Liver mask-guided SAM-enhanced dual-decoder network for landmark segmentation in AR-guided surgery

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Surgical Navigation

Background:

• Accurate liver landmark segmentation is vital for 3D-2D registration in AR-guided laparoscopic surgery.
• Existing methods face challenges with complex anatomical structures, limited datasets, and class imbalance.

Purpose of the Study:

• To enhance liver landmark segmentation performance in AR-guided laparoscopic surgery.
• To address limitations of current segmentation techniques through a novel approach leveraging liver mask prediction.

Main Methods:

• A dual-decoder model was developed, incorporating a pre-trained Segment Anything Model (SAM) encoder.
• One decoder segmenting the liver and another focusing on liver landmarks.
• A liver-guided consistency constraint was implemented to ensure spatial consistency between liver regions and landmarks.

Main Results:

• The proposed method achieved state-of-the-art performance on public laparoscopic liver surgery datasets.
• The dual-decoder framework, enhanced by SAM and consistency constraints, significantly improved segmentation accuracy in complex surgical scenarios.
• Feature entanglement was addressed, leading to more robust landmark segmentation.

Conclusions:

• The SAM-enhanced dual-decoder network with liver-guided consistency constraints provides a robust solution for 2D landmark segmentation in AR-guided laparoscopic surgery.
• The method demonstrates improved accuracy and robustness for intraoperative applications by mutually reinforcing liver mask and landmark segmentation.