Scholarlink

<p>Ultrasound(US) report generation aims to alleviate the burden on clinicians and has emerged as a significant research focus. However, automatically generating clinically accurate and coherent reports from US images remains challenging due to the variability of US images, operator dependence, and the need for expert knowledge to bridge the semantic gap between visual and textual features. Moreover, prior works usually ignore data redundancy common in medical scenarios, where a report corresponds to multiple images, but only those mirroring textual descriptions are valid. We propose a Redundancy-Adaptive Dual Memory Network (RADMN) for knowledge-enhanced ultrasound report. We design a redundancy-adaptive visual encoder with a dual attention mechanism to selectively fuse valid visual information. The Dual Memory Network is designed for more accurate and coherent reports generation. It consists of a static memory module storing fine-grained clinical attributes and a dynamic memory matrix for evolving contextual information. A memory alignment module bridges raw visual features with knowledge-grounded representations and dynamically updates the contextual memory. Finally, a knowledge-injected decoder integrates this global diagnostic knowledge with dynamic contextual guidance to generate reports. The proposed RA-DMN was evaluated on a renal artery ultrasound dataset from Peking Union Medical College Hospital. Experimental results demonstrate that RA-DMN achieves competitive performance compared to state-of-the-art (SOTA) models, including R2Gen, R2GenCMN, and DACG, with scores of 0.722, 0.687, 0.657, 0.631, and 0.799 on BLEU-1, BLEU-2, BLEU-3, BLEU-4, and ROUGE-L, respectively. Compared to the previous best-performing model (DACG), these metrics represent relative improvements of 8.6%, 8.2%, 7.2%, 5.9%, and 2.8%. The RA-DMN effectively addresses key challenges in ultrasound report generation by mitigating data redundancy, bridging the visual-textual semantic gap, and ensuring clinical consistency. The results indicate its potential for generating reliable and clinically valuable reports in practical applications.</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

Redundancy-adaptive dual memory network for knowledge-enhanced ultrasound report generation

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Xin Wang,
  • Guangyi Huang,
  • Yuqing Yang,
  • Peipei Zhang,
  • Hongyan Wang,
  • Jianchu Li,
  • Fujiao Ju,
  • Yujuan Feng

摘要

Ultrasound(US) report generation aims to alleviate the burden on clinicians and has emerged as a significant research focus. However, automatically generating clinically accurate and coherent reports from US images remains challenging due to the variability of US images, operator dependence, and the need for expert knowledge to bridge the semantic gap between visual and textual features. Moreover, prior works usually ignore data redundancy common in medical scenarios, where a report corresponds to multiple images, but only those mirroring textual descriptions are valid. We propose a Redundancy-Adaptive Dual Memory Network (RADMN) for knowledge-enhanced ultrasound report. We design a redundancy-adaptive visual encoder with a dual attention mechanism to selectively fuse valid visual information. The Dual Memory Network is designed for more accurate and coherent reports generation. It consists of a static memory module storing fine-grained clinical attributes and a dynamic memory matrix for evolving contextual information. A memory alignment module bridges raw visual features with knowledge-grounded representations and dynamically updates the contextual memory. Finally, a knowledge-injected decoder integrates this global diagnostic knowledge with dynamic contextual guidance to generate reports. The proposed RA-DMN was evaluated on a renal artery ultrasound dataset from Peking Union Medical College Hospital. Experimental results demonstrate that RA-DMN achieves competitive performance compared to state-of-the-art (SOTA) models, including R2Gen, R2GenCMN, and DACG, with scores of 0.722, 0.687, 0.657, 0.631, and 0.799 on BLEU-1, BLEU-2, BLEU-3, BLEU-4, and ROUGE-L, respectively. Compared to the previous best-performing model (DACG), these metrics represent relative improvements of 8.6%, 8.2%, 7.2%, 5.9%, and 2.8%. The RA-DMN effectively addresses key challenges in ultrasound report generation by mitigating data redundancy, bridging the visual-textual semantic gap, and ensuring clinical consistency. The results indicate its potential for generating reliable and clinically valuable reports in practical applications.