Debiasing method for large vision-language models in the medical domain
WANG Wenhao
HAN Zhongyi
WANG Bin
WU Shenjing
WEI Benzheng
Abstract：To address the bias issues caused by uneven data distribution in large vision-language models(LVLMs)for medical auxiliary diagnosis,we proposed a universal medical debiasing method called bias fairness enhancement(BFE).The effectiveness of BFE in mitigating bias was validated by constructing a benchmark to evaluate bias issues in medical LVLMs.This benchmark coverd bi-nary classification,multi-classification,and open-ended question tasks,while incorporating two parameters that control output ran-domness to ensure the method's generality.Experimental results demonstrated that BFE outperformed other mainstream debiasing meth-ods in the medical LVLMs(LLaVA-Med,SkinGPT).Notably,in open-ended medical question-answering tasks,LLaVA-Med's per-formance improved by 6.3%(cd_beta=0.1)and 6.7%(cd_beta=0.5).These findings indicate that BFE can effectively alleviate bias in medical LVLMs,and provide significant support for improving the accuracy and reliability of medical auxiliary diagnosis.
Keywords：Large vision-language modelsDebiasing algorithmsMedical assisted diagnosisBias issuesContrastive learning
Publication Date：2025-08-30
Online Publishing Date：2026-09-11(First online date of this platform, not the publication date of the document)
Pages：9( 229-237 )