Pairwise attention-enhanced adversarial model for automatic bone segmentation in CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate bone segmentation is crucial for surgical navigation, particularly in screw placement.
• Deep learning has advanced bone segmentation, but challenges remain with irregular shapes and similar features.

Purpose of the Study:

• To develop an effective deep learning model for automatic bone segmentation in computed tomography (CT) images.
• To address the limitations of existing methods in segmenting local bone structures with complex shapes and features.

Main Methods:

• Proposed the pairwise attention-enhanced adversarial model (Pair-SegAM) comprising a segmentation model and a discriminator.
• Improved the discriminator to enhance awareness of target regions and semantic feature parsing.
• Implemented a pairwise structure with attention maps and semantic fusion to refine segmentation and filter unstable regions.

Main Results:

• Evaluated Pair-SegAM on two bone datasets against existing segmentation and adversarial models.
• Demonstrated superior bone segmentation performance and effective generalization capabilities.
• The improved discriminator provided refinement information for accurate bone outline capture.

Conclusions:

• Pair-SegAM offers a more efficient and accurate method for segmenting specific bones from CT images.
• The model shows significant potential for extension to other semantic segmentation applications in medical imaging.