Res-trans networks for lung nodule classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pulmonary nodules on early diagnostic images are key indicators of lung cancer.
• Accurate malignancy estimation of pulmonary nodules is crucial for lung cancer prevention and diagnosis.
• Deep learning, particularly convolutional neural networks, shows promise for pulmonary nodule classification, but diverse nodule sizes pose a challenge.

Purpose of the Study:

• To propose a novel deep learning architecture, Res-trans networks, for classifying pulmonary nodules in computed tomography (CT) scans.
• To address the challenge of classifying lung nodules with diverse sizes (3-30 mm).

Main Methods:

• Designed local and global blocks to extract features capturing long-range pixel dependencies for varied nodule sizes.
• Utilized residual blocks with convolutional operations for local feature extraction.
• Employed transformer blocks with self-attention for global feature extraction.
• Incorporated a sequence fusion block to aggregate and extract sequence feature information, enhancing classification accuracy.

Main Results:

• The Res-trans network was evaluated on the public LIDC-IDRI dataset (1,018 CT scans).
• Tenfold cross-validation demonstrated superior performance compared to leading methods, achieving an Area Under the Curve (AUC) of 0.9628 and Accuracy of 0.9292.

Conclusions:

• A novel network capturing both local and global features was developed for classifying lung nodules in chest CT scans.
• The proposed method exhibits enhanced classification performance, offering potential assistance to radiologists in accurate lung nodule analysis.