DSTED: decoupling temporal stabilization and discriminative enhancement for surgical workflow recognition

Area of Science:

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

Background:

• Surgical workflow recognition is crucial for context-aware assistance and skill assessment in computer-assisted interventions.
• Current methods face challenges with prediction jitter and discriminating ambiguous surgical phases.

Purpose of the Study:

• To develop a stable surgical workflow recognition framework.
• To address prediction jitter and enhance discrimination of ambiguous phases using reliable historical information and uncertainty modeling.

Main Methods:

• A dual-pathway framework, DSTED, was proposed, incorporating Reliable Memory Propagation (RMP) and Uncertainty-Aware Prototype Retrieval (UPR).
• RMP filters and fuses high-confidence historical features for temporal coherence.
• UPR refines ambiguous frame representations by constructing and matching prototypes from high-uncertainty samples.

Main Results:

• The method achieved state-of-the-art performance on AutoLaparo-hysterectomy with 84.36% accuracy and 65.51% F1-score.
• Ablation studies showed complementary gains from RMP (2.19%) and UPR (1.93%).
• Significant reduction in temporal jitter and improvement on challenging phase transitions were confirmed.

Conclusions:

• The dual-pathway design offers a novel paradigm for stable workflow recognition.
• Decoupling temporal consistency and phase ambiguity modeling leads to superior performance and clinical applicability.