A method, system, and storage medium for clinical diagnostic medical visual question answering based on multi-margin collaborative debiasing.

By conducting multi-level analysis of the medical visual question answering model and adding MAM and DCL modules, the multimodal representation was optimized, which solved the bias problem of the model in out-of-distribution clinical diagnosis and improved the accuracy and robustness of medical visual question answering.

CN122135946APending Publication Date: 2026-06-02HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
Filing Date
2026-04-16
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing medical visual question answering models tend to over-rely on text questions in out-of-distribution clinical diagnosis, ignoring key lesion features in medical images, leading to medical language bias and affecting the reliability and accuracy of diagnosis, especially in the diagnosis of rare diseases or complex cases.

Method used

This paper theoretically analyzes the formation mechanism of medical language bias from three aspects: modal gradient imbalance, feature fusion bias, and classifier weight direction bias. The MAM mechanism and DCL module are added to the UpDn model to optimize the multimodal representation of the model and enhance its robustness and accuracy.

Benefits of technology

It effectively alleviates medical language bias, improves the model's prediction accuracy and robustness on out-of-distribution clinical diagnostic data, and enhances diagnostic performance for rare or complex cases.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122135946A_ABST
    Figure CN122135946A_ABST
Patent Text Reader

Abstract

This invention relates to a method, system, and storage medium for clinical diagnostic medical visual question answering based on multi-margin collaborative debiasing. The method includes: Step 1: theoretically analyzing the formation mechanism of medical language bias in clinical diagnosis from the perspectives of modal gradient imbalance, feature fusion bias, and classifier weight direction bias, and providing experimental evidence; Step 2: theoretically proving that the margin mechanism can alleviate medical language bias in clinical diagnosis; Step 3: using UpDn as the baseline model, adding the MAM mechanism and DCL module; Step 4: training the model; adding MAM loss and DCL loss as the overall loss function to the cross-entropy loss to obtain the MMCD model for clinical diagnostic medical visual question answering; Step 5: testing the effect of the trained model on the validation set. The beneficial effects of this invention are: providing theoretical support for the margin mechanism to alleviate medical language bias in clinical diagnosis, enhancing the persuasiveness and credibility of the technology.
Need to check novelty before this filing date? Find Prior Art