance threshold, for example, less than 0.1mm to 0.3mm, it may be a local burr formed by framing rather than a real break. In this embodiment, smoothing or burr removal can be prioritized instead of path completion.

[0029] For a candidate endpoint pair, generate one or more candidate connection paths between them, denoted as . The path can be a straight line, a 3D B-spline curve path, a shortest path based on a probability graph, or a minimum-cost path based on the joint cost of the original grayscale and probability. Imagery support is used to determine whether there is image or probabilistic evidence that can be interpreted as a semicircular canal fluid structure between two endpoints, for each voxel on the path. Read the probability of the semicircular canal or vestibular semicircular canal category of this voxel. And read the liquid structure support values ​​after grayscale normalization of the original image. Path imagery support can be written as: ,in, Indicates the number of voxels along the path. Voxel representation The probability of belonging to the semicircular canal or vestibular semicircular canal category. This represents the fluid structure support value obtained after normalizing the original MRI grayscale. The weight between the probability map and the original grayscale is 0 to 1, preferably 0.5 to 0.8. The engineering significance of this formula is that if, although the two endpoints are not connected by hard labels, the probability map still shows a continuous weak foreground response, or if the original image contains a high-signal pathway similar to the fluid-filled structure of the inner ear, then... The relatively high probability indicates that the break is more likely caused by a missed detection in the actual segmentation. If the front-end model cannot output a probability map, then... Only the original image grayscale is used for support; if the original image has strong noise or low contrast, it can be improved. It relies more on probability plots; if the overall confidence of the probability plot is too low, then reduce... This is then marked as a low-confidence correction in the quality report.

[0030] When calculating the consistency of endpoint direction, let's assume that from the endpoint... Point to endpoint The unit vector is endpoints and The local direction vectors are respectively and Then, directional consistency can be expressed as Where “⋅” represents the vector dot product, The value of is between 0 and 1. The meaning of this formula is: if the endpoints... The local extension direction roughly points to the endpoint. ,but It is positive and the value is large; if the endpoints The local extension direction roughly points back to the endpoint. ,but The value is positive and relatively large. When both conditions are met simultaneously, it indicates that the two endpoints are more like the two ends of a broken, slender tubular structure. This is used in the formula. This is to avoid negative scores canceling out other information when directions are opposite. If the confidence of a certain endpoint direction is low, the direction item related to that endpoint can be set to a neutral value, such as 0.5. If both endpoint directions are unreliable, then direction consistency is not used as the main criterion, and a more stringent distance threshold and path image support are needed for judgment.

[0031] To avoid abnormally large completion paths or connections to structures that should not be connected, this embodiment also calculates local radius consistency and out-of-bounds penalties. Local radius consistency The radius can be estimated from the distance transformation values ​​near the two endpoints and the radius within the neighborhood of the candidate path. If the radii near the two endpoints are close, and the allowed radius on the path is consistent with the endpoint radius, then... Higher; if the radius of one endpoint is significantly larger than that of the other endpoint, or if the path needs to pass through a significantly enlarged region to connect, then Lower. Penalty for crossing the boundary. This indicates whether the path crosses non-target areas, such as cochlear tags, contralateral inner ear tags, high-confidence background regions, areas outside the male boundary, or other prohibited sub-regions. If the path lies entirely within a reasonable search window and does not cross prohibited areas, then... Lower; if the path crosses conflicting labels or goes beyond the reasonable range of the ipsilateral inner ear, then Relatively high.

[0032] Taking into account the above factors, this embodiment calculates a break connection score for each candidate endpoint pair: in, Indicates endpoint pair and The overall rating was automatically completed; , , , , These represent the weights of distance, orientation consistency, image support, radius consistency, and out-of-bounds penalty, respectively. Distance term Indicates two The closer the endpoints are, the higher the distance score; when achieve At that time, the value drops to 0. Indicates whether the endpoints are facing each other; Indicates whether an image or probability support exists on the path; This indicates whether the completed path matches the two ends of the break in terms of pipe diameter; This is a negative term used to penalize paths that traverse non-target structures or unreasonable regions. The weights can be adjusted based on the data source and image quality, typically with distance as a key factor. The value can be between 0.2 and 0.4, for directional weights. A value of 0.2 to 0.4 is acceptable; image support weights are also possible. The value can be between 0.2 and 0.5, representing the radius consistency weight. The value can be between 0.05 and 0.2, representing the weight for out-of-bounds penalties. A value of 0.2 to 0.6 is acceptable. If... Greater than the connection threshold If so, the endpoint pair is considered a candidate for correctable fracture. The value can be 0.55 to 0.85, preferably 0.65 to 0.75. If multiple endpoint pairs share the same endpoint, the pair with the highest score is selected first. If the difference between the highest score and the second highest score is less than 0.05 to 0.15, it indicates that there is ambiguity in the connection relationship. In this embodiment, the connection is not automatically completed, but the suspected break position is marked in the correction mark diagram and manual review is recommended.

[0033] As an optional implementation of this embodiment, for the endpoint pairs that need to be automatically completed, a bounding box corresponding to the line connecting the two endpoints is generated, and the bounding box is expanded outward to obtain a search window. Spatial constraints are applied to the search window based on the ipsilateral inner ear tag region and a preset non-coverable region, ensuring that the search window is located within a preset neighborhood of the ipsilateral inner ear structure, and prohibiting path searches from crossing into the contralateral inner ear tag region, the cochlear high-confidence tag region, the background high-confidence region, or other preset non-coverable regions. Paths within the search window consisting of sequentially adjacent voxels, with their start and end points being the two endpoints that need to be automatically completed, are determined as paths to be evaluated. Wherein, the path to be evaluated is represented as ; Calculate any voxel in the path to be evaluated Basic toll cost ; Calculate the path to be evaluated Cumulative toll costs ,in, voxels in the subregion probability map The probability value of belonging to the semicircular canal or vestibular semicircular canal category; For voxels The support value of the liquid structure is obtained by normalizing the grayscale of the original magnetic resonance image. For voxel boundary risk penalty, when voxel When within the allowed search area, Take the first preset value; when voxels When located in the boundary buffer zone between the allowed search area and the preset uncoverable area, Take a second preset value that is greater than the first preset value; when the voxel When located in a pre-defined uncoverable area, voxels will be... Set as an unusable voxel; To determine based on adjacent path segments and The local curvature penalty is determined by the change in direction between paths; the greater the change in path direction, the greater the local curvature penalty. to Preset non-negative weights; select the cumulative passage cost from the paths to be evaluated through minimum cost path search. The shortest path is used as the center completion path.

[0034] For endpoint pairs identified as correctable breaks, this embodiment generates a completion path within a local search window. The local search window can be obtained by expanding outwards from the bounding box of the endpoint pair connection, with the expansion range set to 1 to 8 voxels, or 0.5 mm to 3 mm. The search window should be limited to a reasonable neighborhood of the ipsilateral inner ear structure, and should not cross into the contralateral inner ear or obviously non-target areas. Path searching is not preferred for simple straight-line connections because the semicircular canals themselves are curved structures, and straight-line completion may traverse unreasonable tissues or create local sharp edges. Therefore, this embodiment preferably uses a minimum-cost path search. For each voxel within the search window... Set path cost: in, Indicates the path passes through voxels The lower the cost, the more preferentially the path passes through that element. This provides probabilistic support for the semicircular canal or vestibular semicircular canal categories in the front-end model; the higher the probability, the better. The lower the value, the easier it is for the path to pass through that voxel. The closer the image's grayscale is to the fluid-like structure of the inner ear, the better. The lower For boundary or boundary crossing penalties, if the voxel is located in the cochlea, the high-confidence background restricted area, the contralateral structure, or other prohibited areas, then... Take the larger value; if the voxel is within the allowed search area, take the smaller value or 0. Curvature penalty is used to limit sudden path reversals or excessively small turning radii. to These are weighting coefficients, which can be adjusted based on the presence or absence of a probability plot, image contrast, and segmentation confidence. If there is no probability plot, remove these coefficients. Item or order If the original image's grayscale is significantly affected by noise, then reduce the grayscale value. If the completed path is close to the cochlea or other easily confused structures, it will improve... This makes the boundary violation penalty stronger. The purpose of this formula is to prioritize the completion path through areas where the model still has a weak response, the original image still has structural support, and the spatial orientation is reasonable, rather than drawing a connecting line unconditionally between the two endpoints.

[0035] As an optional implementation of this embodiment, generating a tubular completion mask based on the local diameter of the endpoints after obtaining the central completion path includes: performing a three-dimensional Euclidean distance transformation considering the triaxial voxel spacing on the binary mask of the semicircular tube candidate region to obtain a distance transformation map; extracting continuous skeleton segments of a preset real spatial arc length from the two endpoints to be automatically completed along the corresponding skeleton branches to the semicircular tube candidate region as the endpoint neighborhoods of the corresponding endpoints; and taking the median of the distance transformation values ​​corresponding to the skeleton voxels in the endpoint neighborhoods of the two endpoints to obtain the local radius of the first endpoint. Local radius of the second endpoint ; Calculate the ratio of the local radii of the two endpoints. When the radius ratio When the radius is not greater than a preset radius consistency threshold, the local radius of the first endpoint is adjusted along the center completion path from the first endpoint to the second endpoint. Local radius of the second endpoint Perform linear interpolation to obtain the local completion radius corresponding to each path position on the center completion path; when the radius ratio When the radius is greater than the preset radius consistency threshold, the local radius of the first endpoint is adjusted. Local radius of the second endpoint The smaller value in the equation is determined as the local completion radius corresponding to each path position on the central completion path; the local completion radius is restricted to a preset real space radius; the local tangent vector of each path position is determined according to the adjacent path points on the central completion path, and the real space circular filling region is determined according to the local completion radius of the corresponding path position in the cross section perpendicular to the local tangent vector; the real space circular filling region is mapped to the voxel coordinate system according to the three-axis voxel spacing, and the union of the filling regions corresponding to each path position on the central completion path is taken to obtain the tubular completion mask.

[0036] In this optional implementation, after obtaining the central completion path, a tubular completion mask is generated based on the local radius near the break endpoints. Specifically, distance transformation values ​​are read within the neighborhood of the two endpoints, and their mean, median, or smaller value is taken as the completion radius. and limit it to to Within the range. 0.2mm to , Desirable to If the difference in radii between the two endpoints is small, the average of the two can be used; if the difference exceeds a preset multiple, such as more than twice, the smaller radius should be used first, or linear interpolation should be performed along the path from one end to the other to avoid abnormal expansion of the completed region in the middle. When generating the completed mask, the center path can be used as the axis to fill the neighborhood with voxels that meet the radius conditions in the local direction perpendicular to the path; alternatively, the center path can be directly expanded into a three-dimensional spherical or ellipsoidal shape, but the expansion kernel should be anisotropically corrected according to the actual voxel spacing, and a fixed voxel radius cannot be simply used.

[0037] As an optional implementation of this embodiment, merging the tubular completion mask into the semicircular canal corresponding label region of the initial subregion label map includes: identifying voxels in the tubular completion mask that overlap with the cochlear label region, the contralateral inner ear label region, the background high-confidence region, or other preset non-coverable regions as conflict voxels; calculating the ratio of the real space volume of the conflict voxels to the total real space volume of the tubular completion mask based on the triaxial voxel spacing to obtain the conflict voxel ratio; when the conflict voxel ratio is not lower than a preset conflict ratio threshold, determining that the center completion path does not meet the preset label constraint, abandoning the automatic completion of the corresponding endpoint pair, and recording the endpoint pair as a suspected broken endpoint pair; when the conflict voxel ratio is lower than the preset conflict ratio threshold, deleting the conflict voxel from the tubular completion mask to obtain the remaining tubular completion mask; determining that the Whether the remaining tubular completion mask is connected to the semicircular canal label areas where the two endpoints to be automatically completed are respectively; when the remaining tubular completion mask is connected to the semicircular canal label areas where the two endpoints are respectively, the remaining tubular completion mask is merged into the corresponding semicircular canal label area of ​​the initial subregion label map, and the label conflict record is stored; when the remaining tubular completion mask is not connected to the semicircular canal label areas where the two endpoints are respectively, the automatic completion of the corresponding endpoint pair is abandoned, and the endpoint pair is recorded as a suspected broken endpoint pair; the newly added connection relationship is checked on the merged semicircular canal label area; when the remaining tubular completion mask, outside the preset allowed connection area, causes the semicircular canal label area to form a new connection relationship with the cochlear label area, the vestibular main area or other designated branches, the automatic completion result of the corresponding endpoint pair is rolled back.

[0038] In this optional implementation, after the completion mask is generated, this embodiment performs a tag conflict check. If the completion mask overlaps with the cochlear tag, the contralateral inner ear tag, a high-confidence background region, or other uncoverable regions, the conflicting voxels are deleted. If the proportion of conflicting voxels to the total volume of the completion mask exceeds a preset upper limit, for example... to If the completion path is deemed unreasonable, this embodiment abandons automatic completion and records the endpoint pair as a suspected break. If the proportion of conflicting voxels is low, for example less than... to If a conflicting voxel is removed, the remaining completion mask can be retained, but the occurrence of a label conflict must be recorded in the quality report. If the completion path causes new abnormal adhesions between the semicircular canals and the cochlea, vestibular body, or other branches, this embodiment will roll back the completion during subsequent review.

[0039] For example, refer to Figure 2a Schematic diagram of semicircular canal structural errors in the inner ear subregion segmentation results. (Reference) Figure 2bA schematic diagram of a break completion mechanism based on centerline and endpoint scoring.

[0040] As an optional implementation of this embodiment, the abnormal connection elimination based on the skeleton set includes: determining the skeleton endpoints and skeleton bifurcation points in the skeleton set to be reviewed as skeleton graph nodes, and determining the continuous skeleton voxel sequence between two adjacent skeleton graph nodes as skeleton branches; for any skeleton branch connected to a skeleton bifurcation point, if the bifurcation degree of the skeleton bifurcation point exceeds a preset bifurcation degree upper limit, and the skeleton length of the skeleton branch is less than a preset abnormal branch length threshold, then the skeleton branch is determined as a first type of abnormal connection candidate branch; skeleton branches whose two end nodes are both skeleton bifurcation points and whose skeleton length is less than a preset bridging length threshold are determined as short bridge candidate branches, wherein, for the short bridge candidate branch, other skeleton branches connected to each of its end nodes but not belonging to the short bridge candidate branch are determined as adjacent branches; when the same end node is connected to multiple adjacent branches, the adjacent branch with the largest skeleton length is determined as the adjacent reference branch corresponding to that end node; for any skeleton branch Calculate the skeleton branches Average category confidence ,in, For skeleton branches The number of skeletal voxels included. The skeleton voxels in the subregion probability map The probability value of belonging to the semicircular canal or vestibular semicircular canal category; the median of the distance transformation values ​​of the skeletal voxels contained in the skeletal branch in the distance transformation map to be verified is taken as the skeletal branch. Local radius of branch ; Calculate the skeleton branches Average liquid structure grayscale support ,in, For skeleton voxels The liquid structure support value is obtained by normalizing the grayscale of the original magnetic resonance image. When the average class confidence of the short bridge candidate branch is lower than the product of the smaller of the average class confidence of the two adjacent reference branches and a preset confidence ratio threshold, the local radius of the short bridge candidate branch is lower than the product of the smaller of the local radius of the two adjacent reference branches and a preset radius ratio threshold, and the average liquid structure grayscale support of the short bridge candidate branch is lower than the product of the smaller of the average liquid structure grayscale support of the two adjacent reference branches and a preset grayscale support ratio threshold, the short bridge candidate branch is identified as a second type of abnormal connection candidate branch. In the label region corresponding to the first type of abnormal connection candidate branch and the second type of abnormal connection candidate branch, voxels with class probability values ​​lower than a preset class probability threshold and liquid structure grayscale support values ​​lower than a preset grayscale support threshold are identified as voxels to be deleted. The voxels to be deleted are then deleted. After deletion, the maximum connected components in the corresponding label regions of the semicircular canals before and after deletion are determined, and the ratio of the skeleton length of the maximum connected component after deletion to the skeleton length of the maximum connected component before deletion is calculated to obtain the main structure retention rate. Connected components that appear after deletion but do not belong to the maximum connected component after deletion are identified as newly added isolated connected components, and the ratio of the sum of the skeleton lengths of the newly added isolated connected components to the skeleton length of the maximum connected component before deletion is calculated to obtain the newly added breakage ratio. When the main structure retention rate is not lower than the preset main structure retention threshold, the newly added breakage ratio is not higher than the preset breakage ratio threshold, and the increase in the number of skeleton endpoints after deletion relative to before deletion does not exceed the preset upper limit of endpoint increase, the deletion result is retained, and a temporary corrected label map is obtained. Otherwise, the deletion operation is rolled back, and the corresponding region is marked as a suspected abnormal connection that has not been automatically corrected.

[0041] In this optional implementation, abnormal connections refer to structures that should not be directly connected being linked together in the initial segmentation result due to probability overflow, local noise, or model misjudgment. Examples include a thin bridge-like connection between the semicircular canals and the cochlea, a short circuit between the anterior and posterior semicircular canals, adhesion between the semicircular canals and background noise segments, or incorrect connections between the left and right structures near the clipping edge. This embodiment identifies abnormal connections through scaffold bifurcation points, short bridge branches, and local confidence. If a bifurcation point appears in a scaffold region, and the number of bifurcations exceeds a reasonable range, or if a short branch connects two main branches but its average probability is significantly lower than that of the two main branches, its average radius is significantly smaller than that of the two main branches, and its grayscale support is insufficient, then this short branch is marked as an abnormal connection candidate. For abnormal connection candidates, this embodiment does not directly delete the entire connected region, but only deletes low-confidence voxels within the neighborhood of the short bridge centerline. The deletion range can be determined based on the distance transformation value of the short bridge region, or it can be set to a local neighborhood of 1 to 2 voxels. After deletion, this embodiment re-executes the connected component analysis. If the deletion restores the main semicircular canal structure from an abnormal short-circuit state to a reasonable branch and does not cause new large-scale fractures, the deletion result is retained. If the main structure is broken into multiple unreasonable small segments after deletion, or the number of endpoints increases abnormally, the deletion operation is rolled back, and the area is marked as a suspected abnormal connection that is not automatically corrected.

[0042] As an optional implementation of this embodiment, the post-correction review of the temporary corrected label map includes: recording the candidate regions of the semicircular canals before correction as... The modified semicircular canal region corresponding to the modified pre-modification semicircular canal candidate region in the temporary modified label map is denoted as... ; Determine the newly added and supplemented areas The newly added and completed regions are calculated based on the three-axis voxel spacing. and the modified anterior semicircular canal candidate region The actual spatial volume; and calculate the proportion of the completed volume. ,in, This indicates that the actual spatial volume of the corresponding region is calculated based on the triaxial voxel spacing; when the padded volume accounts for... When the percentage of the completed volume is not higher than the preset completion volume ratio threshold, the completion result of the corresponding endpoint pair is retained; when the completion volume ratio is... When the result exceeds the preset completion volume ratio threshold, roll back all completion results for the corresponding endpoint pair and output the manual review mark associated with the corresponding endpoint pair.

[0043] In this optional implementation, after completing the completion and removing abnormal connections, this embodiment generates a temporary corrected label graph, denoted as... and to Perform a review after correction. The purpose of the review is to limit the scope of automatic correction and prevent the quality control module from expanding local corrections into unfounded structural reconstructions. The review first checks the number of newly added filler elements. Let the newly added filler area be... The candidate region for the anterior semicircular canal was revised to... Then the completed volume percentage is ,in, This indicates the number of newly added complementary elements. This indicates the number of voxels in the candidate region of the anterior semicircular canal. This represents the proportion of volume change relative to the original candidate region of the semicircular canals. This formula is used to determine whether the correction still falls under the category of local fracture repair. If... Less than the preset upper limit, for example to This indicates that the amount of completion is relatively controllable; if If the upper limit is exceeded, it indicates that the algorithm may be generating structures over a large area. In this case, this embodiment reverts to all or part of the completion results and outputs a manual verification marker. For high-resolution images with good front-end segmentation quality, Upper limit preferred to For low-resolution images or images with significant missing segments in the initial semicircular canal segmentation, the restrictions can be relaxed to... to However, it also requires each completed path to... Above a stricter threshold, such as above 0.75.

[0044] The review of the temporary modified label map after modification also includes: reviewing the candidate regions of the semicircular canals before modification. and the modified semicircular canal region Perform connected component analysis; when the modified semicircular canal region The number of connected components is greater than the number of candidate regions for the modified anterior semicircular canal. The number of connected components, or the modified semicircular canal region. When a new connected component with a skeleton length less than a preset short segment length threshold exists, the newly added completion region is... Perform pipe diameter verification; for the corrected semicircular tube area The binary mask is subjected to a 3D Euclidean distance transformation considering the three-axis voxel spacing to obtain a corrected distance transformation map. The median of the distance transformation values ​​of the skeleton voxels in the neighborhood of each of the two endpoints to be automatically completed is taken to obtain the local radius of the first endpoint. Local radius of the second endpoint And calculate the average radius of the endpoints according to the following formula. The center completion path is read to be located in the newly added completion area. The maximum value of the distance transformation values ​​of each path point within the corrected distance transformation map is determined as the maximum local radius of the completed region. And calculate the radius expansion ratio according to the following formula. ; Determine the local tangent vector of each path position based on the adjacent path points on the center-completed path; For the center-completed path located in the newly added completion region For any path point within the region, calculate the newly added and completed region within the cross section perpendicular to the corresponding local tangent vector. Real spatial cross-sectional area The maximum value among the real spatial cross-sectional areas corresponding to each path point is determined as the maximum cross-sectional area of ​​the completed region. ; Calculate the reference cross-sectional area at the endpoints and cross-sectional area expansion ratio ; When the radius expansion ratio The expansion ratio is higher than the preset radius expansion threshold, or the cross-sectional area expansion ratio. When the cross-sectional area expansion exceeds the preset threshold, the normalized path arc length is used to complete the path along the center. Determine the target local radius at each path location. ,in, , To refine the local completion radius of the corresponding path position before refinement; using the center completion path as the axis, based on the target local radius... Regenerate a refined tubular completion mask, and replace the newly added completion area with the refined tubular completion mask. The refined and corrected semicircular canal region is re-executed with connected component analysis, short segment detection, radius expansion ratio calculation, and cross-sectional area expansion ratio calculation. If the refined and corrected semicircular canal region still has any of the following conditions: an increase in the number of connected components, a new connected component with a skeleton length less than the preset short segment length threshold, a radius expansion ratio higher than the preset radius expansion threshold, or a cross-sectional area expansion ratio higher than the preset cross-sectional area expansion threshold, all completion results for the corresponding endpoint pairs are rolled back, and a manual review mark is output. Otherwise, the refined completion results are retained.

[0045] As an optional implementation of this embodiment, the post-correction review of the temporary corrected label map further includes: resampling along the center completion path according to a preset real space step size with equal arc length to obtain a path point sequence; for any internal path point in the path point sequence... , respectively at the internal path points The first reference path point is selected for the preset path arc length positions before and after. ⁻ and second reference path point ⁺, Calculate vector ⁻− with vector ⁺− The local interior angle between When the local included angle When the angle is less than a preset smoothing angle threshold, the internal path point... Identify a candidate sharp turn path point; establish a preset real-world spatial neighborhood centered on the candidate sharp turn path point. The local image support corresponding to the candidate sharp turn path points is calculated according to the following formula. ,in, The number of voxels in the preset real-world neighborhood. voxels in the subregion probability map The larger of the corresponding semicircular canal category probability and vestibular category probability. For voxels The fluid structure support value is obtained by normalizing the grayscale of the original magnetic resonance image. The preset fusion weights are used; when the local image support corresponding to the candidate sharp turn path point is... If the local image support threshold is lower than the preset local image support threshold, it is determined that the center completion path does not meet the preset smooth shape condition at the candidate sharp turn path point, and the local completion area containing the candidate sharp turn path point is backed up; after backing up the local completion area, centerline extraction, equal arc length resampling, local interior angle calculation, and local image support calculation are re-executed on the remaining completion results; if there are still candidate sharp turn path points in the remaining completion results whose local interior angle is less than the preset smooth angle threshold and whose local image support is lower than the preset local image support threshold, then all completion results of the corresponding endpoint pair are backed up.

[0046] Under normal circumstances, after break completion, the number of connected components in the semicircular canal candidate region should decrease or remain reasonable, the number of isolated short segments should decrease, and the number of main path endpoints should not increase abnormally. If the number of connected components increases after correction, or multiple new segments with lengths less than [a certain value] are added, [failure to correct this will result in a problem]. The short segments indicate that the completion or removal operation may introduce fragmentation errors. In this embodiment, the corresponding operation should be reviewed and the local radius changes should be checked. If the local radius of the completed region exceeds 1.5 to 2.5 times the average radius of the neighborhood of both endpoints, or if the cross-sectional area of ​​the completed region is significantly larger than the cross-sectional area of ​​the original semicircular canal, the completion mask is considered too coarse and should be refined. If the conditions are still not met after refinement, the completion should be abandoned. The path curvature should be checked. If the completed path has an abnormally sharp angle, such as the angle between adjacent path segments being less than 1 / 2, the completion mask is considered too coarse and should be refined. to If there is insufficient image grayscale or probability to support the location, the path is considered not to conform to the basic shape of the semicircular canal's smooth curve, and a rollback or research is performed.

[0047] For all corrections that fail the review, this embodiment handles them according to the reason for failure. If the failure reason is that the completion volume is too large, the newly added completion elements are deleted, the original labels are retained, and the candidate break positions are recorded in the correction marker map. If the failure reason is that there are too many label conflicts, the completion elements that conflict with non-target labels are deleted; if the path is no longer continuous after deletion, the completion is abandoned. If the failure reason is that abnormal connection removal causes the main structure to break, the deleted area is restored. If the failure reason is insufficient image or probability support, automatic correction is not performed, and only a low confidence prompt is output. Through this review and rollback mechanism, this embodiment can improve the structural continuity while avoiding forcibly modifying uncertain areas into seemingly complete structures that lack image evidence.

[0048] Finally, this embodiment outputs a corrected inner ear subregion label map. .exist In the process, except for the confirmed semicircular canal partial completion regions and aberration connection removal regions, the labels of other subregions remain unchanged from the front-end segmentation results to avoid the quality control module affecting non-target regions such as the cochlea. In addition, this embodiment also outputs a correction markup map. The correction marker map is used to record the source and status of each processed voxel, such as newly added completed voxels, deleted voxels, voxels suspected of being broken but not automatically completed, voxels suspected of being abnormally connected but not automatically deleted, and low-confidence review areas. This marker map can be overlaid on the original MRI image, allowing doctors or annotators to quickly see the location of algorithm modifications. Furthermore, this embodiment outputs a structural quality report R, which includes at least the volume of the ipsilateral semicircular canal candidate region, the number of connected components, the number of isolated segments, the number of broken endpoints, the number of candidate endpoint pairs, the number of actual completed paths, the number of abnormal connections removed, the percentage change in volume before and after correction, the maximum completion distance, the completion score, whether label conflicts occurred, and whether manual review is recommended. The structural quality report can be output in text, table, JSON file, or DICOM structured report format. (See reference...) Figure 3 A schematic diagram showing the output of the results of abnormal connection removal and correction.

[0049] This embodiment also allows for various alternative implementations. Centerline extraction can employ a 3D thinning algorithm, or a fast traversal method, Dijkstra's shortest path method, A-star search method, or a centerline tracing method based on a distance field. Breakpoint pairing can be performed using explicit rule scoring or a lightweight classifier; however, even when using a classifier, its input should include at least several of the following: endpoint distance, orientation consistency, path image support, radius consistency, and out-of-bounds penalty, rather than simply providing correction results based on black-box features. Path completion can employ a minimum-cost path, or B-spline curve fitting, probability field-guided deformable paths, graph-cut-based local filling, or morphological reconstruction-based fine-tube completion. Abnormal connection removal can employ short-bridge identification, local topological pruning, or confidence-based branch deletion. None of these alternative methods change the core idea of ​​this embodiment: based on existing inner ear subregion segmentation results, constrained corrections are made to breaks, missing, and abnormal connections through semicircular canal structural continuity constraints, and a verification mechanism prevents over-correction.

[0050] This embodiment can also adjust the triggering conditions according to different application scenarios. In scientific research batch processing scenarios, a relatively aggressive automatic completion strategy can be set, such as lowering the connection threshold and allowing a higher completion volume ratio, to improve the semicircular canal structure integrity rate; in clinical auxiliary diagnosis scenarios, a conservative strategy can be adopted, such as increasing... ,reduce The upper limit is set, and all low-confidence completions are marked for manual review. For cases with poor image quality, incomplete scanning range, or obvious motion artifacts, this embodiment can perform only quality checks without automatic correction, i.e., outputting break locations, abnormal connection locations, and quality scores for manual judgment. For cases where the front-end model has already output multiple independent labels for the semicircular canals, this embodiment can perform endpoint matching and path completion for each semicircular canal separately, and additionally check whether there are unreasonable cross connections between different semicircular canals; for cases where the front-end model only outputs combined labels for the vestibule and semicircular canals, this embodiment can first extract long and thin branches, and then perform overall structural quality control.

[0051] Compared to methods that rely solely on the output of a front-end segmentation model, the key difference in this embodiment lies in the fact that front-end segmentation models typically focus on voxel classification, and the usability of their output is primarily evaluated using metrics such as Dice, IoU, and Hausdorff distance. However, these metrics are insufficient to fully reflect topological issues such as local fractures, missing short segments, or misconnected thin bridges in the semicircular canal. Existing technologies also emphasize that the semicircular canal is a thin and complex structure and enhance model segmentation performance in this region through label-preserving data augmentation, but this approach still falls within the training and model output stages. This embodiment, however, introduces a structural quality control layer after the output results, specifying "whether the semicircular canal is continuous, whether fracture locations can be supported by image completion, whether anomalous connections should be removed, and whether the corrected results still conform to local canal diameter and label boundary constraints" into executable processing rules. Therefore, this embodiment can be used in conjunction with different front-end segmentation models or deployed as an independent quality control module for existing inner ear segmentation software, possessing clear technical boundaries and scalability.

[0052] This embodiment receives a 3D inner ear MRI image to be processed, a front-end segmentation probability map, and an initial subregion label map; then, it extracts semicircular canal candidate regions from the initial subregion label map; subsequently, it performs 3D connected component analysis and centerline extraction on the semicircular canal candidate regions to identify isolated segments, fracture endpoints, abnormal bifurcation points, and suspected pseudo-connection regions; then, it scores endpoint pairs that may belong to both sides of the same semicircular canal fracture by combining endpoint spatial distance, local orientation consistency, image grayscale support, segmentation probability support, local canal diameter consistency, and anatomical boundary violation penalty; when the score meets the set conditions, it generates a constrained completion path and completion mask; for suspected abnormal connection regions, it performs elimination through local low-confidence bridging segment identification and rollback verification; finally, it reviews the corrected semicircular canal structure, confirming that the corrected volume, connectivity, canal diameter changes, and label conflicts are all within the allowable range, and outputs the corrected inner ear subregion label map, corrected marker map, and structural quality report. The above process does not simply involve expanding, eroding, or preserving the maximum connectivity of the segmentation results, but rather establishes a conditional judgment and correction mechanism based on the slender, tortuous, and continuous anatomical features of the semicircular canals.

[0053] This embodiment's technical solution revolves around existing inner ear subregion segmentation results. Instead of requiring model retraining, it uses the connectivity of the slender semicircular canal structure, local diameter, path orientation, image support, and label boundaries as constraints, forming a closed-loop processing flow of "detection—scoring—correction—verification—output." This solution can conditionally correct local breaks and anomalous connections in the semicircular canals while preserving the overall morphology of the original segmentation. This makes the segmentation results not only acceptable at the voxel overlap level but also more stable in terms of the continuity and anatomical rationality of the semicircular canal structure.

[0054] This embodiment enables structured detection and constrained completion of local fractures in the semicircular canals without retraining the front-end model. The completion process utilizes spatial distance, endpoint orientation, probability maps, original grayscale values, and canal diameter consistency simultaneously, providing a more comprehensive basis for correction. It can identify low-confidence aberrations between the semicircular canals and the cochlea, vestibule, or noise fragments, reducing the risk of false connections and adhesions. Quality control results are output through correction marking maps and structural quality reports, facilitating rapid review of abnormal areas by doctors or annotators.

[0055] This embodiment aims to address the issues of easy breakage, missing parts, and misconnections in the slender semicircular canal structure after automatic segmentation of the inner ear subregion, by providing a structural constraint quality control correction method. Specific objectives include: resolving the difficulty in consistently identifying semicircular canal fracture locations through connected component analysis, centerline extraction, and endpoint clustering; addressing the lack of image-based evidence for path completion through joint scoring of distance, direction, grayscale, probability, and radius; resolving the difficulty in accurately removing abnormal adhesions through low-confidence short bridge identification, bifurcation analysis, and rollback verification; and resolving the uncontrollable range of automatic correction through volume, connectivity, diameter, and label conflict verification.

[0056] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0057] According to this embodiment, an inner ear subregion segmentation control system for correcting fractures in the slender structure of the semicircular canal is also provided. This system includes an endpoint pair determination unit, which acquires a three-dimensional inner ear magnetic resonance image, the triaxial voxel spacing corresponding to the three-dimensional inner ear magnetic resonance image, a subregion probability map output by the front-end segmentation, and an initial subregion label map; extracts candidate regions of the semicircular canal to be controlled from the initial subregion label map; performs centerline extraction on the candidate regions of the semicircular canal to obtain a corresponding skeleton set; identifies a skeleton endpoint set based on the skeleton set; filters endpoint pairs that need to be automatically completed from the skeleton endpoint set; and a completion unit is used to automatically complete the pairs that need to be completed. The automatically completed endpoint pairs generate a center completion path; and after obtaining the center completion path, a tubular completion mask is generated based on the local diameter of the endpoints; the tubular completion mask is merged into the semicircular tube corresponding label region of the initial subregion label map to obtain a merged subregion label map; the centerline extraction is re-performed on the semicircular tube corresponding label region in the merged subregion label map to obtain a skeleton set to be reviewed; abnormal connections are removed based on the skeleton set to be reviewed to obtain a temporary corrected label map; a review unit is used to perform a review on the temporary corrected label map after correction; after the review is passed, the corrected subregion label map and the corresponding corrected label map are output.

[0058] According to an embodiment of this invention, this embodiment also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to implement the method described in any of the above embodiments.

[0059] According to an embodiment of this invention, a readable storage medium is also provided, which stores computer instructions that enable a computer to perform the methods described in any of the above embodiments.

[0060] According to an embodiment of this invention, this embodiment also provides a computer program product that, when executed by a processor, can implement the methods described in any of the above embodiments.

[0061] Figure 4 A schematic block diagram of an example electronic device 300 that can be used to implement this embodiment is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0062] like Figure 4 As shown, the electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0063] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0064] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as the object matching method. For example, in some embodiments, the object matching method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of the methods described above may be performed.

[0065] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0066] The program code used to implement the methods of this embodiment can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0067] In the context of this embodiment, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

Claims

1. A method for inner ear subregion segmentation control to correct fractures in the slender structure of the semicircular canals, characterized in that, include: Acquire a three-dimensional inner ear magnetic resonance image, the triaxial voxel spacing corresponding to the three-dimensional inner ear magnetic resonance image, the subregion probability map and the initial subregion label map output by the front-end segmentation; From the initial subregion label map, extract the candidate regions of the semicircular canals to be quality controlled; perform centerline extraction on the candidate regions of the semicircular canals to obtain the corresponding skeleton set; Based on the skeleton set, a skeleton endpoint set is identified; Filter the skeleton endpoint set to obtain the endpoint pairs that need to be automatically completed; For endpoint pairs that need to be automatically completed, generate the center to complete the path; After obtaining the center completion path, a tubular completion mask is generated based on the local diameter of the endpoints; the tubular completion mask is merged into the corresponding label region of the semicircular tube in the initial subregion label map to obtain the merged subregion label map; the centerline extraction is re-performed on the corresponding label region of the semicircular tube in the merged subregion label map to obtain the skeleton set to be reviewed. Based on the skeleton set to be reviewed, abnormal connections are removed to obtain a temporary corrected label map; The temporary modified label image was then modified and reviewed. After the review is approved, the corrected subregion label map and the corresponding corrected label map will be output.

2. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 1, characterized in that, The endpoint pairs that need to be automatically completed are selected from the skeleton endpoint set. For any two endpoints in the set of skeleton endpoints, they are respectively determined as the first endpoint and the second endpoint, and the endpoint pair composed of the first endpoint and the second endpoint is determined as the fracture candidate endpoint pair; a continuous skeleton segment with a preset real spatial arc length is cut from each endpoint along its respective skeleton branch into the semicircular canal candidate region as the endpoint neighborhood of the corresponding endpoint. Based on the spatial coordinates of the skeleton voxels in the neighborhood of each endpoint, the local direction vector of the corresponding endpoint is determined. Then, using the first and second endpoints as path endpoints and the local direction vectors of the two endpoints as direction constraints, a candidate scoring path connecting the two endpoints is generated. ; Distance fit is determined based on the actual spatial distance between the two endpoints. The consistency of direction is determined based on the angle between the local direction vectors of the two endpoints and the direction of the line connecting the two endpoints. ; Along the scoring candidate path Perform path sampling and calculate path image support according to the following formula. ,in, For the scoring candidate path The number of path sampling points on the path voxels in the subregion probability map The probability value of belonging to the semicircular canal or vestibular semicircular canal category. For voxels The fluid structure support value is obtained by normalizing the grayscale of the original magnetic resonance image. Preset fusion weights; A three-dimensional Euclidean distance transformation considering the three-axis voxel spacing is performed on the binary mask of the semicircular canal candidate region to obtain a distance transformation map, wherein the distance transformation value of any foreground voxel in the distance transformation map is the shortest true spatial distance from the foreground voxel to the boundary of the semicircular canal candidate region; The local radius of the first endpoint is obtained by taking the median of the distance transformation values ​​of the skeleton voxels in the neighborhood of the first and second endpoints, respectively. Local radius of the second endpoint And calculate local radius consistency ; According to the scoring candidate path The number of path sampling points falling into the preset uncoverable area, accounting for a certain percentage of the scoring candidate paths. The proportion of the total number of path sampling points determines the boundary violation penalty. The preset non-coverable areas include the cochlear high-confidence label area, the contralateral inner ear label area, the background high-confidence area, and other prohibited completion areas. Calculate the fracture connectivity score of the candidate fracture endpoint pairs. ,in, to Preset non-negative weights; score broken connections. Candidate endpoint pairs with broken connections that are not lower than a preset connection threshold are identified as endpoint pairs that need to be automatically completed.

3. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 1, characterized in that, For endpoint pairs that need to be automatically completed, the generation center completion path includes: For endpoint pairs that need to be automatically completed, a bounding box corresponding to the line connecting the two endpoints is generated, and the bounding box is expanded outward to obtain a search window. The search window is spatially constrained according to the ipsilateral inner ear label area and a preset non-coverable area, so that the search window is located within a preset neighborhood of the ipsilateral inner ear structure, and path search is prohibited from crossing the contralateral inner ear label area, the cochlear high-confidence label area, the background high-confidence area, or other preset non-coverable areas. The paths within the search window that consist of consecutively adjacent voxels, with their start and end points being the two endpoints that need to be automatically completed, are identified as the paths to be evaluated. Wherein, the path to be evaluated is represented as ; Calculate any voxel in the path to be evaluated. Basic toll cost ; Calculate the path to be evaluated Cumulative toll costs ,in, voxels in the subregion probability map The probability value of belonging to the semicircular canal or vestibular semicircular canal category; For voxels The support value of the liquid structure is obtained by normalizing the grayscale of the original magnetic resonance image. For voxel boundary risk penalty, when voxel When within the allowed search area, Take the first preset value; when voxels When located in the boundary buffer zone between the allowed search area and the preset uncoverable area, Take a second preset value that is greater than the first preset value; when the voxel When located in a pre-defined uncoverable area, voxels will be... Set as an unusable voxel; To determine based on adjacent path segments and The local curvature penalty is determined by the change in direction between paths; the greater the change in path direction, the greater the local curvature penalty. to Preset non-negative weights; The cumulative travel cost is selected from the paths to be evaluated through a minimum cost path search. The shortest path is used as the center completion path.

4. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 1, characterized in that, After obtaining the center completion path, generating a tubular completion mask based on the local diameter of the endpoints includes: A three-dimensional Euclidean distance transformation considering the triaxial voxel spacing is performed on the binary mask of the semicircular canal candidate region to obtain the distance transformation map; For each of the two endpoints that need to be automatically completed, a continuous skeleton segment of a preset real spatial arc length is extracted from the skeleton branch to which the corresponding endpoint belongs and taken into the candidate region of the semicircular canal as the endpoint neighborhood of the corresponding endpoint. The local radius of the first endpoint is obtained by taking the median of the distance transformation values ​​of the skeleton voxels in the neighborhood of each endpoint. Local radius of the second endpoint ; Calculate the ratio of the local radii of the two endpoints. ; When the radius ratio When the radius is not greater than a preset radius consistency threshold, the local radius of the first endpoint is adjusted along the center completion path from the first endpoint to the second endpoint. Local radius of the second endpoint Linear interpolation is performed to obtain the local completion radius corresponding to each path position on the central completion path; When the radius ratio When the radius is greater than the preset radius consistency threshold, the local radius of the first endpoint is adjusted. Local radius of the second endpoint The smaller value in is determined as the local completion radius corresponding to each path position on the central completion path; The local completion radius is limited to a preset real space radius; the local tangent vector of each path position is determined according to the adjacent path points on the central completion path, and the real space circular filling area is determined according to the local completion radius of the corresponding path position in the cross section perpendicular to the local tangent vector; Based on the triaxial voxel spacing, the real-space circular filling region is mapped to the voxel coordinate system, and the union of the filling regions corresponding to each path position on the central completion path is taken to obtain the tubular completion mask.

5. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 1, characterized in that, Incorporating the tubular completion mask into the semicircular tube corresponding label region of the initial subregion label map includes: The voxels in the tubular completion mask that overlap with the cochlear tag region, the contralateral inner ear tag region, the background high-confidence region, or other preset non-coverable regions are identified as conflict voxels; based on the triaxial voxel spacing, the ratio of the actual spatial volume of the conflict voxels to the total actual spatial volume of the tubular completion mask is calculated to obtain the conflict voxel ratio. When the proportion of conflicting voxels is not lower than a preset conflict proportion threshold, it is determined that the center completion path does not meet the preset label constraint, the automatic completion of the corresponding endpoint pair is abandoned, and the endpoint pair is recorded as a suspected broken endpoint pair. When the conflict voxel ratio is lower than the preset conflict ratio threshold, the conflict voxel is deleted from the tubular completion mask to obtain the remaining tubular completion mask; it is determined whether the remaining tubular completion mask is connected to the semicircular tube label area where the two endpoints that need to be automatically completed are located. When the remaining tubular completion mask is connected to the semicircular tube label area where the two endpoints are located, the remaining tubular completion mask is merged into the corresponding label area of ​​the semicircular tube in the initial subregion label map, and the label conflict record is stored; when the remaining tubular completion mask is not connected to the semicircular tube label area where the two endpoints are located, the automatic completion of the corresponding endpoint pair is abandoned, and the endpoint pair is recorded as a suspected broken endpoint pair. The newly added connectivity relationship is checked for the label area corresponding to the merged semicircular canals; when the remaining tubular completion mask is outside the preset allowed connection area, and the label area corresponding to the semicircular canals forms a new connectivity relationship with the cochlear label area, the vestibular main area or other designated branches, the automatic completion result of the corresponding endpoint pair is rolled back.

6. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 1, characterized in that, Abnormal connection removal based on the skeleton set includes: The skeleton endpoints and skeleton bifurcation points in the skeleton set to be reviewed are determined as skeleton graph nodes, and the continuous skeleton voxel sequence between two adjacent skeleton graph nodes is determined as skeleton branches. For any skeleton branch connected to a skeleton bifurcation point, if the bifurcation degree of the skeleton bifurcation point exceeds a preset upper limit of bifurcation degree, and the skeleton length of the skeleton branch is less than a preset abnormal branch length threshold, then the skeleton branch is determined as a first type of abnormal connection candidate branch. Skeleton branches whose two end nodes are both skeleton bifurcation points and whose skeleton length is less than a preset bridging length threshold are identified as short bridge candidate branches. For each short bridge candidate branch, other skeleton branches connected to each of its end nodes but not belonging to the short bridge candidate branch are identified as adjacent branches. When the same end node is connected to multiple adjacent branches, the adjacent branch with the largest skeleton length is identified as the adjacent reference branch corresponding to that end node. Perform a three-dimensional Euclidean distance transformation considering the triaxial voxel spacing on the binary mask of the semicircular tube corresponding label region in the merged subregion label map to obtain the distance transformation map to be verified. For any skeleton branch Calculate the skeleton branches Average category confidence ,in, For skeleton branches The number of skeletal voxels included. The skeleton voxels in the subregion probability map The probability value of belonging to the semicircular canal or vestibular semicircular canal category; The median of the distance transformation values ​​of the skeleton voxels contained in the skeleton branch in the distance transformation graph to be verified is taken as the skeleton branch. Local radius of branch ; Calculate the skeleton branches Average liquid structure grayscale support ,in, For skeleton voxels The support value of the liquid structure is obtained by normalizing the grayscale of the original magnetic resonance image. When the average class confidence of the short bridge candidate branch is lower than the product of the smaller of the average class confidence of the two adjacent reference branches and a preset confidence ratio threshold, the local radius of the short bridge candidate branch is lower than the product of the smaller of the local radius of the two adjacent reference branches and a preset radius ratio threshold, and the average liquid structure grayscale support of the short bridge candidate branch is lower than the product of the smaller of the average liquid structure grayscale support of the two adjacent reference branches and a preset grayscale support ratio threshold, the short bridge candidate branch is determined as a second type of abnormal connection candidate branch. Within the label regions corresponding to the first type of abnormal connection candidate branch and the second type of abnormal connection candidate branch, voxels with a category probability value lower than a preset category probability threshold and a liquid structure grayscale support value lower than a preset grayscale support threshold are identified as voxels to be deleted, and the voxels to be deleted are deleted. After deletion, the maximum connected component in the corresponding label area of ​​the semicircular canal before and after deletion is determined, and the skeleton length of the maximum connected component after deletion is calculated to the skeleton length of the maximum connected component before deletion to obtain the main structure retention rate; the connected components that appear after deletion and do not belong to the maximum connected component after deletion are determined as newly added isolated connected components, and the ratio of the sum of the skeleton lengths of the newly added isolated connected components to the skeleton length of the maximum connected component before deletion is calculated to obtain the newly added breakage ratio. When the main structure retention rate is not lower than the preset main structure retention threshold, the newly added fracture ratio is not higher than the preset fracture ratio threshold, and the increase in the number of skeleton endpoints after deletion relative to the number before deletion does not exceed the preset upper limit of endpoint increase, the deletion result is retained and a temporary correction label map is obtained; otherwise, the deletion operation is rolled back and the corresponding area is marked as a suspected abnormal connection that is not automatically corrected.

7. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 1, characterized in that, The post-correction review of the temporary corrected label image includes: The candidate region for the modified anterior semicircular canal is denoted as... The modified semicircular canal region corresponding to the modified pre-modification semicircular canal candidate region in the temporary modified label map is denoted as... ; Determine the newly added and supplemented areas ; The newly added and completed regions are calculated based on the triaxial voxel spacing. and the modified anterior semicircular canal candidate region The actual spatial volume; and calculate the proportion of the completed volume. ,in, This indicates that the actual spatial volume of the corresponding region is calculated based on the triaxial voxel spacing; when the padded volume accounts for... When the percentage of the completed volume is not higher than the preset completion volume ratio threshold, the completion result of the corresponding endpoint pair is retained; when the completion volume ratio is... When the result exceeds the preset completion volume ratio threshold, roll back all completion results for the corresponding endpoint pair and output the manual review mark associated with the corresponding endpoint pair.

8. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 7, characterized in that, The review of the temporary modified label image after modification also includes: The candidate regions of the modified anterior semicircular canal were respectively... and the modified semicircular canal region Perform connected component analysis; when the modified semicircular canal region The number of connected components is greater than the number of candidate regions for the modified anterior semicircular canal. The number of connected components, or the modified semicircular canal region. When a new connected component with a skeleton length less than a preset short segment length threshold exists, the newly added completion region is... Perform pipe diameter verification; For the modified semicircular canal region The binary mask is subjected to a three-dimensional Euclidean distance transformation considering the three-axis voxel spacing to obtain the corrected distance transformation map; For each of the two endpoints that need to be automatically completed, a continuous skeleton segment of a preset real spatial arc length is extracted from the skeleton branch to which the corresponding endpoint belongs and taken into the candidate region of the corrected anterior semicircular canal, as the endpoint neighborhood of the corresponding endpoint. The median of the distance transformation values ​​of the skeleton voxels within the endpoint neighborhoods of the two endpoints to be automatically completed is used to obtain the local radius of the first endpoint. Local radius of the second endpoint And calculate the average radius of the endpoints according to the following formula. ; The center completion path is read from the newly added completion area. The maximum value of the distance transformation values ​​of each path point within the corrected distance transformation map is determined as the maximum local radius of the completed region. And calculate the radius expansion ratio according to the following formula. ; The local tangent vectors of each path position are determined based on the adjacent path points on the center completion path; for the center completion path located in the newly added completion region... For any path point within the region, calculate the newly added and completed region within the cross section perpendicular to the corresponding local tangent vector. Real spatial cross-sectional area The maximum value among the real spatial cross-sectional areas corresponding to each path point is determined as the maximum cross-sectional area of ​​the completed region. ; Calculate the reference cross-sectional area at the endpoints and cross-sectional area expansion ratio ; When the radius expansion ratio The expansion ratio is higher than the preset radius expansion threshold, or the cross-sectional area expansion ratio. When the cross-sectional area expansion exceeds the preset threshold, the normalized path arc length is used to complete the path along the center. Determine the target local radius at each path location. ,in, , The local completion radius for the corresponding path position before refinement; Using the center-completed path as the axis, based on the target local radius Regenerate a refined tubular completion mask, and replace the newly added completion area with the refined tubular completion mask. ; The refined and corrected semicircular canal region is re-executed with connected component analysis, short segment detection, radius expansion ratio calculation, and cross-sectional area expansion ratio calculation. If the refined and corrected semicircular canal region still has any of the following conditions: an increase in the number of connected components, a new connected component with a skeleton length less than the preset short segment length threshold, a radius expansion ratio higher than the preset radius expansion threshold, or a cross-sectional area expansion ratio higher than the preset cross-sectional area expansion threshold, all completion results for the corresponding endpoint pairs are rolled back, and a manual review mark is output. Otherwise, the refined completion results are retained.

9. The inner ear subregion segmentation control method for fracture correction of slender semicircular canal structures according to claim 8, characterized in that, The review of the temporary modified label image after modification also includes: The path point sequence is obtained by resampling along the center completion path according to the preset real space step size with equal arc length; For any internal path point in the path point sequence , respectively at the internal path points The first reference path point is selected for the preset path arc length positions before and after. ⁻ and second reference path point ⁺, Calculate vector ⁻− with vector ⁺− The local interior angle between ; When the local inner angle When the angle is less than a preset smoothing angle threshold, the internal path point... This point has been identified as a candidate sharp turn path point. Establish a preset real-world spatial neighborhood centered on the candidate sharp turn path points. The local image support corresponding to the candidate sharp turn path points is calculated according to the following formula. ,in, The number of voxels in the preset real-world neighborhood. voxels in the subregion probability map The larger of the corresponding semicircular canal category probability and vestibular category probability. For voxels The fluid structure support value is obtained by normalizing the grayscale of the original magnetic resonance image. Preset fusion weights; When the local image support corresponding to the candidate sharp turn path point When the local image support threshold is lower than the preset threshold, it is determined that the center completion path does not meet the preset smooth shape condition at the candidate sharp turn path point, and the local completion area containing the candidate sharp turn path point is backed up. After reverting the local completion area, centerline extraction, equal arc length resampling, local interior angle calculation, and local image support calculation are re-executed on the remaining completion results. If there are still candidate sharp turn path points in the remaining completion results whose local interior angle is less than the preset smooth angle threshold and whose local image support is lower than the preset local image support threshold, then all completion results of the corresponding endpoint pair are reverted.

10. The inner ear subregion segmentation control system for fracture correction of slender semicircular canal structures according to claim 9, characterized in that, include: The endpoint pair determination unit is used to acquire a three-dimensional inner ear magnetic resonance image, the triaxial voxel spacing corresponding to the three-dimensional inner ear magnetic resonance image, the subregion probability map and the initial subregion label map output by the front-end segmentation; extract the semicircular canal candidate regions to be quality controlled from the initial subregion label map; perform centerline extraction on the semicircular canal candidate regions to obtain the corresponding skeleton set; Based on the skeleton set, a skeleton endpoint set is identified; Filter the skeleton endpoint set to obtain the endpoint pairs that need to be automatically completed; The completion unit is used to complete the path for the endpoint pairs that need to be automatically completed; After obtaining the center completion path, a tubular completion mask is generated based on the local diameter of the endpoints; the tubular completion mask is merged into the semicircular tube corresponding label region of the initial subregion label map to obtain the merged subregion label map; the centerline extraction is re-executed on the semicircular tube corresponding label region in the merged subregion label map to obtain the skeleton set to be reviewed; Based on the skeleton set to be reviewed, abnormal connections are removed to obtain a temporary corrected label map; The verification unit is used to perform a verification of the temporary modified label image after modification; After the review is approved, the corrected subregion label map and the corresponding corrected label map will be output.

Images