LRTD: long-range temporal dependency based active learning for surgical workflow recognition.

Area of Science:

• Computer-assisted surgery
• Robotic surgery
• Medical image analysis

Background:

• Surgical workflow recognition in videos is crucial for advancing computer-assisted and robotic surgery.
• Current deep learning methods require extensive, surgeon-annotated datasets, which are costly and time-consuming to create.
• The scarcity of annotated surgical data presents a significant bottleneck for developing effective AI tools.

Purpose of the Study:

• To develop a cost-effective active learning method for surgical video analysis.
• To reduce the reliance on large-scale, manually annotated surgical video datasets.
• To improve the efficiency of training AI models for surgical workflow recognition.

Main Methods:

• Proposed a novel active learning strategy using a non-local recurrent convolutional network.
• Introduced non-local blocks to capture long-range temporal dependency (LRTD) within surgical video frames.
• Developed an intra-clip dependency score to identify and select the most informative video clips for annotation.

Main Results:

• The LRTD-based active learning strategy outperformed state-of-the-art methods on the Cholec80 dataset for surgical workflow recognition.
• Achieved performance comparable to full-data training using only up to 50% of the annotated samples.
• Demonstrated superior clip selection capability by modeling intra-clip dependency.

Conclusions:

• The proposed LRTD-based active learning strategy effectively identifies informative surgical video clips, reducing annotation workload.
• This approach shows significant potential for practical application in clinical settings by minimizing annotation efforts.
• The method offers a promising solution for data-scarce scenarios in surgical video analysis.