[Multi-scale medical image segmentation based on pixel encoding and spatial attention mechanism].

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer-assisted diagnosis

Context:

• U-Net and its variants suffer from single-scale information loss and large parameter size in medical image segmentation.
• Accurate multi-organ segmentation is crucial for clinical diagnosis and treatment planning.
• Existing methods struggle to effectively capture multi-scale features crucial for complex anatomical structures.

Purpose:

• To propose a novel multi-scale medical image segmentation method addressing U-Net limitations.
• To enhance feature extraction by incorporating pixel encoding and spatial attention mechanisms.
• To improve segmentation accuracy, model convergence, and computational efficiency.

Summary:

• A new method utilizes a redesigned Transformer input strategy with pixel encoding for multi-scale feature extraction, incorporating deformable convolutions to accelerate convergence.
• A spatial attention module with residual connections is employed to focus on foreground information in fused feature maps.
• Ablation studies confirm network lightweighting enhances segmentation accuracy and speeds up convergence, achieving a Dice Similarity Coefficient of 77.65 and HD95 of 18.34 on the Synapse dataset.

Impact:

• The proposed algorithm demonstrates improved multi-organ segmentation performance, potentially addressing limitations in current multi-scale medical image segmentation techniques.
• This advancement can aid physicians in diagnosis by providing more accurate segmentation results.
• The method offers a more efficient and accurate solution for medical image analysis, contributing to the field of computer-assisted diagnosis.