Scholarlink

<p>Recent advances in multimodal learning have propelled rapid progress in medical vision-language research. As a fundamental task in this domain, automatic radiology report generation has garnered increasing attention for its potential to alleviate radiologists’ workload and enhance diagnostic efficiency. However, radiology report generation remains challenged by visual-textual data biases caused by sparse abnormal regions in images and dominant normal descriptions in reports, as well as the lack of related clinical expertise essential for accurate and reliable diagnosis. In this paper, we propose a novel Boltzmann-driven Dynamic Annealing Network for Knowledge-guided Radiology Report Generation. Our approach employs a Structure-Semantic Adaptive Intensifier that preserves both fine-grained medical image structure and semantic information. A Dual-view Perception Aggregator effectively captures high-frequency components critical for medical image and integrates diagnostic information from dual views, resulting in a unified global visual representation. We introduce a Dynamic Annealing Sampling assisted Knowledge-Enhancement Module that employs Boltzmann distribution with an annealing mechanism to process global visual features, while integrating medical knowledge with dynamic sparse attention for precise lesion identification. This strategy efficiently leverages medical knowledge for guidance while mitigating semantic interference caused by redundant image backgrounds. Extensive experiments on IU-Xray and MIMIC-CXR datasets demonstrate that our method outperforms state-of-the-art approaches across all evaluation metrics. Our code is available in: <a href="https://github.com/lm089381/BDANet">https://github.com/lm089381/BDANet</a>.</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

Boltzmann-driven dynamic annealing network for knowledge-guided radiology report generation

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Wei Xiao,
  • Man Liu,
  • He Zhu,
  • Jiajia Zhang

摘要

Recent advances in multimodal learning have propelled rapid progress in medical vision-language research. As a fundamental task in this domain, automatic radiology report generation has garnered increasing attention for its potential to alleviate radiologists’ workload and enhance diagnostic efficiency. However, radiology report generation remains challenged by visual-textual data biases caused by sparse abnormal regions in images and dominant normal descriptions in reports, as well as the lack of related clinical expertise essential for accurate and reliable diagnosis. In this paper, we propose a novel Boltzmann-driven Dynamic Annealing Network for Knowledge-guided Radiology Report Generation. Our approach employs a Structure-Semantic Adaptive Intensifier that preserves both fine-grained medical image structure and semantic information. A Dual-view Perception Aggregator effectively captures high-frequency components critical for medical image and integrates diagnostic information from dual views, resulting in a unified global visual representation. We introduce a Dynamic Annealing Sampling assisted Knowledge-Enhancement Module that employs Boltzmann distribution with an annealing mechanism to process global visual features, while integrating medical knowledge with dynamic sparse attention for precise lesion identification. This strategy efficiently leverages medical knowledge for guidance while mitigating semantic interference caused by redundant image backgrounds. Extensive experiments on IU-Xray and MIMIC-CXR datasets demonstrate that our method outperforms state-of-the-art approaches across all evaluation metrics. Our code is available in: https://github.com/lm089381/BDANet.