E-SegNet: E-Shaped Structure Networks for Accurate 2D and 3D Medical Image Segmentation.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Deep Learning Architectures

Background:

• U-shaped architectures are standard for medical image segmentation.
• Current models increase parameters for higher accuracy, limiting real-world use.
• Need for efficient yet accurate segmentation methods.

Purpose of the Study:

• Introduce an E-shaped segmentation framework as an efficient alternative to U-shaped models.
• Develop novel modules for enhanced feature representation.
• Improve medical image segmentation accuracy and reduce computational cost.

Main Methods:

• Proposed an E-shaped framework aggregating multi-scale encoder features for deep integration.
• Introduced a multi-scale large-kernel convolution (MLKConv) for local and global context.
• Developed 2D and 3D E-SegNet models for medical image segmentation.

Main Results:

• E-shaped approach significantly reduces parameters compared to U-shaped models.
• Achieved state-of-the-art (SOTA) performance on multiple 2D and 3D medical image datasets.
• Demonstrated superior accuracy, especially in complex segmentation tasks.

Conclusions:

• The E-shaped framework provides a more efficient and effective approach to medical image segmentation.
• MLKConv module enhances feature representation capabilities.
• The proposed models offer a promising advancement for clinical applications.