Multi-Class Segmentation in 2D Chest CT using Deep Learning

Abstract This study introduces a deep learning-based method for multi-class segmentation of lung and airway structures from 2D chest CT scans, tackling ongoing challenges in radiologic interpretation and automated analysis. Accurate outlining of pulmonary anatomy, especially the detailed airway tree, is crucial for correct diagnosis, treatment planning, and monitoring respiratory diseases. However, traditional and hybrid segmentation methods often struggle to clearly resolve blurred lumen boundaries and the complex branching of airways, especially in diseased conditions. To overcome these issues, we propose an Attention U-Net architecture that incorporates attention mechanisms within the standard encoder–decoder structure. This method improves spatial sensitivity, allowing the network to identify clinically important airway structures effectively. Our model was created and tested using a publicly available thoracic CT dataset from The Cancer Imaging Archive (TCIA), ensuring reproducibility and clinical relevance. Compared to a standard U-Net, our attention-based approach achieved more accurate and balanced segmentation of both lung and airway structures. While U-Net performed well for lung segmentation, the proposed Attention U-Net increased the airway segmentation Dice score from 0.0406 to 0.3891, demonstrating its ability to capture small, complex structures better. These findings show the potential of attention-enhanced deep learning techniques to significantly improve pulmonary segmentation and support their use in clinical radiology workflows and AI-driven diagnostics.