Objective:

To evaluate LensNarrate, an AI system that converts cataract-surgery video into structured operative reports.

Approach:

• System: LensNarrate samples two frames per second, classifies actions into 12 surgical phases, orders them over time, and generates a narrative report. Visual annotations of instruments and eye structures support physician review.
• Evaluation: Researchers tested the system on 20 recordings from Poland and Ukraine and 10 recordings from a hospital in China. Three ophthalmologists, blinded to report source, assessed generated reports against those from Qwen2.5-VL and VideoCLIP.

Key Findings:

• Phase-recognition accuracy was 83.2% in the initial dataset and 71.6% in the Chinese dataset.
• Temporal ordering increased agreement between recognized surgical segments and the reference record from 6.5% to 58.0%.
• LensNarrate performed slightly better than Qwen2.5-VL and VideoCLIP in ophthalmologists’ assessments, but overall clinical-acceptability ratings remained relatively low.
• The system sometimes generated unsupported details, including lens specifications, medication composition, and incision angle.

Interpretation:

The article describes LensNarrate as a potential tool for preparing draft reports, but states that physician review is necessary and that the system is not ready to generate documentation independently.

Limitations:

• Evaluation involved only 30 recordings across the reported datasets.
• Recognition accuracy was lower in the Chinese dataset, where surgical materials and techniques differed from the European recordings.
• The system may generate details that cannot be reliably determined from video.
• The article calls for larger multicenter studies and prospective evaluation of documentation time and omitted details.

Conclusion:

The proposed use is human–AI collaboration: LensNarrate prepares a structured draft, and the surgeon checks, completes, and approves the final report.

Sources:

• Article context provided by the user