An ultra-wide-angle fundus multi-modal medical image intelligent diagnosis and data management system

Abstract

The application relates to the technical field of medical image processing, in particular to an ultra-wide-angle fundus multi-modal medical image intelligent diagnosis and data management system, which comprises an image distortion correction module, a multi-modal fusion analysis module, a dense task scheduling module and an abnormal takeover response module. The system generates an image quality label by extracting distortion gradient, blood vessel curvature change and lesion boundary features, realizes diagnosis weight distribution and task scheduling optimization in combination with the correlation requirements of retinal blood vessel distribution characteristics and clinical indexes, and completes task replacement dispatching and collaborative grouping through an abnormal takeover mechanism. The application can improve the diagnosis accuracy and efficiency of the ultra-wide-angle fundus image, and construct a closed-loop diagnosis control system covering the whole process of identification, matching, adjustment, replacement and grouping.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an intelligent diagnostic and data management system for ultra-wide-angle fundus multimodal medical images. Background Technology

[0002] With the continuous development of medical imaging technology, multimodal medical image data is increasingly widely used in disease diagnosis and health management, especially in the early screening and intelligent diagnosis of fundus diseases. Ultra-wide-angle fundus images, due to their wide coverage and large information content, have become a research hotspot and have demonstrated significant value in the diagnosis of diseases such as retinopathy and diabetic retinopathy. However, existing multimodal medical image processing and data management systems still have significant shortcomings in intelligent diagnosis, data fusion accuracy, and system integration, making it difficult to meet the needs of efficient and accurate fundus disease diagnosis. A search revealed a patent with publication number CN117115045B that proposes a method for improving the quality of medical image data based on internet-based generative artificial intelligence. By introducing generative artificial intelligence and industrial internet technology, it improves traditional methods for improving the quality of medical image data, achieving accurate fusion of multimodal medical image data and deep learning-based noise reduction. However, this technical solution mainly focuses on improving the quality of medical image data and lacks specific optimization for the characteristics of ultra-wide-angle fundus images, such as insufficient support for edge distortion correction and local detail enhancement in ultra-wide-angle images. Furthermore, while its distributed storage and cloud-based data management simplify data sharing, it fails to fully integrate intelligent diagnostic algorithms in fundus image diagnosis scenarios, resulting in room for improvement in the system's diagnostic efficiency and accuracy.

[0003] On the other hand, patent CN119480124B proposes a lung infectious disease prediction system based on multimodal data fusion. Through dynamic fusion and temporal analysis of multimodal data, it improves the accuracy and reliability of lung infectious disease prediction. However, this technical solution mainly focuses on the fusion and analysis of lung images and related clinical data, failing to fully consider the unique characteristics of fundus images, such as the extraction and analysis of key information like retinal vessel distribution and lesion location in ultra-wide-angle fundus images. Furthermore, the system's application of attention mechanisms in multimodal data fusion is relatively general, lacking customized design for fundus image features, which may lead to insufficient specificity and sensitivity of diagnostic results. In addition, the data management module of this solution does not adequately support large-scale storage and efficient retrieval of ultra-wide-angle fundus images, making it difficult to meet the needs of rapid access to massive amounts of image data in real-world medical scenarios.

[0004] The aforementioned issues indicate that existing multimodal medical image processing and data management systems still have significant shortcomings in intelligent diagnosis of ultra-wide-angle fundus images, optimization of data fusion accuracy, and efficient data management. Therefore, there is an urgent need for a multimodal medical image intelligent diagnosis and data management system that can incorporate the characteristics of ultra-wide-angle fundus images to optimize multimodal data fusion algorithms, enhance intelligent diagnostic capabilities, and achieve efficient data management and sharing, thereby meeting the needs of modern medicine for accurate and efficient diagnosis of fundus diseases. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent diagnosis and data management system for ultra-wide-angle fundus multimodal medical images.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent diagnostic and data management system for ultra-wide-angle fundus multimodal medical images, the system comprising:

[0007] The image distortion correction module acquires records of edge distortion distribution, retinal vessel curvature change process and lesion area location information of ultra-wide-angle fundus images, extracts distortion gradient, curvature abnormal points and lesion boundary features, completes image annotation in sequence, defines image quality level based on annotation results, and generates image quality labels.

[0008] The multimodal fusion analysis module calls the image quality label, processes the retinal vessel distribution layer, lesion area segmentation and clinical indicator association requirements, judges the degree of influence of vessel distribution on lesion localization, evaluates the support effect of clinical indicators on lesion classification, matches diagnostic weight boundaries, and outputs a diagnostic weight allocation list.

[0009] The intensive task scheduling module calls the image task relationship in the diagnostic weight allocation list, adjusts high-coherence images to the intensive area according to diagnostic coherence and task intensive area distribution, removes low-coherence images, reconstructs task affiliation, and outputs a task intensive scheduling distribution map.

[0010] The abnormal takeover response module reads the task execution records in the task density scheduling distribution map, identifies image distortion, misjudgment of lesions, and missing data, triggers a takeover request, broadcasts to nearby computing nodes, records feedback and identifies and confirms the response object, and generates an abnormal task replacement assignment record after the replacement is completed.

[0011] As a further embodiment of the present invention, the image quality label includes a distortion correction level identifier, a lesion localization code, and a diagnostic adaptation level; the diagnostic weight allocation list includes a vascular distribution weight coefficient, a lesion segmentation index, and a clinical index association tolerance; the task density scheduling distribution map includes a high-density diagnostic block, a low-density scheduling band, and a coherence distribution parameter; and the abnormal task substitution assignment record includes a substitution task number, a takeover response identifier, and a feedback confirmation time stamp.

[0012] As a further aspect of the present invention, the image distortion correction module includes:

[0013] The distortion distribution extraction submodule acquires the edge distortion distribution record of ultra-wide-angle fundus images, divides continuous region segments in spatial order, identifies the change pattern of distortion gradient in the segments, filters according to the number and distribution pattern of gradient mutations, marks the region segments with distortion feature changes, and generates the distortion anomaly segment annotation results.

[0014] The vascular curvature analysis submodule, based on the spatial region corresponding to the labeling results of the aberration abnormal segments, calls the same segment record in the process of retinal vascular curvature change, sorts out the fluctuation trend and change rhythm of curvature data, filters out segments that have curvature abnormality or fluctuation at the same time as the aberration abnormal segments, summarizes them to form an abnormal behavior chain of vascular curvature, and generates a set of curvature fluctuation related segments.

[0015] The quality label generation submodule calls the set of curvature fluctuation related segments and the annotation results of the distortion abnormal segments, compares them with the spatial identifiers of the interruption points recorded in the lesion area location information, screens the spatial intersections in the three-dimensional records, identifies the image numbers with multiple abnormal manifestations, and classifies the abnormal combinations according to the preset image quality standards to generate image quality labels.

[0016] As a further aspect of the present invention, the multimodal fusion analysis module includes:

[0017] The vascular distribution weight recognition submodule acquires vascular distribution records in the retinal region where the image is located. Based on the spatial range and depth information marked in each record, it calls the label content related to vascular distribution in the image quality label of the corresponding image, establishes a one-to-one correspondence between vascular distribution records and lesion location, matches them according to the combination of vascular distribution intensity level and adaptive description in lesion location label, marks the influential weight segments and organizes their corresponding image numbers, and generates vascular distribution weight matching results.

[0018] The lesion segmentation judgment submodule calls the image number involved in the blood vessel distribution weight matching result, obtains the clinical indicator label of its position segment in the lesion segmentation target area, and compares it with the segmentation node data in the lesion distribution layer in the image area to identify the intersection range of unstable clinical indicator nodes and lesion points, sorts and classifies the set of nodes subject to dual restrictions in the segmentation, and generates a list of lesion segmentation restriction node distribution.

[0019] The diagnostic list generation submodule obtains the lesion classification capability level content in the quality label of each image number corresponding to each node in the lesion segmentation restriction node distribution list, and filters out the image numbers and corresponding task names that can be diagnosed within the accuracy tolerance range in the restricted area according to the classification accuracy requirements set for diagnosis. The task numbers and executable images are combined and mapped one by one to generate a diagnostic weight allocation list.

[0020] As a further aspect of the present invention, the matching based on the combination of the vascular distribution intensity level and the adaptive description in the lesion location label specifically involves: comparing the real-time vascular width data and real-time vascular branch angle data in the vascular distribution record with the preset vascular distribution threshold and lesion location adaptive description in the lesion location label; when the real-time vascular width data exceeds the vascular distribution threshold or the real-time vascular branch angle data does not conform to the lesion location adaptive description, the matching is completed.

[0021] The identification of unstable clinical indicators specifically involves: identifying the segmentation nodes in the clinical indicator labels where the fluctuation range of the indicator is higher than a preset fluctuation range threshold or the deviation of the indicator from the mean exceeds a preset deviation threshold as unstable clinical indicator nodes.

[0022] The process of identifying the intersection range between unstable clinical indicator nodes and lesion points specifically involves: using each lesion point in the lesion distribution layer as the geometric center, calculating and generating a dynamic safety buffer area based on the image resolution and the area of ​​the lesion region, and determining the unstable clinical indicator nodes that fall within the dynamic safety buffer area as nodes within the intersection range.

[0023] As a further aspect of the present invention, the intensive task scheduling module includes:

[0024] The coherence extraction submodule calls the task and image combination information in the diagnostic weight allocation list to obtain the continuous diagnostic records of each image. According to the frequency of continuous repetition and time order of the task number in the diagnostic record, it marks the diagnostic segments with close task connection, and numbers and groups the combinations with interval jumps between diagnostic segments to establish the image diagnostic coherence level sorting and generate the diagnostic coherence level distribution result.

[0025] The task-dense association submodule extracts the spatial density data of the current corresponding task based on the image number in the diagnostic coherence level distribution result. According to the spatial overlap area of ​​the task number and the task cluster density classification rules, it forms a combination structure of high coherence level images and high-density task areas. At the same time, it marks the degree of separation between the task corresponding to the low coherence level image and the high-density area in spatial location, and generates an image task matching combination list.

[0026] The regional scheduling update submodule calls the scheduling mapping information in the image task matching combination list, reassigns high-coherence-level images to task-centered regional path nodes, repositions low-coherence-level images to task-peripheral regional nodes, and maps and updates the adjustment trajectory of all task numbers in spatial distribution, completing the latest correspondence between task numbers and image numbers in all regions, and generating a task density scheduling distribution map.

[0027] As a further aspect of the present invention, the abnormal takeover response module includes:

[0028] The anomaly identification submodule reads the task execution process of each image in the task density scheduling distribution map, extracts the image distortion signal, lesion misjudgment behavior and data missing records in the task stage in chronological order, sorts them according to the order in which the records appear, determines whether there are two or more consecutive abnormal label combinations, marks the corresponding task number and image number, and generates an abnormal combination task label table.

[0029] The response feedback recording submodule collects the quality labels of other images in the same scheduling area based on the task number of the image already marked in the abnormal combination task labeling table, broadcasts the task takeover request to these images and records the feedback content, extracts the acceptance field and reception time from the feedback response, identifies the image number that has confirmed the takeover intention and sorts them in ascending order of response time, and generates the task takeover response sorting result.

[0030] The task replacement generation submodule calls the first response image number and the corresponding task number in the task takeover response sorting result, updates the execution image field of the task in the task density scheduling distribution map, records the adjustment node of the target task and the takeover image in the scheduling configuration, summarizes all completed replacement task information, and generates an abnormal task replacement assignment record.

[0031] As a further aspect of the present invention, the extraction of image distortion signals during the task phase specifically involves: comparing the pixel grayscale values ​​of the image with preset standard grayscale values ​​in real time; and generating the image distortion signal when the deviation between the pixel grayscale values ​​and the standard grayscale values ​​remains greater than a first grayscale threshold for a preset duration.

[0032] The lesion misjudgment behavior during the extraction task stage is specifically as follows: obtain the lesion region segmentation result of the image and calculate the segmentation error. When the segmentation error is continuously higher than the expected segmentation error generated based on task planning and image data, and the difference between the two exceeds a preset error threshold, the lesion misjudgment behavior is generated.

[0033] The process of extracting missing data records during the task phase specifically involves: determining whether the data integrity flag of the image has been received within a preset time period; and generating the missing data record if the number of consecutive loss of the data integrity flag exceeds a preset packet loss threshold.

[0034] As a further aspect of the present invention, the system further includes:

[0035] The image collaborative grouping module reads the abnormal task replacement assignment record, screens the task overlap trajectory and diagnosis synchronization record, identifies collaborative feature images and groups them, rearranges their task paths and order, integrates them into the cloud for unified scheduling, and outputs the image collaborative grouping structure diagram.

[0036] The image collaborative grouping structure diagram includes collaborative diagnostic grouping numbers, path overlap measurement indicators, and joint scheduling time axis.

[0037] As a further aspect of the present invention, the image collaborative grouping module includes:

[0038] The trajectory integration submodule reads the access order and corresponding diagnostic trajectory of each task execution image in the abnormal task replacement assignment record, extracts the frequency of occurrence of image task number, the overlapping position points and time segments of task areas in the trajectory path, sorts out the access times of each image task and the number of path intersection nodes, and summarizes the image numbers within the same time period for diagnosis, generating a list of task trajectory intersection features.

[0039] The collaborative screening module determines whether the image numbers and path overlaps recorded in the task trajectory intersection feature list simultaneously possess the characteristics of multiple task overlap records and diagnosis time overlap. The image numbers that meet the conditions are grouped and classified, marked as grouping objects, and a corresponding task number set structure is established for each group to generate an image task collaborative grouping structure.

[0040] The scheduling and reconfiguration submodule calls the task set corresponding to each group in the image task collaborative grouping structure, reorders the task numbers in each group according to the original task order, and uniformly specifies the corresponding execution image number as the scheduling entry in the sorting structure, updates the corresponding task node order in the platform scheduling arrangement structure, and maps the updated content to the regional task distribution layer to generate an image collaborative grouping structure diagram.

[0041] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0042] In this invention, a multidimensional quality label is generated by extracting the distortion gradient, vascular curvature changes, and lesion boundary features of ultra-wide-angle fundus images, establishing a diagnostic usability quantification mechanism. Combining the retinal vascular distribution characteristics, lesion region segmentation, and clinical indicator correlation requirements, task interference assessment, lesion segmentation identification, and diagnostic weight allocation are achieved. The scheduling distribution structure is adjusted based on the image diagnostic coherence and task spatial density. Task substitution and assignment response are realized through an abnormal takeover mechanism. Image reorganization scheduling is completed based on path overlap and diagnostic coordination time periods. Thus, a closed-loop diagnostic control system covering the entire process of identification, matching, adjustment, substitution, and grouping is constructed. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the overall system flow of the present invention.

[0044] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0046] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0047] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0048] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0049] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0050] Please see Figure 1This invention provides a technical solution: an intelligent diagnostic and data management system for ultra-wide-angle fundus multimodal medical images, the system comprising:

[0051] The image distortion correction module acquires records of edge distortion distribution, retinal vessel curvature change process and lesion area location information of ultra-wide-angle fundus images, extracts distortion gradient, curvature abnormal points and lesion boundary features, completes image annotation in sequence, defines image quality level based on annotation results, and generates image quality labels.

[0052] The multimodal fusion analysis module calls image quality labels, processes retinal vessel distribution layers, lesion region segmentation and clinical indicator association requirements, judges the degree of influence of vessel distribution on lesion localization, evaluates the support effect of clinical indicators on lesion classification, matches diagnostic weight boundaries, and outputs a diagnostic weight allocation list.

[0053] The intensive task scheduling module calls the image task relationships in the diagnostic weight allocation list, adjusts highly coherent images to the intensive area according to diagnostic coherence and task intensive area distribution, removes low coherent images, reconstructs task affiliation, and outputs a task intensive scheduling distribution map.

[0054] The abnormal takeover response module reads the task execution records in the task density scheduling distribution map, identifies image distortion, misjudgment of lesions and missing data, triggers takeover requests, broadcasts to nearby computing nodes, records feedback and identifies and confirms the response object, and generates an abnormal task replacement assignment record after the replacement is completed.

[0055] The image collaborative grouping module reads abnormal task replacement assignment records, screens overlapping task trajectories and diagnostic synchronous records, identifies collaborative feature images and groups them, rearranges their task paths and sequences, integrates them into the cloud for unified scheduling, and outputs an image collaborative grouping structure diagram.

[0056] Image quality labels include distortion correction level identifiers, lesion localization codes, and diagnostic adaptation levels. The diagnostic weight allocation list includes vascular distribution weight coefficients, lesion segmentation indicators, and clinical indicator correlation tolerances. The task density scheduling distribution map includes high-density diagnostic blocks, low-density scheduling bands, and coherence distribution parameters. Abnormal task substitution assignment records include substitution task numbers, takeover response identifiers, and feedback confirmation timestamps. The image collaborative grouping structure diagram includes collaborative diagnostic grouping numbers, path overlap measurement indicators, and joint scheduling timelines.

[0057] Please see Figure 2 The image distortion correction module includes:

[0058] The distortion distribution extraction submodule acquires the edge distortion distribution record of ultra-wide-angle fundus images, divides continuous region segments in spatial order, identifies the change pattern of distortion gradient in the segments, filters according to the number and distribution pattern of gradient mutations, marks the region segments with distortion feature changes, and generates the distortion anomaly segment annotation results.

[0059] Obtain a record of the edge distortion distribution in an ultra-wide-angle fundus image (Image-001). This record consists of a set of regions covering the edge (defined as the distance from the image center). Greater than 0.8 times the total radius The distortion vector field of the region) First, the edge region is divided into 36 segments, each 10 degrees from the image center, in spatial order. to Then, the distortion gradient within each segment is calculated. Distortion gradient Defined as the distortion vector within the segment Length of the module Relative to radial distance rate of change For fragments (40-50 degree range), with 5 sampling points inside. to of The values ​​(unit: pixels) are [2.1, 2.3, 2.8, 3.1, 3.0], corresponding to... The values ​​(unit: mm) are [8.0, 8.2, 8.4, 8.6, 8.8], calculated as follows: The value sequence (unit: pixels / mm) is [ , , , The range is defined as [1.0, 2.5, 1.5, -0.5]. Subsequently, a selection process is performed based on the number and distribution of gradient mutations, with a gradient mutation threshold. The settings are based on the average gradient calculated from a standard test image library (containing 100 distortion-free images). (0.4 pixels / mm) and standard deviation (0.3 pixels / mm), setting pixels / mm, "gradient abrupt change" is defined as (1.3 pixels / mm) events, in of In the sequence [1.0, 2.5, 1.5, -0.5], both 2.5 and 1.5 are greater than 1.3, therefore... There are two gradient mutations. The selection rule is defined as: when the number of gradient mutations within a segment is... Greater than or equal to 2, or exist When a single mutation of (2.6 pixels / mm) occurs, the fragment is selected. of If the filter criteria are met, mark it as such. For regions exhibiting distortion characteristics, perform the same operation on all 36 segments, for example... of It is also marked, generating the annotation results for the aberration fragments.

[0060] The vascular curvature analysis submodule, based on the spatial region corresponding to the labeling results of the aberration abnormal segments, calls the same segment records in the process of retinal vascular curvature change, sorts out the fluctuation trend and change rhythm of curvature data, filters out segments that have curvature abnormality or fluctuation at the same time as the aberration abnormal segments, summarizes them to form the abnormal behavior chain of vascular curvature, and generates a set of curvature fluctuation related segments.

[0061] Based on the spatial region corresponding to the anomaly fragment annotation results, i.e. and By calling up the same segment of recordings during the change in retinal vessel curvature, and extracting the spatial location falling within the vascular database of Image-001,... All vascular segments within the region, such as blood vessels and Obtain its curvature data (unit: ), exist Curvature records within the region The result is [0.08, 0.09, 0.11, 0.25, 0.10]. Next, we will analyze... The fluctuation trend and rhythm of change are calculated as the first difference of the curvature sequence. , The rhythm of change is calculated as follows absolute value , Subsequently, segments with aberrations were selected ( (Segments exhibiting curvature anomalies or fluctuations, curvature anomaly threshold) With curvature fluctuation threshold The settings are based on the central area of ​​image Image-001 ( The average curvature of blood vessels (0.07) ) and its standard deviation (0.03) ),set up ,set up "Curvature anomaly" is defined as (0.16) "Curvature fluctuation" is defined as (0.09) ),examine The sequence, with a value of 0.25 greater than 0.16, satisfies the curvature anomaly; check... The sequence values ​​0.14 and 0.15 are both greater than 0.09, satisfying the curvature oscillation condition. lie in (abnormal fragment), and Simultaneously satisfying both curvature anomaly and curvature fluctuation, therefore Belonging to The segments are filtered out, and finally all segments that meet the criteria are summarized, such as... and If all conditions are met, an abnormal behavior chain of vascular curvature is formed, generating a set of curvature fluctuation-related segments.

[0062] The quality label generation submodule calls the curvature fluctuation associated fragment set and the distortion abnormal fragment annotation results, compares the spatial identifier of the interruption point recorded by the lesion area location information, screens the spatial intersection in the three-dimensional records, identifies the image number with multiple abnormal manifestations, and classifies the abnormal combination according to the preset image quality standards to generate image quality labels.

[0063] The curvature fluctuation associated fragment set and the distortion anomalous fragment annotation results are retrieved, and the spatial identifiers of the interruption points in the lesion area location information are compared with those. The lesion location record of image Image-001 is then retrieved, which includes known lesions (such as microaneurysms). Bleeding points The spatial coordinates and processing status of the data are recorded and displayed. lie in The region is in a processing status of "Breakpoint_SegFail" (segmentation failure breakpoint). lie in The region is set to "Normal". Next, the spatial intersection of the three-dimensional records (distortion, curvature, and lesion breakpoints) is screened. Dimension 1 (distortion) is found to exist. Dimension 2 (curvature) exists Dimension 3 (lesion breakpoint) exists Calculate the intersection The result is Subsequently, image IDs with multiple anomalous features (i.e., segments with spatial overlap) were identified. Image Image-001 was identified because... The presence of the anomalous combination was identified, and finally, the anomalous combinations were classified and graded according to a preset image quality standard. This standard (see Table 1) defines the quality level corresponding to different anomalous combinations. Image-001 in The region exhibits three types of abnormalities simultaneously: distortion, curvature fluctuation, and lesion breakpoints, which meets the definition of "Rule 3".

[0064] Table 1: Image Quality Standard Grading Table

[0065]

[0066] As shown in Table 1, the standard classifies images based on anomaly combinations across three dimensions. Image-001 conforms to rule 3 and is rated as a level 3 quality image, generating an image quality label.

[0067] The multimodal fusion analysis module includes:

[0068] The vascular distribution weight recognition submodule acquires vascular distribution records in the retinal region where the image is located. Based on the spatial range and depth information marked in each record, it calls the label content related to vascular distribution in the image quality label of the corresponding image, establishes a one-to-one correspondence between vascular distribution records and lesion location, matches them according to the combination of vascular distribution intensity level and adaptive description in lesion location label, marks the influential weight segments and organizes their corresponding image numbers, and generates vascular distribution weight matching results.

[0069] Acquire a record of the vascular distribution in the retinal region where the image (Image-001) is located. This record is derived from an OCT-A scan and includes spatial extent, depth, vessel width, and branching angles. For example, record... [Image ID: Image-001, Spatial Range:] Depth: 60 Real-time blood vessel width: Real-time vascular branch angle: °], based on records Spatial range ( ) and depth (60 This involves calling the image quality labels (from paragraph 3) related to blood vessel distribution in image Image-001, i.e. The quality level of the region was QL-3 (severely distorted), and then a vascular distribution record was established. With lesion localization ( ,lie in The one-to-one correspondence is determined by matching the vascular distribution intensity level with the adaptive description in the lesion localization label. The location tag contains a preset analysis adaptability description: [Preset blood vessel distribution threshold] Adaptive description of lesion localization The matching process specifically involves: […]. In (18) )and (15) Perform numerical comparison, and at the same time... (50°) and Compare ([20°, 40°]) and perform a judgment: when or When the match is complete, in this example, (18) (greater than) (15) ),at the same time (50°) does not meet the requirements The description of ([20°, 40°]) satisfies the matching condition. Finally, the influential weight segments are labeled and their corresponding image numbers are organized, and the influential weights are determined. The calculation is based on the degree of mismatch, and the weight settings are based on... ,in , ,calculate fragments Labeled weight Generate the weighted matching results for blood vessel distribution.

[0070] The lesion segmentation judgment submodule calls the image number involved in the blood vessel distribution weight matching result, obtains the clinical indicator label of its location segment in the lesion segmentation target area, and compares it with the segmentation node data in the lesion distribution layer in the image area to identify the intersection range of unstable clinical indicator nodes and lesion points. It sorts and classifies the set of nodes subject to dual restrictions in the segmentation and generates a list of lesion segmentation restriction nodes.

[0071] The image number involved in the blood vessel distribution weight matching result is called, namely Image-001 (because...). With a fragment weight of 0.52, its position within the target lesion segmentation region is obtained. The clinical indicator labels within the location segment, for example Segmentation nodes within the region The label is: [Indicator: Retinal nerve fiber layer (RNFL) thickness, Real-time value: 120] Mean: 95 Fluctuation range: 18 ], comparing the segmentation node data in the lesion distribution layer within the imaging area ( ) and lesion point ( Geometric center coordinates The location of the clinical indicator is used to identify the intersection range between unstable nodes and lesion points. First, unstable nodes of clinical indicators are identified, and a preset fluctuation range threshold is set. Deviation from preset threshold The settings are based on the statistical analysis of indicators from the same region in 1000 healthy samples, taking the 98th percentile of their fluctuation range as... (10) ), take its 98th percentile of deviation from the mean as (20) ),examine : Indicator fluctuation range (18) (Higher than) (10) ), indicator deviation from the mean ( (Higher than) (20) ), The nodes were identified as having unstable clinical indicators. Secondly, the overlap range was identified, based on the lesion point. As the geometric center, based on image resolution (5) (per pixel) and area of ​​the lesion region Calculate the radius of the generated dynamic safety buffer area using (900 pixels squared). ,calculate ,in The safety factor is set at 0.8, based on experimental verification, to ensure that the buffer area can cover 99% of the segmentation uncertainty. That is, the buffer area is Center 4.8 Radius range, measurement (coordinate )and (coordinate distance ,because (3.5) (less than) (4.8) ), Nodes falling within the dynamic safety buffer area are identified as being within the intersection range, and finally... (Due to the instability of the RNFL index and) (Dual restrictions in the buffer zone) are sorted and categorized to generate a list of disease segmentation restriction nodes.

[0072] The diagnosis list generation submodule obtains the lesion classification capability level content in the quality label of each image number corresponding to each node in the lesion segmentation restriction node distribution list, and filters out the image numbers and corresponding task names that can be diagnosed within the accuracy tolerance range in the restricted area according to the classification accuracy requirements set for diagnosis. It then maps the task numbers and executable images one by one to generate a diagnosis weight allocation list.

[0073] According to the list of nodes restricting lesion segmentation (e.g., ... For the image number (Image-001) corresponding to the image, obtain the lesion classification capability level from its quality label. The label for Image-001 is QL-3 (severe distortion), which corresponds to the lesion classification capability level. "Low" (expected segmentation accuracy loss) ), in accordance with the classification accuracy requirements set for the diagnosis, assuming the current diagnostic task Precision requirements set for (fine segmentation of microaneurysms) 90% (i.e., allowable loss) ), filter out those that can be found within the restricted area ( Maintain the image number and corresponding task name for diagnosis within the accuracy tolerance range. The settings are based on Reduce by a fixed percentage, for example (i.e., allowable loss) Compare Image-001's ability levels. (>15%) and tolerance range (allowable loss) Image-001 Image-001 was excluded because it did not fall within the tolerance range and therefore could not perform the task. There is also an image, Image-002, which has restricted nodes. The corresponding regional quality label is QL-1, capability level. For "high" ( ), (<5%) satisfies Precision tolerance range ( Therefore, Image-002 was selected, and finally, the task number and executable image were mapped one by one. (Image-002), generate a list of diagnostic weight assignments.

[0074] The intensive task scheduling module includes:

[0075] The coherence extraction submodule calls the task and image combination information in the diagnostic weight allocation list to obtain the continuous diagnostic records of each image. According to the frequency of continuous repetition and time order of the task number in the diagnostic record, it marks the diagnostic segments with close task connection, and numbers and groups the combinations with interval jumps between diagnostic segments to establish the image diagnostic coherence level ranking and generate the diagnostic coherence level distribution results.

[0076] Retrieve task and image combination information from the diagnostic weight allocation list, for example ( Image-002), ( Image-002), ( (Image-003), retrieve the continuous diagnostic records of Image-002, which are a sequence of timestamps and task numbers: [10:01:00, [10:01:30,] [10:02:00,] [10:04:00,] [10:04:30,] According to the frequency of consecutive repetitions of task numbers in the diagnostic record ( Repeat twice. Repeat once. (Repeat twice) In chronological order, mark diagnostic segments with closely connected tasks, and set a threshold for the time interval between closely connected segments. The time interval is set to 45 seconds (calculated based on the system's average task switching overhead of 30 seconds + 1.5 times the standard deviation of 10 seconds). In the sequence [10:01:00, [10:01:30,] (Interval 30s < 45s), [10:01:30, [10:02:00,] (Interval 30s < 45s), [10:04:00, [10:04:30,] All segments with intervals of 30s to 45s were marked as tightly connected. Then, combinations with interval jumps between diagnostic segments were numbered and grouped, with an interval jump threshold set. Set to 60 seconds ( +25-second buffer), [10:02:00, [10:04:00,] The interval is 120 seconds. Since 120 seconds is greater than 60 seconds, a jump occurs here. Therefore, the record of Image-002 is divided into two groups: Group1:([10:01:00, [10:01:30,] [10:02:00,] Group2:([10:04:00, [10:04:30,] Finally, a ranking of the consistency levels of imaging diagnosis was established, and the consistency levels were determined. Through consistent scoring calculate, Image-002 Consistency level (see Table 2) definition: Level A (High) Grade B (Medium) Grade C (Low) Image-002 received a score of 0.9375, which is classified as a C grade.

[0077] Table 2: Classification of Image Diagnostic Continuity

[0078]

[0079] (Note: Image-003 has no redirection, and the average redirection time is calculated as 0, resulting in an infinite score. The system will classify it as Grade A with the highest score of 8.0.)

[0080] As shown in Table 2, calculations were performed on all images in the list, with Image-003 classified as Grade A and Image-004 as Grade B, generating a consistent diagnostic grade distribution result.

[0081] The task-dense association submodule extracts the spatial density data of the current corresponding task based on the image number in the diagnostic coherence level distribution results. According to the spatial overlap area of ​​the task number and the task cluster density classification rules, it forms a combination structure of high coherence level images and high-density task areas. At the same time, it marks the degree of separation between the task corresponding to the low coherence level image and the high-density area in spatial location, and generates an image task matching combination list.

[0082] Based on the image numbers in the diagnostic coherence level distribution results, namely Image-003 (Level A), Image-004 (Level B), and Image-002 (Level C), the spatial density data of their current corresponding tasks is extracted. Image-003 corresponds to the task... Located in the region Image-004 corresponds to the task Located in the region Image-002 corresponds to the task Located in the region Retrieve regional task density data: Includes tasks [ , , , There are 4 tasks in total. Includes tasks [ There is one task in total. Based on the spatial overlap of task numbers and the task clustering density classification rules, the spatial overlap area... By calculating each task bounding box intersection Obtain, cluster density Defined as the number of tasks within a region, density classification rule set: high density for medium density for low density for Therefore, (4 tasks) are located in a high-density area. (One task) is a low-density area. Then, high-coherence-level imagery is combined with the high-density task area to form a composite structure. Image-003 (Level A) is currently in... (High density), forming a combination (Image-003, Match: High-High), while also marking the task corresponding to the low coherence level image (Image-002, Level C). In terms of spatial location, it is related to high-density areas ( The degree of separation, since Image-002 has already been... In, its degree of separation The value is 0, if there is another Image-005 (Class C) in (Low density), then calculate center and center Euclidean distance ,For example , , Pixels, generate a list of image task matching combinations.

[0083] The regional scheduling update submodule calls the scheduling mapping information in the image task matching combination list, reassigns high-coherence-level images to the task set area path nodes, repositions low-coherence-level images to the task periphery area nodes, and maps and updates the adjustment trajectory of all task numbers in spatial distribution, completing the latest correspondence between task numbers and image numbers in all areas, and generating a task density scheduling distribution map.

[0084] Invoke the scheduling mapping information from the image task matching combination list, for example (Image-003, ,Match:High-High),(Image-002, Match:Low-High), and Image-005, Match:Low-Low, Sep:500), redirecting the high-coherence image (Image-003) to the task-focused area. The path node of Image-003, the original scheduling node of Image-003 is It has now been updated to Core path nodes The low-coherence-level images (Image-002 and Image-005) were repositioned to nodes in the mission's outer perimeter. Image-002 (Level C) is currently located in... (High density), its path nodes from ( (Inner) Repositioned to the outer area nodes Image-005 (Class C) has been... (Low density, considered as the periphery), maintain its nodes. Without changing the settings, the adjustment trajectories of all task numbers in the spatial distribution are then mapped and updated, recording the adjustment trajectory of Image-002. :[Start: @ End: @ ],Task (Originally Image-002) The execution image field of the task has been cleared and is awaiting reallocation. (Originally Image-003) The execution image field remains Image-003, but the node is updated to... Finally, the latest mapping between task numbers and image numbers for all regions was completed, Image-003 @ Image-002 NULL@ Image-005 @ Generate a task density scheduling distribution map.

[0085] The exception takeover response module includes:

[0086] The anomaly identification submodule reads the task execution process of each image in the task density scheduling distribution map, extracts the image distortion signal, lesion misjudgment behavior and data missing records in the task stage in chronological order, sorts them according to the order in which the records appear, determines whether there are two or more consecutive abnormal label combinations, marks the corresponding task number and image number, and generates an abnormal combination task label table.

[0087] Read the task execution process for each image in the task density scheduling distribution map, for example, Image-003 executes the task. (exist The process involves extracting image distortion signals, lesion misjudgment behaviors, and missing data records in chronological order within the task phase. Monitoring begins at timestamp 11:01:05. First, image distortion signals are extracted using preset standard grayscale values. The first grayscale threshold is 120 (based on image library statistics). The settings are based on 2 standard deviations (25), Preset duration At 11:01:05.100, with a processing time of 300 milliseconds (based on 3 frames of image processing time), the pixel grayscale value of Image-003 was detected. The deviation value is 180. 60 is greater than (50), this state lasted for 400 milliseconds (greater than) (300 milliseconds), so at 11:01:05.500, the image distortion signal is generated. (Distortion), secondly, extract the misjudgment behavior of the lesion and obtain Calculate the segmentation error based on the segmentation results. (Use the Dice coefficient to compare with the true value, error = 1 - Dice), Task Expected segmentation error for Image-003 (Based on the ability level of paragraph 6) The preset error threshold is 0.05. Set to 0.03 (based on) (60% of the total), at 11:01:05.600 (frame 1). Difference > (0.03), at 11:01:05.700 (frame 2) Difference > (0.03), the segmentation error is continuously higher than the expected segmentation error and the difference exceeds the error threshold, so a lesion misjudgment behavior is generated at 11:01:05.700. (Misjudgment), again, extract missing data records, with a preset time period. The preset threshold for packet loss is 50 milliseconds. The error occurred 3 times (based on network stability testing, 99.9% of packet losses do not exceed 3 times). Between 11:01:05.800 and 11:01:06.000, data integrity flags were lost 4 times consecutively (within a 50ms period), and these 4 losses were greater than... (3 times), so a data missing record was generated at 11:01:06.000. (DataLoss), sorts records according to their order of appearance: [ (11:01:05.500), (11:01:05.700), [11:01:06.000], determine whether there are two or more consecutive abnormal label combinations. and If the occurrences are consecutive in time (interval of 200ms, less than the 500ms anomaly detection time window), and the condition is met, the corresponding task number is marked. The image number (Image-003) is used to generate an anomaly combination task annotation table.

[0088] The response feedback recording submodule collects the quality labels of other images in the same scheduling area based on the task number of the image already marked in the abnormal combination task labeling table, broadcasts the task takeover request to these images and records the feedback content, extracts the acceptance field and reception time from the feedback response, identifies the image number that has confirmed the takeover intention and sorts them in ascending order by response time, and generates the task takeover response sorting result.

[0089] Based on the task number of the image (Image-003) already marked in the abnormal combination task labeling table ( ), collect data from the same scheduling area ( Quality labels for other images within the area, query the scheduling distribution map. There were no other images in the area (Image-002 had been removed), so the search area was expanded to the adjacent area. Find Image-002, retrieve its quality label as QL-1, and its capability level as "high," then search... Image-005 was found, with a quality label of QL-2 and a capability level of "Medium". A broadcast task takeover request was initiated. Broadcast to Image-002 and Image-005 at 11:01:06.100 and record the feedback content. Image-002 feedback: [Receive time: 11:01:06.250, Acceptance: Yes, Load: 20%], Image-005 feedback: [Receive time: 11:01:06.220, Acceptance: Yes, Load: 60%]. Extract the acceptance field (both Yes) and reception time (Image-002: 11:01:06.250, Image-005: 11:01:06.220) from the feedback responses. Identify the image numbers Image-002 and Image-005 that have confirmed their willingness to take over, sort them in ascending order of response time, and calculate the response time. millisecond, The sorting result is: [Image-005(120ms), Image-002(150ms)], generating the sorting result of the task takeover response.

[0090] The task replacement generation submodule calls the first response image number and the corresponding task number in the task takeover response sorting result, updates the execution image field of the task in the task density scheduling distribution map, records the adjustment nodes of the target task and the takeover image in the scheduling configuration, summarizes all completed replacement task information, and generates an abnormal task replacement assignment record.

[0091] The first response image number (Image-005) in the task takeover response sorting results is compared with the corresponding task number (Image-005). ), update the task in the task density scheduling distribution map ( The execution image field of the record was originally [Image-003]. @ [Image-005] @ Meanwhile, the original Image-005 record [Image-005] @ [Image-005] has been updated. ( , )@ The scheduling node for image Image-005 starts from... Adjust to Record the target tasks of the alternative takeover ( The adjustment node in the scheduling configuration for the receiving image (Image-005), i.e., the node from... Change to And summarize all completed replacement task information to form a record: [Timestamp:11:01:07.000,Task: ,OriginalImage:Image-003,TakeoverImage:Image-005,NodeAdjust: to If other alternatives occur subsequently, such as [Timestamp:11:05:02.000,Task:] ,OriginalImage:Image-010,TakeoverImage:Image-005,NodeAdjust: to If any of these are found, they will be aggregated together to generate an abnormal task replacement assignment record.

[0092] The image collaborative grouping module includes:

[0093] The trajectory integration submodule reads the access order and corresponding diagnostic trajectory of each task execution image in the abnormal task replacement assignment record, extracts the frequency of image task number occurrence, overlapping position points and time segments of task areas in the trajectory path, sorts out the access times of each image task and the number of path intersection nodes, and summarizes the image numbers within the same time period for diagnosis, generating a list of task trajectory intersection features.

[0094] Read the access order of each task execution image (Image-005) in the abnormal task replacement dispatch record (first takeover). Second takeover ) and corresponding diagnostic trajectory ( , Extract the frequency of image task numbers, Image-005 appears 2 times, extract the overlapping locations and time segments of task areas in the trajectory path, and the two takeover tasks of Image-005 ( Ultimately, they all point to The overlapping points are The regions and time segments are respectively (From 11:01:07) and (Starting at 11:05:02), the number of times each image task was accessed (Image-005: 2 times) and the number of path intersection nodes were compiled (Image-005: 1 node, i.e.) (and summarize the image numbers within the same time period for diagnosis, and set the time period.) It lasts for 5 minutes. (11:01:07) and (11:05:02) falls within two different 5-minute time slots (11:00-11:05, 11:05-11:10), assuming that Image-008 took over the task. From 11:02:30 to Therefore, the images collected between 11:00 and 11:05 are numbered [Image-005, Image-008]. Based on this, Image-005 and Image-008 are... If nodes (path intersections) are active within the same time period (11:00-11:05), a list of task trajectory intersection features is generated.

[0095] The collaborative screening module determines whether the image numbers and path overlaps recorded in the task trajectory intersection feature list simultaneously possess the characteristics of multiple task overlap records and diagnosis time overlap. The image numbers that meet the conditions are grouped and classified, marked as grouping objects, and a corresponding task number set structure is established for each group to generate an image task collaborative grouping structure.

[0096] Based on the image numbers (Image-005, Image-008) and path overlap counts recorded in the task trajectory intersection feature list (Image-005: 2 accesses), Image-008: 1st access To determine whether multiple tasks overlap in recording and diagnosis time coincide simultaneously, the criteria are: imaging... Number of accesses And images Number of accesses or images With images There is path overlap (both are) And the times overlap (both 11:00-11:05), Image-005's (based on The summary, assuming Also occurred at 11:04), Image-008 The first condition (both greater than 1) is not met, but the second condition (path overlap) is not. If the time overlaps (11:00-11:05), the images that meet this condition (Image-005, Image-008) are grouped and classified to create groups. ,mark For grouping objects, and for Establish a corresponding task number set structure, where the task corresponding to Image-005 is { , The task corresponding to Image-008 is { }, The task set is { , , Generate a collaborative grouping structure for image tasks.

[0097] The scheduling and reconfiguration submodule calls the task set corresponding to each group in the image task collaborative grouping structure, reorders the task numbers in each group according to the original task order, and uniformly specifies the corresponding execution image number as the scheduling entry in the sorting structure, updates the corresponding task node order in the platform scheduling arrangement structure, and maps the updated content to the regional task distribution layer to generate an image collaborative grouping structure diagram.

[0098] Calling each group in the image task collaborative grouping structure ( The corresponding task set { , , },right Task numbers within the group follow the original task order (this order is based on the priority or timestamp at the time of task creation, assuming the order is as follows). > > Reorder, the sorting result is [ , , Within the sorting structure, a corresponding execution image number is uniformly designated as the scheduling entry point. The selection criterion for the entry image is based on the image's overall capability level: Image-005 is QL-2 (medium), and Image-008 is QL-1 (high). Image-008 is selected as the entry image. The scheduling entry point updates the corresponding task node order in the platform's scheduling arrangement structure, and the original scheduling structure... It was executed in a distributed manner, and has now been updated to: [Group: ,EntryImage:Image-008,TaskSequence:( , , [),BackupImage:Image-005], and map the updated content to the region task distribution layer, Update region layer, mark For the collaborative region, its task list ( )Depend on (Main Image Image-008, Backup Image Image-005) are responsible for executing and generating the image collaborative grouping structure diagram.

[0099] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart diagnostic and data management system for ultra-wide-angle fundus multimodal medical images, characterized in that, The system includes: The image distortion correction module acquires records of edge distortion distribution, retinal vessel curvature change process and lesion area location information of ultra-wide-angle fundus images, extracts distortion gradient, curvature abnormal points and lesion boundary features, completes image annotation in sequence, defines image quality level based on annotation results, and generates image quality labels. The multimodal fusion analysis module calls the image quality label, processes the retinal vessel distribution layer, lesion area segmentation and clinical indicator association requirements, judges the degree of influence of vessel distribution on lesion localization, evaluates the support effect of clinical indicators on lesion classification, matches diagnostic weight boundaries, and outputs a diagnostic weight allocation list. The intensive task scheduling module calls the image task relationship in the diagnostic weight allocation list, adjusts high-coherence images to the intensive area according to diagnostic coherence and task intensive area distribution, removes low-coherence images, reconstructs task affiliation, and outputs a task intensive scheduling distribution map. The abnormal takeover response module reads the task execution records in the task density scheduling distribution map, identifies image distortion, misjudgment of lesions and data missing situations, triggers takeover requests, broadcasts to nearby computing nodes, records feedback and identifies and confirms response objects, and generates abnormal task replacement assignment records after the replacement is completed. The image collaborative grouping module reads the abnormal task replacement assignment record, screens the task overlap trajectory and diagnosis synchronization record, identifies collaborative feature images and groups them, rearranges their task paths and order, integrates them into the cloud for unified scheduling, and outputs the image collaborative grouping structure diagram. The image collaborative grouping structure diagram includes collaborative diagnostic grouping numbers, path overlap measurement indicators, and joint scheduling time axis.

2. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 1, characterized in that: The image quality label includes distortion correction level identifier, lesion localization code and diagnostic adaptation level; the diagnostic weight allocation list includes vascular distribution weight coefficient, lesion segmentation index and clinical index association tolerance; the task density scheduling distribution map includes high-density diagnostic blocks, low-density scheduling bands and coherence distribution parameters; and the abnormal task substitution assignment record includes substitution task number, takeover response identifier and feedback confirmation time stamp.

3. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 1, characterized in that, The image distortion correction module includes: The distortion distribution extraction submodule acquires the edge distortion distribution record of ultra-wide-angle fundus images, divides continuous region segments in spatial order, identifies the change pattern of distortion gradient in the segments, filters according to the number and distribution pattern of gradient mutations, marks the region segments with distortion feature changes, and generates the distortion anomaly segment annotation results. The vascular curvature analysis submodule, based on the spatial region corresponding to the labeling results of the aberration abnormal segments, calls the same segment record in the process of retinal vascular curvature change, sorts out the fluctuation trend and change rhythm of curvature data, filters out segments that have curvature abnormality or fluctuation at the same time as the aberration abnormal segments, summarizes them to form an abnormal behavior chain of vascular curvature, and generates a set of curvature fluctuation related segments. The quality label generation submodule calls the set of curvature fluctuation related segments and the annotation results of the distortion abnormal segments, compares them with the spatial identifiers of the interruption points recorded in the lesion area location information, screens the spatial intersections in the three-dimensional records, identifies the image numbers with multiple abnormal manifestations, and classifies the abnormal combinations according to the preset image quality standards to generate image quality labels.

4. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 1, characterized in that, The multimodal fusion analysis module includes: The vascular distribution weight recognition submodule acquires vascular distribution records in the retinal region where the image is located. Based on the spatial range and depth information marked in each record, it calls the label content related to vascular distribution in the image quality label of the corresponding image, establishes a one-to-one correspondence between vascular distribution records and lesion location, matches them according to the combination of vascular distribution intensity level and adaptive description in lesion location label, marks the influential weight segments and organizes their corresponding image numbers, and generates vascular distribution weight matching results. The lesion segmentation judgment submodule calls the image number involved in the blood vessel distribution weight matching result, obtains the clinical indicator label of its position segment in the lesion segmentation target area, and compares it with the segmentation node data in the lesion distribution layer in the image area to identify the intersection range of unstable clinical indicator nodes and lesion points, sorts and classifies the set of nodes subject to dual restrictions in the segmentation, and generates a list of lesion segmentation restriction node distribution. The diagnostic list generation submodule obtains the lesion classification capability level content in the quality label of each image number corresponding to each node in the lesion segmentation restriction node distribution list, and filters out the image numbers and corresponding task names that can be diagnosed within the accuracy tolerance range in the restricted area according to the classification accuracy requirements set for diagnosis. The task numbers and executable images are combined and mapped one by one to generate a diagnostic weight allocation list.

5. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 4, characterized in that: The matching based on the combination of vascular distribution intensity level and adaptive description in lesion location label is specifically as follows: the real-time vascular width data and real-time vascular branch angle data in the vascular distribution record are compared with the preset vascular distribution threshold and lesion location adaptive description in the lesion location label. When the real-time vascular width data exceeds the vascular distribution threshold or the real-time vascular branch angle data does not conform to the lesion location adaptive description, the matching is completed. The identification of unstable clinical indicators specifically involves: identifying the segmentation nodes in the clinical indicator labels where the fluctuation range of the indicator is higher than a preset fluctuation range threshold or the deviation of the indicator from the mean exceeds a preset deviation threshold as unstable clinical indicator nodes. The process of identifying the intersection range between unstable clinical indicator nodes and lesion points specifically involves: using each lesion point in the lesion distribution layer as the geometric center, calculating and generating a dynamic safety buffer area based on the image resolution and the area of ​​the lesion region, and determining the unstable clinical indicator nodes that fall within the dynamic safety buffer area as nodes within the intersection range.

6. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 1, characterized in that, The intensive task scheduling module includes: The coherence extraction submodule calls the task and image combination information in the diagnostic weight allocation list to obtain the continuous diagnostic records of each image. According to the frequency of continuous repetition and time order of the task number in the diagnostic record, it marks the diagnostic segments with close task connection, and numbers and groups the combinations with interval jumps between diagnostic segments to establish the image diagnostic coherence level sorting and generate the diagnostic coherence level distribution result. The task-dense association submodule extracts the spatial density data of the current corresponding task based on the image number in the diagnostic coherence level distribution result. According to the spatial overlap area of ​​the task number and the task cluster density classification rules, it forms a combination structure of high coherence level images and high-density task areas. At the same time, it marks the degree of separation between the task corresponding to the low coherence level image and the high-density area in spatial location, and generates an image task matching combination list. The regional scheduling update submodule calls the scheduling mapping information in the image task matching combination list, reassigns high-coherence-level images to task-centered regional path nodes, repositions low-coherence-level images to task-peripheral regional nodes, and maps and updates the adjustment trajectory of all task numbers in spatial distribution, completing the latest correspondence between task numbers and image numbers in all regions, and generating a task density scheduling distribution map.

7. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 1, characterized in that, The abnormal takeover response module includes: The anomaly identification submodule reads the task execution process of each image in the task density scheduling distribution map, extracts the image distortion signal, lesion misjudgment behavior and data missing records in the task stage in chronological order, sorts them according to the order in which the records appear, determines whether there are two or more consecutive abnormal label combinations, marks the corresponding task number and image number, and generates an abnormal combination task label table. The response feedback recording submodule collects the quality labels of other images in the same scheduling area based on the task number of the image already marked in the abnormal combination task labeling table, broadcasts the task takeover request to these images and records the feedback content, extracts the acceptance field and reception time from the feedback response, identifies the image number that has confirmed the takeover intention and sorts them in ascending order of response time, and generates the task takeover response sorting result. The task replacement generation submodule calls the first response image number and the corresponding task number in the task takeover response sorting result, updates the execution image field of the task in the task density scheduling distribution map, records the adjustment node of the target task and the takeover image in the scheduling configuration, summarizes all completed replacement task information, and generates an abnormal task replacement assignment record.

8. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 7, characterized in that: The image distortion signal extraction task stage specifically involves: comparing the pixel grayscale value of the image with a preset standard grayscale value in real time; and generating the image distortion signal when the deviation between the pixel grayscale value and the standard grayscale value is greater than a first grayscale threshold for a preset duration. The lesion misjudgment behavior during the extraction task stage is specifically as follows: obtain the lesion region segmentation result of the image and calculate the segmentation error. When the segmentation error is continuously higher than the expected segmentation error generated based on task planning and image data, and the difference between the two exceeds a preset error threshold, the lesion misjudgment behavior is generated. The process of extracting missing data records during the task phase specifically involves: determining whether the data integrity flag of the image has been received within a preset time period; and generating the missing data record if the number of consecutive loss of the data integrity flag exceeds a preset packet loss threshold.

9. The ultra-wide-angle fundus multimodal medical image intelligent diagnosis and data management system according to claim 1, characterized in that, The image collaborative grouping module includes: The trajectory integration submodule reads the access order and corresponding diagnostic trajectory of each task execution image in the abnormal task replacement assignment record, extracts the frequency of occurrence of image task number, the overlapping position points and time segments of task areas in the trajectory path, sorts out the access times of each image task and the number of path intersection nodes, and summarizes the image numbers within the same time period for diagnosis, generating a list of task trajectory intersection features. The collaborative screening module determines whether the image numbers and path overlaps recorded in the task trajectory intersection feature list simultaneously possess the characteristics of multiple task overlap records and diagnosis time overlap. The image numbers that meet the conditions of multiple task overlap records and diagnosis time overlap are grouped and classified, marked as grouping objects, and a corresponding task number set structure is established for each group to generate an image task collaborative grouping structure. The scheduling and reconfiguration submodule calls the task set corresponding to each group in the image task collaborative grouping structure, reorders the task numbers in each group according to the original task order, and uniformly specifies the corresponding execution image number as the scheduling entry in the sorting structure, updates the corresponding task node order in the platform scheduling arrangement structure, and maps the updated content to the regional task distribution layer to generate an image collaborative grouping structure diagram.

Images