Edge-Guided Multi-Scale Frequency Attention Network for Gastrointestinal Cancer Image Segmentation

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of gastrointestinal tumors is crucial for clinical decision-making in computer-aided diagnostic systems.
• Existing segmentation models struggle with the diverse morphology and fuzzy boundaries of gastrointestinal tumors, leading to suboptimal performance.
• There is a need for advanced segmentation techniques to improve feature recognition and adaptability in complex pathological contexts.

Purpose of the Study:

• To develop an advanced segmentation network, EGMFA-Net, specifically designed for gastrointestinal tumor segmentation.
• To address the challenges posed by diverse tumor morphologies and fuzzy boundaries in medical images.
• To enhance the accuracy and robustness of computer-aided diagnostic systems through improved image segmentation.

Main Methods:

• Designed an edge-guided multi-scale frequency attention network (EGMFA-Net) incorporating a Kernel Adaptive Enhancement Module (KAEM) and a Frequency-domain Self-attention Module (FDSA).
• KAEM adaptively adjusts feature extraction kernels based on lesion morphology, optimizing feature representation.
• FDSA aggregates multi-scale features in the frequency domain to achieve global receptive fields while preserving high-frequency details.

Main Results:

• EGMFA-Net demonstrated state-of-the-art performance across eight diverse medical image benchmark datasets (SEED, Kvasir, ClinicDB, ColonDB, ETIS, BKAI, CVC-300, and Synapse).
• The proposed modules effectively enhanced recognition of varied morphology regions and improved adaptability to complex pathological contexts.
• Experimental results indicate superior accuracy and robustness compared to existing gastrointestinal tumor segmentation methods.

Conclusions:

• EGMFA-Net offers a significant advancement in gastrointestinal tumor segmentation, outperforming current state-of-the-art methods.
• The network's architecture, particularly KAEM and FDSA, effectively handles challenges associated with tumor morphology and boundary ambiguity.
• This work contributes to improving the accuracy of computer-aided diagnostic systems, potentially leading to better clinical decisions and treatments.