Scholarlink

Background <p>The current standard of subjective assessment by radiologists for lymphovascular invasion (LVI) in colorectal cancer (CRC) using CT images often falls short in diagnostic accuracy. This study introduces an advanced CT-based prediction model aimed at providing support for radiologists’ assessment to accurately diagnose LVI.</p> Methods <p>We conducted a retrospective analysis of 1409 patients with pathologically confirmed CRC from four institutions. Radiomics features were extracted from tumor areas on CT images, and Deep Residual Shrinkage Networks with Channel-wise Thresholds (DRSN-CW) algorithm was utilized to build prediction model. We assessed the model’s impact on enhancing radiologists’ diagnostic performance and employed Shapley Additive Explanation (SHAP) to interpret the influence of key features on predictions.</p> Results <p>The prediction model achieved strong prediction performance with AUCs of 0.896 (95% CI: 0.866–0.923), 0.849 (0.782–0.908), 0.845 (0.782–0.901) and 0.799 (0.709–0.881) in the training and validation cohorts. Crucially, when informed by the model, radiologists demonstrated a significant improvement in diagnosing LVI. SHAP analysis provided detailed insights into the model’s decision-making process, enhancing its clinical relevance. We also observed that patients predicted to be LVI-negative by the model had significantly longer overall survival (OS) compared to those LVI-positive (training cohort, <i>p</i> = 0.012; internal validation cohort, <i>p</i> = 0.046).</p> Conclusions <p>This study introduces a CT-based prediction model that significantly enhances radiologists’ ability to accurately diagnose LVI in CRC. By improving diagnostic accuracy and demonstrating the association between LVI predictions and OS, the model provides a valuable tool for clinical decision-making, with the potential to improve patient outcomes.</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

Improving radiologists’ diagnostic accuracy for lymphovascular invasion in colorectal cancer: insights from a multicenter CT-based study

  • 期刊论文

  • 发表日期：

    2025年4月10日

查看全文
  • Wenjun Diao,
  • Kaiqi Hou,
  • Xiaobo Chen,
  • Chaokang Han,
  • Suyun Li,
  • Zhishan Wang,
  • Ruxin Xu,
  • Jiayi Liao,
  • Liuyang Yang,
  • Ruozhen Gu,
  • Ge Zhang,
  • Zaiyi Liu,
  • Yanqi Huang

摘要

Background

The current standard of subjective assessment by radiologists for lymphovascular invasion (LVI) in colorectal cancer (CRC) using CT images often falls short in diagnostic accuracy. This study introduces an advanced CT-based prediction model aimed at providing support for radiologists’ assessment to accurately diagnose LVI.

Methods

We conducted a retrospective analysis of 1409 patients with pathologically confirmed CRC from four institutions. Radiomics features were extracted from tumor areas on CT images, and Deep Residual Shrinkage Networks with Channel-wise Thresholds (DRSN-CW) algorithm was utilized to build prediction model. We assessed the model’s impact on enhancing radiologists’ diagnostic performance and employed Shapley Additive Explanation (SHAP) to interpret the influence of key features on predictions.

Results

The prediction model achieved strong prediction performance with AUCs of 0.896 (95% CI: 0.866–0.923), 0.849 (0.782–0.908), 0.845 (0.782–0.901) and 0.799 (0.709–0.881) in the training and validation cohorts. Crucially, when informed by the model, radiologists demonstrated a significant improvement in diagnosing LVI. SHAP analysis provided detailed insights into the model’s decision-making process, enhancing its clinical relevance. We also observed that patients predicted to be LVI-negative by the model had significantly longer overall survival (OS) compared to those LVI-positive (training cohort, p = 0.012; internal validation cohort, p = 0.046).

Conclusions

This study introduces a CT-based prediction model that significantly enhances radiologists’ ability to accurately diagnose LVI in CRC. By improving diagnostic accuracy and demonstrating the association between LVI predictions and OS, the model provides a valuable tool for clinical decision-making, with the potential to improve patient outcomes.