Mask then classify: multi-instance segmentation for surgical instruments.

Area of Science:

• Robotics
• Computer Vision
• Medical Imaging

Background:

• Accurate surgical instrument detection and segmentation are crucial for minimally invasive robotic surgery.
• Existing semantic segmentation methods lack instance-level information and precise instrument type classification.

Purpose of the Study:

• To develop a novel instance segmentation method for surgical instruments.
• To overcome limitations of semantic segmentation in differentiating instrument instances and types.
• To improve the accuracy and informativeness of surgical instrument segmentation.

Main Methods:

• A novel instance segmentation approach is proposed, involving pixel-wise masking of instrument instances before classification.
• An encoder-decoder network is utilized for extracting instrument instances.
• Instrument priors from surgical robots are incorporated to enhance performance.

Main Results:

• The method achieved over 18% improvement compared to the state-of-the-art on the 2017 endoscopic vision challenge dataset.
• An ablation study demonstrated a 10% increase in performance over semantic segmentation methods.
• The approach provides precise segmentation masks and instance-level information.

Conclusions:

• A novel instance segmentation method for surgical instruments has been presented, outperforming previous semantic segmentation techniques.
• The method offers more informative instance-level details while maintaining precise segmentation.
• Incorporating robotic instrument priors further boosts performance.