A hybrid CNN feature model for pulmonary nodule malignancy risk differentiation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pulmonary nodule malignancy risk differentiation is a critical challenge in computer-aided diagnosis (CADx).
• Existing CADx methods often rely on local features and statistical analysis, limiting their comprehensive assessment.
• Convolutional Neural Networks (CNNs) show promise for target recognition by simulating human neural networks.

Purpose of the Study:

• To develop and evaluate a hybrid CADx model integrating global and local features for pulmonary nodule malignancy risk differentiation.
• To compare the performance of newly proposed CNN models against existing methods using a large public dataset.

Main Methods:

• Utilized the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset, the largest public database for lung imaging.
• Developed and compared three types of CNN models, including two novel multi-channel CNN architectures.
• Assessed model performance based on the differentiation capacity of extracted features and receiver operating characteristic (ROC) curve analysis.

Main Results:

• The multi-channel CNN model demonstrated superior discrimination in differentiating pulmonary nodule malignancy risk.
• The proposed CADx scheme using the multi-channel CNN model significantly outperformed a previous 3D texture feature analysis method.
• The area under the receiver operating characteristic curve (AUC) improved from 0.9441 to 0.9702 with the new model.

Conclusions:

• A hybrid model incorporating global and local features via a multi-channel CNN offers enhanced accuracy for pulmonary nodule malignancy risk assessment.
• This approach represents a significant advancement in computer-aided diagnosis for lung cancer screening.
• The developed CADx scheme shows potential for improving early and accurate lung cancer detection.