EISegNet: Enhancing Instrument Segmentation Network via Dual-View Disparity Estimation.

Area of Science:

• Computer Vision
• Robotics
• Medical Imaging

Background:

• Accurate segmentation of endoscopic instruments is crucial for robot-assisted surgery, enabling precise navigation and automation.
• Existing monocular methods struggle with instrument segmentation due to complex environments, instrument-tissue similarity, and lighting variations.
• The distinct depth distributions of instruments, often overlooked, offer a key feature for improved segmentation.

Purpose of the Study:

• To develop an advanced framework, EISegNet, for enhanced endoscopic instrument segmentation.
• To leverage multi-task learning by combining instrument segmentation with disparity estimation.
• To improve the robustness and generalization of segmentation methods in diverse surgical scenarios.

Main Methods:

• Proposed EISegNet, a multi-task framework integrating instrument segmentation and auxiliary disparity estimation.
• Implemented an asymmetric cross-attention mechanism for feature fusion between segmentation and disparity tasks.
• Adapted stereo disparity estimation for dual-view depth estimation and incorporated a Gaussian-weighted loss function for edge feature emphasis.

Main Results:

• Achieved a 5.97% increase in Intersection over Union (IoU) for instrument segmentation.
• Demonstrated superior accuracy and generalization across extensive cross-dataset experiments.
• Showcased promising performance in qualitative evaluations on clinical datasets.

Conclusions:

• EISegNet effectively enhances endoscopic instrument segmentation accuracy by incorporating depth and edge information.
• The multi-task framework and novel loss function improve performance in challenging surgical conditions.
• The method shows significant potential for advancing surgical automation and safety in real-world clinical applications.