DAN-PD: Domain adaptive network with parallel decoder for polyp segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopic polyp segmentation is crucial for colorectal cancer prevention.
• Deep learning methods face challenges due to scarce labeled data and domain variability in endoscopic images.
• Efficient utilization of existing labeled data is critical for advancing polyp segmentation.

Purpose of the Study:

• To propose a novel method, DAN-PD, based on the Vision Transformer for automatic semantic segmentation of polyps.
• To address the challenges of data scarcity and domain variation in polyp segmentation.
• To improve the generalization and performance of polyp segmentation models.

Main Methods:

• Developed DAN-PD, a Vision Transformer-based model incorporating a Teacher Parallel Encoder (TPE) and Domain-Aware Parallel Decoder (DAPD).
• Integrated Unsupervised Domain Adaptation (UDA) methods and adversarial learning strategies into the polyp segmentation task.
• Conducted four transfer learning experiments using three public polyp image datasets.

Main Results:

• The proposed DAN-PD method outperformed existing methods in all conducted experiments.
• The model achieved state-of-the-art performance in polyp segmentation across different datasets.
• Demonstrated the effectiveness of UDA and adversarial learning for polyp segmentation.

Conclusions:

• DAN-PD offers an effective solution for polyp segmentation, addressing key limitations of current deep learning approaches.
• The method shows significant potential for improving the accuracy and reliability of automated polyp detection in clinical settings.
• Highlights the importance of domain adaptation and adversarial learning for robust medical image analysis.