点击蓝字 关注我们
译文：
END
往期推荐
喜欢此内容的人还喜欢
阅读到这，期待您的点赞、分享，“在看”
硕士生宋向康关于放射学报告生成的研究被 EAAI 期刊录用再创佳绩 近日，实验室硕士生宋向康关于放射学报告生成的研究成果被人工智能领域国际知名期刊《Engineering Applications of Artificial Intelligence》接收。该期刊专注于人工智能在工程领域的应用研究，是中科院分区TOP期刊，最新影响因子8，被CCF列为C类推荐期刊。本工作提出了一种用于放射学报告生成的混合专家框架（MOE-RRG），通过阶段-视图混合专家模块（SV-MoE）实现了阶段感知的纵向建模，并结合分层前缀模块（HP-QF）强化了细粒度的图文语义对齐，显著提升了模型在复杂临床场景下的上下文适应能力和描述准确性。 题目：Mixture of experts for radiology report generation 摘要：Radiology Report Generation is becoming a key tool for boosting efficiency in medical imaging. In clinical practice, semantic shifts across visit stages and variation introduced by imaging views can influence the narrative style and content of clinical reports. However, existing methods often overlook the influence of these variables on report generation. Although some approaches attempt to leverage longitudinal visit and view information to improve report quality, they often rely on simple concatenation, lacking layerwise interaction and hierarchical semantic alignment, which limits their ability to generate reports at a fine-grained level. To address these issues, we propose Mixture of experts for radiology report generation to achieve stage-aware longitudinal modeling and hierarchical semantic alignment. Specifically, we present a stage-view mixture of experts module that routes with a joint stage-view key, guiding individual expert toward their designated clinical scenarios. Furthermore, we introduce a hierarchical prefix module that integrates auxiliary textual cues and image features into a shared token space to generate informative prefix representations, which are injected into every decoder layer to guide report generation and reinforce fine-grained image-text alignment. Experiments on two public datasets demonstrate that our model achieves state-of-the-art performance across both linguistic and clinical efficacy evaluation criteria, exhibiting superior contextual adaptability and descriptive precision.Our code is available athttps://github.com/kongkong935/MOE-RRG 放射报告生成正成为提升医学影像效率的关键工具。在临床实践中，就诊阶段的语义变化以及影像视图带来的变化会影响临床报告的叙述风格和内容。然而，现有方法常常忽视这些变量对报告生成的影响。尽管一些方法试图利用纵向就诊和视图信息提升报告质量，但它们往往依赖简单的串接，缺乏层次交互和层级语义对齐，这限制了其在细粒度生成报告的能力。为解决这些问题，我们提出专家混合生成方案，以实现阶段感知的纵向建模和层级语义对齐。具体来说，我们提出了一个专家阶段视图混合模块，通过联合阶段视图键引导专家，引导专家前往其指定的临床场景。此外，我们引入了一个分层前缀模块，将辅助文本线索和图像特征整合到共享的令牌空间中，生成信息性前缀表示，这些前缀表示被注入每个解码层，指导报告生成并强化细粒度的图像-文本对齐。在两个公开数据集上的实验表明，我们的模型在语言学和临床效能评估标准上均达到了最先进的性能，展现出卓越的上下文适应性和描述精确度。代码链接：https://github.com/kongkong935/MOE-RRG
