Scholarlink

Background <p>Leptomeningeal metastasis (LM) in non-small cell lung cancer (NSCLC) is a terminal complication with limited prognostic tools, especially after failure of tyrosine kinase inhibitor treatment. This study aimed to develop and validate a multi-modal model to predict overall survival (OS) in LM patients. The model utilizes deep learning (DL) and radiomic features extracted from T1-weighted contrast-enhanced MRI, combined with clinical variables.</p> Methods <p>We retrospectively enrolled 174 NSCLC patients with LM confirmed via cerebrospinal fluid cytology. Regions of interest were manually segmented on MRI, and features were extracted following IBSI guidelines. Robust radiomic features were selected using reproducibility metrics and LASSO-Cox regression. Deep learning signatures were generated using pre-trained convolutional networks with transfer learning and Cox proportional hazards loss. An integrated prognostic model was constructed by combining significant clinical, radiomic, and DL-based predictors. Performance was assessed using the C-index, time-dependent ROC, calibration curves, and decision curve analysis.</p> Results <p>The integrated model demonstrated superior predictive performance, achieving a C-index of 0.709 (95% CI: 0.628–0.790) in the training cohort and 0.629 (0.499–0.759) in validation, outperforming clinical, radiomic, and DL-only models. Time-dependent AUC reached 0.807 at 18&#xa0;months. Calibration and decision curves confirmed clinical utility and accuracy. Intrathecal chemotherapy was identified as a significant prognostic factor (HR = 0.654, <i>p</i> &lt; 0.05).</p> Conclusion <p>The multi-modal model integrating DL, radiomics, and clinical data shows potential for improving OS prediction in NSCLC patients with LM, providing a preliminary reference for personalized prognostic assessment and therapeutic strategy development.</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

A survival prediction model for leptomeningeal metastasis patients with non-small cell lung cancer based on deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月28日

查看全文
  • Kai Jin,
  • Yuanyuan Qu,
  • Sheng Li,
  • Caixing Sun,
  • Lin Wang

摘要

Background

Leptomeningeal metastasis (LM) in non-small cell lung cancer (NSCLC) is a terminal complication with limited prognostic tools, especially after failure of tyrosine kinase inhibitor treatment. This study aimed to develop and validate a multi-modal model to predict overall survival (OS) in LM patients. The model utilizes deep learning (DL) and radiomic features extracted from T1-weighted contrast-enhanced MRI, combined with clinical variables.

Methods

We retrospectively enrolled 174 NSCLC patients with LM confirmed via cerebrospinal fluid cytology. Regions of interest were manually segmented on MRI, and features were extracted following IBSI guidelines. Robust radiomic features were selected using reproducibility metrics and LASSO-Cox regression. Deep learning signatures were generated using pre-trained convolutional networks with transfer learning and Cox proportional hazards loss. An integrated prognostic model was constructed by combining significant clinical, radiomic, and DL-based predictors. Performance was assessed using the C-index, time-dependent ROC, calibration curves, and decision curve analysis.

Results

The integrated model demonstrated superior predictive performance, achieving a C-index of 0.709 (95% CI: 0.628–0.790) in the training cohort and 0.629 (0.499–0.759) in validation, outperforming clinical, radiomic, and DL-only models. Time-dependent AUC reached 0.807 at 18 months. Calibration and decision curves confirmed clinical utility and accuracy. Intrathecal chemotherapy was identified as a significant prognostic factor (HR = 0.654, p < 0.05).

Conclusion

The multi-modal model integrating DL, radiomics, and clinical data shows potential for improving OS prediction in NSCLC patients with LM, providing a preliminary reference for personalized prognostic assessment and therapeutic strategy development.