EDSRNet: An Enhanced Decoder Semantic Recovery Network for 2D Medical Image Segmentation

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning, with deep learning methods widely adopted.
• Current encoder-decoder models face challenges with semantic information gaps and ineffective global-local feature interaction during decoding.

Purpose of the Study:

• To propose a novel Enhanced Decoder Semantic Recovery Network to overcome limitations in current deep learning-based medical image segmentation.
• To improve the accuracy and effectiveness of semantic recovery in medical image segmentation.

Main Methods:

• Introduced the Multi-Level Semantic Fusion (MLSF) module to fuse features at various levels, weighted by semantic gaps.
• Employed Multiscale Spatial Attention (MSSA) and Cross Convolution Channel Attention (CCCA) modules for richer feature extraction.
• Designed the Global-Local Semantic Recovery (GLSR) module to enhance semantic recovery by considering both global and local information.

Main Results:

• The proposed model demonstrated significant improvements in Intersection over Union (IoU) on public datasets (BUSI, CVC-ClinicDB, Kvasir-SEG).
• Achieved performance gains of 0.81% on BUSI, 0.85% on CVC-ClinicDB, and 1.98% on Kvasir-SEG compared to existing methods.
• Significantly enhanced the performance of 2D medical image segmentation.

Conclusions:

• The Enhanced Decoder Semantic Recovery Network effectively addresses semantic information gaps and improves feature interaction in medical image segmentation.
• The proposed method offers a robust solution for accurate medical image segmentation, providing technical support for future advancements in the field.