ERSegDiff: a diffusion-based model for edge reshaping in medical image segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Medical image segmentation is vital for precise clinical analysis.
• Current CNN and attention-based models struggle with accurate segmentation at region edges.
• Refining initial segmentation without altering data or architecture is key.

Purpose of the Study:

• To propose ERSegDiff, a diffusion model for reshaping segmentation borders.
• To enhance the accuracy of pathological area segmentation.
• To improve edge segmentation in medical images.

Main Methods:

• Utilized diffusion models to fit the distribution of target edge areas.
• Trained the diffusion model to modify initial segmentation edges.
• Incorporated prior knowledge into the diffusion model for accurate edge probability simulation.
• Introduced an edge concern module with attention mechanisms for feature weighting.

Main Results:

• ERSegDiff improved Dice scores by 3%-4% on COVID-19 lung segmentation.
• ERSegDiff improved Dice scores by 2%-4% on ISIC-2018 skin cancer segmentation.
• Achieved state-of-the-art results compared to mainstream networks like swinUNETR.

Conclusions:

• Diffusion models can significantly enhance medical image segmentation accuracy.
• ERSegDiff effectively refines segmentation borders, leading to more precise pathological area identification.
• The proposed method offers a robust solution for improving segmentation edge accuracy.