DPGNet: A Boundary-Aware Medical Image Segmentation Framework Via Uncertainty Perception.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Existing methods struggle with precise boundary delineation of complex anatomical structures.
• Deep learning models often require extensive annotated data and can be computationally intensive.

Purpose of the Study:

• To introduce DPGNet, an adaptive deep learning model for precise medical image segmentation.
• To emulate expert perception of intricate anatomical edges using a novel approach.
• To improve the balance between performance and computational efficiency in segmentation models.

Main Methods:

• A three-stage progressive refinement strategy: global context, hierarchical feature enhancement, and local boundary delineation.
• A novel Edge Difference Attention (EDA) module to quantify boundary uncertainties without explicit supervision.
• A lightweight, transformer-based architecture for computational efficiency.

Main Results:

• DPGNet demonstrated consistent superiority over state-of-the-art methods on diverse medical image datasets.
• Achieved high accuracy in boundary refinement, validated by Boundary-IoU and HD95 metrics.
• Significantly lower computational overhead with 25.51 M parameters compared to existing models.

Conclusions:

• DPGNet offers a highly accurate and computationally efficient solution for medical image segmentation.
• The model provides explicit uncertainty boundary maps, aiding clinicians in identifying ambiguous regions.
• DPGNet enhances diagnostic precision and facilitates more accurate clinical segmentation outcomes.