Multi-scale fusion semantic enhancement network for medical image segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Medical Image Segmentation (MIS) is crucial for clinical diagnosis and treatment.
• Transformer models excel at global context but struggle with local feature dependencies.
• Existing methods need improvement for accurate segmentation of detailed medical images.

Purpose of the Study:

• To develop an advanced network, MFSE-Net, for improved endoscopic image segmentation.
• To enhance the capture of global information and the detailing of local features in medical images.
• To address the limitations of Transformer models in handling local feature dependencies.

Main Methods:

• MFSE-Net employs a dual-encoder architecture: PVTv2 for global features and CNNs for local details.
• Incorporates Large-kernel Grouped Deformable Attention (LGDA) for noise filtering and semantic extraction.
• Utilizes Multi-Layered Cross-attention Fusion (MLCF) to merge semantic and spatial information.
• Features a Parallel Semantic Enhancement (PSE) module and Scale Aware Module (SAM) in the decoder for boundary and global information recovery.

Main Results:

• MFSE-Net demonstrates superior performance compared to state-of-the-art (SOTA) methods.
• Achieved significant improvements in segmentation accuracy on renal tumor and polyp datasets.
• Effectively captures both global context and local details for precise segmentation.

Conclusions:

• MFSE-Net offers a robust solution for medical image segmentation challenges.
• The proposed network architecture effectively enhances segmentation precision and detail.
• MFSE-Net shows significant potential for clinical applications in diagnosis and treatment planning.