首页 > AI前沿 > Multimodal Error Detection in Chest X-ray Reports: A Large Language Model Reading Image Evidence

Multimodal Error Detection in Chest X-ray Reports: A Large Language Model Reading Image Evidence

arXiv自然语言 2026-11-20 01:46 16 阅读 查看原文
用户需要我把一段纯文本的新闻正文排版成语义化的HTML片段。让我仔细分析这段文本,看看哪些部分需要用哪些HTML标签来包裹。 文本内容是关于Rad-Linter系统的技术描述,主要讲它如何检查胸部X光报告与图像派生证据的一致性。 分析结构: 1. 第一段介绍Rad-Linter的基本功能和工作原理 2. 第二段描述评估方法 3. 第三段讲实验结果和对比 4. 最后一段是结论 让我考虑如何划分: - 整体可以分成几个部分,用小标题来划分?但用户说"长文用

/

划分小节",这段文本不算特别长,但逻辑上有几个清晰的层次。 - 看内容,第一段是总体介绍,后面是评估和结果,再是结论。可以用小标题来划分小节。 让我再仔细读一下文本: 第一部分(总体介绍):"Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores." 这部分是系统和方法的总体介绍,可以作为一个小节。 第二部分(评估方法):"Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy." 这是评估方法,可以作为一个小节。 第三部分(实验结果):"On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, including unmentioned and negated findings; larger differences occur on the remaining item types, which require interpreting how findings are described. The same language model without fine-tuning remains near 50% accuracy. On a second set of 181 quads that also includes reports on which the automatic labelers disagree, pair accuracy is 93.4% versus 75.3%." 这是具体实验结果,可以作为小节。 第四部分(结论):"These controlled results show that, given the same image-derived scores, a trained language model detects report-score inconsistencies more accurately than the tested rule baseline, with the advantage concentrated in interpreting how findings are worded." 这是结论。 现在考虑标签使用: - 用

来划分小节标题,比如"系统概述"、"评估方法"、"实验结果"、"结论"——但等等,用户说"不得增删改内容",我不能添加原文没有的标题文字。所以不能自己编小标题。 重新考虑:用户说"长文用

/

划分小节",但如果原文没有小标题,我是否需要强行加?规则4说"保留原文所有信息,不得增删改内容"。添加小标题算不算增删内容? 仔细看规则:用户说"长文用

/

划分小节",但同时又说"不得增删改内容"。这里可能有冲突。但如果我不加标题,整篇就只是几个

段落。让我再想想。 其实这段文本虽然逻辑清晰,但没有明显的小标题。为了遵守"不得增删改内容"的硬性规则,我应该只使用

来包裹自然段落,而不强行添加原文没有的小标题。 让我看看哪些句子需要强调。用户说"重要句子可用"。比如结论部分和关键数据结果可以加。 让我划分自然段落: 段落1:Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores. 段落2:Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy. 段落3:On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, including unmentioned and negated findings; larger differences occur on the remaining item types, which require interpreting how findings are described. The same language model without fine-tuning remains near 50% accuracy. On a second set of 181 quads that also includes reports on which the automatic labelers disagree, pair accuracy is 93.4% versus 75.3%. 段落4:These controlled results show that, given the same image-derived scores, a trained language model detects report-score inconsistencies more accurately than the tested rule baseline, with the advantage concentrated in interpreting how findings are worded. 这样4个自然段落,每个用

包裹。没有列表、引用或代码内容。 关于的使用,我可以在关键数据和结论处加: - "98.5% pair accuracy" 和 "80.5%" 这些关键数据 - "94.4%" 和 "56.5%" - "93.4% versus 75.3%" - 结论中的关键陈述 不过用户说"重要句子可用",不是必须。我应该自然地加,不要过度。让我选几处关键的: - 第一段中"checks their consistency with the report"可以不加,保持克制 - 第三段的实验数据比较重要 - 第四段的结论很重要 让我最终决定: 段落1不加strong,因为是方法介绍 段落2中的"either input alone permits only 50% accuracy"可以加,因为是关键设定 段落3中的数据和结果加strong 段落4整个结论加strong 不过要注意不要过度使用。让我再斟酌一下。用户说的是"重要句子可用",我选择几个真正重要的句子来加。 段落3中:"Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline" 这整句很重要,可以加strong。"classifies all four pairs correctly in 94.4% of quads versus 56.5%"也可以。 "On a second set of 181 quads that also includes reports on which the automatic labelers disagree, pair accuracy is 93.4% versus 75.3%." 也可以加。 段落4整段是结论,可以加strong。 让我把结果写出来。注意要保留所有原文信息,不能增删改。 最终HTML:

Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores.

Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy.

On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, including unmentioned and negated findings; larger differences occur on the remaining item types, which require interpreting how findings are described. The same language model without fine-