Less Is More: Surgical Phase Recognition From Timestamp Supervision.

Area of Science:

• Computer-assisted surgery
• Medical image analysis
• Machine learning in healthcare

Background:

• Surgical phase recognition is crucial for computer-assisted surgery.
• Current methods rely on expensive, time-consuming full video annotations.
• This limits the scalability and practicality of these systems.

Purpose of the Study:

• To introduce timestamp supervision for surgical phase recognition, reducing annotation costs.
• To propose a novel method, Uncertainty-Aware Temporal Diffusion (UATD), for effective training with timestamp annotations.
• To demonstrate the efficacy of timestamp supervision and UATD in achieving competitive performance.

Main Methods:

• Developed timestamp supervision, requiring surgeons to mark only one timestamp per phase.
• Proposed Uncertainty-Aware Temporal Diffusion (UATD) to iteratively diffuse timestamp labels to confident neighboring frames.
• Evaluated UATD's performance against full supervision methods.

Main Results:

• Timestamp annotation reduced annotation time by 74% compared to full annotation.
• UATD achieved competitive results, comparable to full supervision methods, with significantly lower annotation costs.
• The study found that fewer, more discriminative pseudo-labels can outperform extensive, ambiguous labels.
• UATD demonstrated effectiveness as a plug-and-play module for refining phase boundary labels.

Conclusions:

• Timestamp supervision offers a cost-effective alternative to full annotation for surgical phase recognition.
• The proposed UATD method effectively leverages timestamp annotations to train high-performing models.
• This approach has the potential to improve the development and deployment of computer-assisted surgery systems.