Method and device for recognizing fundus diseases based on prototype memory bank and VLM

By constructing a fundus disease identification method based on prototype memory and visual-language model, the problem of data scarcity and lack of medical knowledge in the differential diagnosis of multiple types of fundus diseases by deep learning models is solved, and efficient identification and robust diagnosis of complex fundus diseases are achieved.

CN122392896APending Publication Date: 2026-07-14HUNAN UNIV OF CHINESE MEDICINE
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HUNAN UNIV OF CHINESE MEDICINE
Filing Date
2026-05-21
Publication Date
2026-07-14

AI Technical Summary

Technical Problem

Existing deep learning models suffer from problems such as data scarcity, lack of medical knowledge modeling, computational resources, and overfitting risk when facing the differential diagnosis of various types of fundus diseases, especially performing poorly in long-tailed distributions and complex cases.

Method used

We employ a method for identifying fundus diseases based on prototype memory and visual-language model (VLM). By constructing a visual-language contrastive learning network and combining an alignment mechanism based on evidence consistency and a prototype vector updated by momentum, we explicitly model noise and uncertainty in cross-modal matching, establish category semantic anchors and sample-level memory queues, and achieve cross-batch hard case mining and long-tail information compensation.

Benefits of technology

It improves the model's ability to identify complex fundus diseases, enhances diagnostic robustness and generalization, and can effectively identify multiple types of fundus diseases, especially providing intelligent screening and assisted diagnosis solutions under long-tail distribution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122392896A_ABST
    Figure CN122392896A_ABST
Patent Text Reader

Abstract

The application relates to an eye fundus disease recognition method and device based on a prototype memory bank and a VLM. The method establishes stable category semantic anchors by constructing momentum-updated prototype vectors for each disease category. By introducing a memory queue that preferentially retains rare categories and high-uncertainty samples, cross-batch difficult example mining and long-tail information compensation are achieved. Meanwhile, in combination with an evidence-consistency-based visual-language alignment mechanism, noise and uncertainty in the cross-modal matching process are explicitly modeled, thereby enhancing the model's representation ability and diagnosis robustness for complex eye fundus diseases. The method can effectively improve the recognition performance of multiple types of eye fundus diseases under a long-tail distribution, and provides a solution with good generalization ability and application prospect for intelligent screening and auxiliary diagnosis of complex eye fundus diseases.
Need to check novelty before this filing date? Find Prior Art