An Anomaly Self-Supervised Representation To Classify Malignant Lung Nodules.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Deep Learning for Medical Diagnosis

Background:

• Lung nodules are key biomarkers for lung cancer diagnosis, but their analysis from CT scans is challenging due to variability.
• Current computer-aided methods struggle with imbalanced datasets, where malignant nodules are rare (<10%).

Purpose of the Study:

• To develop a robust deep self-supervised learning method for classifying lung nodules.
• To address the challenge of imbalanced data in lung nodule malignancy assessment.

Main Methods:

• Utilized a one-class learning scheme leveraging the abundance of benign nodules.
• Employed deep self-supervised representation learning to extract spatial, local, and attention features.
• Classified malignant nodules as anomalies based on reconstruction errors.

Main Results:

• Achieved an Area Under the Curve (AUC) of 93.20% on the LIDC-IDRI dataset.
• Demonstrated competitive performance against state-of-the-art methods.
• Highlighted the capability of self-supervised learning in capturing complex nodule features.

Conclusions:

• Self-supervised representations offer a promising alternative for generalizable lung nodule characterization.
• This methodology shows potential for seamless transfer into clinical practice for improved lung cancer diagnosis.
• The anomaly detection approach effectively handles the rarity of malignant cases.