An Efficient Swin Transformer-Based Contextual Feature-Guided Enhancment for Lung Segmentation

Abstract

Accurate lung segmentation from CT scans is essential for diagnosing and monitoring respiratory diseases. This proposal introduces a novel approach that combining the Swin Transformer in combination with Contextual Feature-Guided Attention Refinement for enhanced lung segmentation. This framework uses the Swin Transformer to capture local and global context, with attention dynamically refined by context-specific cues like anatomical landmarks and saliency maps. These cues steer attention toward key lung structures, such as nodules or airways, while suppressing focus on irrelevant regions like background noise. The enhanced representation is then fed into an improved U-Net++ architecture, which performs segmentation operations with its nested skip connections and adaptive attention mechanism, allowing for more accurate and robust segmentation, even in the presence of image quality variations and complex lung morphologies. The proposed method excelled in segmenting the Finding and Measuring Lungs in CT Data dataset achieving high scores across multiple metrics (Dice: 0.978, accuracy: 0.998) and demonstrating strong image quality. These results indicate strong potential for clinical and diagnostic applications, offering more precise and automated identification of lung abnormalities and improving the efficiency of medical image analysis.