A multi-disease joint risk assessment method and system based on an evidence chain of fundus structure and lesions

Abstract

The present application belongs to the technical field of medical image analysis and fundus screening, and relates to a multi-disease joint risk assessment method and system based on fundus structure and lesion evidence chain. In view of the problems that the existing fundus screening is independently output according to a single disease and the evidence basis is scattered, the method performs structure positioning or segmentation and abnormal candidate region detection or segmentation on the interpretable fundus image, constructs an evidence chain representation composed of anatomic structure anchor points, abnormal candidate nodes, spatial relationships between nodes, candidate sources and evidence integrity markers, and based on the evidence chain representation, at least two non-diagnostic screening prompt results of a diabetic retinopathy related feature prompt, an age-related macular degeneration related feature prompt and a glaucoma related structural abnormality prompt are jointly output, and screening prompt sorting, evidence items, abnormal candidate region markers and review prompts are generated. The present application can be used for multi-disease joint screening prompt and result interpretation of the fundus screening platform.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis, fundus screening, evidence chain organization, and non-diagnostic screening indication, and particularly to a multi-disease joint risk assessment method and system based on fundus structure and lesion evidence chain. Background Technology

[0002] Existing fundus screening algorithms typically make independent judgments based on a single disease, such as identifying only features related to diabetic retinopathy, age-related macular degeneration, or glaucoma-related structural abnormalities. While this approach can accomplish individual tasks, different tasks often generate features and output results independently, making it difficult to form a unified and traceable evidence on the same fundus image.

[0003] In practical scenarios of grassroots screening, physical examination screening, and remote image interpretation, operators need to obtain multiple disease screening prompts, evidence, prompt ranking, and follow-up prompts simultaneously after a single data collection. Especially in handheld non-mydriatic fundus screening scenarios, the interpretable image resources acquired in a single session are quite valuable. If multiple independent models are still used to output unrelated results, it will not only lead to fragmented interpretive information but also make it difficult to support subsequent ophthalmological examination prompts, follow-up prompts, and result traceability.

[0004] Existing technologies that simply concatenate structural localization results, anomaly detection results, weights, and confidence scores for multi-task output easily degenerate into ordinary medical image feature fusion or conventional multi-task models, failing to fully reflect the spatial constraints between anatomical structures such as the optic disc, macula, blood vessels, optic cup, and optic disc in fundus images and abnormal candidate regions. Therefore, a method and system are needed that can organize abnormal candidate evidence using anatomical structural anchor points and output multi-disease non-diagnostic screening suggestions based on the evidence chain path. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-disease joint risk assessment method and system based on the evidence chain of fundus structure and lesions. By establishing anatomical structure anchor points in the same interpretable fundus image, abnormal candidate regions are spatially bound to the anatomical structure anchor points to form an evidence chain representation including candidate source, image quality source, and evidence integrity marker. Multiple non-diagnostic screening prompts are generated based on different evidence sub-chains to improve the consistency and traceability of fundus screening result interpretation.

[0006] To achieve the above objectives, the present invention employs the following technical solution: acquiring fundus images and their frame-level quality metadata that have been determined to be readable by end-side quality gating; performing structural localization or segmentation on the fundus images to generate anatomical structure anchor points such as the optic disc, macula, blood vessels, optic cup, and optic disc; performing abnormal candidate region detection or segmentation on the fundus images to generate abnormal candidate nodes such as hemorrhage candidates, exudation candidates, microaneurysm candidates, and cotton wool spot candidates; organizing the spatial relationship, candidate source, image quality source, and evidence integrity of the abnormal candidate nodes and anatomical structure anchor points into an evidence chain representation; generating at least two types of non-diagnostic screening prompts from diabetic retinopathy-related feature prompts, age-related macular degeneration-related feature prompts, and glaucoma-related structural abnormality prompts based on different evidence sub-chains, and performing joint sorting and interpretation output.

[0007] In this invention, the evidence chain representation is not simply a combination of multiple features, nor is it a conventional weighting based solely on node weights or confidence levels. Instead, it uses fundus anatomical structures such as the optic disc, macula, blood vessels, optic cup, and optic disc as anchor points to record the spatial positional relationship of abnormal candidate regions relative to the central macular region, the central optic disc region, the vascular arc, and the posterior pole region, and calls different evidence sub-chains according to different screening prompt types.

[0008] Compared with existing technologies, the present invention has at least the following beneficial effects: First, it reuses a unified chain of evidence in the same interpretable fundus image, reducing the problem of inconsistent interpretations from multiple independent models; Second, it enables the screening prompt results to have a traceable evidence path through the spatial binding of anatomical structure anchor points and abnormal candidate regions; Third, it reduces the impact of single abnormal candidate errors on the ranking of multiple disease prompts through evidence integrity marking and image quality source recording; Fourth, it is suitable for primary care screening, physical examination screening, and remote image reading assistance scenarios through non-diagnostic screening prompts and re-examination prompts. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0010] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0011] Figure 3 This is a schematic diagram illustrating the construction of the evidence chain in this invention.

[0012] Figure 4 This is a schematic diagram illustrating the output of the present invention. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0014] Example 1: System Overall Structure like Figure 1 As shown, the system of the present invention includes an image input module 1, a structure localization module 2, a structure segmentation module 3, an anomaly candidate detection module 4, an evidence chain construction module 5, a multi-disease screening prompt module 6, a result interpretation module 7, and a joint output module 8. The image input module 1 receives fundus images and their frame-level quality metadata that have been determined to be interpretable by end-side quality gating. The structure localization module 2 and the structure segmentation module 3 generate anatomical anchor points for structures such as the optic disc, macula, blood vessels, optic cup, and optic disc. The anomaly candidate detection module 4 generates abnormal candidate regions such as hemorrhage candidates, exudation candidates, microaneurysms candidates, and cotton wool spot candidates. The evidence chain construction module 5 constructs an evidence chain representation. The multi-disease screening prompt module 6 generates at least two types of non-diagnostic screening prompt results. The result interpretation module 7 and the joint output module 8 output interpretation and ranking results.

[0015] Example 2: Method Flow like Figure 2 As shown, firstly, fundus images and their frame-level quality metadata that are deemed interpretable through end-side quality gating are acquired; then, the optic disc, macula, blood vessels, or optic cup / disc structures are located or segmented to generate anatomical structure anchor points; subsequently, abnormal candidate regions such as hemorrhage candidates, exudation candidates, microaneurysm candidates, or cotton wool spot candidates are detected or segmented; then, the abnormal candidate regions are spatially bound to the anatomical structure anchor points to construct an evidence chain representation; then, at least two types of non-diagnostic screening prompts are generated based on different evidence sub-chains; finally, joint sorting is performed, and evidence entries, abnormal candidate region markers, quality impact descriptions, and re-examination prompts are generated.

[0016] Example 3: Constructing the Chain of Evidence like Figure 3 As shown, structural feature set 9 and abnormal candidate feature set 10 are jointly entered into evidence chain representation 11. The evidence chain representation 11 includes anatomical structure anchor points, abnormal candidate nodes, spatial relationships between abnormal candidate nodes and anatomical structure anchor points, candidate sources, image quality sources, and evidence integrity markers, and further inputs multi-disease screening prompt results 12 to form multi-disease joint screening prompt results.

[0017] In one optional implementation, structural feature set 9 includes the optic disc center location, optic disc boundary, macular center location, optic cup boundary, optic cup-optic disc ratio candidate value, location of major vascular arcs, and vascular bifurcation characteristics; abnormal candidate feature set 10 includes the number of hemorrhage candidates, exudate candidate area, number of microaneurysms candidate, cotton wool patch candidate location, and candidate confidence level. The evidence chain representation 11 further records the distance of each abnormal candidate region to the macular center, its relative orientation to the optic disc center, its spatial adjacency with the vascular arc, and its distribution density in different anatomical regions.

[0018] Example 4: Multi-disease screening prompts In one implementation, diabetic retinopathy-related features are generated based on the number of candidate microaneurysms, the number of candidate hemorrhages, the area of ​​candidate exudates, the distance of abnormal candidate regions relative to the macular center, and vascular abnormality candidate information; age-related macular degeneration-related features are generated based on abnormal candidate regions in and around the macular center, pigmentary heterogeneity candidate regions, exudative candidate regions, and central structural change candidate information; glaucoma-related structural abnormalities are generated based on the optic cup-disc ratio candidate value, optic disc margin candidate abnormalities, peripapillary structural candidate abnormalities, vascular course candidate abnormalities, and optic disc morphology candidate abnormalities.

[0019] The multi-disease screening prompt module 6 can uniformly map the above results into screening prompt levels, prompt scores, or prompt confidence levels. These screening prompt levels, prompt scores, or prompt confidence levels are used to indicate whether further ophthalmological examination or re-examination is needed, and are not considered as clinical diagnostic conclusions.

[0020] Example 5: Explanation Output like Figure 4 As shown, the screening prompt results 13, the evidence item list 14, and the abnormal candidate areas and structural annotations 15 are jointly input into the interpretation output 16. The interpretation output 16 further forms the review prompts and priority results 17. The interpretation output 16 includes each screening prompt result, the screening prompt ranking, the description of the abnormal candidate areas, the main evidence items, the evidence chain path, and the quality impact description.

[0021] For example, when an image simultaneously contains candidate exudative regions near the macula and a high number of candidate microaneurysms, the system can prioritize features related to diabetic retinopathy. Similarly, when the optic cup-to-disc ratio is high and there are candidate abnormalities in the disc-edge relationship, the system can prioritize glaucoma-related structural abnormalities. Through these methods, primary care screening personnel or remote image readers can more intuitively understand the rationale behind the combined suggestive results.

[0022] Example 6: Joint Output Strategy In one implementation, the joint output module 8 sorts the screening prompts from high to low confidence. In another implementation, the joint output module 8 prioritizes abnormal candidate regions involving the central macular region and then sorts them based on the completeness of the evidence chain and the degree of impact on image quality. In yet another implementation, the system outputs the non-diagnostic screening prompts for multiple diseases for the same image simultaneously, but gives a prominent mark to the priority review items.

Claims

1. A multi-disease joint risk assessment method based on evidence chains of fundus structure and lesions, characterized in that, The process includes the following steps: S101, acquiring fundus images deemed readable by end-side quality gating and their frame-level quality metadata, wherein the frame-level quality metadata includes at least the evaluation results of sharpness, reflectivity, exposure, coverage, and motion stability; S102, performing structural localization or segmentation on the readable fundus images to generate anatomical anchor points that include at least the optic disc region, the macular region, and vascular structures; S103, performing abnormal candidate region detection or segmentation on the readable fundus images to obtain at least two types of abnormal candidate regions; S104, binding the abnormal candidate regions to their spatial relationships relative to the macular center, optic disc center, vascular arc, or posterior pole region, and combining this with the candidate source and image quality source. S105. Construct an evidence chain representation with evidence integrity markers; S106. Based on the evidence chain representation, call different evidence sub-chains to generate at least two types of non-diagnostic screening prompts from diabetic retinopathy-related feature prompts, age-related macular degeneration-related feature prompts, and glaucoma-related structural abnormality prompts; S107. Jointly sort the at least two types of non-diagnostic screening prompts according to the evidence chain integrity, confidence of abnormal candidate regions, whether they involve the central macular region, whether they involve optic disc structural abnormalities, and the degree of impact on image quality; S108. Output the evidence entries, abnormal candidate region markers, structural segmentation results, quality impact descriptions, and re-examination prompts corresponding to the non-diagnostic screening prompts.

2. The method according to claim 1, characterized in that, The chain of evidence includes anatomical anchor points, abnormal candidate nodes, spatial relationships between abnormal candidate nodes and anatomical anchor points, candidate sources, image quality sources, and evidence integrity markers; the spatial relationships include at least the positional relationships of abnormal candidate nodes relative to the macula center, optic disc center, or vascular arc.

3. The method according to claim 1, characterized in that, The structural localization or segmentation includes at least optic disc localization and macular localization, and further includes at least one of optic cup-optic disc segmentation, vascular segmentation, or posterior pole region determination.

4. The method according to claim 1, characterized in that, The abnormal candidate region detection or segmentation includes at least two of the following: hemorrhage candidate detection, exudation candidate detection, microaneurysm candidate detection, and cotton wool spot candidate detection.

5. The method according to claim 1, characterized in that, The diabetic retinopathy-related features are generated based at least on the number of microaneurysm candidates, the number of hemorrhage candidates, the area of ​​exudation candidates, the distance of abnormal candidate regions relative to the macula center, and vascular abnormality candidate information.

6. The method according to claim 1, characterized in that, The age-related macular degeneration-related feature suggestions are generated based on abnormal candidate regions, uneven pigmentation candidate regions, exudative candidate regions, and structural change candidate information in the central and surrounding macular regions.

7. The method according to claim 1, characterized in that, The glaucoma-related structural abnormality indications are generated based on candidate values ​​for the optic cup-disc ratio, candidate abnormalities for the optic disc margin, candidate abnormalities for the periplasmic structure, candidate abnormalities for the course of blood vessels, and candidate abnormalities for the optic disc morphology.

8. The method according to claim 1, characterized in that, The joint sorting and interpretation output includes generating a sorted list of screening prompts, a chain of evidence path, a list of evidence entries, markers for abnormal candidate regions, structural segmentation results, and re-examination prompts; wherein, the non-diagnostic screening prompt results are used to indicate whether further ophthalmological examination is needed and are not used as clinical diagnostic conclusions.

9. A multi-disease joint risk assessment system based on the evidence chain of fundus structure and lesions, characterized in that, The system includes an image input module (1), a structure localization module (2), a structure segmentation module (3), an anomaly candidate detection module (4), an evidence chain construction module (5), a multi-disease screening prompt module (6), a result interpretation module (7), and a joint output module (8). The image input module (1) is used to receive interpretable fundus images and frame-level quality metadata. The structure localization module (2) and the structure segmentation module (3) are used to generate anatomical structure anchor points. The anomaly candidate detection module (4) is used to generate anomaly candidate regions. The evidence chain construction module (5) is used to construct the evidence chain representation as described in any one of claims 1 to 8. The multi-disease screening prompt module (6) is used to generate at least two types of non-diagnostic screening prompt results. The result interpretation module (7) and the joint output module (8) are used to output evidence entries, anomaly candidate region markers, structure segmentation results, screening prompt ranking, and re-examination prompts.

10. The system according to claim 9, characterized in that, The evidence chain construction module (5) includes a structural anchor point management unit, an abnormal candidate node management unit, a spatial relationship binding unit, a quality impact marking unit, and an evidence integrity marking unit; the joint output module (8) adjusts the screening prompt sorting according to the evidence integrity marking and the quality impact marking.

Images