BiFUSE-UNet: a novel dual-encoder enhanced TransUNet for polyp segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Polyps are abnormal mucosal tissue growths with potential for malignant transformation.
• Accurate segmentation of polyps in medical images is challenging due to morphological heterogeneity, blurry boundaries, and data limitations.
• Early detection and intervention for polyps can prevent malignant progression.

Purpose of the Study:

• To develop an advanced deep learning model for accurate polyp segmentation in medical images.
• To address the challenges of polyp segmentation, including appearance variability and boundary ambiguity.
• To improve the detection of subtle and early-stage polyps with atypical morphology.

Main Methods:

• Introduction of BiFuse-UNet, a novel dual-encoder TransUNet architecture with a bidirectional fusion mechanism.
• Integration of a multi-axis attention module (MaxViT block) for multi-scale feature extraction and long-range dependency capture.
• Incorporation of a Merge fusion module with Coordinate Attention and SaE block for enhanced local feature perception.
• Utilization of a Biformer module to balance global semantics and local details for improved segmentation.

Main Results:

• BiFuse-UNet achieved state-of-the-art segmentation performance on the Kvasir-SEG dataset.
• The model outperformed seven existing methods in Dice coefficient, Recall, Precision, and IoU metrics.
• BiFuse-UNet effectively handled morphological heterogeneity, blurry boundaries, and background interference.

Conclusions:

• BiFuse-UNet demonstrates superior accuracy in pixel-level polyp segmentation.
• The proposed architecture offers a robust solution for challenging polyp segmentation tasks.
• This advancement contributes to more reliable early detection and diagnosis of polyps.