Scholarlink

Background <p>Accurate preoperative identification of visceral pleural invasion (VPI) is crucial for surgical planning and lymph node dissection in patients with early-stage lung adenocarcinoma. This study aims to develop and independently validate a predictive model for VPI in clinical stage T1 invasive lung adenocarcinoma, based on intratumoral and peritumoral radiomic features. Additionally, it initiates a preliminary exploration into the interpretability of the optimal model.</p> Methods <p>This retrospective analysis gathered imaging and clinical data from 316 patients diagnosed with potentially pleural-invading invasive lung adenocarcinoma based on computed tomography (CT) images across three medical institutions. A total of 109 patients from Center 1 and 99 patients from Center 2 were randomly assigned to a training set (146 patients) and a validation set (62 patients), in a 7:3 ratio, respectively. And 108 patients from Center 3 constituted an independent external test set. Feature selection was executed using univariate and multivariate analyses, Spearman correlation, minimal redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO) techniques, to develop of clinical CT semantics, intratumoral radiomics, peritumoral radiomics, and combined intratumoral-peritumoral radiomics models. The influential features of the optimal model were subsequently ranked and evaluated using the SHapley additive explanation (SHAP).</p> Results <p>The intratumoral model, exhibiting the lowest Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), indicated a superior model fitting in the training set. It also showed the highest Matthews Correlation Coefficient (MCC) and Diagnostic Odds Ratio (DOR) in both validation and external test sets, demonstrating the best predictive capacity. The Area Under the Curve (AUC) values for the intratumoral model across training, validation, and external test sets were 0.837 (95% CI: 0.812–0.860), 0.825 (95% CI: 0.794–0.858), and 0.819 (95% CI: 0.791–0.845) respectively; these values surpassed those of the peritumoral, combined models, and clinical CT semantic models. SHAP analysis results showed Small-Area-Low-Gray-Level-Emphasis emerges as the pivotal feature.</p> Conclusion <p>Intratumoral radiomic features possess significant predictive value for VPI in clinical stage T1 invasive lung adenocarcinoma, aiding physicians in formulating precise clinical strategies.</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

Preoperative prediction of visceral pleural invasion in clinical stage T1 invasive lung adenocarcinoma using intratumoral and peritumoral radiomics: a multicenter development, validation, and comparison study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月28日

查看全文
  • Hao Dong,
  • Zebin Yang,
  • Yonggang Qiu,
  • Fangfang Qiu,
  • Xinbin Wang,
  • Zihao Xie,
  • Junjie Yang,
  • Leyi Shou,
  • Xiaojun Guan,
  • Xiaodan Ye,
  • Xiaojun Xu

摘要

Background

Accurate preoperative identification of visceral pleural invasion (VPI) is crucial for surgical planning and lymph node dissection in patients with early-stage lung adenocarcinoma. This study aims to develop and independently validate a predictive model for VPI in clinical stage T1 invasive lung adenocarcinoma, based on intratumoral and peritumoral radiomic features. Additionally, it initiates a preliminary exploration into the interpretability of the optimal model.

Methods

This retrospective analysis gathered imaging and clinical data from 316 patients diagnosed with potentially pleural-invading invasive lung adenocarcinoma based on computed tomography (CT) images across three medical institutions. A total of 109 patients from Center 1 and 99 patients from Center 2 were randomly assigned to a training set (146 patients) and a validation set (62 patients), in a 7:3 ratio, respectively. And 108 patients from Center 3 constituted an independent external test set. Feature selection was executed using univariate and multivariate analyses, Spearman correlation, minimal redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO) techniques, to develop of clinical CT semantics, intratumoral radiomics, peritumoral radiomics, and combined intratumoral-peritumoral radiomics models. The influential features of the optimal model were subsequently ranked and evaluated using the SHapley additive explanation (SHAP).

Results

The intratumoral model, exhibiting the lowest Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), indicated a superior model fitting in the training set. It also showed the highest Matthews Correlation Coefficient (MCC) and Diagnostic Odds Ratio (DOR) in both validation and external test sets, demonstrating the best predictive capacity. The Area Under the Curve (AUC) values for the intratumoral model across training, validation, and external test sets were 0.837 (95% CI: 0.812–0.860), 0.825 (95% CI: 0.794–0.858), and 0.819 (95% CI: 0.791–0.845) respectively; these values surpassed those of the peritumoral, combined models, and clinical CT semantic models. SHAP analysis results showed Small-Area-Low-Gray-Level-Emphasis emerges as the pivotal feature.

Conclusion

Intratumoral radiomic features possess significant predictive value for VPI in clinical stage T1 invasive lung adenocarcinoma, aiding physicians in formulating precise clinical strategies.