Scholarlink

<p>The generation of radiology reports represents a significant advancement in medical AI, enhancing diagnostic accuracy and supporting clinical decision-making. However, existing methods often suffer from poor cross-modal alignment, data bias, and inadequate representation of rare abnormalities. To address these challenges, we propose memory-driven weighted contrastive learning (MWCL), a novel radiology report generation framework that integrates three synergistic modules: (1) an adaptive feature refinement Module (AFRM) for denoising and enhancing abnormal regions; (2) a memory-driven alignment module (MDAM) that employs a dynamic memory bank to selectively reinforce visual-textual associations; and (3) a weighted contrastive learning (WCL) mechanism that leverages auxiliary similarity signals to optimize feature separation while minimizing false negatives. MWCL is the first to jointly optimize feature refinement, alignment, and contrastive separation within a unified architecture. Extensive evaluations on benchmark datasets demonstrate that the MWCL achieves substantial performance gains, outperforming R2GenCNM by 10.85% in BLEU-1, 16.12% in BLEU-2, 21.80% in BLEU-3, 24.84% in BLEU-4, 8.12% in METEOR, 12.36% in ROUGE-L, and 30.99% in CIDEr. These improvements confirm the effectiveness of the MWCL in generating high-quality and clinically meaningful reports.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mwcl: Memory-driven and mapping alignment with weighted contrastive learning for radiology reports

  • 期刊论文

  • 发表日期：

    2025年10月31日

查看全文
  • Muhammad Usman,
  • Ziwei Ma,
  • Zhang Yijia

摘要

The generation of radiology reports represents a significant advancement in medical AI, enhancing diagnostic accuracy and supporting clinical decision-making. However, existing methods often suffer from poor cross-modal alignment, data bias, and inadequate representation of rare abnormalities. To address these challenges, we propose memory-driven weighted contrastive learning (MWCL), a novel radiology report generation framework that integrates three synergistic modules: (1) an adaptive feature refinement Module (AFRM) for denoising and enhancing abnormal regions; (2) a memory-driven alignment module (MDAM) that employs a dynamic memory bank to selectively reinforce visual-textual associations; and (3) a weighted contrastive learning (WCL) mechanism that leverages auxiliary similarity signals to optimize feature separation while minimizing false negatives. MWCL is the first to jointly optimize feature refinement, alignment, and contrastive separation within a unified architecture. Extensive evaluations on benchmark datasets demonstrate that the MWCL achieves substantial performance gains, outperforming R2GenCNM by 10.85% in BLEU-1, 16.12% in BLEU-2, 21.80% in BLEU-3, 24.84% in BLEU-4, 8.12% in METEOR, 12.36% in ROUGE-L, and 30.99% in CIDEr. These improvements confirm the effectiveness of the MWCL in generating high-quality and clinically meaningful reports.