Multi-Modal Feature Fusion-Based Multi-Branch Classification Network for Pulmonary Nodule Malignancy Suspiciousness...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early detection of malignant pulmonary nodules via chest computed tomography (CT) is crucial for lung cancer treatment and reducing mortality.
• Existing diagnostic methods often overlook the value of integrating clinical radiological structured data with unstructured CT data.
• Accurate classification of pulmonary nodules is essential for timely and effective patient management.

Purpose of the Study:

• To develop and evaluate a multi-modal fusion network for improved detection and classification of pulmonary nodules.
• To incorporate both structured radiological data and unstructured 3D CT patch data for enhanced diagnostic accuracy.
• To differentiate between benign and malignant pulmonary nodules by fusing multi-modal features.

Main Methods:

• A multi-modal fusion network was constructed, utilizing structured features derived from radiological data (9 features).
• A multi-branch attention mechanism network employing 3D ECA-ResNet was designed for processing 3D CT patch unstructured data, incorporating multi-layer feature fusion for multi-scale representation.
• Multi-modal fusion combined structured and unstructured data for the final classification of nodules.

Main Results:

• The proposed network achieved high performance in classifying benign and malignant pulmonary nodules.
• Achieved the highest accuracy at 94.89%, sensitivity at 94.91%, and F1-score at 94.65%.
• Demonstrated a low false positive rate of 5.55%, indicating robust diagnostic capability.

Conclusions:

• The developed multi-modal fusion network effectively classifies pulmonary nodules for clinical diagnosis.
• This approach enhances the accuracy and reliability of lung cancer diagnosis through integrated data analysis.
• The findings suggest a promising advancement in computer-aided diagnosis for pulmonary nodule classification.