Multimodal Deep Learning Integrating Tumor Radiomics and Mediastinal Adiposity Improves Survival Prediction in...

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Non-small cell lung cancer (NSCLC) prognosis is challenging due to tumor heterogeneity.
• Adipose tissue may influence oncological outcomes.
• Current prognostic models for NSCLC require enhancement.

Purpose of the Study:

• To develop a multimodal prognostic model for NSCLC patients.
• To integrate deep learning (DL)-derived computed tomography (CT) imaging biomarkers with mediastinal adiposity metrics.
• To predict postoperative survival in NSCLC.

Main Methods:

• Retrospective analysis of 702 surgically resected NSCLC patients.
• Extraction of tumor radiomic features using DenseNet121 convolutional neural network.
• Quantification of mediastinal fat area (MFA) via semiautomated segmentation.
• Feature-level fusion of DL tumor features and MFA for a multimodal model.
• Performance evaluation using C-index and ROC analysis.
• Risk stratification using Kaplan-Meier survival analysis.

Main Results:

• The DL-based tumor model achieved high predictive performance for disease-free survival (DFS) and overall survival (OS).
• Integration of MFA with DL features significantly enhanced predictive performance (higher C-indices for OS and DFS).
• Kaplan-Meier analysis showed significant survival differences between high- and low-risk groups.

Conclusions:

• Multimodal fusion of DL radiomics and mediastinal adiposity metrics improves postoperative survival prediction in NSCLC.
• This approach demonstrates superior prognostic capability compared to unimodal methods.
• The developed model offers a promising tool for enhanced NSCLC patient stratification.