3D multi-view squeeze-and-excitation convolutional neural network for lung nodule classification.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology Diagnostics

Background:

• Early lung cancer screening is vital for patient survival and recovery.
• Computer-aided diagnosis (CAD) systems assist clinicians in early lung cancer detection.
• Existing 2D multi-view methods struggle with lung nodule spatial heterogeneity and view variability.

Purpose of the Study:

• To propose a novel three-dimensional multi-view convolutional neural network (3D MVCNN) framework.
• To address spatial heterogeneity and view variability in lung nodule characterization.
• To enhance the accuracy of computer-aided diagnosis for lung nodules using a squeeze-and-excitation (SE) module.

Main Methods:

• Extraction of 3D multi-view lung nodule samples using spatial sampling.
• Development of a 3D CNN to extract abstract features.
• Construction of a 3D MVSECNN model by integrating a SE module into the 3D MVCNN framework for feature fusion.

Main Results:

• Achieved 96.04% accuracy and 98.59% sensitivity in binary classification on the LIDC-IDRI dataset.
• Attained 87.76% accuracy in ternary classification, outperforming state-of-the-art methods.
• Demonstrated a significantly higher consistency score (0.948) between model predictions and pathological diagnosis compared to clinician annotations.

Conclusions:

• The proposed 3D MVSECNN method effectively captures lung nodule spatial heterogeneity and addresses multi-view variability.
• The model provides more accurate benign-malignant lung nodule classification for auxiliary diagnosis.
• This advancement holds significant importance for assisting clinicians in accurate clinical diagnosis and improving patient outcomes.