Scholarlink

<p>Radiologists can miss subtle secondary findings relevant to tumor staging and management. Large reasoning models (LRMs) may mitigate this by improving interpretive completeness and transparency, yet systematic evaluation remains lacking. We studied 900 multicenter oncologic cases from three Chinese hospitals to compare an LRM’s reasoning processes with its conclusion-only format and two non-reasoning models. Three senior radiologists assessed diagnostic errors and qualitative attributes, and a human-in-the-loop study with six radiologists evaluated workflow-related effects. Cross-language generalization was tested using an English MIMIC-Cancer-90 cohort. Reasoning processes showed the fewest missed or misclassified errors and the highest ratings for comprehensiveness, explainability, and unbiasedness, though with reduced conciseness. Performance dropped when only conclusions were used, and non-reasoning models underperformed across metrics. Improvements were consistent across cancer types, modalities, institutions, and languages. Reader studies confirmed greater perceived completeness and reasoning clarity, especially among juniors, while revealing workflow costs requiring optimization for clinical use.</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

Diagnostic and interpretive gains from reasoning over conclusions with a large reasoning model in radiology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月31日

查看全文
  • Ruixin Wang,
  • Jinghang Wang,
  • Yisong Wang,
  • Chao Zheng,
  • Sihong Huang,
  • Xiaohui Liu,
  • Guoping Tan,
  • Wei Zhao,
  • Zhiyuan Wang,
  • Shaoliang Peng,
  • Jun Liu

摘要

Radiologists can miss subtle secondary findings relevant to tumor staging and management. Large reasoning models (LRMs) may mitigate this by improving interpretive completeness and transparency, yet systematic evaluation remains lacking. We studied 900 multicenter oncologic cases from three Chinese hospitals to compare an LRM’s reasoning processes with its conclusion-only format and two non-reasoning models. Three senior radiologists assessed diagnostic errors and qualitative attributes, and a human-in-the-loop study with six radiologists evaluated workflow-related effects. Cross-language generalization was tested using an English MIMIC-Cancer-90 cohort. Reasoning processes showed the fewest missed or misclassified errors and the highest ratings for comprehensiveness, explainability, and unbiasedness, though with reduced conciseness. Performance dropped when only conclusions were used, and non-reasoning models underperformed across metrics. Improvements were consistent across cancer types, modalities, institutions, and languages. Reader studies confirmed greater perceived completeness and reasoning clarity, especially among juniors, while revealing workflow costs requiring optimization for clinical use.