CN-122023932-A - Method and system for classifying myopic macular degeneration fundus color photographic images

CN122023932ACN 122023932 ACN122023932 ACN 122023932ACN-122023932-A

Abstract

The invention discloses a method and a system for classifying myopic macular degeneration fundus color photograph images, and belongs to the technical field of image classification. The method comprises the steps of obtaining fundus color illumination to be analyzed, preprocessing the fundus color illumination to extract an effective visual field area and reduce image acquisition difference interference, inputting the preprocessed image into a deep learning classification network, and outputting a META-PM classification result corresponding to myopic macular degeneration. The invention fuses the global structure and the local focus evidence simultaneously on the premise of not depending on accurate focus labeling, and explicitly models the ordered hierarchical relationship, thereby improving the robustness of the cross-equipment and the adaptability of long tail distribution.

Inventors

CHI WEI
YANG WEIHUA
HAN HONGZHE
YANG JUAN
WANG XIAOYAN
JI YUKE

Assignees

深圳市眼科医院(深圳市眼病防治研究所)

Dates

Publication Date: 20260512
Application Date: 20260212

Claims (9)

1. A method for classifying myopic macular degeneration fundus color photograph images, which is characterized by comprising the following steps: Acquiring fundus color illumination to be analyzed; preprocessing the fundus color illumination to extract an effective visual field area and reduce image acquisition difference interference; Inputting the preprocessed image to a deep learning classification network, the deep learning classification network comprising: a global branch for encoding the whole image and extracting global features; A local branch, which is used for coding a plurality of local areas extracted from the image, and carrying out weighted aggregation on a plurality of local features through an attention mechanism to obtain local aggregation features; Fusing the global features and the local aggregation features to obtain fused features; And inputting the fusion characteristics to an ordered grading output head, and explicitly modeling ordered progressive relations among different lesion grades through the ordered grading output head so as to output META-PM grading results corresponding to the myopic macular lesions.
2. A method of classifying myopic macular degeneration fundus illumination images as claimed in claim 1, wherein said preprocessing of said fundus illumination comprises: Positioning an effective field circular area in the fundus illumination; Performing boundary expansion based on the circle center and the radius of the circular area of the effective visual field, and cutting to obtain a square interested area image containing the complete effective visual field; and scaling the square region-of-interest image to obtain a standardized image.
3. The method of claim 2, wherein locating the effective field of view circular region in the fundus illumination comprises locating by at least one of thresholding, color channel constraints, and morphological operations.
4. A method of classifying myopic macular degeneration fundus illumination images as claimed in claim 2, wherein said preprocessing of said fundus illumination further comprises: and carrying out illumination correction and color intensity normalization processing on the standardized image.
5. A method of classifying myopic macular degeneration fundus illumination images as claimed in claim 1, wherein in said local branches, extracting a plurality of local areas from said images comprises: setting a plurality of candidate center points in the image; superimposing random perturbations on each candidate center point; and cutting the position after disturbance as a center according to a fixed size to obtain a plurality of local area image blocks.
6. A method of classifying myopic macular degeneration fundus illumination images as claimed in claim 1, wherein in said local branches, weighting and aggregating a plurality of local features by an attention mechanism comprises: Calculating the attention weight corresponding to each local feature; According to the attention weight, carrying out weighted summation on the local features to obtain the local aggregation feature; And obtaining explanatory information indicating the critical lesion area according to the attention weight.
7. The method of claim 1, wherein said fusing said global features with said local aggregate features comprises fusing via a mapping layer comprising at least one of a fully connected layer, a normalized layer, or a discard layer.
8. A myopic macular degeneration fundus illumination image classification method as claimed in claim 1, wherein said deep learning classification network is trained in a training phase with an ordered regression loss function configured to penalize prediction errors inconsistent with clinical classification distances, wherein in the training phase, according to sample distribution of each category in training data, adaptive weights are set for different threshold subtasks in said ordered regression loss function for relieving category imbalance.
9. A system for classifying a myopic macular degeneration fundus color photograph image using the method of any one of claims 1 to 8, comprising: The image acquisition module is used for acquiring fundus color photographs to be analyzed; The preprocessing module is used for preprocessing the fundus color illumination so as to extract an effective visual field area and reduce image acquisition difference interference; The classification module inputs the preprocessed image to a deep learning classification network, wherein the deep learning classification network comprises: a global branch for encoding the whole image and extracting global features; A local branch, which is used for coding a plurality of local areas extracted from the image, and carrying out weighted aggregation on a plurality of local features through an attention mechanism to obtain local aggregation features; Fusing the global features and the local aggregation features to obtain fused features; And inputting the fusion characteristics to an ordered grading output head, and explicitly modeling ordered progressive relations among different lesion grades through the ordered grading output head so as to output META-PM grading results corresponding to the myopic macular lesions.

Description

Method and system for classifying myopic macular degeneration fundus color photographic images Technical Field The invention relates to the technical field of image classification, in particular to a method and a system for classifying myopic macular degeneration fundus color photographic images. Background Pathological myopia (Pathological Myopia, PM) and its associated myopic maculopathy (Myopic Maculopathy, MM) are one of the important causes of irreversible central vision loss. The international grading system META-PM is used for classifying fundus color illumination into five grades of C0 (no MM), C1 (leopard-shaped fundus), C2 (diffuse chorioretinal atrophy), C3 (zebra-shaped chorioretinal atrophy) and C4 (macular atrophy) for disease screening, follow-up and grading management. The conventional MM classification mainly depends on manual film reading of fundus special doctors, has the problems of poor subjective consistency, insufficient basic medical resources, low screening efficiency for large-scale high-myopia crowd, and the like, and simultaneously reduces the robustness of the traditional computer vision and simple deep learning model due to black/background interference, dark corners and exposure differences caused by the difference of fundus color illumination acquisition equipment and illumination conditions. In addition, META-PM belongs to ordered classification tasks (C0 < C1< C2< C3< C4), confusion of adjacent grades is more common, cross-grade erroneous judgment inconsistent with clinical classification distance is easy to occur only according to common multi-classification training, and performance is easy to degrade under long tail distribution (few high-grade samples). Therefore, there is a need for an MM automatic grading method and system that can fuse global structure and local lesion evidence simultaneously, explicitly model ordered grading relationships, and improve cross-device robustness and long tail distribution adaptability without relying on accurate lesion labeling. Disclosure of Invention In view of the above, the present invention provides a method and a system for classifying myopic macular degeneration fundus color photograph images, which are used for solving the technical problems existing in the prior art. In order to achieve the above purpose, the present invention adopts the following technical scheme: a method for classifying myopic macular degeneration fundus color photograph images, comprising: Acquiring fundus color illumination to be analyzed; preprocessing the fundus color illumination to extract an effective visual field area and reduce image acquisition difference interference; Inputting the preprocessed image to a deep learning classification network, the deep learning classification network comprising: a global branch for encoding the whole image and extracting global features; A local branch, which is used for coding a plurality of local areas extracted from the image, and carrying out weighted aggregation on a plurality of local features through an attention mechanism to obtain local aggregation features; Fusing the global features and the local aggregation features to obtain fused features; And inputting the fusion characteristics to an ordered grading output head, and explicitly modeling ordered progressive relations among different lesion grades through the ordered grading output head so as to output META-PM grading results corresponding to the myopic macular lesions. Further, the preprocessing of the fundus color photograph includes: Positioning an effective field circular area in the fundus illumination; Performing boundary expansion based on the circle center and the radius of the circular area of the effective visual field, and cutting to obtain a square interested area image containing the complete effective visual field; and scaling the square region-of-interest image to obtain a standardized image. Further, the locating the effective field of view circular region in the fundus color photograph comprises locating by at least one of thresholding, color channel constraints and morphological operations. Further, the preprocessing the fundus color photograph further includes: and carrying out illumination correction and color intensity normalization processing on the standardized image. Further, in the local branch, extracting a plurality of local areas from the image includes: setting a plurality of candidate center points in the image; superimposing random perturbations on each candidate center point; and cutting the position after disturbance as a center according to a fixed size to obtain a plurality of local area image blocks. Further, in the local branch, the weighting and aggregating the plurality of local features through the attention mechanism includes: Calculating the attention weight corresponding to each local feature; According to the attention weight, carrying out weighted summation on the local features to obtain the local aggregati