Objectives This study aimed to explore a multiple-instance learning (MIL) framework incorporating radiomics features and deep learning representations to predict the invasiveness of ground-glass nodules (GGNs) in lung adenocarcinoma (LUAD) using preoperative CT. Materials and methods We retrospectively analyzed 1247 GGNs from 1182 LUAD patients across six hospitals, and divided them into training, validation and three test sets. According to postoperative pathological findings, the data were further classified into invasive and non-invasive subgroups. Five kinds of predictive models were developed: radiomics models, 3D deep learning models, 2.5D deep learning models, deep learning-based MIL (MIL-DL) models, and deep learning and radiomics-based MIL (MIL-DL-Rad) models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis (DCA). Results The MIL-DL-Rad model with the ExtraTrees classifier exhibited superior and consistent performance across all sets, achieving AUCs of 0.936, 0.881, 0.868, 0.926, and 0.918 in training, validation and external test sets. In contrast, the AUC performance of MIL-DL and radiomics models was relatively unstable. The calibration curve and DCA indicated that the integrated model achieved favorable predictive efficiency and clinical predictive benefits. Conclusions The MIL-DL-Rad model showed better overall performance for invasiveness prediction of GGNs in LUAD patients, providing a novel perspective on feature fusion that can contribute to more accurate preoperative predictions in clinical practice. Critical relevance statement Multi-instance learning integrating deep learning and radiomics enhances the prediction of ground-glass nodule (GGN) invasiveness and is expected to provide optimal preoperative clinical decision-making for lung adenocarcinoma patients. Key Points <UnorderedList Mark="Bullet"> <ItemContent> Ground-glass nodules invasiveness directly influences surgical strategies and prognosis. </ItemContent> <ItemContent> Multiple-instance learning framework integrates radiomics and deep learning features. </ItemContent> <ItemContent> Integrated model achieves superior accuracy and consistency in invasiveness prediction. </ItemContent> </UnorderedList> Graphical Abstract

Integrating CT-based radiomics and deep learning for invasive prediction of ground-glass nodules in lung adenocarcinoma: a multicohort study

期刊论文
出版模式：

开放获取
发表日期：

2025年12月8日

查看全文

Hai Du,
Jing Shen,
Feng Chen,
Kaifeng Wang,
Lili Qin,
Yijiang Hu,
Yue Xiao,
Xiulin Wang,
Jianlin Wu

摘要

Objectives

This study aimed to explore a multiple-instance learning (MIL) framework incorporating radiomics features and deep learning representations to predict the invasiveness of ground-glass nodules (GGNs) in lung adenocarcinoma (LUAD) using preoperative CT.

Materials and methods

We retrospectively analyzed 1247 GGNs from 1182 LUAD patients across six hospitals, and divided them into training, validation and three test sets. According to postoperative pathological findings, the data were further classified into invasive and non-invasive subgroups. Five kinds of predictive models were developed: radiomics models, 3D deep learning models, 2.5D deep learning models, deep learning-based MIL (MIL-DL) models, and deep learning and radiomics-based MIL (MIL-DL-Rad) models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis (DCA).

Results

The MIL-DL-Rad model with the ExtraTrees classifier exhibited superior and consistent performance across all sets, achieving AUCs of 0.936, 0.881, 0.868, 0.926, and 0.918 in training, validation and external test sets. In contrast, the AUC performance of MIL-DL and radiomics models was relatively unstable. The calibration curve and DCA indicated that the integrated model achieved favorable predictive efficiency and clinical predictive benefits.

Conclusions

The MIL-DL-Rad model showed better overall performance for invasiveness prediction of GGNs in LUAD patients, providing a novel perspective on feature fusion that can contribute to more accurate preoperative predictions in clinical practice.

Critical relevance statement

Multi-instance learning integrating deep learning and radiomics enhances the prediction of ground-glass nodule (GGN) invasiveness and is expected to provide optimal preoperative clinical decision-making for lung adenocarcinoma patients.

Key Points

Ground-glass nodules invasiveness directly influences surgical strategies and prognosis.

Multiple-instance learning framework integrates radiomics and deep learning features.

Integrated model achieves superior accuracy and consistency in invasiveness prediction.

Graphical Abstract