Automatic segmentation of colon, small intestine, and duodenum based on scale attention network

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automatic segmentation of abdominal organs like the colon, small intestine, and duodenum presents challenges due to significant organ scale variability.
• Existing methods often rely on hierarchical structures for multi-scale feature extraction, but efficient exploitation remains an area for improvement.

Purpose of the Study:

• To develop a novel method for efficiently exploiting multi-scale features to improve the segmentation accuracy of abdominal organs.
• To introduce a Scale Attention Module (SAM) that adaptively recalibrates multi-scale features based on their importance.

Main Methods:

• A Scale Attention Network (SANet) was constructed by integrating the proposed Scale Attention Module (SAM) into a segmentation model.
• The SAM adaptively recalibrates multi-scale features by modeling their importance scores using pooled representations and a lightweight network.
• Multi-scale features are fused via pixel-wise summation before being fed into the segmentation head.

Main Results:

• SANet demonstrated superior performance in segmenting the colon, small intestine, and duodenum compared to established models like UNet and Deeplabv3p.
• Achieved high Dice Similarity Coefficients (DSCs) of 84.06% for colon, 76.79% for small intestine, and 61.68% for duodenum.
• The SAM module provided improvements of 0.83-2.71 points in DSC with comparable or fewer parameters than other attention mechanisms.

Conclusions:

• The Scale Attention Network (SANet) effectively addresses the scale-variability problem in abdominal organ segmentation.
• SANet offers an efficient approach to exploit multi-scale features, leading to enhanced segmentation performance and accuracy.