Accurate Segmentation of Surgical Instruments via Spectral-Attentive Contextual Interaction Network.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotic Surgery

Background:

• Surgical instrument segmentation is vital for robotic surgery but faces challenges like complex backgrounds and low contrast.
• Existing deep learning models struggle with fine edges and global context in surgical instrument segmentation.

Purpose of the Study:

• To introduce the Spectral-attentive Contextual Interaction Network (SCI-Net) for improved surgical instrument segmentation.
• To enhance accuracy and robustness in segmenting surgical instruments within complex surgical environments.

Main Methods:

• Proposed SCI-Net with a Global Context Aggregation Module (GCAM) for coarse localization.
• Introduced a Spectral-enhanced Feature Module (SFM) for frequency-domain feature enhancement.
• Designed a Scale-aware Dilation Module (SDM) for adaptive multi-scale feature integration and boundary refinement.

Main Results:

• SCI-Net demonstrated superior performance on multiple public surgical instrument segmentation datasets.
• The proposed modules (GCAM, SFM, SDM) effectively addressed limitations of existing segmentation models.
• Experimental results confirmed significant improvements over state-of-the-art methods.

Conclusions:

• SCI-Net offers a robust solution for surgical instrument segmentation in robotic surgery.
• The network's innovative modules enhance feature representation and boundary segmentation accuracy.
• This work advances the capabilities of computer-assisted robotic surgical systems.