GESur_Net: attention-guided network for surgical instrument segmentation in gastrointestinal endoscopy

Area of Science:

• Medical Imaging
• Computer Vision
• Robotic Surgery

Background:

• Accurate surgical instrument segmentation is crucial for robotic autonomous surgical navigation.
• Existing methods struggle with edge/center attention and low-level feature utilization, impacting segmentation accuracy.

Purpose of the Study:

• To propose a lightweight network, GESur_Net, for improved surgical instrument segmentation in gastrointestinal (GI) endoscopy.
• To address limitations in attention and feature detail usage in current segmentation models.

Main Methods:

• Introduced the pixel data aggregation (PDA) mechanism for feature channel importance analysis.
• Developed the skip connection attention (SK_A) block to focus on critical instrument regions.
• Implemented the global guidance attention (GGA) block to integrate semantic and detailed features.
• Created the Gastrointestinal Endoscopic Instrument (GEI) dataset.

Main Results:

• GESur_Net demonstrated increased segmentation accuracy.
• The proposed model outperformed state-of-the-art segmentation models on both the GEI and Kvasir-instrument datasets.
• The network effectively fuses high-level semantic information with low-level detailed features.

Conclusions:

• The proposed GESur_Net effectively enhances surgical instrument segmentation accuracy in GI endoscopy.
• The novel PDA, SK_A, and GGA mechanisms contribute to improved feature analysis and fusion.
• The GEI dataset provides a valuable resource for future research in this domain.