A multimodal radiomics model to predict disease-free survival in resected non-small cell lung cancer: integrating clinicopathology, dual-energy CT, and deep learning features

Read full text Summarise with VIRAT
Authors: Weiming Zhao, Tingting Li, Xilong Zhou, Wenjing Fan, Yaqi Wang, Huanhuan Zhou, Shuhua Li, Juan Wang, Yihong Gu, Zongyu Xie, Fang Su
Cited by: 0
References: 37

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Background: Predicting disease-free survival (DFS) after surgery for non-small cell lung cancer (NSCLC) is crucial for personalized treatment decisions. However, traditional clinical models have limited predictive efficacy. Moreover, the combined value of dual-energy computed tomography (DECT) quantitative parameters, radiomic features, and deep learning (DL) in assessing NSCLC prognosis remains unclear. This study aimed to develop and validate a multimodal machine learning model using feature-level fusion to improve the accuracy of predicting postoperative disease recurrence in NSCLC. Methods: We retrospectively included 171 NSCLC patients who underwent surgical resection and were pathologically confirmed between January 2019 and October 2023. All patients underwent dual-phase DECT scanning preoperatively. We extracted four types of features from multimodal (I) DECT quantitative parameters; (II) clinicopathological features; (III) radiomic features based on PyRadiomics; and (IV) deep features extracted from a pre-trained three-dimensional (3D) DenseNet121 model. We applied a four-step cascading feature selection pipeline to identify the most predictive feature subset. Subsequently, single-modal predictive models were built using seven machine learning algorithms-logistic regression (LR), K-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), extreme random trees (ExtraTrees), naive Bayes, and extreme gradient boosting (XGBoost)-with hyperparameters optimized via five-fold cross-validation. The fusion model employed a feature-level pre-fusion strategy, synchronously training with the same seven algorithms to maintain comparability. The optimal model for each modality was selected based on the area under the receiver operating characteristic (ROC) curve (AUC). Finally, a comprehensive performance comparison was conducted between the selected optimal fusion model and each single-modal champion model and model interpretation. Results: The Fusion-ExtraTrees model constructed based on 12 optimal features showed great performance on the validation set, with an AUC of 0.889, significantly outperforming the single-modal champion models Clinic-LR, Rad-LR, and DECT-RF (all DeLong test P

Fields of Study & Subjects

Radiomics and Machine Learning in Medical Imaging Lung Cancer Diagnosis and Treatment AI in cancer detection

Publication details

Journal / source
Journal of Thoracic Disease
Volume
18
Issue
3
Pages
227–227
Publication date
1 March 2026
Language
EN