Data augmented lung cancer prediction framework using the nested case control NLST cohort.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Supervised learning for lung cancer screening is hindered by limited labeled medical images.
• Data augmentation is a promising technique to address data scarcity in medical AI.
• The application of data augmentation in lung cancer screening requires further investigation.

Purpose of the Study:

• To analyze state-of-the-art data augmentation techniques for lung cancer binary prediction.
• To evaluate the efficiency of various data augmentation methods in the context of lung cancer screening.
• To identify optimal data augmentation strategies for deep learning models used in lung cancer detection.

Main Methods:

• Utilized the National Lung Screening Trial (NLST) cohort (253 participants) with non-contrast CT scans.
• Processed CT scans into 3D volumes and applied five basic (online) and two generative (offline) data augmentation methods.
• Evaluated ten state-of-the-art (SOTA) 3D deep learning models for lung cancer prediction.

Main Results:

• Performance improvement varied significantly based on the augmentation approach.
• Cutmix yielded the highest average performance gains: 1.07% accuracy, 3.29% F1 score, and 1.19% AUC.
• MobileNetV2 with simple augmentation achieved the best AUC (0.8719), a 7.62% improvement; MED-DDPM rebalanced data and improved predictions.

Conclusions:

• Data augmentation effectiveness is highly model-dependent, emphasizing the need for careful method selection.
• Traditional augmentation methods can outperform SOTA online approaches in stability and performance.
• Findings provide insights for developing and integrating data-augmented deep learning tools for lung cancer screening.