Clinical Report: AI in the Operating Room

Overview

LensNarrate is an AI system designed to convert cataract surgery video into structured operative-report drafts. In initial evaluations, it identified surgical phases with 83.2% accuracy in European recordings and 71.6% in a separate Chinese dataset; ophthalmologists rated its reports slightly better than comparison models, but overall clinical acceptability remained low. The source describes physician review as necessary before any report is finalized.

Background

Cataract surgery is among the most frequently performed surgical procedures worldwide, and operative reports are commonly prepared after surgery from the surgeon’s memory and a standard template. The article notes that this approach can omit brief procedural stages, misstate their sequence, and produce inconsistencies across similar operations. LensNarrate was developed to use unedited surgical video to identify procedural phases and generate a chronological narrative. Its evaluation also addresses whether performance transfers across hospitals using different instruments, materials, and techniques.

Data Highlights

Key Findings

• LensNarrate samples two frames per second, classifies actions into 12 core cataract-surgery phases, orders the results temporally, filters short-lived classification changes, and generates a narrative report.
• Additional AI models highlight instruments and ocular structures in the video so a physician can check the visual basis of the generated description.
• In 20 recordings from Poland and Ukraine, phase-recognition accuracy was 83.2%; temporal ordering increased agreement between recognized segments and the reference record from 6.5% to 58.0%.
• In 10 recordings from a Chinese hospital, phase-recognition accuracy was 71.6%; the article reports differences in instruments, materials, and surgical-technique details between the Chinese and European datasets.
• Three blinded ophthalmologists rated LensNarrate slightly better than Qwen2.5-VL and VideoCLIP, while overall clinical-acceptability ratings remained relatively low.
• The system sometimes generated details not reliably discernible from video, including implanted-lens model and power, medication composition, and incision angle; the article states that physician review is required.

Clinical Implications

The article presents LensNarrate as a tool for preparing a structured draft, with the surgeon verifying accuracy, adding missing information, and approving the final documentation. It identifies larger multicenter studies and prospective evaluation of documentation time and omitted details as next steps; the reported results do not establish those effects.

Conclusion

LensNarrate demonstrated automated cataract-surgery phase recognition and video-based report generation, but performance varied across datasets and report acceptability remained limited. The source describes human–AI collaboration, rather than independent report generation, as the current model for use.

Related Resources & Content

1. LensNarrate study authors, Communications Medicine, 2026 — Video-to-report generation for cataract surgery using procedurally grounded vision–language models
2. MDSpire News, n.d. — AI in Surgery: Debate Highlights Benefits, Gaps
3. MDSpire News, n.d. — AI could turn every surgical patient into an 'information donor'
4. Frontiers in Surgery, 2026 — Utilizing Artificial Intelligence to Enhance Precision in Surgical Oncology During the Perioperative Phase
5. reuters — AI-Generated Operative Notes More Accurate
6. Video-to-report generation for cataract surgery using procedurally grounded vision–language models
7. Provider Data Catalog: Measures and Current Data Collection Periods
8. Guidance on AI-enhanced surgical practice: a Delphi consensus on ontology, data, implementation and evaluation
9. Video-to-report generation for cataract surgery using procedurally grounded vision–language models | Communications Medicine