Deep learning-based lung cancer classification of CT images

Area of Science:

• Medical imaging
• Artificial intelligence in oncology
• Computer-aided diagnosis

Background:

• Lung cancer is a major global health concern, with early diagnosis crucial for survival.
• Current radiological methods for lung nodule classification face challenges with accuracy and high false-positive rates.
• Advanced diagnostic tools are needed to improve the classification of benign and malignant lung nodules in CT images.

Purpose of the Study:

• To introduce DCSwinB, a novel deep learning model for enhanced lung nodule classification in CT images.
• To improve the accuracy and efficiency of distinguishing between benign and malignant lung nodules.
• To leverage advanced deep learning architectures for better feature extraction and diagnostic performance.

Main Methods:

• Developed DCSwinB, a dual-branch classifier based on Swin-Tiny Vision Transformer (ViT).
• Integrated CNNs for local feature extraction and Swin Transformer for global feature extraction.
• Employed a Conv-MLP module to capture long-range dependencies in 3D CT images.
• Pretrained the model on LUNA16 and LUNA16-K datasets and evaluated using ten-fold cross-validation.

Main Results:

• DCSwinB achieved high performance metrics: 90.96% accuracy, 90.56% recall, 89.65% specificity, and 0.94 AUC.
• The model outperformed established models like ResNet50 and Swin-T in lung nodule classification.
• Demonstrated enhanced feature representation and optimized computational efficiency.

Conclusions:

• DCSwinB shows significant potential in improving the accuracy and reliability of lung nodule classification.
• The model can assist radiologists in reducing diagnostic errors, leading to earlier intervention.
• This deep learning approach offers a promising advancement for improved lung cancer diagnosis and patient outcomes.