3D CNN with Visual Insights for Early Detection of Lung Cancer Using Gradient-Weighted Class Activation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• 3D Convolutional Neural Networks (CNNs) leverage 3D context for lung nodule detection in Digital Imaging and Communications in Medicine (DICOM) images.
• Gradient Class Activation aids in interpreting fine-grained features and model behavior, crucial for clinical trust.

Purpose of the Study:

• To explore lung nodule classification using an improvised 3D AlexNet with a lightweight architecture.
• To enhance model interpretability for clinicians and radiologists through Gradient-weighted Class Activation.

Main Methods:

• Implemented a lightweight 3D AlexNet incorporating a multiview network strategy.
• Performed binary classification (benign vs. malignant) on computed tomography (CT) images from LUNA 16 and DIaCOM databases.
• Utilized 10-fold cross-validation for robust performance evaluation.

Main Results:

• Achieved a superior classification accuracy of 97.17% on the LUNA 16 dataset.
• Demonstrated effectiveness compared to existing classification algorithms and on low-dose CT images.
• Gradient Class Activation provided visual explanations for model decisions.

Conclusions:

• The proposed lightweight 3D AlexNet architecture is highly effective for lung nodule classification.
• The model's interpretability through Gradient Class Activation enhances clinical adoption and trust.
• This approach offers a promising advancement in AI-assisted radiological diagnosis.