An international team of researchers from China, Poland, and Ukraine has presented LensNarrate (1) — an artificial intelligence system that automatically converts cataract surgery recordings into structured medical reports. The study results were published in Communications Medicine, part of the Nature publishing portfolio.

Cataract surgery is one of the most frequently performed surgical procedures in the world. Although the procedure is often recorded on video, the operative report is usually prepared only after the procedure is completed, based on the surgeon’s memory and a standard template. This can lead to short stages of the procedure being omitted, their sequence being inaccurately reconstructed, and inconsistencies in how similar operations are documented.

LensNarrate analyzes unedited footage, identifies the actions being performed, and assigns them to 12 core phases of cataract surgery. These include, among others, corneal incision, injection of viscoelastic substance, capsulorhexis, hydrodissection, phacoemulsification, removal of residual lens cortex, intraocular lens implantation, and wound sealing.

The system samples two frames per second from the recording and, using a model that combines image analysis with language generation, describes the actions visible in them. It then arranges the results in temporal order, removes spurious, short-lived changes in classification, and converts them into a coherent narrative of the surgery’s course. Additional AI models highlight surgical instruments and ocular structures within the image, allowing the physician to verify the visual basis of the generated description.

The system was initially tested on 20 surgical recordings from Poland and Ukraine. LensNarrate correctly identified the surgical phase in 83.2% of the frames analyzed. Temporal ordering of the results proved particularly important: it increased agreement between the recognized surgical segments and the reference record from 6.5% to 58.0%.

The system was then evaluated on ten videos from Ningbo Eye Hospital in China. These recordings differed from the European material in, among other things, the instruments, materials, and surgical-technique details used. In this dataset, phase-recognition accuracy was 71.6%, showing that differences between hospitals remain a significant challenge for medical artificial intelligence.

The automatically generated reports were also evaluated by three ophthalmologists who did not know which system had produced each document. LensNarrate achieved slightly better results than the comparison models, Qwen2.5-VL and VideoCLIP. However, overall clinical-acceptability ratings remained relatively low, meaning the solution is not yet ready to independently generate medical documentation.

One of the most significant problems was the system’s tendency to generate details that cannot be reliably read from the video, such as the model and power of the implanted lens, the composition of the medication used, or the exact angle of the incision. The system sometimes repeated information that occurred frequently in the training data, even when it did not correspond to the procedure being analyzed. This shows that a well-worded, AI-generated report always requires review by a physician.

LensNarrate should currently be regarded as a promising tool to support the surgeon. In the future, similar systems could prepare preliminary operative reports, facilitate quality analysis of procedures, and support the training of young ophthalmologists. The next steps should include larger multicenter studies and prospective evaluation of whether using the system actually shortens documentation time and reduces the number of omitted details.

The most realistic model for using this technology remains human–AI collaboration: the system analyzes the recording and prepares a structured draft report, while the surgeon verifies its accuracy, fills in any missing data, and approves the final documentation.

These results form part of the work of the Polish Artificial Intelligence in Ophthalmology Consortium, established in 2024. The consortium brings together leading Polish clinical, academic, and technology centers, as well as international partners, creating a collaborative platform for the development, validation, and implementation of artificial intelligence methods for the diagnosis and treatment of eye diseases. One of the consortium’s main objectives is to conduct multicenter clinical studies on the application of artificial intelligence, build high-quality imaging databases, and develop safe, reliable clinical decision-support tools in line with the highest scientific and regulatory standards. In 2025 and 2026, the consortium took part in other highly significant studies, the results of which were published in Nature Medicine (2, 3). Consortium’s leader and coordinator is the Foundation for Ophthalmology Development.

I invite all institutions and researchers interested in research on AI in ophthalmology to collaborate.

Andrzej Grzybowski (ae.grzybowski@gmail.com)