Self-Supervised Transfer Learning Based on Domain Adaptation for Benign-Malignant Lung Nodule Classification on...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Spatial heterogeneity in lung nodules is crucial for malignancy assessment in lung cancer diagnosis.
• 3D CT volumes offer richer discriminative information than 2D images for nodule analysis.
• Deep learning models struggle with limited labeled data for capturing 3D nodule features.

Purpose of the Study:

• To develop a self-supervised transfer learning framework for benign-malignant lung nodule classification.
• To improve the capture of 3D discriminative features from limited labeled lung nodule CT data.
• To address the challenges of deep learning in analyzing 3D nodule volumes.

Main Methods:

• A self-supervised transfer learning based on domain adaptation (SSTL-DA) 3D CNN framework was developed.
• Adaptive slice selection (ASS) was employed for pre-processing to reduce noise in CT samples.
• A self-supervised network learned robust image representations, followed by domain adaptation for classification.

Main Results:

• The SSTL-DA method achieved 91.07% accuracy on the LIDC-IDRI dataset.
• An Area Under the Curve (AUC) of 95.84% was obtained, indicating strong classification performance.
• The model demonstrated competitive results compared to existing state-of-the-art approaches.

Conclusions:

• The proposed SSTL-DA framework effectively classifies benign and malignant lung nodules.
• This approach offers a robust solution for analyzing 3D lung nodule CT data with limited samples.
• SSTL-DA shows significant potential for improving lung cancer diagnosis through advanced AI techniques.