3D multi-view convolutional neural networks for lung nodule classification.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Deep Learning for Oncology

Background:

• Accurate lung nodule classification is crucial for early lung cancer detection and patient management.
• Two-dimensional Convolutional Neural Networks (2D CNNs) have shown promise but may not fully capture complex spatial information.
• Multi-view strategies can enhance the performance of CNNs by integrating information from different perspectives.

Purpose of the Study:

• To investigate the efficacy of 3D Multi-View Convolutional Neural Networks (MV-CNN) for lung nodule classification.
• To compare different MV-CNN architectures (chain, directed acyclic graph) including 3D Inception and 3D Inception-ResNet.
• To evaluate the 'multi-view-one-network' strategy against the 'one-view-one-network' strategy.

Main Methods:

• Implementation of 3D MV-CNN with chain and directed acyclic graph architectures (3D Inception, 3D Inception-ResNet).
• Utilized the 'multi-view-one-network' strategy for all models.
• Conducted binary (benign/malignant) and ternary (benign/primary malignant/metastatic malignant) classification on LIDC-IDRI CT images using 10-fold cross-validation.

Main Results:

• 3D MV-CNN models significantly outperformed 2D MV-CNN models, demonstrating the benefit of 3D spatial context.
• A 3D Inception network achieved superior results with an error rate of 4.59% for binary and 7.70% for ternary classification.
• The 'multi-view-one-network' strategy yielded lower error rates compared to the 'one-view-one-network' strategy.

Conclusions:

• 3D MV-CNNs are highly effective for lung nodule classification, leveraging comprehensive spatial information.
• The 'multi-view-one-network' strategy enhances classification performance by integrating multiple views.
• Advanced 3D architectures like 3D Inception show significant potential for improving diagnostic accuracy in lung nodule detection.