Image segmentation network based on enhanced dual encoder.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Medical Imaging

Background:

• Convolutional Neural Networks (CNNs) face limitations in capturing global context, while Transformers struggle with scalability and local feature extraction.
• Hybrid network architectures integrating CNNs and Transformers offer a promising direction to leverage the strengths of both.
• Medical image segmentation requires precise delineation of local details and understanding of global context for accurate diagnosis.

Purpose of the Study:

• To propose an enhanced dual encoder network (EDE-Net) for improved medical image segmentation.
• To integrate convolutional kernels and pyramid transformer structures for comprehensive feature extraction.
• To develop an efficient feature fusion module for balancing local and global information.

Main Methods:

• The proposed EDE-Net employs parallel convolutional kernels and pyramid transformer structures in the encoder for feature extraction.
• A novel phase-based iterative feature fusion (PIFF) module is introduced to fuse local details and global features at each downsampling stage.
• The PIFF module assigns distinct weights to local and global features to enhance foreground pixel classification and lesion edge delineation.

Main Results:

• EDE-Net demonstrated superior performance in medical image segmentation tasks on the GlaS and MoNuSeg datasets.
• The hybrid approach effectively captured both local image details and global semantic information.
• The PIFF module significantly improved the network's ability to delineate fine lesion edges.

Conclusions:

• The EDE-Net, integrating convolution and pyramid transformers, offers a powerful solution for medical image segmentation.
• The proposed PIFF module enhances feature fusion, leading to more accurate segmentation results.
• EDE-Net outperforms existing CNN-based and transformer-based methods, highlighting its potential for clinical applications.