Deep Multi-path Network Integrating Incomplete Biomarker and Chest CT Data for Evaluating Lung Cancer Risk

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Radiomics

Background:

• Clinical data elements (CDEs), blood markers, and chest computed tomography (CT) imaging are used for lung cancer risk assessment.
• These data modalities offer complementary information but are often incomplete in real-world clinical settings.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework, the multi-path multi-modal missing network (M3Net), for integrating multi-modal data for lung cancer risk prediction.
• To address the challenge of missing data across different modalities (CDEs, biomarkers, CT images).

Main Methods:

• Proposed M3Net, a multi-path neural network designed to integrate CDEs, biomarkers, and CT image features.
• The network fuses features from individual modalities in a second stage for integrated prediction, allowing end-to-end training and prediction with single or multiple modalities.
• Evaluated on datasets from the Consortium for Molecular and Cellular Characterization of Screen-Detected Lesions (MCL) project, including cross-validation and external validation.

Main Results:

• Combining multiple data modalities significantly improved prediction performance compared to single modalities in both cross-validation and external validation.
• The M3Net framework demonstrated the ability to effectively integrate subjects with missing data, contributing to the model's discriminatory power (p < 0.05).

Conclusions:

• The M3Net framework offers an effective approach for integrating diverse data types (imaging and non-imaging) for lung cancer risk prediction.
• Accounting for missing data across modalities enhances the predictive performance of lung cancer risk assessment models.