Related Experiment Video
Updated: Sep 2, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Development and Validation of a Triple-classification Multimodal Model Based on Intratumoral and Peritumoral
1Northern Jangsu People's Hospital Affiliated to Yangzhou University, Jiangsu, China, 225001.
Objective:
To develop a multimodal model for predicting ground-glass nodule (GGNs) invasiveness, enabling accurate differentiation of pre-invasive lesions (PIL), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IA).
Methods:
The retrospective study included pathologically confirmed GGNs between May 2018 and May 2025. Univariate and multivariate logistic regression analysis were performed to screen for independent predictive factors. Semi-automated segmentation of intratumoral region (ITR) and two peritumoral regions (PTRs) scaling strategies were used. Radiomics features from ITR and PTRs were selected via Least Absolute Shrinkage and Selection Operator (LASSO), and random forest (RF), support vector machine (SVM), and logistic regression (LR) models were constructed. Four progressive integration models were further created: Conventional, Conventional + follow-up, Conventional + follow-up + ITR, and Conventional + follow-up + ITR + PTR.
Results:
517 GGNs were included in the training and internal test cohorts, and 91 GGNs from the external test cohort. Maximum diameter, follow-up composite parameter, vacuole, and air bronchus sign were identified as independent predictive factors. Both PTR construction strategies yielded optimal models (PTR3/ PTR6 for fixed-size, PTR50%/ PTR75% for diameter-ratio). The RF exhibited the highest stability. As feature dimensions expanded from conventional features to follow-up information, ITR, and PTR, the AUC exhibited a progressive increase. The full multimodal model with highest AUCs of 0.849 (training cohort), 0.795 (internal test cohort), and 0.838 (external test cohort).
Conclusion:
The proposed multimodal predictive triple-classification model, integrating conventional, follow-up, ITR, and PTR, may act as a non-invasive tool for preoperative evaluation of GGNs invasiveness.
Advances In Knowledge:
The study first investigated PTR model via the tumor diameter - ratio method and compare it with the fixed-size method. A follow-up composite parameter (Follow-up duration × Follow-up changes status) was used to quantified GGN dynamic changes. The novel multimodal model, integrating conventional, follow-up, ITR, and PTR radiomics features, outperformed single-modal approaches in differentiating PIL, MIA, and IA.