Medical image automatic analysis and layout system and method based on artificial intelligence large model

Abstract

This invention discloses an automatic medical image analysis and layout system and method based on an artificial intelligence large model, relating to the field of image recognition technology. The method comprises the following steps: acquiring fine-sampling reconstruction sequences and coarse-sampling reconstruction sequences from the same original acquisition; forming stable anomaly cores based on threshold perturbation to maintain relationships and determining cross-reconstruction topological conflicts; establishing a conditional local cross-reconstruction mapping model using real paired images spatially separated from the anomalies to be judged; fixing conditional gating states within the effective domain and applying finite gray-level perturbations; determining single-source connection attenuation pseudo-splitting, multi-source hybrid pseudo-fusion, or unconfirmed physical attribution based on changes in connection states; forming physical anomaly groups based on cross-reconstruction correspondences and attribution results; and using these physical anomaly groups to limit the analysis objects of the artificial intelligence large model and the layout of medical images. The key technical point is that this invention can improve the accuracy of physical attribution of cross-reconstruction anomaly objects and image layout.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to an automatic medical image analysis and typesetting system and method based on a large artificial intelligence model. Background Technology

[0002] With the development of medical imaging digitization and artificial intelligence image analysis technology, examinations such as chest CT and abdominal CT usually generate reconstructed images with different slice thicknesses or spatial sampling conditions from the same scan data. The medical image analysis system then detects, segments, and spatially correlates abnormal image regions to assist in subsequent image analysis and key image organization.

[0003] Taking the analysis of small nodules in chest CT as an example, the same nodule can span multiple slices in a thin-slice reconstructed image. Its edge grayscale changes, internal low-density areas, and fine connections with adjacent tissues can be presented as a relatively complete three-dimensional voxel structure. When the same scan data forms a thicker reconstructed image, the range of voxels involved in the formation of a single-layer image along the slice direction changes, and the local pixel grayscale and region boundaries also change accordingly. Existing medical image processing usually establishes the correspondence between different reconstructed images based on the patient's spatial coordinates, the location of abnormal areas, and the region overlap relationship, and combines three-dimensional connected region analysis to determine candidate abnormal objects. This type of processing can meet most medical image analysis scenarios with stable boundaries and clear spatial separation relationships.

[0004] In small and irregularly shaped anomalous regions, changes in reconstruction conditions may further alter the connectivity structure of the anomalous region in the image. For example, a real nodule with narrow connecting parts may appear as a single continuous three-dimensional region in a thinner image, while in a thicker image, the voxel corresponding to the connecting part may appear as two spatially close separated regions after the surrounding tissue participates in grayscale formation.

[0005] In another scenario, two small anomalies that are adjacent to each other and originally have independent boundaries may appear as a continuous region in a thicker reconstructed image. In this case, the locations of the anomaly regions in each reconstructed image can still maintain a good spatial correspondence. Therefore, based solely on the spatial registration results, the location of the region center, or the analysis of a single connected region, it is impossible to directly indicate the true number of anomalies represented by the one-to-many or many-to-one correspondence between image regions.

[0006] If subsequent large-scale artificial intelligence models directly use the connected regions in each reconstructed image as the analysis object, the same physical anomaly may form multiple candidate objects, and multiple adjacent physical anomalies may also form a single candidate object, further affecting the selection of key medical images and layout organization. Therefore, it is necessary to establish an image analysis mechanism that can characterize the spatial evolution relationship of the changes in the connected structure of the abnormal image region under different reconstruction conditions, so as to provide an image-level basis for determining the physical quantity of abnormal objects. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides the following technical solution: An automated medical image analysis and typesetting system based on a large-scale artificial intelligence model includes: The medical image processing module acquires the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence from the same original acquisition. It forms a stable abnormal core based on the preservation relationship under threshold perturbation, establishes a cross-reconstruction correspondence based on the patient's spatial correspondence, cross-combines the threshold states of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence to determine the cross-reconstruction topological conflict, and selects paired images that are separated from the spatial abnormality to be judged in the neighborhood of the cross-reconstruction topological conflict to form a cross-reconstruction reference set. The cross-reconstruction mapping module establishes a conditional local cross-reconstruction mapping model from fine-sampling reconstruction sequence to coarse-sampling reconstruction sequence based on the cross-reconstruction reference set, and determines the effective scope based on local image conditions. The attribution verification module determines the critical connection region and critical bridging region for one-to-many and many-to-one cross-reconstruction topological conflicts, respectively. It determines the conditional gating state based on the original fine-sampling reconstruction sequence, applies finite gray-level perturbation to the image region corresponding to the fine-sampling reconstruction sequence, and determines the attribution result based on the connection state change of the region corresponding to the coarse-sampling reconstruction sequence. The attribution result is one of the following: single-source connection attenuation pseudo-splitting, multi-source hybrid pseudo-fusion, or physical attribution unconfirmed. The analysis and layout module establishes group relationships and separation relationships based on cross-reconstruction correspondence and attribution results. Through relationship propagation and conflict verification, physical anomaly groups are formed. The physical anomaly groups limit the analysis objects of the artificial intelligence big model, and medical image layouts are formed based on the medical images of the physical anomaly groups.

[0008] This invention also discloses a method for automatic analysis and typesetting of medical images based on a large artificial intelligence model, including: The fine-sampling reconstruction sequence and coarse-sampling reconstruction sequence of the same original acquisition are obtained. A stable abnormal core is formed according to the preservation relationship under threshold perturbation. A cross-reconstruction correspondence is established based on the patient spatial correspondence. The threshold states of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence are cross-combined to determine the cross-reconstruction topological conflict. Paired images that are separated from the spatial abnormality to be judged are selected in the neighborhood of the cross-reconstruction topological conflict to form a cross-reconstruction reference set. A conditional local cross-reconstruction mapping model from fine-sampling reconstruction sequence to coarse-sampling reconstruction sequence is established based on the cross-reconstruction reference set, and the effective scope is determined based on local image conditions. For one-to-many and many-to-one cross-reconstruction topological conflicts, critical connection regions and critical bridging regions are determined respectively. The conditional gating state is determined based on the original fine-sampling reconstruction sequence. A limited gray-level perturbation is applied to the image region corresponding to the fine-sampling reconstruction sequence. The attribution result is determined based on the connection state change of the region corresponding to the coarse-sampling reconstruction sequence. The attribution result is one of the following: single-source connection attenuation pseudo-splitting, multi-source hybrid pseudo-fusion, or physical attribution unconfirmed. Based on the cross-reconstruction correspondence and attribution results, same-group and separation relationships are established. Through relationship propagation and conflict verification, physical anomaly groups are formed. The physical anomaly groups limit the analysis objects of the artificial intelligence large model, and medical image layouts are formed based on the medical images of the physical anomaly groups.

[0009] This invention provides a system and method for automatic analysis and typesetting of medical images based on a large artificial intelligence model, which has the following beneficial effects: 1. This invention addresses the one-to-many and many-to-one correspondence of abnormal regions that occur under different reconstruction conditions in the same original acquisition. It first determines the truly stable cross-reconstruction topological conflicts by maintaining the relationship of stable abnormal cores under threshold perturbation and verifying cross-reconstruction joint stability. This avoids misjudging single segmentation threshold fluctuations, local boundary changes, or accidental connectivity states as changes in the number of physical anomalies. It establishes a stable basis for determining the attribution of abnormal objects in different reconstruction sequences from the image topology level.

[0010] 2. Based on this, the present invention utilizes real paired images of the current patient acquired in the same session and spatially separated from the abnormality to be judged to establish a conditional local cross-reconstruction mapping model. Within the effective domain supported by the real image conditions, the conditional gating state is fixed, and finite gray-level perturbation is implemented around the critical connection region or critical bridging region. Based on the change in the target connection state caused by the perturbation after cross-reconstruction mapping, it distinguishes between single-source connection attenuation pseudo-splitting and multi-source hybrid pseudo-fusion. Thus, the apparent splitting and apparent fusion caused by different reconstruction conditions are transformed into cross-reconstruction attribution verification with local source contribution basis, improving the reliability of the determination of the number of abnormal physical quantities.

[0011] 3. This invention transforms cross-reconstruction correspondence and attribution verification results into same-group relationships, separation relationships, and physical anomaly groups. The same physical anomaly group simultaneously defines the analysis objects of the artificial intelligence big model and the abnormal objects in the medical image layout. This allows the separation performance of the same physical anomaly in different reconstruction sequences to be uniformly grouped, and the appearance fusion of different physical anomalies caused by reconstruction conditions to be distinguishable. This suppresses the repeated analysis and erroneous merging of abnormal objects from the source, as well as the resulting deviations in the selection of key medical images and layout organization. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall processing flow of the present invention; Figure 2 This is a schematic diagram of the processing flow of the cross-reconstruction anomaly attribution verification module of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0014] Please see Figure 1 and Figure 2 This embodiment provides an automatic medical image analysis and layout system based on a large artificial intelligence model, including a medical image processing module, a cross-reconstruction mapping module, an attribution verification module, and an analysis and layout module.

[0015] The processing objects are the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence formed from the same original acquisition. The fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence are named only according to the relative density of the physical sampling grid in the patient's space. The system determines the stable cross-reconstruction topological conflict, local cross-reconstruction response and cross-reconstruction abnormal object attribution in turn based on the differences in one-to-many and many-to-one connection states in the abnormal regions of the two sets of reconstruction sequences. The resulting physical abnormality group limits the analysis objects of the artificial intelligence large model and the medical image layout. The physical abnormality group only represents the set of cross-reconstruction abnormal objects confirmed by image evidence and does not directly represent the clinical diagnostic conclusion.

[0016] The medical image processing module forms a stable anomaly core, cross-reconstruction correspondence, cross-reconstruction topological conflict, and cross-reconstruction reference set. Based on this, the cross-reconstruction mapping module establishes a conditional local cross-reconstruction mapping model from fine-sampling reconstruction sequence to coarse-sampling reconstruction sequence and determines the effective scope. The attribution verification module applies limited gray-level perturbation to the corresponding source image region under fixed-condition gating and determines the attribution result based on the connection state change of critical connection region or critical bridging region. The analysis and layout module establishes same-group relationships, separation relationships, and physical anomaly groups based on the cross-reconstruction correspondence and attribution results, and uses the same physical anomaly group identifier to limit the analysis objects of the artificial intelligence large model and the abnormal objects in the medical image layout.

[0017] In this embodiment, "same original acquisition" refers to two sets of reconstruction sequences formed from the same original acquisition data, with consistent examination acquisition identifiers, reconstruction association information, and patient reference coordinates. It does not include sequences formed by re-acquisition after different enhancement phases, repeated scans, or significant patient movement. The current examination is limited to one processing process of the original acquisition and its derived reconstruction sequences. It starts after the common acquisition source of the two sets of reconstruction sequences is confirmed and a fixed input snapshot is formed, and ends after the physical abnormality group, unconfirmed physical attribution objects, and medical image layout are output. Historical true paired reconstruction data are obtained from previous examinations that are consistent with the current anatomical location and fine sampling and coarse sampling reconstruction category combination. Only true paired reconstruction data with verifiable common original acquisition sources, complete patient spatial relationships, and abnormal object correspondence confirmed by existing quality review records are retained. Historical data is divided into parameter determination data and independent verification data according to the examination level. Image blocks from the same examination only enter one set. Historical data is only used to pre-calibrate the admission threshold, allowable domain, and error benchmark, and does not participate in the formation of the current patient's physical abnormality group.

[0018] The anomalies to be judged include stable anomaly cores that form current cross-reconstruction topological conflicts and abnormal transition regions that connect or separate the corresponding stable anomaly cores within a range of consecutive effective threshold combinations; their patient spatial extent is determined by the union of the corresponding abnormal connected regions within a range of consecutive effective threshold combinations, and the stable anomaly core is formed by the position where the threshold maintenance amount reaches the stable core admission threshold, and subsequent steps do not redefine it.

[0019] The conditionalized local cross-reconstruction mapping model determines parameters using real paired images in the current patient cross-reconstruction reference set. It conditionally invokes the finite-support local spatial response based on the local image of the original fine-sampling reconstruction sequence and calculates the target prediction response at the corresponding location in the coarse-sampling reconstruction sequence. The effective scope is used to limit the model's invocation location in the patient space. The conditional gating state is used to determine the local responses of the conditional subdomains invoked at locations already within the effective scope, along with their combined weights. The parameter roles of the effective scope and the conditional gating state are kept separate. The model exit validation allowance domain is used to determine whether the model as a whole is qualified to enter the effective scope for computation. The target-side prediction error threshold is used to determine whether the target response caused by finite gray-level perturbations exceeds the model's own prediction fluctuation. The model exit validation allowance domain, effective scope, and target-side prediction error threshold are used for model-level admission, location-level admission, and response distinguishability judgment, respectively. They are used independently and do not substitute for each other. The conditional gating state is determined by the undisturbed original fine-sampling reconstruction sequence and is fixed within the same attribution validation process.

[0020] Each processing stage adopts a step-by-step admission process. If the preceding object, data, or model does not meet the corresponding objective admission conditions, subsequent attribution verification based on the results of that preceding stage will be stopped. The formed image evidence and the reason for exit will be saved, and the physical attribution will remain unconfirmed. The unconfirmed state will not generate a definite same-group relationship, separation relationship, or a new physical anomaly group.

[0021] The patient spatial correspondence is established based on the patient's spatial location, image plane orientation, pixel physical spacing, and slice physical location in medical digital imaging and communication images. The voxel positions of the two sets of reconstructed sequences are converted to the same patient physical coordinate system. This process preserves the original grayscale values, and subsequent physical coverage, boundary distance, connection position, and separation gap are all calculated in this coordinate system.

[0022] The parameters are divided into current inspection parameters and pre-calibrated parameters. The local response parameters of the conditional local cross-reconstruction mapping model are determined by the cross-reconstruction reference set of the current inspection. The stable core admission threshold, cross-reconstruction corresponding admission condition, cross-reconstruction topological conflict admission threshold, model validation allowable domain, critical region preservation admission threshold, baseline topological replay admission threshold, and topological inversion preservation admission threshold are determined by parameters in historical real paired reconstruction data to form candidate values, and then corrected on independent validation data. Parameters involving quantile values ​​are all calculated and distributed according to inspection-level samples. A large number of adjacent voxels in the same inspection are not directly regarded as independent historical samples. When the number of historical inspections is insufficient to stably form the adopted quantile position, it is not supplemented by the current anomaly data to be judged, debugging records, or empirical constants. The corresponding parameters are considered to be incompletely calibrated. The physical attribution is not confirmed in this process. The preferred range is used as the parameter candidate range. The preferred value is selected within the corresponding range and fixed before the current inspection begins. When there is no value in the preferred range that simultaneously meets the corresponding admission condition, the corresponding processing branch remains unconfirmed in physical attribution and does not obtain a determined attribution result by further relaxing the threshold.

[0023] The medical image processing module obtains the DICOM image corresponding to the original acquisition from the medical image storage system or medical imaging equipment, and confirms the fine sampling reconstruction sequence and the coarse sampling reconstruction sequence based on the examination acquisition identifier, reconstruction association information and patient reference coordinates. The common acquisition source is only confirmed when the examination acquisition identifier, reconstruction association information and patient reference coordinates correspond to each other and can be traced back to the same source acquisition record; if any source field is missing, conflicts with each other or cannot complete the patient spatial correspondence, the cross-reconstruction topology conflict judgment will not be entered.

[0024] After confirming the common acquisition source, the image instance list, pixel data, and patient spatial geometry information of the two sets of reconstructed sequences are used to form the current examination input snapshot. The image instances are sorted according to the patient spatial coordinates of the slice normal, and the pixel data is checked to be decodeable, the spatial location is not repeated, and the distance between adjacent instances is consistent with the physical sampling relationship recorded in the sequence. Based on this, the common effective patient spatial coverage of the two sets of reconstructed sequences in this input snapshot is determined. When there are undecodeable instances, repeated spatial locations, abnormal distances, or geometric information conflicts, the corresponding patient spatial range is recorded as an incomplete area of ​​input data. After the input snapshot is formed, the current processing will not include the image instances that arrive later. The later-arriving instances need to form a new input snapshot and re-execute the corresponding examination. The anomalies to be judged, cross-reconstruction reference images, and target verification ranges formed subsequently are checked at their formation stages to see if they are within the common effective patient spatial coverage. When any necessary range intersects with the incomplete area of ​​input data, the corresponding processing branch remains unconfirmed in terms of physical attribution.

[0025] Two sets of reconstruction sequences are processed using the same medical image segmentation model that has been deployed and validated in the fine-sampling reconstruction category and the coarse-sampling reconstruction category, respectively, to obtain an abnormal probability field with values ​​ranging from 0 to 1. Deployment validation is a routine model validation before the segmentation model goes online and does not participate in the cross-reconstruction object attribution calculation in this implementation method; the current processing only receives the output of the segmentation model that has been deployed and validated in the corresponding reconstruction category.

[0026] The medical image processing module forms a stable abnormal core based on the preservation relationship under threshold perturbation. Threshold perturbation intervals are established for the abnormal probability fields of fine-sampling reconstruction sequences and coarse-sampling reconstruction sequences, respectively. Each threshold perturbation interval is centered on the abnormal segmentation benchmark threshold of the corresponding reconstruction category, and multiple discrete threshold states are set on both sides of the benchmark threshold. The number of threshold states for each set of reconstruction sequences is preferably 9 to 15, with 11 being preferred. The boundary of the threshold perturbation interval is determined by the historical verification data of the corresponding reconstruction category: taking the abnormal subject region that has been quality-checked under the benchmark threshold as a reference, the threshold is adjusted step by step along the threshold decreasing direction and increasing direction. When further adjustment will make the abnormal subject no longer continuously tracked by the connected region containing the same stable subject, the last threshold that can still be continuously tracked is used as the boundary of the corresponding direction. Discrete states are set at equal threshold intervals within the interval. 9 to 15 states can form multiple independent perturbation positions on both sides of the benchmark threshold. 11 states are preferred to balance the resolution of threshold changes and the repeatability of adjacent states.

[0027] Abnormal connected regions are formed under each threshold state, and abnormal candidate trajectories are established based on the patient spatial inclusion relationship of abnormal connected regions under adjacent threshold states. In this implementation, the same three-dimensional adjacency rule is uniformly used to form abnormal connected regions within the current examination. The three-dimensional adjacency rule is consistent with the connectivity determination rule used when deploying the abnormal region extraction model. The medical image processing module and the attribution verification module both follow this rule. The adjacency rule is not switched during the current examination. For each patient spatial location in the abnormal candidate trajectory, the number of threshold states retained at that location within the threshold perturbation interval is counted. Then, the number of retained states is divided by the total number of threshold states in the reconstructed sequence to obtain the threshold retention amount between 0 and 1.

[0028] Regions whose threshold retention reaches the stable core admission threshold are defined as stable abnormal cores. The stable core admission threshold is preferably between 0.80 and 0.95, with 0.90 being the most preferred. In the historical parameter determination data, the proportion of abnormal subjects that are incorrectly excluded after quality review and the proportion of connection positions that only appear in a few threshold states are incorrectly included in the stable core are calculated for each candidate threshold. The candidate threshold with the smaller sum of the two error proportions is taken as the priority candidate. When there are multiple candidate values, the higher threshold is taken to prioritize the exclusion of transient connections. The final candidate value is then reviewed on independent validation data. If the validation result still satisfies the same error direction, it is taken as the fixed stable core admission threshold for the current reconstruction category. Regions whose threshold retention does not reach the stable core admission threshold but still have abnormal responses near the baseline threshold are retained as segmentation unstable regions and are not used as independent physical abnormal objects in the cross-reconstruction quantity judgment. When an abnormal candidate cannot form a stable abnormal core on any reconstruction side, the original medical image, patient spatial location, and physical attribution are saved as unconfirmed and directly transferred to the analysis and layout module.

[0029] The medical image processing module establishes cross-reconstruction correspondences based on patient spatial correspondences. After stable abnormality cores are formed in both the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence, for each stable abnormality core in the fine-sampling reconstruction sequence, a search is conducted within its physical neighborhood in the patient space for stable abnormality cores in the coarse-sampling reconstruction sequence. When a candidate stable abnormality core overlaps with the current stable abnormality core after geometric tolerance expansion, it enters the candidate set. The geometric tolerance is initially based on the actual physical sampling scale of the two reconstruction sequences and is corrected in the historical parameter determination data by simultaneously using stable one-to-one abnormality correspondences verified by patient spatial consistency and verified spatially adjacent independent abnormality correspondences: from the candidate tolerances that can cover the residual positional deviation of stable one-to-one abnormality correspondences, the smallest tolerance that still keeps spatially adjacent independent abnormalities separated is selected and fixed for the current reconstruction category combination after verification on independent validation data. If the historical data cannot form a tolerance that simultaneously covers the residual positional deviation of stable one-to-one abnormality correspondences and keeps spatially adjacent independent abnormalities separated, the cross-reconstruction correspondence parameters are considered to be incompletely calibrated, and the current examination does not form a definite cross-reconstruction correspondence.

[0030] The bidirectional physical coverage is calculated for the combination of stable abnormal cores from fine-sampling reconstructed sequences and stable abnormal cores from coarse-sampling reconstructed sequences. First, the physical volume of the stable abnormal core from the fine-sampling reconstructed sequence is used as the denominator to calculate the proportion of the volume covered by the coarse-sampling combination after geometric tolerance expansion. Then, the physical volume of the coarse-sampling combination is used as the denominator to calculate the proportion of the volume covered by the stable abnormal core from the fine-sampling reconstructed sequence after geometric tolerance expansion. The smaller of the two coverage proportions is taken as the bidirectional physical coverage. This calculation uses the overlapping volume of voxels corresponding to the physical units in the patient space, rather than directly comparing the number of voxels in different grids.

[0031] Simultaneously, the bidirectional boundary distances of stable abnormal core combinations are calculated. For each of the two directions, the nearest patient spatial distance from one stable abnormal core boundary point to the boundary of the other combination is calculated, and the 90th to 97.5th percentile values ​​are taken for all boundary point distances in that direction, preferably the 95th percentile. The larger of the two directional higher percentile distances is then taken as the bidirectional boundary distance. When the bidirectional physical coverage reaches the bidirectional coverage admission threshold and the bidirectional boundary distance does not exceed the upper limit of the cross-reconstruction boundary distance, the corresponding combination is considered a valid candidate. The bidirectional coverage admission threshold is preferably 0.70 to 0.85, preferably 0.80, and its candidate values ​​are derived from historical stable one-to-one abnormal... The bidirectional coverage distribution corresponding to the verified spatially adjacent independent anomalies is determined together, and values ​​that can maintain true correspondence and exclude accidental overlap are selected on the independent validation data. The upper limit of the cross-reconstruction boundary distance is formed by the bidirectional boundary distance distribution of historically stable one-to-one anomaly correspondences, and the verified spatially adjacent independent anomaly correspondences are used as exclusion samples for correction. The quantile level is preferably 90% to 97.5%, preferably 95%. If the historical data cannot form a boundary distance upper limit that maintains a stable one-to-one correspondence and excludes spatially adjacent independent anomalies, the cross-reconstruction correspondence parameter is considered to be incompletely calibrated, and the current check does not form a definite cross-reconstruction correspondence relationship.

[0032] When multiple valid candidate combinations exist, the combination with the fewest stable anomaly core members is selected first. If the number of members is the same, the combination with the larger bidirectional physical coverage is selected. If the bidirectional physical coverage still cannot distinguish them, the combination with the smaller bidirectional boundary distance is selected to form the stable anomaly core combination with the fewest members. Then, the same reverse search is performed from the coarse sampling reconstruction sequence to the fine sampling reconstruction sequence. When a single stable anomaly core is formed in both the forward and reverse directions, it is saved as a stable one-to-one cross-reconstruction correspondence. When there is a case where one stable anomaly core corresponds to multiple stable anomaly cores, one-to-many and many-to-one cross-reconstruction correspondence candidates are formed.

[0033] The candidate relationships obtained from the reverse search are checked for consistency with the forward search results. If the stable anomaly core of the same fine-sampling reconstruction sequence corresponds to one stable anomaly core of the coarse-sampling reconstruction sequence in the forward search, and the stable anomaly core of the coarse-sampling reconstruction sequence also uniquely corresponds to the stable anomaly core of the current fine-sampling reconstruction sequence in the reverse search, then the correspondence is saved as a stable one-to-one cross-reconstruction correspondence. If multiple stable anomaly cores are required to cover any direction in order to satisfy the bidirectional physical coverage and boundary distance conditions, then the corresponding one-to-many and many-to-one candidates are retained, and the nearest center distance is not used to replace the combined search.

[0034] The medical image processing module determines cross-reconstruction topological conflicts by cross-combining the threshold states of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence. It pairs each threshold state of the fine-sampling reconstruction sequence with each threshold state of the coarse-sampling reconstruction sequence to form a 2D threshold combination domain. For each threshold combination, it redetermines the stable abnormal core members actually contained in the current abnormal connected region and the corresponding cross-reconstruction direction, while the stable abnormal core itself remains unchanged.

[0035] When the stable anomaly core members and their corresponding cross-reconstruction directions under a certain threshold combination are consistent with the current one-to-many or many-to-one candidate relationships, the threshold combination is recorded as a valid combination. Adjacency relationships are established for combinations based on the threshold axis of the fine-sampling reconstruction sequence or the threshold axis of the coarse-sampling reconstruction sequence, with one adjacent threshold position. A continuous combination region containing the most valid combinations is searched from the valid combinations. The number of valid combinations contained in this continuous combination region is divided by the total number of combinations in the 2D threshold combination domain to obtain the cross-reconstruction topological conflict preservation value. If the cross-reconstruction topological conflict preservation value reaches the cross-reconstruction topological conflict admission threshold, and the continuous combination region contains… When a baseline threshold combination consisting of two sets of reconstruction sequence baseline thresholds is used, cross-reconstruction topological conflicts are determined. The cross-reconstruction topological conflict admission threshold is preferably 0.6 to 0.8, and more preferably 0.7. Historical parameter determination data are used to statistically analyze the retention of the number of verified stable cross-reconstruction conflicts and the apparent conflicts that only change with the unilateral threshold. For each candidate threshold, the sum of the stable conflict error exclusion ratio and the apparent conflict error acceptance ratio is calculated. The candidate value with the smaller total error is selected as the priority candidate. When candidate values ​​are tied, the higher threshold is selected. The final value is verified on independent validation data according to the same rules and kept fixed in the current inspection.

[0036] Complex many-to-many correspondences are first checked according to the stable abnormal core members to see if they can be split into one-to-many and many-to-one sub-relationships that do not share stable abnormal cores. If they can be split, cross-reconstruction topological conflicts are formed respectively. If they cannot be split, the stable abnormal cores, patient spatial correspondences and original medical images of the participating relationships are retained, and the attribution result is determined as physical attribution unconfirmed. This branch does not enter the parameter determination of the conditional local cross-reconstruction mapping model.

[0037] After cross-reconstruction topological conflicts are identified, relevant abnormal connected regions within the range of continuous effective threshold combinations are mapped to the patient space and their union is taken to form a conflict region union. From the boundary of this union, the reference search range is expanded outward within the same anatomical tissue until the minimum number of qualified true paired images are obtained or the current anatomical tissue boundary is reached. The local conflict space is jointly determined by the conflict region union and the actual reference search range, and is only used to limit the reference data source of the current conflict and the subsequent model space range. The conflict region union should be completely within the common effective patient space coverage of the current input snapshot. When it intersects with an incomplete input data region, the current cross-reconstruction topological conflict remains physically unconfirmed and does not continue to form a cross-reconstruction reference set for parameter determination.

[0038] The cross-reconstruction reference set consists of real paired images of the current patient and the original images acquired in the same session. The candidate reference image and the abnormality to be judged are kept separate in the patient space, and other abnormality candidates formed under stable abnormal cores, unstable segmentation regions, and minimum allowable threshold conditions are excluded. When the cross-reconstruction topological conflict is located near the anatomical boundary, the candidate reference image and the cross-reconstruction topological conflict are located on the same anatomical boundary side. The difference between the patient space distance from the center of the candidate reference image to the anatomical boundary and the distance from the center of the cross-reconstruction topological conflict background image block to the same boundary is calibrated according to the boundary side matching error range in the historical real paired reference data. Image blocks that exceed this range are not used as boundary region references.

[0039] For each pair of candidate real images, local image blocks of fine-sampling reconstruction sequence and local image blocks of coarse-sampling reconstruction sequence are formed within the same patient space. The patient space range of the local image block is determined according to the actual sampling interval of the coarse-sampling reconstruction sequence after the current cross-reconstruction topological conflict and its anatomical tissue boundary are determined, and remains consistent during the reference search process of the same cross-reconstruction topological conflict. If the image block cannot provide the adjacent samples required for gradient calculation or crosses the current anatomical boundary, the candidate reference pair is excluded.

[0040] Four local image conditions were calculated for local image blocks in both the fine-sampling and coarse-sampling reconstruction sequences: The local gray-level dynamic range was calculated by subtracting the 10th quantile from the 90th quantile within the block; the principal spatial gradient scale was calculated based on the actual physical pixel spacing, using the gradients in three patient spatial directions and obtained from the principal response of the structure tensor; the principal gradient direction was taken from the principal eigenvector of the structure tensor, and the absolute value of its dot product with the unit vector normal of the coarse-sampling reconstruction sequence slice was calculated as the directional relationship; the local high-frequency fluctuation level was calculated by subtracting the local mean of the same fixed patient spatial neighborhood from the original gray level, and then subtracting the 10th quantile from the 90th quantile of the high-frequency residual. The % quantile is obtained. The four conditions of the fine-sampling reconstruction sequence form the source-side local image conditions, and the four conditions of the coarse-sampling reconstruction sequence form the target-side local image conditions. The source-side local image conditions and the target-side local image conditions together constitute paired local image conditions. When the main gradient scale is lower than the resolvable lower limit determined by the minimum resolvable change in gray level and the actual sampling interval, the corresponding image block is recorded as having no stable principal gradient direction. When only one side of the candidate reference pair and the conflicting background has a stable principal gradient direction, an effective reference pair is not formed. When neither the candidate reference pair nor the conflicting background has a stable principal gradient direction, the direction component does not participate in the conditional distance comparison.

[0041] Subsequently, the absolute difference between the candidate reference pair and the cross-reconstruction topological conflict background is calculated for each effective conditional component, and normalized using the difference between the 90th and 10th quantile values ​​of that conditional component in the same anatomical tissue. A positive dimensional stability term is added to the normalization denominator, which is the larger of 1% of the robust dynamic range and the minimum resolvable positive change of the corresponding conditional component. The maximum value of the normalized difference of the effective conditional components is taken as the reference conditional distance. The reference conditional admission threshold is preferably 0.25 to 0.4, preferably 0.3. Its specific value is determined by gradually widening the reference conditional distance through historical parameter determination data and observing the independent residuals. It is then verified by independent validation data. True paired images whose reference conditional distance does not exceed the admission threshold are included in the cross-reconstruction reference set.

[0042] The minimum number of true paired images is determined through independent holdout validation. True paired images that do not overlap are gradually included according to the reference condition distance from near to far. After each inclusion, a new local cross-reconstruction mapping model is formed, and the three types of holdout residuals and changes in local spatial response parameters are recorded. When more reference pairs are included, and the holdout residuals and parameter changes both fall within the stable fluctuation range formed by historical repeated sampling, the minimum number of reference pairs that can meet the requirements for model building and independent validation before entering the stable platform is determined as the minimum number. This number is preferably 12 to 30 pairs, with 18 pairs being preferred. After the conditional subdomain parameters are determined, each conditional subdomain should also be able to form an identifiable parameter response matrix and retain at least one pair of true paired images that did not participate in the determination of the parameters of that subdomain for independent validation. If this is not met, adjacent conditional subdomains are merged. When the number of qualified reference pairs within the current anatomical tissue range is less than the minimum number, the relevant cross-reconstruction topological conflicts remain unconfirmed in terms of physical attribution and are not supplemented with images from other patients.

[0043] Each pair of images across the reconstructed reference set preserves the common patient spatial location and local image conditions, and ensures that the reference image blocks do not overlap. When the reference search reaches the current anatomical tissue boundary but still does not have enough reference data to reach the minimum number of references, the search stops and the insufficient reference data status is recorded.

[0044] The medical image processing module passes the cross-reconstruction topological conflict, local conflict space, and cross-reconstruction reference set to the cross-reconstruction mapping module, and retains the anomaly probability field, threshold perturbation interval, stable anomaly core, continuous effective threshold combination range, and stable one-to-one cross-reconstruction correspondence for subsequent steps to call.

[0045] The cross-reconstruction mapping module establishes a conditional local cross-reconstruction mapping model from fine-sampling reconstruction sequence to coarse-sampling reconstruction sequence only for cross-reconstruction topological conflicts that have already formed a cross-reconstruction reference set. This model is limited to the local response relationship in the current patient, current examination, and current conflict neighborhood, and its prediction results are only used as intermediate calculation results for attribution verification.

[0046] The fine-sampled reconstructed sequence image blocks across the reconstruction reference set are used as source reference images, and the coarse-sampled reconstructed sequence image blocks corresponding to the patient space are used as target reference images. The model parameters are determined only by the cross-reconstruction reference set. The local image conditions used to select local responses are calculated by the original fine-sampled reconstructed sequence. The local image conditions of the coarse-sampled reconstructed sequence are only used for parameter calibration and output verification. The anomalies to be judged do not participate in the determination of model parameters.

[0047] Conditional subdomains are divided based on the normalized differences in local image conditions of the source reference image. In one implementation, each real paired reference image block is first treated as an independent initial subdomain. Following the same dimensional normalization method used when forming the cross-reconstruction reference set, the normalized difference of each source-side conditional component between any two subdomain conditional anchor points is calculated, and the maximum value is taken as the distance between subdomains. The two subdomains with the smallest distance are then merged sequentially. After merging, a new conditional anchor point is formed using the median value of each conditional component. The maximum normalized difference from each reference pair in the new subdomain to this conditional anchor point is calculated in the same way. Merging in this direction is stopped when the maximum normalized difference exceeds the reference condition admission threshold if merging continues. Each final conditional subdomain is considered a valid conditional subdomain only if the parameter identifiable condition is met and an independent reference pair can be retained. If the condition cannot be met, merging continues with the nearest adjacent subdomain in terms of conditional distance. If the condition still cannot be met after merging, the current cross-reconstruction topological conflict remains unconfirmed in terms of physical attribution.

[0048] Each conditional subdomain establishes a finitely supported local spatial response, which is formed by combining one basic response basis function and several correction response basis functions. The total number of basis functions is preferably 3 to 6, specifically the minimum number that can be identified by parameters and independently verified. No fixed single values ​​are set across patients. The physical support range of the basic response basis function is initially based on the actual sampling scale of the coarse-sampled reconstruction sequence and is limited to the range that can be completely covered by the cross-reconstruction reference image. Non-negative candidate weights are first formed within this support, and then weights with equality constraints are corrected to make the sum of all coefficients equal to 1 and the first moments of the three patient spatial directions all equal to 0. The correction response basis functions are used to describe the finite axial mixing changes, in-plane smoothing changes, and edge compensation changes near the basic response. Each candidate correction response is formed within the same physical support, and then the constant component and the first moment components of the three patient spatial directions are removed to make the sum of their coefficients equal to 0 and the first moment equal to 0.

[0049] When parameters are determined, each real source reference image block is processed by the basic response basis function and the correction response basis function to form a parameter response sequence, which is then combined with the real target reference image corresponding to the patient space to form parameter determination data. When dimensionless transformation is required to improve numerical conditions, all basis function responses within the same reference pair adopt the same source-side scale, and the target response adopts the same target-side scale. Different scales are not used for different basis function responses. After obtaining the combination coefficients, the original basis function definition scale is restored, and the local spatial response is formed again. The sum of coefficients, first moment, negative sidelobe, and physical support constraints are re-examined. The combination coefficients are obtained by minimizing the target prediction residual on the parameter determination data. The absolute gray-level correspondence is determined by continuous monotonically piecewise linear gray-level mapping.

[0050] When the local spatial response is used to characterize the finite sharpening response, the existence of limited negative sidelobes in the local spatial response is allowed. The dimensionless negative sidelob ratio is formed by dividing the sum of the absolute values ​​of all negative coefficients by the sum of all positive coefficients. This ratio is greater than or equal to 0 and less than 1. The upper limit of the sidelobes is formed by the model in the historical real paired reconstruction data that can simultaneously satisfy the allowable domain of edge position residuals and contrast-preserving residuals. The smaller upper limit that meets the independent validation requirements is preferred. No fixed values ​​are set across devices. The patient space support range of each basis function does not exceed the physical range that the smallest effective reference image block in the cross-reconstruction reference set can completely cover. The model parameters are determined only when the negative sidelobes, physical support and parameter identifiability conditions are all satisfied.

[0051] The local responses of each correction response basis function applied to the real source reference image are used to form a parameter identifiable matrix. First, it is checked whether the effective columns used to determine the combination coefficients are linearly independent. Then, the maximum singular value and minimum singular value of the matrix are calculated. When the maximum singular value is greater than 0, the minimum singular value is divided by the maximum singular value to obtain the identifiable proportion. When the identifiable proportion reaches a predetermined identifiable lower limit, the corresponding parameters are retained. When the identifiable lower limit is not reached, the correction response basis functions that contribute less to the prediction of the real target image are removed and recalculated. When the identifiable lower limit is still not reached, adjacent conditional subdomains are merged.

[0052] The lower limit of parameter identifiability is determined by the stability of repeated sampling of historical real paired reconstructed data. For the parameter identifiability matrix formed by each repeated sampling, the ratio of the minimum singular value to the maximum singular value is calculated, and the change of the corresponding combination coefficients relative to the complete parameter determination set is calculated. The singular value ratio is between 0 and 1, preferably the lowest ratio that makes the change of the combination coefficients in the independent verification data fall within the stable fluctuation range of historical repeated sampling, and is fixed before the current inspection begins. No fixed value is set across devices. When the maximum singular value is 0, the parameter is directly determined to be unidentifiable. When it is necessary to reduce the number of basis functions, the correction response basis functions are temporarily removed one by one and the independent verification residuals are recalculated. Priority is given to permanently removing the correction response basis functions that minimize the change of grayscale prediction residuals and keep the residuals at the edge positions and the residuals of contrast retention within their respective allowable domains. If the identifiability condition cannot be met after reducing the number of basis functions, adjacent conditional subdomains are merged.

[0053] The cross-reconstruction mapping module determines the conditional gating state based on the original fine-sampling reconstruction sequence without finite gray-level perturbation. For each target location in the local conflict space, it calculates the normalized conditional distance between its local image condition and each conditional anchor point: the difference between each conditional component is calculated to form the same robust dynamic range and dimensional stability term when crossing the reconstruction reference set, and then the square root of the sum of squares of each normalized difference is taken.

[0054] The conditional radius of each conditional subdomain is determined by the normalized conditional distance distribution from the true state of the subdomain to the conditional anchor point. It is corrected by independently maintaining the true reference state, ensuring effective conditional support for the true state of the subject within the conditional subdomain and excluding local spatial responses corresponding to states exceeding the conditional radius. When the normalized conditional distance is less than the conditional radius, a distance ratio of 0 to 1 is formed by dividing the normalized conditional distance by the conditional radius. The value of the tight support conditional function is calculated by subtracting the fourth power of this distance ratio from 1, then multiplying by four times the distance ratio plus 1. When the distance ratio reaches 1, the function value is 0. The tight support conditional function value of each conditional subdomain is divided by the sum of all tight support conditional function values ​​at the current target position to obtain the local response combination weight. When all true state distances within a conditional subdomain are 0, the minimum positive conditional distance calculated from the minimum resolvable change of each conditional component is used as the lower limit of the radius. If a positive radius still cannot be formed, the subdomains are merged into the nearest effective conditional subdomain.

[0055] The weights of the local response combination are taken from 0 to 1. The sum of the weights of the local response combination at the same effective location is 1. When the tight support condition function values ​​of all conditional subdomains at the current target location are 0, it is determined that the original local image conditions at this location exceed the coverage of the cross-reconstruction reference set, and the conditionalized local cross-reconstruction mapping model does not form an effective response at this location.

[0056] For the target location, based on the local response combination weights determined by the original fine-sampling reconstruction sequence, the fine-sampling reconstruction sequence image data within the actual support range of the local spatial response of each conditional subdomain is read. Each local spatial response acts on the fine-sampling reconstruction sequence to form a local response value. The local response value is then multiplied by the corresponding local response combination weights and summed to obtain the conditional spatial response of the current target location. This calculation directly reads the support range of the source image in the patient's physical space and does not perform uniform resampling on the entire set of fine-sampling reconstruction sequences.

[0057] A continuous, monotonic, piecewise linear grayscale mapping is established using real source reference images and target reference images across the reconstruction reference set. The grayscale nodes are sorted from low to high according to the source-side conditionalized spatial response and are set within the coverage area of ​​the real reference response. The number of nodes is limited to a condition that each adjacent segment has real paired reference data and that independent output verification still satisfies the three types of residual allowable domains. Adding more nodes stops when it fails to improve output verification or causes a segment to lack real reference data. For each input node, the robust median value of the adjacent real conditionalized spatial response is read as the input grayscale, and the robust median value of the grayscale of the corresponding target reference image in the patient space is read as the output grayscale. Output nodes are subjected to non-decreasing constraints in the order of input nodes. When identical input nodes appear, adjacent nodes are merged. If there are fewer than two valid different input nodes, the current conditional subdomain does not form a valid grayscale mapping. The conditionalized spatial response of the current target location is located between two adjacent input nodes. Linear interpolation is performed on the corresponding two output grayscale values ​​according to their relative positions to the two input nodes to obtain the target prediction response at the corresponding position of the coarsely sampled reconstruction sequence.

[0058] The conditionalized local cross-reconstruction mapping model stores the actual slope of each segment of the piecewise linear grayscale mapping. When the current conditionalized spatial response is within the coverage of the true reference grayscale node and the absolute value of the segment slope reaches the minimum resolvable transfer slope, the location is allowed to enter a limited grayscale perturbation; otherwise, the physical attribution remains unconfirmed. The minimum resolvable transfer slope is formed only by effective grayscale segments from historical independent validation data that simultaneously pass the grayscale prediction residual, edge position residual, and contrast-preserving residual allowable domain. The lowest absolute value of the slope that can stably distinguish the source response change from the target grayscale change is taken as the calibration candidate, and then verified by independent validation data. The current anomaly to be judged and segments that have not passed the exit validation do not participate in lowering this lower limit.

[0059] After the model parameters are formed, the real paired images that did not participate in the determination of the corresponding parameters are used for hold-out verification. Each conditional subdomain is only held out pairwise if its parameter identifiability matrix still meets the identifiability condition after excluding the hold-out reference pair. If it is no longer identifiable after exclusion, it is first merged with the adjacent conditional subdomain with the closest conditional distance and the parameters are re-determined. If at least one independent hold-out reference pair cannot be formed after merging or the parameters are still not identifiable, the current conditional subdomain does not have independent verification conditions. The corresponding cross-reconstruction topology conflict remains unconfirmed in terms of physical ownership. The identifiability of the parameters is rechecked for each hold-out, and the hold-out image does not participate in the determination of the model parameters in this step.

[0060] For each prediction result, the grayscale prediction residual, edge position residual, and contrast preservation residual are calculated. The grayscale prediction residual is calculated pixel-by-pixel as the absolute difference between the predicted target grayscale and the true target grayscale. For the absolute difference within a block, a pre-calibrated high quantile value is taken, and then divided by the difference between the 90th and 10th quantile grayscale values ​​of the true target image block. When the dynamic range of the true target grayscale is 0, the predicted block cannot achieve effective grayscale normalization, indicating insufficient validation of the current conditional subdomain. The edge position residual uses the same patient spatial gradient calculation method as the local image conditions, and the same fixed gradient ridge is used for both the predicted and true images. Rule selection: First, non-maximum suppression is performed along the gradient direction. Then, the main gradient ridges are retained according to the pre-calibrated high quantile threshold of the gradient magnitude based on the historical real paired reconstruction data. The nearest patient spatial distances from the ridge point in the predicted image to the ridge point in the real image and from the ridge point in the real image to the ridge point in the predicted image are calculated respectively. The same high quantile value is taken for the distances in the two directions, and the larger one is selected. The residuals are compared and the absolute value of the difference between the gray levels of the 90th and 10th quantiles of the predicted image block and the difference between the corresponding gray level spans of the real image block is taken. Then, the residuals are divided by the gray level span of the real image block. The three types of residuals are compared with the corresponding allowable domains respectively and are not combined into a single weighted score.

[0061] The three types of residual allowable regions are formed independently by taking out historical real paired reconstruction data at the inspection level. For each historical inspection, the three types of residual summary values ​​are first obtained in the same way as the current model, and then the residual distribution is formed among different historical inspections. The upper limit of the allowable region adopts the high quantile level of 90% to 97.5%, preferably the 95th quantile value. The current model forms three types of summary values ​​for all currently taken out images with the same caliber. When all three items do not exceed the corresponding upper limit of the allowable region, the model passes the verification. If any item exceeds the upper limit, no valid model output is formed, and no compensation is made by changing the current anomaly to be judged, expanding the reference condition admission threshold, or increasing the model's degrees of freedom.

[0062] For each valid conditional subdomain, the target-side prediction error threshold is calculated using the paired real images that were not involved in the determination of the conditional subdomain parameters. The absolute gray level difference between the predicted target image and the real target image is obtained pixel by pixel. The high quantile value of 90% to 97.5% is taken for each held block, preferably the 95th quantile value. Then, the median value of the high quantile error of each held block in the same conditional subdomain is taken. The obtained target-side prediction error threshold maintains the original dimension of the target gray level in the coarse sampling reconstruction sequence. This threshold is only used to determine whether the target response generated by the limited gray level perturbation exceeds the normal prediction fluctuation of the model itself. It does not participate in the model parameter fitting, conditional subdomain division or effective scope determination. Conditional subdomains that cannot obtain an independent held prediction error threshold are not included in the attribution verification.

[0063] After the model passes the validation, the conditional support quantity is calculated position by position in the local conflict space, which is the maximum value of the conditional function value of each conditional subdomain. The lower limit of conditional support admission is determined by the low quantile value of the conditional support quantity of the independent true state. The low quantile level is preferably 5% to 15%, preferably 10%. When the conditional support quantity of the current target position reaches the lower limit of admission, and the support range of the fine-sampling reconstruction sequence image required by the model calculation is completely within the local conflict space, the position is determined as the effective scope.

[0064] The effective scope is formed point by point according to the patient's spatial location, and can be composed of multiple local regions that are separate from each other. If there is a lack of real local image conditions across the reconstruction reference set between two adjacent effective regions, the intermediate position is not automatically included in the effective scope through mathematical interpolation. The attribution verification module continues to process only when the target critical position is in the effective scope and the fine-sampling reconstruction sequence image support required by the model is complete.

[0065] The cross-reconstruction mapping module transmits the conditional local cross-reconstruction mapping model, effective scope, target-side prediction error threshold, and conditional gating state to the attribution verification module. The target critical position only enters attribution verification when the conditional support reaches the lower limit of admission, the position belongs to the effective scope, the required source image support is completely located in the local conflict space, and the corresponding grayscale mapping segment slope reaches the minimum resolvable transmission slope. If any condition is not met, the physical attribution remains unconfirmed. After entering attribution verification, the local image conditions, conditional subdomains, conditional distances, local response combination weights, and local spatial responses all retain the original fine-sampling reconstruction sequence state. The perturbed local image conditions are only used to determine whether it is still within the original conditional subdomain support boundary and are not used to switch conditional subdomains or recalculate combination weights.

[0066] The attribution verification module calls the cross-reconstruction topology conflict, the range of continuous effective threshold combinations, the stable anomaly core, the conditional local cross-reconstruction mapping model, the effective scope, and the target-side prediction error threshold. It then projects the range of continuous effective threshold combinations onto the threshold axis of the fine-sampling reconstruction sequence and the threshold axis of the coarse-sampling reconstruction sequence, respectively, to form stable threshold states on both sides.

[0067] For one-to-many cross-reconstruction topological conflicts, the attribution verification module determines the critical connectivity region. Based on the patient spatial location of different stable abnormal cores in the coarse-sampling reconstruction sequence, it forms mutually distinguishable corresponding regions in the continuous abnormalities of the fine-sampling reconstruction sequence. For each fine-sampling side stability threshold state, according to the currently uniform and fixed three-dimensional adjacency rules, it searches for candidate paths connecting different corresponding regions within the current abnormal connected region. For each candidate path, the minimum abnormal probability on the path is taken as the connectivity bottleneck. First, the maximum connectivity bottleneck among the candidate paths is obtained, and then the paths whose connectivity bottleneck and the maximum connectivity bottleneck do not exceed the allowable deviation of the connectivity bottleneck are retained. The bottleneck allowable deviation is set based on the dimension of the abnormal probability from 0 to 1, preferably 0.03 to 0.1, and more preferably 0.05. It is corrected by the bottleneck fluctuation under the repeated threshold state in the historical stable topological conflict samples. Then, the shortest patient space physical length is determined in the retained path, and the equivalent path with a physical length not exceeding the shortest length multiplied by 1 plus the path allowable elongation ratio is retained. The path allowable elongation ratio is preferably 0.05 to 0.15, and more preferably 0.1. Its value is determined based on the fluctuation of the actual connection path length under the repeated threshold state. The bottleneck allowable deviation and the path allowable elongation ratio are applied to the abnormal probability difference and the relative path length, respectively, and are not used interchangeably.

[0068] The connection corridor formed by the preserved path is mapped to the patient space. For each location within the connection corridor, the number of times that location enters the connection candidate in the fine-sampling side stable threshold state is counted. This number is then divided by the total number of fine-sampling side stable threshold states to obtain the connection participation quantity. Simultaneously, in the coarse-sampling side stable threshold state, the proportion of the same patient space location continuously existing in the separation gap between different target abnormalities is counted. When both the connection participation quantity and the target separation gap retention ratio reach the critical region retention admission threshold, the corresponding connected region is determined as the critical connection region. The critical region retention admission threshold is preferably 0.7 to 0.9, preferably 0.8. Historical parameter determination data are used to form verified stable critical regions and occasional connection regions that only appear in a few threshold states. The retention ratio of the two types of regions is calculated. For each candidate threshold, the sum of the error rejection ratio of the stable critical region and the error acceptance ratio of the occasional connection region is calculated. The candidate value with the smaller total error is selected, and the higher threshold is selected when they are equal. Finally, the data is verified by independent validation data.

[0069] For multiple-to-one cross-reconstruction topological conflicts, the attribution verification module determines the critical bridging region. Based on the patient spatial location of different stable abnormal cores in the fine-sampling reconstruction sequence, it obtains corresponding regions in the continuous abnormalities of the coarse-sampling reconstruction sequence that overlap with the patient spatial tolerance range of each stable abnormal core but do not overlap with each other. Each corresponding region maintains a unique correspondence with one stable abnormal core in the fine-sampling reconstruction sequence and serves as the target corresponding region for that stable abnormal core. When the critical bridging region does not belong to any target corresponding region and cannot form mutually distinguishable target corresponding regions, the physical attribution of the current cross-reconstruction topological conflict remains unconfirmed.

[0070] For each stable threshold state of the coarse sampling side, the abnormal connected regions of the current coarse sampling reconstruction sequence are constructed into a 3D adjacency graph, with abnormal voxels as nodes. Adjacency edges are established between nodes according to the unified and fixed 3D adjacency rules of the current examination. Nodes in regions corresponding to different targets are set as reserved nodes. Among other nodes, node cuts that separate regions corresponding to different targets are searched. The first cost of a node cut is the sum of the physical volumes of the patient space corresponding to the contained nodes. When there are node cuts with the same minimum volume, the positive margin obtained by subtracting the current threshold from the abnormal probability of each candidate node is calculated and summed. The node cut with the smaller total margin is selected as the priority node cut.

[0071] For each patient's spatial location, the number of times it enters the equivalent minimum node cut candidate in the coarse sampling side stable threshold state is counted. This number is then divided by the total number of coarse sampling side stable threshold states to obtain the bridging participation amount. When the bridging participation amount reaches the critical region maintaining admission threshold, and when the corresponding region is excluded from the set of connectable nodes in the 3D adjacency graph under the baseline threshold state of the coarse sampling reconstruction sequence, forming the same separation relationship as the fine sampling reconstruction sequence, the region is identified as the critical bridging region. When there are multiple parallel bridging pathways, the critical bridging region consists of multiple local regions that jointly block the parallel bridging pathways. When there are multiple equivalent regions that are separate from each other and whose main bridging mechanism cannot be determined, the attribution result is determined as physical attribution unconfirmed.

[0072] When multiple node cuts exist in the same stable threshold state on the coarse sampling side with the same minimum physical volume and sum of abnormal response margins, the patient space union of these node cuts is used as the equivalent bridging candidate range for that threshold state. If the bridging participation reaches the critical region and the region that maintains the admission threshold still cannot block all parallel bridging pathways under the baseline threshold state, the current multi-pair one-span reconstruction topological conflict remains unconfirmed in terms of physical attribution; the search will not continue by expanding the critical bridging region, modifying the stable abnormal core, or changing the preceding threshold.

[0073] After the critical connection region or critical bridging region is formed, baseline topology replay is performed first. The target verification range is determined with the separation gap or critical bridging region of the corresponding coarse sampling reconstruction sequence as the center. Processing continues only if the range is completely within the effective scope and the fine sampling reconstruction sequence image support required for model calculation is completely within the local conflict space; otherwise, the physical attribution remains unconfirmed.

[0074] The original fine-sampling reconstruction sequence without disturbance is input into the conditional local cross-reconstruction mapping model. The conditional gating state determined by the original fine-sampling reconstruction sequence is maintained to obtain the baseline prediction image. The baseline prediction anomaly probability field is formed only when the conditional spatial response at the corresponding position of the baseline prediction image is within the coverage of the real reference gray-level node, the gray-level mapping segment slope reaches the minimum resolvable transfer slope, and the conditional subdomain has passed the three types of residual verification. Otherwise, the physical attribution remains unconfirmed. For each stable threshold state on the coarse sampling side, the connection state of the corresponding region in the baseline prediction anomaly probability field and the real coarse-sampling reconstruction sequence anomaly probability field are compared. The number of threshold states with consistent connection states is counted and then divided by the total number of stable threshold states on the coarse sampling side to obtain the baseline topology replay rate.

[0075] When the baseline topology replay rate reaches the baseline topology replay admission threshold, and the baseline threshold of the coarse-sampled reconstruction sequence is within the maximum continuous threshold range where the predicted connectivity state and the actual connectivity state are consistent, finite gray-scale perturbation is entered. The baseline topology replay admission threshold is preferably 0.75 to 0.9, and preferably 0.8. In the historical parameter determination data, local mappings that can continuously reproduce the true coarse-sampled connectivity state near the baseline threshold are taken as stable replay samples, and mappings that only coincidentally match the discrete threshold or are not within the continuous consistency range of the baseline threshold are taken as unstable replay samples. The sum of the stable replay error rejection ratio and the unstable replay error acceptance ratio is calculated for the candidate threshold. The candidate value with the smaller total error is taken and verified on independent validation data. When the baseline topology replay rate does not reach the admission threshold, or the baseline threshold of the coarse-sampled reconstruction sequence is not within the maximum continuous threshold range, the attribution result remains unconfirmed, and the baseline topology replay result and the corresponding medical image are saved.

[0076] After baseline topology replay is successful, the local response sensitivity of the source location to the target critical location is calculated under fixed-condition gating. For each target critical location, the combined weights of the local responses of each effective conditional subdomain determined by the original fine-sampling reconstruction sequence are read first. Then, for one source location to be evaluated, the discrete spatial kernel coefficients of each local spatial response at the corresponding relative position are read according to the patient spatial relative displacement between the source location and the target critical location. Each spatial kernel coefficient is multiplied by the corresponding local response combined weight and summed to obtain the spatial transfer amount from the source location to the current target critical location. Finally, the actual slope of the piecewise linear grayscale mapping on the segment where the baseline response is located is read. The spatial transfer amount is multiplied by the slope to obtain the local response sensitivity of the source location to the target critical location. When the source location is not within the physical support range of any invoked local spatial response, its spatial transfer amount is 0. The local response sensitivity is only used to screen candidate source locations and determine the initial response direction. The baseline slope is not used for linear extrapolation of larger perturbations. The final determination of pseudo-splitting and pseudo-fusion is based on the target predicted response and topology state recalculated by the fully conditionalized local cross-reconstruction mapping model for each effective perturbation level.

[0077] Simultaneously, the local prediction error threshold of the current target position is obtained by multiplying the target-side prediction error threshold of each effective condition subdomain and the weight of the original local response combination. After applying a finite gray-level perturbation to the candidate source region, the absolute change of the target position relative to the baseline prediction response is calculated. When the absolute change is greater than the local prediction error threshold, it is recorded as a distinguishable source response. When the response direction conforms to the corresponding pseudo-splitting or pseudo-fusion verification direction, but the absolute change is not greater than the local prediction error threshold, the absence of significant change is not interpreted as the source region not contributing, and the attribution result remains unconfirmed.

[0078] The attribution verification module determines the conditional gating state based on the original fine-sampling reconstruction sequence and applies a finite gray-level perturbation. The mathematical range of the perturbation ratio is 0 to 1. The actual replay uses a discrete perturbation ratio from low to high. The perturbation step size is preferably 0.05 to 0.10, with 0.05 being the preferred value. Smaller step sizes are used to avoid crossing the first effective connection change position, while larger step sizes are only used when the verification data shows that the connection state changes smoothly with the perturbation. For each position in the candidate source region, the gray-level difference is obtained by subtracting the original gray-level from the target gray-level. Then, the gray-level difference is multiplied by the current perturbation ratio, and the result is added to the original gray-level to form the current perturbation gray-level. Each perturbation level uses the local image conditions, conditional distance, local response combination weights, and local spatial response determined by the original fine-sampling reconstruction sequence. The conditional subdomain is not reselected. The perturbed local image conditions are only used to check whether the current perturbation is still within the effective range of the original conditional subdomain.

[0079] The maximum effective perturbation ratio is determined by the first failure condition that appears in the current validation chain. No fixed optimal value is set across patients; the perturbation ratio is adjusted level by level, and the following checks are made sequentially: whether the perturbation grayscale is still within the grayscale range between the current patient's true local background and the stable abnormal subject; whether the perturbed local image conditions are still within the support range of the original conditional subdomain; whether the conditional spatial response of the target critical position is still within the range of the true reference grayscale node; whether the grayscale mapping segment slope reaches the minimum resolvable transfer slope; whether the image support of the fine-sampling reconstruction sequence required for model calculation is completely within the local conflict space; and whether the target validation range is completely within the effective scope. When any failure condition appears for the first time, the perturbation ratio immediately preceding the failure level that still satisfies all conditions is taken as the maximum effective perturbation ratio. When the first positive perturbation level, i.e., when the failure condition is triggered, the maximum effective perturbation ratio is set to 0, and the physical attribution remains unconfirmed. Subsequent pseudo-splitting or pseudo-fusion determinations only use connection state changes not exceeding the maximum effective perturbation ratio; changes exceeding this boundary are not used as the basis for determining attribution.

[0080] For multiple-to-one-span reconstruction topological conflicts, a local source region facing the critical bridging region is determined within the stable abnormal core of each fine-sampling reconstruction sequence. First, a local background sampling area is selected in the same type of anatomical tissue that maintains spatial separation from other abnormal candidates around the corresponding stable abnormal core. The robust median value of the grayscale of the local background sampling area is used as the local background reference value. The direction of grayscale change of the current source position within the stable abnormal core towards the local background reference value is used as the attenuation direction. The response of each source position to each target position in the critical bridging region is calculated using local response sensitivity along the attenuation direction. Only source positions whose absolute value of the target response exceeds the local prediction error threshold and whose response direction points to the weakened bridging are retained. According to the three-dimensional adjacency rules that are currently uniformly fixed in the examination, adjacent retained positions are merged to form a bridging source contribution area. The bridging source contribution area is limited within the corresponding stable abnormal core and does not expand outside the stable abnormal core.

[0081] Finite attenuation and joint finite attenuation are performed on the bridging source contribution regions from different stable anomaly cores respectively. The joint finite attenuation uses the minimum of the maximum effective perturbation ratios of each bridging source contribution region in the combination as the common perturbation upper limit. Each perturbation level forms the predicted response and anomaly probability field of the corresponding region of the coarse sampling reconstruction sequence under fixed condition gating. The connection relationship of different corresponding regions is checked in the stable threshold state on the coarse sampling side. When the stable anomaly main body of different corresponding regions still exists, the critical bridging region loses its bridging effect, and the separation position is still located in the original critical bridging region and no substitution fracture is formed at a position far away from the critical bridging region, the current threshold state is recorded as the effective bridging reversal state.

[0082] Divide the number of effective bridging inversion states by the total number of stable threshold states on the coarse sampling side to obtain the topology inversion preservation ratio of the current source contribution combination under the current perturbation ratio. When the topology inversion preservation ratio reaches the topology inversion preservation admission threshold and the maximum continuous effective threshold range includes the baseline threshold of the coarse sampling reconstruction sequence, the current source contribution combination is recorded as an effective combination. The topology inversion preservation admission threshold is preferably 0.75 to 0.9, preferably 0.8. Its historical calibration follows the same check-level data isolation method as the baseline topology replay admission threshold, and separately calculates the preservation ratio of repeatable bridging inversion and occasional bridge breakage, and selects the candidate value with the smaller sum of the two error ratios.

[0083] For valid combinations, they are first compared from the fewest to the most spatially separated source contribution regions, and the combination with the fewest members is retained. For combinations with the same number of members, the perturbation ratios at which they first reach the stable bridging reversal condition in the common perturbation ratio from low to high are compared, and the combination with the smaller minimum effective perturbation ratio is retained. If there are still multiple combinations containing anomalous members from different sources with the same number of members and the same minimum effective perturbation ratio, it is considered that there are multiple equivalent formation mechanisms, and the current assignment result remains unconfirmed in terms of physical assignment. The unique retained combination is then verified one by one for its true sub-combinations. If any true sub-combination can form the same stable bridging reversal, the same screening is performed on that true sub-combination until the smallest multi-source contribution combination that cannot be reduced further is obtained.

[0084] The minimum multi-source contribution combination must contain at least two spatially separated bridging source contribution regions from stable abnormal cores of different fine-sampling reconstruction sequences, and the joint attenuation must be within the range of the maximum effective perturbation ratio. When the minimum multi-source contribution combination satisfies the constraints of spatial separation, number of members, and maximum effective perturbation ratio, the attribution result is determined to be a multi-source mixed pseudo-fusion. The attribution verification module simultaneously saves the patient spatial correspondence between the critical bridging region, the minimum multi-source contribution combination, each bridging source contribution region, and the abnormal region of the fine-sampling reconstruction sequence and the corresponding region of the coarse-sampling reconstruction sequence, so that the analysis and layout module can establish the same-group relationship and the separation relationship.

[0085] The minimum multi-source contribution combination only forms a separation relationship for the abnormal regions of the fine-sampling reconstruction sequence that actually enter the combination; the abnormal regions that do not enter the combination in the same many-to-one cross-reconstruction topological conflict continue to maintain their original cross-reconstruction correspondence state.

[0086] For one-to-many cross-reconstruction topological conflicts, a local background sampling area is selected in the periphery of the critical connection region and in the same type of anatomical tissue that is spatially separated from the abnormal candidate. The robust median value of the gray level of the local background sampling area is used as the local background reference value. At the ends of the critical connection region that are connected to different stable abnormal subjects, the low quantile and high quantile values ​​of the stable abnormal gray level are obtained from the corresponding stable abnormal core to form the stable abnormal gray level range of each end. The low quantile level is preferably 5% to 15%, preferably 10%; the high quantile level is preferably 85% to 95%, preferably 90%. Both quantile ranges are corrected by the gray level fluctuation under the same stable abnormal core repeated threshold state in the historical validation data. The gray level range with the same abnormal contrast direction relative to the local background reference value as the critical connection region is retained. Then, the common intersection of the stable abnormal gray level ranges of each end is taken to form a common reachable gray level range. When the common reachable gray level range is empty, the attribution result remains that the physical attribution is not confirmed.

[0087] When a common reachable grayscale range exists, the common reachable boundary value closest to the current grayscale state of the critical connection region is selected as the target grayscale. For each source location in the critical connection region, the difference between the target grayscale and the original grayscale is used as the direction of change. The predicted response of the source location to the separation gap of the coarse sampling reconstruction sequence is calculated using the local response sensitivity under the fixed condition gating state. Only when the predicted response within the separation gap forms a continuous region in the patient space, the response direction points to the abnormal connection recovery, and the absolute value of the response exceeds the local prediction error threshold, does the connection recovery replay begin. If any of the three conditions is not met, the physical attribution remains unconfirmed, and the search is not continued by changing the target grayscale direction.

[0088] The grayscale of the critical connection region is adjusted step by step within the range of 0 to the maximum effective perturbation ratio. Each perturbation level maintains the conditional gating state. The coarse sampling reconstruction sequence is re-formed to predict the anomaly probability field. The connection status of the original separated anomaly region is checked in the stable threshold state on the coarse sampling side. When the stable anomaly body still exists, the new connection passes through the separation gap corresponding to the critical connection region, the connection position is consistent with the distinguishable response region, and no alternative connection is formed at a position far away from the critical connection region, the current threshold state is recorded as the effective connection recovery state.

[0089] Divide the number of effective connection recovery states by the total number of stable threshold states on the coarse sampling side to obtain the connection recovery and maintenance ratio under the current perturbation ratio. When the connection recovery and maintenance ratio reaches the topology inversion maintenance admission threshold and the maximum continuous effective threshold range includes the baseline threshold of the coarse sampling reconstruction sequence, record the current perturbation ratio. Search from low to high perturbation ratio and take the perturbation ratio that first meets the stable connection recovery condition as the minimum effective perturbation ratio. When the minimum effective perturbation ratio is within the range of the maximum effective perturbation ratio, the assignment result is determined to be a single-source connection decay type pseudo-split, and save the critical connection region and its correspondence with the continuous abnormal region of the fine sampling reconstruction sequence and the separation abnormal region of the coarse sampling reconstruction sequence.

[0090] The finite gray-level perturbation of the single-source connection decay pseudo-split only applies to the critical connection region. The remaining fine-sampled reconstructed sequence images required for the model to calculate the response in this region support the preservation of the original gray level. The perturbation prediction image is only used as an intermediate result for attribution verification, and the medical image layout only references the real medical images in the current examination.

[0091] When the following conditions occur: the critical region cannot be stably formed, the baseline topology replay fails, the target validation range or source image support exceeds the limit, the grayscale mapping segment slope is lower than the minimum resolvable transfer slope, the candidate source response does not exceed the local prediction error threshold, the required connection change exceeds the maximum effective perturbation ratio, the stable abnormal subject disappears before the connection change, the connection change occurs outside the critical region, the minimum multi-source contribution combination cannot be established, or the commonly reachable grayscale range is empty, the attribution result remains physically unconfirmed. Each unconfirmed branch stops relying on the subsequent deterministic attribution judgment of the corresponding failure condition, and does not obtain a definite result by relaxing the previous admission parameters, switching the condition gating state, or introducing other patient data.

[0092] When the physical attribution is not confirmed, the established stable abnormal core, cross-reconstruction topological conflict, critical region candidate, source region candidate, corresponding real medical image and exit reason are saved and transferred to the analysis and layout module. The admission threshold, cross-reconstruction reference set, effective scope, conditional gating state and maximum effective perturbation ratio determined in the previous stage are not adjusted due to the unconfirmed result.

[0093] The analysis and layout module establishes grouping and separation relationships based on cross-reconstruction correspondence and attribution results. In this embodiment, a unilateral abnormal object refers to an abnormal region in a certain set of reconstruction sequences that has formed a stable abnormal core and has a unique patient spatial location identifier. In the case of multiple-to-one cross-reconstruction topological conflicts, the corresponding region of the coarse-sampled reconstruction sequence forms a unilateral abnormal object only when it can be uniquely associated with the stable abnormal core of the fine-sampled reconstruction sequence through patient spatial correspondence. Each unilateral abnormal object represents only one locatable abnormal region in a set of reconstruction sequences and is not merged across reconstructions. Critical connection regions, critical bridging regions, bridging source contribution regions, and minimum multi-source contribution combinations are all used as relational evidence and are not used as physical abnormal objects on their own.

[0094] Based on the stable one-to-one cross-reconstruction correspondence, the anomalous regions of the fine-sampling reconstruction sequence are grouped with the corresponding anomalous regions of the coarse-sampling reconstruction sequence. Based on the single-source connection attenuation pseudo-splitting attribution results, the continuous anomalous regions of the fine-sampling reconstruction sequence are grouped with the separated anomalous regions of the coarse-sampling reconstruction sequence actually interpreted by the attribution results. Based on the multi-source hybrid pseudo-fusion attribution results, each anomalous region of the fine-sampling reconstruction sequence is grouped with its corresponding region of the coarse-sampling reconstruction sequence. Furthermore, a separation relationship is established between anomalous regions from stable anomalous cores of different fine-sampling reconstruction sequences in the minimum multi-source contribution combination. Critical bridging regions maintain shared bridging evidence attributes.

[0095] Unconfirmed physical attribution records only retain the unconfirmed status, anomalous area, and source of evidence, without directly establishing new grouping or separation relationships. When an unconfirmed anomalous object forms a cross-reconstruction association through other independent and confirmed grouping relationships, it only participates in the formation of candidate physical anomalous groups based on that independent grouping relationship, and the original unconfirmed record continues to be retained and stored separately from that independent relationship.

[0096] Before transferring grouped relationships, check whether the same object pair has both grouped and separate relationships. If there is a direct conflict, keep the evidence of the involved objects and relationships as "physical ownership unconfirmed" and do not use the number of evidence, the confidence level of the large artificial intelligence model, or the order of relationship generation to eliminate the conflict. Only when there is no direct conflict will the transfer of grouped relationships be performed.

[0097] After excluding direct relationship conflicts, a candidate set is established starting from each unilaterally anomalous object with a definite group relationship. Other unilaterally anomalous objects that have a definite group relationship with any unilaterally anomalous object in the candidate set but have not yet entered the candidate set are successively retrieved and added until no new objects are added in one round of retrieval. The resulting set forms a candidate physical anomaly group. Unilaterally anomalous objects lacking definite group relationship support remain physically unconfirmed and do not form a confirmed physical anomaly group due to the absence of a separation relationship. After the candidate physical anomaly group is formed, a pairwise check is performed to see if a definite separation relationship exists between the unilaterally anomalous objects within the group. If no definite separation relationship exists, it is identified as a physical anomaly group. If a definite separation relationship exists, the objects involved in the conflict path remain physically unconfirmed. Each confirmed physical anomaly group is configured with a unique physical anomaly group identifier within the current check, and the anomaly region, the source of the group relationship, and the source of the separation relationship within the group are saved.

[0098] In coarse-sampled reconstruction sequence continuous abnormalities corresponding to multi-source hybrid pseudo-fusion, the corresponding regions that have established a group relationship with the abnormal regions of fine-sampled reconstruction sequences are included in the corresponding physical abnormality group; the critical bridging region is jointly referenced by the relevant physical abnormality groups and maintains the shared bridging evidence attribute. The coarse-sampled reconstruction sequence transition region that cannot be uniquely associated with a single stable abnormality core of a fine-sampled reconstruction sequence is only displayed as shared evidence and is not assigned to any confirmed physical abnormality group, nor is it forcibly assigned based on the spatial distance of the nearest patient.

[0099] When the same object participates in multiple cross-reconstruction topology conflicts, objects that can be identified as having continuous associations within the same group and no separation conflicts are classified into the same physical anomaly group; when the same object has both group and separation conflicts, its physical affiliation remains unconfirmed.

[0100] The analysis and layout module selects real medical images from the current examination based on the abnormal and critical regions corresponding to the physical abnormality group. It establishes a set of evidence elements for each physical abnormality group, where stable abnormal subjects, critical connection regions, critical bridging regions, and corresponding abnormal regions involved in grouping or separation relationships are each used as independent evidence elements. The patient space enclosing range of each evidence element is saved. Candidate medical images are obtained only from the fine-sampling reconstruction sequence and coarse-sampling reconstruction sequence that form the physical abnormality group, without calling other examinations or generative completion images.

[0101] For each 3D evidence element, calculate its patient space overlap area with each real slice of the corresponding reconstructed sequence, and select the real slice with the largest overlap area as the candidate main cross-section for that evidence element. If multiple slices have the same maximum overlap area, they are all retained as candidates. In the candidate slices, the projection range of the evidence element on the image plane should be completely within the original image range. When used to display connection or separation relationships, it is also required that the projections of abnormal regions and critical regions involved in the relationship can be displayed together in the same real image. If one image cannot cover both ends of the relationship and the critical region at the same time, real images that can cover the main cross-section of the corresponding evidence element are selected respectively, and a complementary display relationship is established through the same patient space coordinates.

[0102] Among all candidate real medical images, the image that covers the largest number of uncovered evidence elements is first selected and added to the key medical image set. When there are candidate images with the same number of coverages, the image with the larger total overlap area of ​​the patient space of the covered evidence elements is selected first. This process is repeated until each evidence element is covered by at least one real image. Then, the selected images are deleted one by one. Images that still cover all evidence elements after deletion are removed as redundant images. When the existing real medical images cannot fully display a certain necessary evidence or relationship, the state of incomplete evidence display is recorded, and the existing physical abnormality group and object relationship remain unchanged.

[0103] After physical anomaly groups are formed, an independent AI large-scale model input record is established for each physical anomaly group. The input record uses fixed structured fields, including at least the physical anomaly group identifier, the anomaly region identifier within the group, the established status of same-group and separation relationships, shared bridging evidence references, and key medical images. Each field directly reads the object and image identifiers already formed by the analysis and layout module, without being inferred again by the AI ​​large-scale model. The structured fields are serialized as read-only input before calling the AI ​​large-scale model. The large-scale model output interface does not include fields for adding physical anomaly group identifiers, modifying same-group relationships, or modifying separation relationships.

[0104] This implementation does not perform training, fine-tuning, or parameter updates on the large AI model. The large AI model, as an existing inference component, only generates image representations, differences in representations of different reconstruction sequences, structural features, and key medical image descriptions within the boundaries of the corresponding physical anomaly group. After the large model outputs back, the analysis and layout module associates the corresponding text according to the input physical anomaly group identifier. If the output text contains the number of newly added objects, object merging, or object splitting conclusions, only the medical image descriptions within the group that can correspond one-to-one with the input physical anomaly group identifier are retained, and the relationships of newly added objects are not written back to the same group relationship, separation relationship, or physical anomaly group.

[0105] For objects with unconfirmed physical attribution, an independent large model input record is established, which only records the corresponding real medical image, the location of the unilateral abnormal object, cross-reconstruction topological conflict, candidate evidence of the formed critical region, and the status of unconfirmed physical attribution. The object identity and object quantity fields are fixed as unconfirmed physical attribution and do not provide the large artificial intelligence model with a certain number of physical abnormal groups. The output of the large model is only used as a local medical image description and is not used to generate a certain group relationship or separation relationship.

[0106] Stable one-to-one cross-reconstruction correspondence and the same physical anomaly group formed by single-source connection attenuation pseudo-splitting are displayed in the same page unit, showing the corresponding anomalous regions on both sides and related critical connection evidence; different physical anomaly groups formed by multi-source hybrid pseudo-fusion are set in separate page units, and the critical bridging region and related coarse sampling reconstruction sequence context are displayed in a shared bridging evidence unit. The shared bridging evidence unit does not form new physical anomaly groups.

[0107] When a single two-dimensional medical image cannot fully cover shared bridging evidence and related abnormal areas, multiple complementary real medical images are used for joint display; when the existing real medical images still cannot fully cover the necessary evidence, the incomplete evidence display status is recorded and no supplementary medical structure is generated.

[0108] The abnormal objects in the AI ​​large model analysis objects and medical image layouts use the same physical anomaly group identifier. After the layout is formed, the physical anomaly group identifier, abnormal region identifier within the group, relationship status, and shared bridging evidence reference in the large model input records are checked item by item to see if they are consistent with the layout reference. If the identifier or reference is inconsistent, only the corresponding large model input records are regenerated, key medical images are reselected, or the layout is adjusted. The calculation of stable anomaly cores, cross-reconstruction topological conflicts, conditional local cross-reconstruction mapping models, or attribution results are not re-executed. When the existing real medical images cannot supplement the necessary evidence, the incomplete evidence display state is retained, and the physical anomaly group and its relationship are not modified.

[0109] The final output includes confirmed physical anomaly groups, unidentified physical attributions of objects, analysis results within the large artificial intelligence model group, key medical image sets, and medical image layouts.

[0110] Cross-reconstruction topological conflicts are constrained by multi-threshold joint stability, cross-reconstruction correspondences are constrained by bidirectional patient space coverage and boundary distance, and conditional local cross-reconstruction mapping models are constrained by real paired images in the current patient fixed input snapshot, constrained local spatial responses, and independent output validation. Finite gray-scale perturbations are jointly constrained by fixed conditional gating states, effective scope, target-side prediction error thresholds, and maximum effective perturbation ratios. Physically abnormal groups are formed by determining same-group relationships, separation relationships, and conflict checks. The patient space range, input data version, and current examination time boundary of each object are fixed in the input stage. Subsequent steps only call objects and models that have been formed and approved in the previous stage. When any input is missing, source conflicts occur, conditional support is insufficient, model validation fails, source responses are indistinguishable, or object relationships conflict, the physical attribution remains unconfirmed, and low-confidence results are not written back to determine object relationships.

[0111] In alternative implementations, the conditional subdomains can employ distance-based clustering, hierarchical clustering, or local image conditional interval grouping; independent hold-out validation can employ pairwise hold-out, hierarchical hold-out, or cross-validation that meets the same data isolation requirements; and the local spatial response can employ discrete three-dimensional basis functions, separable spatial basis functions, or finite-support convolutional templates. All alternative methods maintain the current patient's real reference data limitation, response total normalization, unchanged response center, finite support of patient space, and fixed conditional gating processing boundaries.

[0112] The anomaly region extraction model, the large-scale artificial intelligence model, and the layout rendering component are implemented using existing engineering methods that match the system's data interface. The anomaly region extraction model has completed deployment and verification of two types of reconstruction sequences before entering this system. The large-scale artificial intelligence model only receives input from the structured physical anomaly groups formed by the analysis and layout module. The layout rendering component performs conventional graphic and text layout according to the key medical image set and physical anomaly group identifiers. The cross-reconstruction attribution results of this implementation do not depend on the internal training methods or rendering algorithms of the existing components.

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A medical image automatic analysis and typesetting system based on a large artificial intelligence model, characterized in that, include: The medical image processing module acquires the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence from the same original acquisition. It forms a stable abnormal core based on the preservation relationship under threshold perturbation, establishes a cross-reconstruction correspondence based on the patient's spatial correspondence, cross-combines the threshold states of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence to determine the cross-reconstruction topological conflict, and selects paired images that are separated from the spatial abnormality to be judged in the neighborhood of the cross-reconstruction topological conflict to form a cross-reconstruction reference set. The cross-reconstruction mapping module establishes a conditional local cross-reconstruction mapping model from fine-sampling reconstruction sequence to coarse-sampling reconstruction sequence based on the cross-reconstruction reference set, and determines the effective scope based on local image conditions. The attribution verification module determines the critical connection region and critical bridging region for one-to-many and many-to-one cross-reconstruction topological conflicts, respectively. It determines the conditional gating state based on the original fine-sampling reconstruction sequence, applies finite gray-level perturbation to the image region corresponding to the fine-sampling reconstruction sequence, and determines the attribution result based on the connection state change of the region corresponding to the coarse-sampling reconstruction sequence. The attribution result is one of the following: single-source connection attenuation pseudo-splitting, multi-source hybrid pseudo-fusion, or physical attribution unconfirmed. The analysis and layout module establishes group relationships and separation relationships based on cross-reconstruction correspondence and attribution results. Through relationship propagation and conflict verification, physical anomaly groups are formed. The physical anomaly groups limit the analysis objects of the artificial intelligence big model, and medical image layouts are formed based on the medical images of the physical anomaly groups.

2. The medical image automatic analysis and typesetting system based on a large artificial intelligence model according to claim 1, characterized in that, The specific steps for forming a stable anomaly core based on the preservation relationship under threshold perturbation are as follows: Abnormal regions are formed within the threshold perturbation intervals of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence, respectively. Abnormal candidate trajectories are formed based on the continuous correspondence between abnormal regions under adjacent threshold states. Calculate the threshold maintenance value at each position in the abnormal candidate trajectory, and determine the region where the threshold maintenance value reaches the stable core admission threshold as the stable abnormal core.

3. The automatic medical image analysis and typesetting system based on a large artificial intelligence model according to claim 2, characterized in that, The specific operation for establishing cross-reconstruction correspondence based on patient spatial correspondence and determining cross-reconstruction topological conflicts by cross-combining the threshold states of fine-sampling reconstruction sequences and coarse-sampling reconstruction sequences is as follows: The combination of stable anomaly cores with the fewest members is determined based on the physical coverage and boundary correspondence of the stable anomaly cores. Cross-reconstruction topological conflicts are determined by cross-combining the threshold states of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence, and by ensuring that the stable core members of the anomaly and the cross-reconstruction correspondence remain consistent within a range of continuous threshold combinations including the baseline thresholds of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence. A cross-reconstruction reference set is formed by selecting real paired images that are spatially separated from the anomalies to be judged and whose local image conditions match in the cross-reconstruction topological conflict neighborhood.

4. The automatic medical image analysis and typesetting system based on a large artificial intelligence model according to claim 3, characterized in that, The specific steps for establishing a conditional local cross-reconstruction mapping model based on the cross-reconstruction reference set are as follows: Based on the local spatial responses corresponding to different local image conditions formed by real paired images in the cross-reconstruction reference set, a conditional local cross-reconstruction mapping model is established under the constraints of normalizing the total response and keeping the response center unchanged. The model parameters are determined by the cross-reconstruction reference set.

5. The medical image automatic analysis and typesetting system based on a large artificial intelligence model according to claim 4, characterized in that, The specific operation for determining the effective scope based on local image conditions is as follows: The conditional local cross-reconstruction mapping model is validated using real paired images independently taken from the cross-reconstruction reference set. The effective scope is determined by the position of the local image condition within the condition range covered by the cross-reconstruction reference set, and the image range of the fine-sampling reconstruction sequence required by the model calculation is within the cross-reconstruction topological conflict neighborhood. The conditional gating state is determined based on the original fine-sampling reconstruction sequence.

6. The automatic medical image analysis and typesetting system based on a large artificial intelligence model according to claim 5, characterized in that, The specific steps for determining the critical connectivity region are as follows: For a one-to-many cross-reconstruction topological conflict, the connection range of the fine-sampling reconstruction sequence region corresponding to different abnormal regions in the coarse-sampling reconstruction sequence is determined in the fine-sampling reconstruction sequence across the stable threshold state, and the part that continuously corresponds to the separation gap of the coarse-sampling reconstruction sequence is determined as the critical connection region.

7. The automatic medical image analysis and typesetting system based on a large artificial intelligence model according to claim 5, characterized in that, The specific steps for determining the critical bridging region are as follows: For multiple-to-one cross-reconstruction topological conflicts, the bridging range of regions in the coarse-sampled reconstruction sequence corresponding to different abnormal regions in the fine-sampled reconstruction sequence that maintain the cross-stable threshold state in the coarse-sampled reconstruction sequence is determined. The region that forms the same separation relationship as the fine-sampled reconstruction sequence after excluding abnormal connectivity is determined as the critical bridging region.

8. The automatic medical image analysis and typesetting system based on a large artificial intelligence model according to claim 7, characterized in that, The specific steps for determining the conditional gating state based on the original fine-sampling reconstruction sequence and applying a finite gray-level perturbation to determine the attribution result are as follows: The undisturbed fine-sampled reconstruction sequence is input into the conditionalized local cross-reconstruction mapping model to form a predicted connectivity state within the stable threshold state range of the coarse-sampled reconstruction sequence. When the predicted connection status and the actual connection status reach the preset consistency condition, the conditional gating state is maintained within the effective scope and a limited grayscale perturbation is applied; if the preset consistency condition is not reached, it is determined that the physical attribution is not confirmed. For a one-to-many span reconstruction topological conflict, when the critical connection region corresponds to the stable recovery of the connection through the separation gap, it is determined to be a single-source connection attenuation type pseudo-split. For multiple pairs of one-span reconstruction topological conflicts, joint perturbation is performed on the anomalous regions of the spatially separated fine-sampling reconstruction sequence to determine the minimum multi-source contribution combination that can stably disconnect the critical bridging region, and it is determined to be a multi-source hybrid pseudo-fusion. If neither the single-source connection attenuation type pseudo-splitting determination condition nor the multi-source hybrid type pseudo-fusion determination condition is met, the physical attribution is determined to be unconfirmed.

9. The automatic medical image analysis and typesetting system based on a large artificial intelligence model according to claim 8, characterized in that, The specific steps for establishing same-group and separation relationships and forming physical anomaly groups based on cross-reconstruction correspondence and attribution results are as follows: Based on the cross-reconstruction correspondence, establish a group relationship for the stable one-to-one cross-reconstruction corresponding abnormal regions; Establish a group relationship between the continuous abnormal regions of the fine-sampling reconstruction sequence corresponding to the single-source connection attenuation pseudo-splitting and the separated abnormal regions of the coarse-sampling reconstruction sequence. Establish a group relationship between the anomalous regions of the fine-sampling reconstruction sequence corresponding to the multi-source hybrid pseudo-fusion and the corresponding regions of the coarse-sampling reconstruction sequence, and establish a separation relationship between the anomalous regions of different fine-sampling reconstruction sequences corresponding to the minimum multi-source contribution combination. Abnormal regions that are continuously associated through the same group relationship are grouped into the same candidate physical anomaly group. When the candidate physical anomaly group passes the separation relationship conflict check, it is determined to be a physical anomaly group. When the conflict check fails, it is determined to be a physical attribution unconfirmed. Both the objects analyzed by the large-scale artificial intelligence model and the abnormal objects in the medical image layout are identified by physical anomaly groups. Medical images are selected to form the medical image layout based on the abnormal regions, critical connection regions, and critical bridging regions corresponding to the physical anomaly groups.

10. A method for automatic analysis and typesetting of medical images based on a large-scale artificial intelligence model, using the automatic analysis and typesetting system for medical images based on a large-scale artificial intelligence model as described in any one of claims 1-9, characterized in that, include: The fine-sampling reconstruction sequence and coarse-sampling reconstruction sequence of the same original acquisition are obtained. A stable abnormal core is formed according to the preservation relationship under threshold perturbation. A cross-reconstruction correspondence is established based on the patient spatial correspondence. The threshold states of the fine-sampling reconstruction sequence and the coarse-sampling reconstruction sequence are cross-combined to determine the cross-reconstruction topological conflict. Paired images that are separated from the spatial abnormality to be judged are selected in the neighborhood of the cross-reconstruction topological conflict to form a cross-reconstruction reference set. A conditional local cross-reconstruction mapping model from fine-sampling reconstruction sequence to coarse-sampling reconstruction sequence is established based on the cross-reconstruction reference set, and the effective scope is determined based on local image conditions. For one-to-many and many-to-one cross-reconstruction topological conflicts, critical connection regions and critical bridging regions are determined respectively. The conditional gating state is determined based on the original fine-sampling reconstruction sequence. A limited gray-level perturbation is applied to the image region corresponding to the fine-sampling reconstruction sequence. The attribution result is determined based on the connection state change of the region corresponding to the coarse-sampling reconstruction sequence. The attribution result is one of the following: single-source connection attenuation pseudo-splitting, multi-source hybrid pseudo-fusion, or physical attribution unconfirmed. Based on the cross-reconstruction correspondence and attribution results, same-group and separation relationships are established. Through relationship propagation and conflict verification, physical anomaly groups are formed. The physical anomaly groups limit the analysis objects of the artificial intelligence large model, and medical image layouts are formed based on the medical images of the physical anomaly groups.

Images