Feature-guided attention network for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• U-Net and its variants are effective for medical image segmentation but suffer from background noise in shallow encoder features.
• Semantic gaps between encoder and decoder features in U-Net hinder performance due to direct skip-connections.

Purpose of the Study:

• To propose a novel feature-guided attention network for medical image segmentation.
• To address U-Net limitations by integrating a cross-level attention filtering module (CAFM) and an attention-guided upsampling module (AUM).

Main Methods:

• The Attention-Guided Upsampling Module (AUM) filters background noise in low-level encoder features using high-level feature maps.
• The Cross-Level Attention Filtering Module (CAFM) reduces the semantic gap between encoder and decoder features.
• The AUM guides decoder feature upsampling with encoder features, improving semantic alignment.

Main Results:

• The proposed method achieved superior Intersection over Union (IoU), dice, accuracy, and sensitivity across four diverse medical imaging datasets (coronary artery, retinal vessels, skin lesions, retinal edema).
• Outperformed state-of-the-art methods like CA-Net and SCS-Net in key segmentation metrics.
• Demonstrated high computational efficiency with fewer parameters (2.43 M) compared to existing models.

Conclusions:

• The feature-guided attention network achieves state-of-the-art performance in medical image segmentation.
• The CAFM effectively filters background noise, and the AUM bridges the semantic gap between encoder and decoder.
• The model offers high computational efficiency, making it a practical solution for medical image analysis.