Spatio-Temporal Causal Transformer for Multi-Grained Surgical Phase Recognition.

Area of Science:

• Computer Science
• Medical Informatics
• Artificial Intelligence

Background:

• Automatic surgical phase recognition is crucial for optimizing clinical workflows.
• Challenges include similar appearances between phases and variable phase durations.
• Existing methods struggle with the complexity of surgical procedures.

Purpose of the Study:

• To develop an effective method for online surgical phase recognition with varying granularity.
• To address the limitations of current approaches in handling complex surgical data.

Main Methods:

• A spatio-temporal transformer model was proposed.
• A spatial transformer was utilized to capture global spatial dependencies.
• A temporal transformer with a dual pyramid pattern was employed to handle variable phase durations.

Main Results:

• The method achieved 91.4% accuracy on the Cholec80 dataset (7 coarse-grained phases).
• It reached 84.2% accuracy on the CATARACTS2020 dataset (19 fine-grained phases).
• The model has an efficient parameter count of 24.5 million.

Conclusions:

• The proposed spatio-temporal transformer demonstrates superior performance in surgical phase recognition.
• This approach offers a robust solution for both coarse and fine-grained surgical phase identification.
• The method provides a promising tool for surgical workflow analysis and optimization.