An improved U-NET3+ with transformer and adaptive attention map for lung segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate lung segmentation in CT scans is crucial for diagnosing and monitoring respiratory diseases.
• Existing methods may struggle with variations in lung morphology, image artifacts, and low-contrast regions.

Purpose of the Study:

• To introduce a novel hybrid architecture, Adaptive Attention U-Net (AA-U-Net), for high-precision lung segmentation.
• To enhance segmentation accuracy and robustness by combining U-Net3+ and Transformer-based attention mechanisms.

Main Methods:

• Developed a hybrid model integrating U-Net3+ for multi-scale feature extraction and a Transformer with an adaptive attention mechanism.
• The adaptive attention mechanism dynamically focuses on critical image regions based on contextual relationships.
• Evaluated the model on benchmark CT datasets.

Main Results:

• The proposed AA-U-Net achieved superior performance, surpassing state-of-the-art methods.
• Key performance metrics include IoU of 0.984, Dice coefficient of 0.989, MIoU of 0.972, and HD95 of 1.22 mm.
• Demonstrated enhanced accuracy, sensitivity, and generalization capability.

Conclusions:

• The AA-U-Net is a highly effective tool for clinical lung segmentation.
• Its hybrid architecture offers improved robustness and precision for medical image analysis.
• The adaptive attention mechanism addresses challenges in segmenting complex lung structures.