CGBA-Net: context-guided bidirectional attention network for surgical instrument segmentation.

Area of Science:

• Robotics
• Computer Vision
• Medical Imaging

Background:

• Automatic surgical instrument segmentation is vital for robotic-aided surgery.
• Existing encoder-decoder networks struggle with complex scenes and uneven illumination, leading to segmentation errors.
• Irrelevant feature fusion in current methods increases misclassification.

Purpose of the Study:

• To propose a novel network, CGBANet, for accurate automatic surgical instrument segmentation.
• To address challenges posed by complex surgical scenes and uneven illumination.
• To enhance feature selection for improved segmentation accuracy.

Main Methods:

• Introduced the context-guided bidirectional attention network (CGBANet).
• Developed the guidance connection attention (GCA) module to filter irrelevant low-level features.
• Integrated a bidirectional attention (BA) module within GCA to capture local and global dependencies.

Main Results:

• CGBANet demonstrated superior performance in multiple instrument segmentation tasks on public datasets (EndoVis 2018, cataract surgery).
• Experimental results confirmed CGBANet outperforms state-of-the-art methods.
• Ablation studies validated the effectiveness of the proposed GCA and BA modules.

Conclusions:

• The proposed CGBANet significantly increases the accuracy of multiple instrument segmentation.
• The network accurately classifies and segments surgical instruments.
• The developed modules effectively provide instrument-related features, improving segmentation outcomes.