Pretraining on Chronic Lung Inflammatory Disease Datasets to Enhance Indeterminant Lung Cancer Classification using...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Lung cancer is a leading cause of death, with low-dose CT screening facing adoption barriers.
• Accurate diagnosis of indeterminate pulmonary nodules is challenging, impacting clinical decisions.
• Chronic lung diseases like COPD share inflammatory features relevant to nodule classification.

Purpose of the Study:

• To improve pulmonary nodule classification by pretraining masked autoencoders (MAE) on the COPDGene dataset.
• To leverage chronic inflammatory lung disease features for enhanced nodule diagnosis.
• To optimize MAE performance through various masking strategies for improved image biomarker extraction.

Main Methods:

• Utilized masked autoencoders (MAE) pretraining on the COPDGene dataset, focusing on chronic obstructive pulmonary disease (COPD) features.
• Explored different masking strategies to enhance network attention on parenchymal and perinodular regions.
• Evaluated classification performance against self-pretraining and supervised learning on the National Lung Cancer Screening Trial (NLST) dataset.

Main Results:

• Pretraining with random masking (r-masking) on COPDGene achieved superior nodule classification performance.
• Achieved a sensitivity of 88.79%, specificity of 86.27%, and an AUC of 0.931.
• Demonstrated improved performance compared to self-pretraining and supervised learning on NLST.

Conclusions:

• Pretraining MAE on chronic disease datasets like COPDGene enhances pulmonary nodule classification.
• MAE-based approaches show significant potential for improving nodule diagnosis in clinical settings.
• Leveraging self-supervised learning on relevant datasets is crucial for advancing lung cancer screening.