DHR²-Net: A Dual-Hierarchical Relational Reasoning Framework for Surgical Action Triplet Recognition

Area of Science:

• Computer-assisted surgery
• Surgical robotics
• Medical image analysis

Background:

• Surgical action triplet recognition is crucial for computer-assisted surgery.
• Current methods struggle with abstract operations and inter-entity relationships.

Purpose of the Study:

• To propose a novel Dual-Hierarchical Relational Reasoning framework (DHR^2-Net) for improved surgical action triplet recognition.
• To address limitations in recognizing abstract surgical operations and modeling inter-entity relationships.

Main Methods:

• DHR^2-Net explicitly models verbs as semantic interactions between instruments and targets.
• Dual-hierarchical relational reasoning constructs first-order and high-order dependencies.
• An inter-sample high-order relational reasoning module (IHR^2) enhances feature representation.
• An auxiliary actor identity preservation module captures identity and relational semantics.

Main Results:

• DHR^2-Net demonstrates superior performance in fine-grained component recognition and interaction modeling.
• The framework achieves robust accuracy in surgical action triplet recognition.
• Experimental results on CholecT45 and CholecTriplet2021 datasets validate the proposed method.

Conclusions:

• DHR^2-Net effectively models complex relationships in surgical actions.
• The proposed method shows strong clinical interpretability and potential for practical deployment.
• This framework advances the field of computer-assisted intervention systems.