A multi-attention deep architecture to stratify lung nodule malignancy from CT scans

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of cancer mortality, with nodules detected via low-dose CT scans being key indicators.
• Current nodule characterization for malignancy is subjective, leading to moderate expert agreement and challenges in stratification.
• Accurate malignancy stratification of lung nodules is crucial for effective patient management and treatment planning.

Purpose of the Study:

• To develop and validate a deep multi-attention strategy for classifying lung nodules into four malignancy degrees.
• To enhance the objectivity and accuracy of lung nodule characterization using artificial intelligence.
• To provide a tool that supports clinical decision-making in lung cancer diagnosis.

Main Methods:

• A deep multi-attention architecture was designed to process volumetric nodule regions.
• The model learns multi-scale saliency maps, focusing on malignancy patterns like lobulated, textural, and spiculated features.
• Extensive validation was performed, analyzing attention features and correlating them with radiological findings.

Main Results:

• The proposed multi-attention approach achieved an Area Under the Curve (AUC) of 85.35% in multi-classification.
• A mean AUC of 82.90% was obtained using a one-vs-all validation methodology.
• The results demonstrate competitive performance against current state-of-the-art methods.

Conclusions:

• The developed deep multi-attention architecture effectively stratifies lung nodules by malignancy stage.
• The approach shows strong generalization performance, indicating potential for real-world clinical application.
• This AI strategy can support nodule stratification and feature classification, improving diagnostic accuracy.