Real-time instance segmentation of surgical instruments using attention and multi-scale feature fusion

Area of Science:

• Medical image analysis
• Computer-assisted surgery
• Deep learning applications

Background:

• Accurate real-time tracking of surgical instruments is vital for minimally invasive computer-assisted surgery.
• Challenges include complex surgical environments and the need for both accuracy and speed in model design.
• Deep learning offers potential for learning complex surgical environments and instrument placements.

Purpose of the Study:

• To develop a light-weight, single-stage instance segmentation model for faster and more accurate surgical instrument segmentation.
• To improve model accuracy through data augmentation and optimal anchor localization.
• To achieve real-time performance while maintaining high accuracy for surgical instrument tracking.

Main Methods:

• Proposed a light-weight single-stage instance segmentation model.
• Integrated a convolutional block attention module to enhance inference speed and accuracy.
• Employed data augmentation and optimal anchor localization strategies for further accuracy improvements.
• Utilized the ROBUST-MIS challenge dataset comprising over 10,000 frames of surgical tools.

Main Results:

• Achieved over 44% improvement on the MI_DSC metric and 39% on the MI_NSD metric compared to top teams in the ROBUST-MIS challenge.
• Demonstrated real-time performance exceeding 60 frames-per-second with competitive variants of the proposed approach.
• Outperformed previous top team performances in the ROBUST-MIS challenge.

Conclusions:

• The proposed model offers a significant advancement in real-time surgical instrument segmentation, balancing speed and accuracy.
• This approach enhances surgeon navigation and patient safety in computer-assisted surgeries.
• The model sets a new benchmark for performance in surgical instrument segmentation challenges.