Scholarlink

Purpose <p>The invasive diagnostic methods for distinguishing between benign and malignant pulmonary nodules may pose various risks. The purpose of this study is to develop a multimodal deep learning model that integrates image features of multi-view nodules and mediastinal fat to non-invasively diagnose pulmonary nodules.</p> Methods <p>This study reviewed 1407 patients with pathologically confirmed pulmonary nodules from three centers. First, the 3D nodules were decomposed into 2D slices from nine different angles for feature extraction, while the image features of the mediastinal fat were extracted using a deep learning model. Finally, the nodule features and mediastinal fat features were input into a support vector machine for accurate prediction.</p> Results <p>Adding mediastinal fat features to the model improved predictive performance compared to using only nodule features. This resulted in area under the receiver operating characteristic curves of 0.926, 0.947, and 0.914 for the internal test and two external test cohorts, respectively. The combination of mediastinal fat and nodules exhibited superior predictive performance based on the net reclassification index (NRI) results in the internal test cohort (NRI = 0.173), external test cohort 1 (NRI = 0.114), and external test cohort 2 (NRI = 0.109).</p> Conclusion <p>The results showed that the multimodal deep learning model had a certain predictive capability for pulmonary nodules, and mediastinal fat could serve as an important biomarker for predicting the benignity and malignancy of pulmonary nodules.</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 Deep Learning-Based Multi-perspective Pulmonary Nodule Classification Incorporating Mediastinal Fat: A Multicenter Study

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Shidi Miao,
  • Yuyang Jiang,
  • Shikai Mu,
  • Zhenghui Xiong,
  • Haipeng Jin,
  • Ruitao Wang,
  • Zengyao Liu,
  • Qiujun Wang,
  • Xuemei Ding

摘要

Purpose

The invasive diagnostic methods for distinguishing between benign and malignant pulmonary nodules may pose various risks. The purpose of this study is to develop a multimodal deep learning model that integrates image features of multi-view nodules and mediastinal fat to non-invasively diagnose pulmonary nodules.

Methods

This study reviewed 1407 patients with pathologically confirmed pulmonary nodules from three centers. First, the 3D nodules were decomposed into 2D slices from nine different angles for feature extraction, while the image features of the mediastinal fat were extracted using a deep learning model. Finally, the nodule features and mediastinal fat features were input into a support vector machine for accurate prediction.

Results

Adding mediastinal fat features to the model improved predictive performance compared to using only nodule features. This resulted in area under the receiver operating characteristic curves of 0.926, 0.947, and 0.914 for the internal test and two external test cohorts, respectively. The combination of mediastinal fat and nodules exhibited superior predictive performance based on the net reclassification index (NRI) results in the internal test cohort (NRI = 0.173), external test cohort 1 (NRI = 0.114), and external test cohort 2 (NRI = 0.109).

Conclusion

The results showed that the multimodal deep learning model had a certain predictive capability for pulmonary nodules, and mediastinal fat could serve as an important biomarker for predicting the benignity and malignancy of pulmonary nodules.