TPDPM: Text promptable diffusion probabilistic model for referring surgical instrument segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Robotics

Background:

• Referring surgical instrument segmentation (RSIS) is crucial for context-aware computer-assisted interventions.
• Existing RSIS methods rely on discriminative per-pixel classification.
• Diffusion probabilistic models show promise in segmentation tasks.

Purpose of the Study:

• To propose a novel Text Promptable Diffusion Probabilistic Model (TPDPM) for RSIS.
• To formulate RSIS as a mask generation task conditioned on images and text.
• To enable target-specific segmentation and uncertainty assessment in robotic surgery.

Main Methods:

• Developed TPDPM, a generative model for RSIS.
• Formulated RSIS as a mask generation task using a reverse diffusion process.
• Utilized ensembling of network predictions for accuracy and uncertainty estimation.

Main Results:

• TPDPM achieved superior RSIS performance compared to state-of-the-art methods.
• Demonstrated competitive generalization capabilities across diverse datasets.
• Successfully handled ambiguous instrument boundaries through stochastic sampling.

Conclusions:

• TPDPM offers a novel and effective approach to RSIS.
• The method enhances interactive capabilities for robot-assisted surgery.
• TPDPM provides accurate segmentation and uncertainty assessment for surgical instruments.