A Timeseries-based Multimodal Deep Learning Approach for Lung Nodule Growth Prediction.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiology and Oncology

Background:

• Accurate monitoring of lung nodule growth is crucial for patient outcomes and clinical decisions.
• Lung nodules, though often benign, require careful surveillance to detect potentially malignant changes.

Purpose of the Study:

• To develop and validate a Multimodal Deep Learning Approach for enhanced lung nodule growth prediction.
• To integrate time-series CT image data with patient demographics and nodule-specific features for improved predictive accuracy.

Main Methods:

• A Multimodal Deep Learning framework was developed using CT image sequences, demographics, and nodule features from Far Eastern Memorial Hospital.
• Model performance was evaluated using Accuracy, Precision, Sensitivity, F1-score, and Area Under the Curve (AUC).
• The repeat frame strategy within the framework demonstrated optimal performance.

Main Results:

• The Multimodal Deep Learning framework significantly outperformed traditional machine learning and unimodal models.
• The repeat frame strategy achieved high performance metrics: 0.929 accuracy, 0.878 precision, 0.908 sensitivity, 0.878 F1-score, and 0.977 AUC.
• Statistical analysis (paired t-test) confirmed significant improvements (p < 0.05) over baseline models.

Conclusions:

• The developed Multimodal Deep Learning model effectively integrates diverse data types for superior lung nodule growth prediction.
• This approach offers a reliable tool for clinical decision support in lung nodule management, potentially improving patient care.
• Deep learning techniques show transformative potential in advancing critical healthcare applications like lung nodule surveillance.